chore(deps): update dependencies

Let Anthropic-compatible providers override the reasoning model id used in signed thinking metadata.

Test: swift test --filter 'Anthropic-compatible provider tags native thinking with wrapper identity'

.github/workflows/ci.yml

@@ -16,10 +16,10 @@ jobs:
     runs-on: ${{ matrix.os }}
     strategy:
       matrix:
-        os: [macos-latest, ubuntu-22.04]
+        os: [macos-15, ubuntu-22.04]
     steps:
     - name: Checkout
-      uses: actions/checkout@v4
+      uses: actions/checkout@v7
 
 
     - name: Fetch Commander dependency
@@ -37,7 +37,7 @@ jobs:
         swift-version: '6.2.1'
 
     - name: Cache Swift Package Manager
-      uses: actions/cache@v4
+      uses: actions/cache@v6
       with:
         path: |
           .build
@@ -52,7 +52,8 @@ jobs:
     - name: Run Tests (Unit Tests Only)
       run: |
         if [[ "${{ runner.os }}" == "Linux" ]]; then
-          swift test --filter TachikomaTests --skip "OpenAIAudioProviderTests" --skip "ProviderEndToEndTests"
+          # Several test suites mutate process-wide env/profile state.
+          swift test --no-parallel --filter TachikomaTests --skip "OpenAIAudioProviderTests" --skip "ProviderEndToEndTests"
         else
           swift test --filter TachikomaTests
         fi
@@ -79,10 +80,10 @@ jobs:
     continue-on-error: true
     strategy:
       matrix:
-        os: [macos-latest]
+        os: [macos-15]
     steps:
     - name: Checkout
-      uses: actions/checkout@v4
+      uses: actions/checkout@v7
 
     - name: Fetch Commander dependency
       shell: bash
@@ -120,7 +121,7 @@ jobs:
     # Disabled for now: simulator SDK downloads are slow/flaky and not needed for core validation.
     if: false
     name: Apple Platforms (${{ matrix.platform }})
-    runs-on: macos-latest
+    runs-on: macos-15
     needs: test
     strategy:
       matrix:
@@ -130,12 +131,12 @@ jobs:
 
   lint:
     name: SwiftLint
-    runs-on: macos-latest
+    runs-on: macos-15
     # SwiftLint only runs on macOS
     
     steps:
     - name: Checkout
-      uses: actions/checkout@v4
+      uses: actions/checkout@v7
 
     - name: Install SwiftLint
       run: brew install swiftlint

.github/workflows/cross-platform.yml

@@ -27,7 +27,7 @@ jobs:
     name: Ubuntu 22.04 LTS
     container: swift:6.2
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v7
       - name: Fetch Commander dependency
         shell: bash
         run: git clone --depth 1 https://github.com/steipete/Commander.git ../Commander
@@ -48,14 +48,15 @@ jobs:
           export OPENAI_API_KEY="${OPENAI_API_KEY:-test-key}"
           export ANTHROPIC_API_KEY="${ANTHROPIC_API_KEY:-test-key}"
           SKIP_FLAGS="--skip ProviderEndToEndTests"
-          swift test $SKIP_FLAGS
+          # Several test suites mutate process-wide env/profile state.
+          swift test --no-parallel $SKIP_FLAGS
 
   test-linux-ubuntu-24:
     runs-on: ubuntu-24.04
     name: Ubuntu 24.04 LTS
     container: swift:6.2
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v7
       - name: Fetch Commander dependency
         shell: bash
         run: git clone --depth 1 https://github.com/steipete/Commander.git ../Commander
@@ -76,7 +77,8 @@ jobs:
           export OPENAI_API_KEY="${OPENAI_API_KEY:-test-key}"
           export ANTHROPIC_API_KEY="${ANTHROPIC_API_KEY:-test-key}"
           SKIP_FLAGS="--skip ProviderEndToEndTests"
-          swift test $SKIP_FLAGS
+          # Several test suites mutate process-wide env/profile state.
+          swift test --no-parallel $SKIP_FLAGS
 
   # Optional: Build release artifacts
   build-release:
@@ -84,7 +86,7 @@ jobs:
     runs-on: ubuntu-latest
     if: github.ref == 'refs/heads/main'
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v7
       
       - name: Create release info
         run: |
@@ -99,7 +101,7 @@ jobs:
           echo "Total: 4 platform configurations tested in parallel" >> release-info.txt
       
       - name: Upload release info
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v7
         with:
           name: cross-platform-validation
           path: release-info.txt

.github/workflows/lint.yml

@@ -13,10 +13,10 @@ concurrency:
 jobs:
   swiftlint:
     name: SwiftLint
-    runs-on: macos-latest
+    runs-on: macos-15
     steps:
     - name: Checkout
-      uses: actions/checkout@v4
+      uses: actions/checkout@v7
 
     - name: Fetch Commander dependency
       shell: bash
@@ -28,7 +28,7 @@ jobs:
         xcode-version: '26.1'
 
     - name: Cache SwiftLint
-      uses: actions/cache@v4
+      uses: actions/cache@v6
       with:
         path: |
           ~/.mint
@@ -66,10 +66,10 @@ jobs:
 
   swiftformat:
     name: SwiftFormat
-    runs-on: macos-latest
+    runs-on: macos-15
     steps:
     - name: Checkout
-      uses: actions/checkout@v4
+      uses: actions/checkout@v7
 
     - name: Fetch Commander dependency
       shell: bash
@@ -81,7 +81,7 @@ jobs:
         xcode-version: '26.1'
 
     - name: Cache SwiftFormat
-      uses: actions/cache@v4
+      uses: actions/cache@v6
       with:
         path: |
           ~/.mint
@@ -108,10 +108,10 @@ jobs:
 
   swift6-compatibility:
     name: Swift 6 Compatibility Check
-    runs-on: macos-latest
+    runs-on: macos-15
     steps:
     - name: Checkout
-      uses: actions/checkout@v4
+      uses: actions/checkout@v7
 
     - name: Fetch Commander dependency
       shell: bash
@@ -158,7 +158,7 @@ jobs:
     runs-on: macos-15
     steps:
     - name: Checkout
-      uses: actions/checkout@v4
+      uses: actions/checkout@v7
 
     - name: Fetch Commander dependency
       shell: bash

.github/workflows/test.yml

@@ -14,7 +14,7 @@ jobs:
   # macOS tests with all Apple platforms
   test-apple-platforms:
     name: Test Apple Platforms
-    runs-on: macos-latest
+    runs-on: macos-15
     strategy:
       fail-fast: false
       matrix:
@@ -25,7 +25,7 @@ jobs:
 
     steps:
     - name: Checkout
-      uses: actions/checkout@v4
+      uses: actions/checkout@v7
 
     - name: Fetch Commander dependency
       shell: bash
@@ -37,7 +37,7 @@ jobs:
         xcode-version: '26.1'
 
     - name: Cache Swift Package Manager
-      uses: actions/cache@v4
+      uses: actions/cache@v6
       with:
         path: .build
         key: ${{ runner.os }}-spm-${{ hashFiles('**/Package.resolved') }}
@@ -75,7 +75,7 @@ jobs:
 
     steps:
     - name: Checkout
-      uses: actions/checkout@v4
+      uses: actions/checkout@v7
 
     - name: Fetch Commander dependency
       shell: bash
@@ -87,7 +87,7 @@ jobs:
         swift-version: ${{ matrix.swift-version }}
 
     - name: Cache Swift Package Manager
-      uses: actions/cache@v4
+      uses: actions/cache@v6
       with:
         path: .build
         key: ${{ runner.os }}-spm-${{ matrix.swift-version }}-${{ hashFiles('**/Package.resolved') }}
@@ -105,7 +105,8 @@ jobs:
           echo "OPENAI_API_KEY missing; skipping OpenAIAudioProviderTests"
           SKIP_FLAGS="$SKIP_FLAGS --skip OpenAIAudioProviderTests"
         fi
-        swift test --verbose $SKIP_FLAGS
+        # Several test suites mutate process-wide env/profile state.
+        swift test --no-parallel --verbose $SKIP_FLAGS
       env:
         OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
         ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
@@ -118,10 +119,10 @@ jobs:
   # Package validation
   validate-package:
     name: Validate Swift Package
-    runs-on: macos-latest
+    runs-on: macos-15
     steps:
     - name: Checkout
-      uses: actions/checkout@v4
+      uses: actions/checkout@v7
 
     - name: Fetch Commander dependency
       shell: bash
@@ -150,11 +151,11 @@ jobs:
   # Integration tests with real APIs (optional, requires secrets)
   integration-tests:
     name: Integration Tests
-    runs-on: macos-latest
+    runs-on: macos-15
     if: github.event_name == 'push' && github.ref == 'refs/heads/main'
     steps:
     - name: Checkout
-      uses: actions/checkout@v4
+      uses: actions/checkout@v7
 
     - name: Fetch Commander dependency
       shell: bash

.swiftlint.yml

@@ -107,8 +107,8 @@ disabled_rules:
 
 # Rule configurations
 file_length:
-  warning: 1000
-  error: 2000
+  warning: 2000
+  error: 2500
   ignore_comment_only_lines: true
 
 function_parameter_count:
@@ -137,8 +137,8 @@ trailing_comma:
   mandatory_comma: true
 
 type_body_length:
-  warning: 800
-  error: 1200
+  warning: 1800
+  error: 2200
 
 type_name:
   min_length:

CHANGELOG.md

@@ -2,6 +2,32 @@
 
 All notable changes to the Tachikoma project will be documented in this file.
 
+## [Unreleased]
+
+### Added
+- Added first-class OpenAI `chat-latest` support with parsing aliases, Responses API routing, model capabilities, and usage estimates.
+- Added first-class MiniMax support with the `MiniMax-M2.7` catalog models, `MINIMAX_API_KEY` / `MINIMAX_BASE_URL` configuration, bearer-token Anthropic-compatible transport, model parsing shortcuts, usage estimates, and provider tests.
+- Added explicit LM Studio model shortcuts such as `lmstudio` and `lmstudio/openai/gpt-oss-120b` so local provider selections no longer fall through to Ollama custom IDs.
+
+### Changed
+- Refreshed the first-class model catalog to current provider IDs: OpenAI GPT-5.5/5.4, Claude Fable 5/Opus 4.8/Opus 4.7/Sonnet 4.6/Haiku 4.5, Gemini 3.1, Mistral latest aliases, Groq current production IDs, and xAI Grok 4.3/4.20.
+- Added explicit `claude-fable-5` support with 1M context, 128K max output, signed-thinking replay, refusal handling, and non-streaming generation; `LanguageModel.default` remains `claude-opus-4-8`, while `LanguageModel.defaultStreaming` now uses streaming-safe `gpt-5.5`.
+- Removed stale direct model support for retired or non-canonical IDs including GPT-5.1/5.2/pseudo-thinking models, deprecated Claude Sonnet/Opus 4 snapshots, Grok 2/3/4-fast rows, old Groq Llama/Mixtral/Gemma aliases, stale Mistral aliases, and invalid LM Studio `current`.
+
+### Fixed
+- OpenAI `gpt-5-chat-latest` now preserves its distinct model identity, appears in model listings, and applies GPT-5 parameter filtering instead of being rewritten to `chat-latest`.
+- SwiftPM consumers now resolve Commander from the package URL instead of accidentally inheriting a sibling local checkout.
+- Ollama model parsing now preserves explicit custom vision model IDs such as `qwen2.5vl:3b` instead of falling back to `llama3.3` (#16).
+- Auth resolution now snapshots environment-ignore state consistently, preventing parallel tests and concurrent callers from falling back to stored OpenRouter credentials when an environment override is present.
+- SwiftPM consumers can now resolve Commander remotely instead of requiring a local `../Commander` checkout. Thanks @malpern.
+- Custom OpenAI-compatible and Anthropic-compatible providers now honor per-provider `options.apiKey` values from profile config. Thanks @381181295.
+- Google/Gemini request encoding now sends tool results as documented user `functionResponse` turns and merges consecutive same-role contents before calling the API. Thanks @hsrvc.
+- Google/Gemini tool schemas now drop orphan `required` entries before request encoding so Gemini accepts simplified MCP tool definitions. Thanks @bcharleson.
+- OpenAI Responses API providers now resolve shared OAuth/API-key credentials instead of requiring `OPENAI_API_KEY` directly.
+- Credential loading now only maps exact API-key credential names to provider API keys, so OAuth access/refresh tokens no longer overwrite configured OpenAI or Anthropic keys.
+- Custom OpenAI-compatible and Anthropic-compatible providers now forward configured proxy headers to request calls.
+- Anthropic base URL environment overrides now survive Swift release builds, so `ANTHROPIC_BASE_URL` can route requests through local proxies. Thanks @shraderdm.
+
 ## [0.2.0] - 2026-04-28
 
 ### Added
@@ -9,10 +35,10 @@ All notable changes to the Tachikoma project will be documented in this file.
 - Azure provider unit tests using URLProtocol stubs to verify path, query, and auth header construction.
 
 ### Changed
-- Added OpenAI's GPT-5.1 family (flagship/mini/nano) throughout the model enums, selectors, provider factories, capability registry, pricing tables, docs, and test suites. GPT aliases (`gpt`, `gpt-5`, `gpt-4o`) now normalize to `.openai(.gpt51)` so downstream apps inherit the new default seamlessly.
+- Added OpenAI's GPT-5.1 family (flagship/mini/nano) throughout the model enums, selectors, provider factories, capability registry, pricing tables, docs, and test suites. GPT aliases (`gpt`, `gpt-5`) now normalize to supported GPT-5 models so downstream apps inherit the new default seamlessly.
 - Expanded xAI Grok support to the full November 2025 catalog (`grok-4-fast-*`, `grok-code-fast-1`, `grok-2-*`, `grok-vision-beta`, etc.), updated the CLI shortcuts so `grok` now maps to `grok-4-fast-reasoning`, and refreshed selectors, provider parsers, capability tables, and docs snippets to match the official API lineup.
 - Google/Gemini support now targets the Gemini 2.5 family exclusively (`gemini-2.5-pro`, `gemini-2.5-flash`, `gemini-2.5-flash-lite`), with updated model selectors, parsers, docs, and pricing tables; older 1.5/2.0 IDs are no longer recognized.
-- Removed deprecated OpenAI reasoning models (`o1`, `o1-mini`, `o3`, `o3-mini`) in favour of the GPT‑5 family plus `o4-mini`, updating enums, provider factories, capability tables, prompts, and documentation metadata accordingly.
+- Removed deprecated OpenAI reasoning models (`o1`, `o1-mini`, `o3`, `o3-mini`, `o4-mini`) in favour of the GPT‑5 family, updating enums, provider factories, capability tables, prompts, and documentation metadata accordingly.
 - Google/Gemini integration now uses the documented `x-goog-api-key` header with `alt=sse` streaming, adds fallbacks for `GOOGLE_API_KEY` / `GOOGLE_APPLICATION_CREDENTIALS`, and hardens the SSE decoder so live tests succeed consistently.
 - Pruned Anthropic model support to the Claude 4.x line (Opus 4, Sonnet 4 / 4.5, Haiku 4.5) to match current API availability and reduce maintenance burden.
 - `TachikomaConfiguration` now loads credentials first and lets environment variables override them so operators can supersede stored settings without editing credentials files.
@@ -61,8 +87,8 @@ All notable changes to the Tachikoma project will be documented in this file.
 
 #### Provider Support
 - **OpenAI Provider**: Complete integration with dual API support
-  - Chat Completions API for standard models (GPT-4o, GPT-4.1)
-  - Responses API for reasoning models (o3, o4 series)
+  - Chat Completions API for standard custom models
+  - Responses API for GPT-5 models
   - Automatic API selection based on model capabilities
   - Parameter filtering for reasoning models
   - Full streaming support for both APIs
@@ -70,7 +96,7 @@ All notable changes to the Tachikoma project will be documented in this file.
 
 - **Anthropic Provider**: Native Claude API integration
   - Support for Claude 4 (Opus, Sonnet) with thinking modes
-  - Claude 3.5/3.7 series compatibility
+  - Claude 4.x series compatibility
   - Content block handling for multimodal inputs
   - System prompt separation
   - Server-Sent Events streaming

Examples/AI-CLI/README.md

@@ -25,7 +25,7 @@ ai-cli "What is the capital of France?"
 ai-cli --model claude "Explain quantum computing"
 
 # Stream the response
-ai-cli --stream --model gpt-4o "Write a short story"
+ai-cli --stream --model gpt-5.5 "Write a short story"
 ```
 
 ## Parameters
@@ -35,7 +35,7 @@ ai-cli --stream --model gpt-4o "Write a short story"
 | `-m, --model <MODEL>` | Specify the AI model to use |
 | `--api <chat\|responses>` | For OpenAI models: select API type (default: responses for GPT-5) |
 | `-s, --stream` | Stream the response in real-time |
-| `--thinking` | Show reasoning process (O3, O4, GPT-5 - note: API currently doesn't expose actual reasoning) |
+| `--thinking` | Show GPT-5 reasoning process (note: API currently doesn't expose actual reasoning) |
 | `--verbose, -v` | Show detailed debug output |
 | `--config` | Show current configuration and API key status |
 | `--help, -h` | Show help message |
@@ -60,30 +60,25 @@ Add to your shell profile (`~/.zshrc`, `~/.bashrc`) for persistence.
 ## Supported Models
 
 ### OpenAI
-- **GPT-5 Series**: `gpt-5`, `gpt-5-mini`, `gpt-5-nano`
-- **O-Series**: `o3`, `o3-mini`, `o3-pro`, `o4-mini`
-- **GPT-4**: `gpt-4.1`, `gpt-4.1-mini`, `gpt-4o`, `gpt-4o-mini`, `gpt-4-turbo`
-- **Legacy**: `gpt-3.5-turbo`
+- **GPT-5 Series**: `gpt-5.5`, `gpt-5.4`, `gpt-5.4-mini`, `gpt-5.4-nano`, `gpt-5`, `gpt-5-mini`, `gpt-5-nano`
 
 ### Anthropic
-- **Claude 4**: `claude-opus-4-1-20250805`, `claude-sonnet-4-20250514`
-- **Claude 3.7**: `claude-3-7-sonnet`
-- **Claude 3.5**: `claude-3-5-opus`, `claude-3-5-sonnet`, `claude-3-5-haiku`
+- **Claude 4.x**: `claude-opus-4-7`, `claude-opus-4-5`, `claude-opus-4-1-20250805`, `claude-sonnet-4-6`, `claude-sonnet-4-5-20250929`, `claude-haiku-4-5`
 
 ### Google
-- **Gemini 2.5**: `gemini-2.5-pro`, `gemini-2.5-flash`, `gemini-2.5-flash-lite`
+- **Gemini**: `gemini-3.1-pro-preview`, `gemini-3.1-flash-lite`, `gemini-3-flash-preview`, `gemini-2.5-pro`, `gemini-2.5-flash`, `gemini-2.5-flash-lite`
 
 ### Others
-- **Mistral**: `mistral-large-2`, `mistral-small`, `codestral`
-- **Groq**: `llama-3.1-70b`, `llama-3.1-8b`, `mixtral-8x7b`
-* **Grok**: `grok-4-0709`, `grok-4-fast-reasoning`, `grok-4-fast-non-reasoning`, `grok-code-fast-1`, `grok-3`, `grok-3-mini`, `grok-2-1212`, `grok-2-vision-1212`, `grok-2-image-1212`
-- **Ollama** (local): `llama3.3`, `llava`, `codellama`, any installed model
+- **Mistral**: `mistral-large-latest`, `mistral-medium-latest`, `mistral-medium-3-5`, `mistral-small-latest`, `open-mistral-nemo-2407`, `codestral-latest`
+- **Groq**: `openai/gpt-oss-120b`, `openai/gpt-oss-20b`, `llama-3.3-70b-versatile`, `llama-3.1-8b-instant`
+- **Grok**: `grok-4.3`, `grok-4.20-0309-reasoning`, `grok-4.20-0309-non-reasoning`
+- **Ollama** (local): `llama3.3`, `llava`, any installed model
 
 ### Model Shortcuts
-- `claude` → claude-opus-4-1-20250805
-- `gpt` → gpt-4.1
-- `gemini` → gemini-2.5-flash
-- `grok` → grok-4-fast-reasoning
+- `claude` → claude-opus-4-7
+- `gpt` → gpt-5.5
+- `gemini` → gemini-3.1-pro-preview
+- `grok` → grok-4.3
 - `llama` → llama3.3
 
 ## Examples
@@ -98,7 +93,7 @@ ai-cli --stream --model gpt-5 "Explain the theory of relativity"
 
 # API selection for OpenAI
 ai-cli --model gpt-5 --api chat "Use Chat Completions API"
-ai-cli --model o3 --api responses "Use Responses API"
+ai-cli --model gpt-5.5 --api responses "Use Responses API"
 
 # Debug mode
 ai-cli --verbose --model opus "Debug this request"

Examples/AI-CLI/Sources/AI-CLI.swift

@@ -58,7 +58,7 @@ struct AICLI {
             if let modelString = config.modelString {
                 model = try ModelSelector.parseModel(modelString)
             } else {
-                model = .openai(.gpt51) // Default to GPT-5.1
+                model = .openai(.gpt55)
             }
         } catch {
             print("❌ Error parsing model: \(error)")
@@ -190,7 +190,7 @@ struct AICLI {
             -m, --model <MODEL>     Specify the AI model to use
             --api <API>            For OpenAI models: 'chat' or 'responses' (default: responses for GPT-5)
             -s, --stream           Stream the response (partial support)
-            --thinking             Show reasoning/thinking process (O3, O4, GPT-5 via Responses API)
+            --thinking             Show reasoning/thinking process (GPT-5 via Responses API)
             --verbose, -v          Show detailed debug output
             --config               Show current configuration and exit
             -h, --help             Show this help message
@@ -202,7 +202,7 @@ struct AICLI {
 
             # Use specific models
             ai-cli --model claude "Explain quantum computing"
-            ai-cli --model gpt-4o "Describe this image"
+            ai-cli --model gpt-5.5 "Describe this image"
             ai-cli --model grok "Tell me a joke"
             ai-cli --model llama3.3 "Help me debug this code"
 
@@ -213,53 +213,51 @@ struct AICLI {
             # Streaming responses
             ai-cli --stream --model claude "Write a short story"
 
-            # Show thinking process (reasoning models)
-            ai-cli --thinking --model gpt-5-thinking "Solve this logic puzzle"
+            # Show thinking process
+            ai-cli --thinking --model gpt-5.5 "Solve this logic puzzle"
             ai-cli --thinking --model gpt-5 "Complex reasoning task"
 
         PROVIDERS & MODELS:
 
         OpenAI:
           • gpt-5, gpt-5-pro, gpt-5-mini, gpt-5-nano (GPT-5 series, August 2025)
-          • gpt-5-thinking, gpt-5-thinking-mini, gpt-5-thinking-nano
-          • gpt-4.1, gpt-4.1-mini, o4-mini (GPT-4.1 / reasoning)
-          • gpt-4o, gpt-4o-mini (Multimodal)
-          • gpt-4-turbo (Legacy)
+          • gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-5.4-nano
 
         Anthropic:
-          • claude-opus-4-1-20250805, claude-sonnet-4-20250514 (Claude 4)
-          • claude-3-7-sonnet (Claude 3.7)
-          • claude-3-5-opus, claude-3-5-sonnet, claude-3-5-haiku (Claude 3.5)
+          • claude-opus-4-7, claude-opus-4-5, claude-opus-4-1-20250805
+          • claude-sonnet-4-6, claude-sonnet-4-5-20250929, claude-haiku-4-5
 
         Google:
+          • gemini-3.1-pro-preview, gemini-3.1-flash-lite
+          • gemini-3-flash-preview
           • gemini-2.5-pro (reasoning, thinking support)
           • gemini-2.5-flash, gemini-2.5-flash-lite
 
         Mistral:
-          • mistral-large-2, mistral-large, mistral-small
-          • mistral-nemo, codestral
+          • mistral-large-latest, mistral-medium-latest, mistral-medium-3-5
+          • mistral-small-latest, open-mistral-nemo-2407, codestral-latest
 
         Groq (Ultra-fast):
-          • llama-3.1-70b, llama-3.1-8b
-          • mixtral-8x7b, gemma2-9b
+          • openai/gpt-oss-120b, openai/gpt-oss-20b
+          • llama-3.3-70b-versatile, llama-3.1-8b-instant
+          • meta-llama/llama-4-maverick-17b-128e-instruct
 
         Grok (xAI):
-          • grok-4-0709, grok-4-fast-reasoning, grok-4-fast-non-reasoning
-          • grok-code-fast-1, grok-3, grok-3-mini
-          • grok-2-1212, grok-2-vision-1212, grok-2-image-1212 (Vision)
+          • grok-4.3
+          • grok-4.20-0309-reasoning, grok-4.20-0309-non-reasoning
 
         Ollama (Local):
           • llama3.3, llama3.2, llama3.1 (Recommended)
           • llava, bakllava (Vision models)
-          • codellama, mistral-nemo, qwen2.5
-          • deepseek-r1, command-r-plus
+          • mistral-nemo, qwen2.5
+          • deepseek-r1:8b, deepseek-r1:671b, command-r-plus
           • Custom: any-model:tag
 
         SHORTCUTS:
-          • claude, opus → claude-opus-4-1-20250805
-          • gpt, gpt4 → gpt-4.1
-          • grok → grok-4-fast-reasoning
-          • gemini → gemini-2.5-flash
+          • claude, opus → claude-opus-4-7
+          • gpt → gpt-5.5
+          • grok → grok-4.3
+          • gemini → gemini-3.1-pro-preview
           • llama, llama3 → llama3.3
 
         API KEYS:
@@ -376,6 +374,8 @@ struct AICLI {
         case .openai: .openai
         case .anthropic: .anthropic
         case .google: .google
+        case .minimax: .minimax
+        case .minimaxCN: .minimaxCN
         case .mistral: .mistral
         case .groq: .groq
         case .grok: .grok
@@ -406,6 +406,14 @@ struct AICLI {
             print("# Legacy names still supported:")
             print("export GOOGLE_API_KEY='gk-your-key-here'")
             print("Get your key at: https://aistudio.google.com/apikey")
+        case .minimax:
+            print("Set your MiniMax API key:")
+            print("export MINIMAX_API_KEY='your-key-here'")
+        case .minimaxCN:
+            print("Set your MiniMax China API key:")
+            print("export MINIMAX_CN_API_KEY='your-key-here'")
+            print("# or reuse the global MiniMax key name:")
+            print("export MINIMAX_API_KEY='your-key-here'")
         case .mistral:
             print("Set your Mistral API key:")
             print("export MISTRAL_API_KEY='your-key-here'")
@@ -473,7 +481,7 @@ struct AICLI {
 
         let supportsThinking = self.isReasoningModel(model) && actualApiMode != .chat
         if config.showThinking, !supportsThinking {
-            print("⚠️  Note: --thinking only works with O3, O4, and GPT-5 models via Responses API")
+            print("⚠️  Note: --thinking only works with GPT-5 models via Responses API")
         }
 
         if case let .openai(openaiModel) = model, actualApiMode == .chat {
@@ -585,15 +593,15 @@ struct AICLI {
     static func isReasoningModel(_ model: LanguageModel) -> Bool {
         guard case let .openai(openaiModel) = model else { return false }
         switch openaiModel {
-        case .o4Mini,
-             .gpt5,
+        case .gpt55,
+             .gpt54,
+             .gpt54Mini,
+             .gpt54Nano,
              .gpt5Pro,
-             .gpt5Mini,
-             .gpt5Nano,
-             .gpt5Thinking,
-             .gpt5ThinkingMini,
-             .gpt5ThinkingNano,
-             .gpt5ChatLatest:
+             .gpt5Mini:
+            return true
+        case .gpt5,
+             .gpt5Nano:
             return true
         default:
             return false
@@ -672,7 +680,7 @@ struct AICLI {
                     }
                 }
 
-                // If no summary, try content array (for O3/O4)
+                // If no summary, try content array.
                 if reasoningText == nil || reasoningText?.isEmpty == true {
                     if let contentArray = output["content"] as? [[String: Any]] {
                         let reasoningParts = contentArray.compactMap { item -> String? in
@@ -801,23 +809,19 @@ struct AICLI {
         switch model {
         case let .openai(openaiModel):
             switch openaiModel {
-            case .gpt5: return nil // Pricing TBD
-            case .gpt5Mini: return nil // Pricing TBD
-            case .gpt5Nano: return nil // Pricing TBD
-            case .gpt4o:
+            case .gpt55, .gpt54, .gpt5:
                 inputCostPer1k = 0.005
-                outputCostPer1k = 0.015
-            case .gpt4oMini:
-                inputCostPer1k = 0.000_15
-                outputCostPer1k = 0.0006
+                outputCostPer1k = 0.020
+            case .gpt54Mini, .gpt54Nano, .gpt5Mini, .gpt5Nano:
+                return nil // Pricing TBD
             default: return nil
             }
         case let .anthropic(anthropicModel):
             switch anthropicModel {
-            case .opus4, .opus4Thinking:
+            case .opus47, .opus45, .opus4:
                 inputCostPer1k = 0.015
                 outputCostPer1k = 0.075
-            case .sonnet4, .sonnet4Thinking:
+            case .sonnet46, .sonnet45:
                 inputCostPer1k = 0.003
                 outputCostPer1k = 0.015
             case .haiku45:
@@ -825,6 +829,9 @@ struct AICLI {
                 outputCostPer1k = 0.004
             default: return nil
             }
+        case .minimax:
+            inputCostPer1k = 0.0003
+            outputCostPer1k = 0.0012
         default:
             return nil
         }

Examples/Advanced/TachikomaExamples.swift

@@ -56,18 +56,18 @@ extension TachikomaExamples {
         print("Creating provider-specific models with type safety:")
 
         // OpenAI models
-        let gpt4o = Model.openai(.gpt4o)
-        let gpt41 = Model.openai(.gpt41)
+        let gpt55 = Model.openai(.gpt55)
+        let gpt54 = Model.openai(.gpt54)
         let gpt5Mini = Model.openai(.gpt5Mini)
 
         // Anthropic models
         let opus4 = Model.anthropic(.opus4)
-        let sonnet4 = Model.anthropic(.sonnet4)
+        let sonnet46 = Model.anthropic(.sonnet46)
         let haiku45 = Model.anthropic(.haiku45)
 
         // Grok models
-        let grok4 = Model.grok(.grok4FastReasoning)
-        let grok2Vision = Model.grok(.grok2Vision)
+        let grok43 = Model.grok(.grok43)
+        let grokReasoning = Model.grok(.grok420Reasoning)
 
         // Ollama models
         let llama33 = Model.ollama(.llama3_3)
@@ -75,17 +75,17 @@ extension TachikomaExamples {
 
         // Custom endpoints
         let openRouter = Model.openRouter(modelId: "anthropic/claude-sonnet-4.5")
-        let customAPI = Model.openaiCompatible(modelId: "gpt-4", baseURL: "https://api.azure.com")
+        let customAPI = Model.openaiCompatible(modelId: "gpt-5.5", baseURL: "https://api.azure.com")
 
         let models = [
-            gpt4o,
-            gpt41,
+            gpt55,
+            gpt54,
             gpt5Mini,
             opus4,
-            sonnet4,
+            sonnet46,
             haiku45,
-            grok4,
-            grok2Vision,
+            grok43,
+            grokReasoning,
             llama33,
             llava,
             openRouter,
@@ -100,9 +100,9 @@ extension TachikomaExamples {
 
         // Model capabilities
         print("\nModel capabilities:")
-        print("  • Vision support: \(gpt4o.supportsVision)")
+        print("  • Vision support: \(gpt55.supportsVision)")
         print("  • Tool support: \(opus4.supportsTools)")
-        print("  • Streaming support: \(sonnet4.supportsStreaming)")
+        print("  • Streaming support: \(sonnet46.supportsStreaming)")
     }
 }
 
@@ -192,7 +192,7 @@ extension TachikomaExamples {
         // Record some usage
         try? await tracker.recordUsage(
             operation: .generation,
-            model: "gpt-4o".lowercased(),
+            model: "gpt-5.5".lowercased(),
             inputTokens: 100,
             outputTokens: 50,
             cost: 0.003,
@@ -273,9 +273,9 @@ extension TachikomaExamples {
 
         // Test different provider creations (these will fail without API keys, which is expected)
         let models: [(String, Model)] = [
-            ("OpenAI GPT-4o", .openai(.gpt4o)),
+            ("OpenAI GPT-5.5", .openai(.gpt55)),
             ("Anthropic Opus 4", .anthropic(.opus4)),
-            ("Grok 4 Fast", .grok(.grok4FastReasoning)),
+            ("Grok 4.3", .grok(.grok43)),
             ("Ollama Llama 3.3", .ollama(.llama3_3)),
         ]
 
@@ -319,7 +319,7 @@ extension String {
 extension Model {
     var supportsVision: Bool {
         switch self {
-        case .openai(.gpt4o), .grok(.grok2Vision), .ollama(.llava):
+        case .openai(.gpt55), .ollama(.llava):
             true
         default:
             false

Examples/Agent-CLI/README.md

@@ -7,7 +7,7 @@ An advanced AI agent command-line interface with conversation support, MCP tool
 - **Multi-turn Conversations**: Maintain context across multiple interactions
 - **MCP Tool Support**: Connect to Model Context Protocol servers for extended capabilities
 - **Status Bar UI**: Real-time status updates with spinners and progress indicators
-- **Thinking Display**: Show reasoning process for models that support it (O3, GPT-5)
+- **Thinking Display**: Show reasoning process for models that support it (GPT-5)
 - **Interactive Mode**: Continuous conversation with history management
 - **Multiple Output Formats**: Normal, JSON, or Markdown formatting
 - **Tool Visualization**: See tool calls with timing and results
@@ -50,7 +50,7 @@ In interactive mode:
 ```bash
 agent-cli --model gpt-5 "Explain quantum computing"
 agent-cli --model claude "Write a haiku"
-agent-cli --model o3 --thinking "Solve this step by step: ..."
+agent-cli --model gpt-5-thinking --thinking "Solve this step by step: ..."
 ```
 
 ### With MCP Tools
@@ -171,13 +171,10 @@ agent-cli --mcp-server "db -- npx @modelcontextprotocol/server-postgres postgres
 ## Supported Models
 
 ### OpenAI
-- GPT-5 series: `gpt-5`, `gpt-5-mini`, `gpt-5-nano`
-- O-series: `o3`, `o3-mini`, `o3-pro`, `o4-mini`
-- GPT-4: `gpt-4.1`, `gpt-4o`, `gpt-4-turbo`
+- GPT-5 series: `gpt-5.5`, `gpt-5.4`, `gpt-5`, `gpt-5-mini`, `gpt-5-nano`
 
 ### Anthropic
-- Claude 4: `opus-4`, `sonnet-4`
-- Claude 3.5: `claude`, `sonnet`, `haiku`
+- Claude 4.x: `claude-opus-4.7`, `claude-sonnet-4.6`, `claude-haiku-4.5`
 
 ### Others
 - Google: `gemini-2.5-pro`, `gemini-2.5-flash`, `gemini-2.5-flash-lite`
@@ -209,7 +206,7 @@ agent-cli --mcp-server "db -- npx @modelcontextprotocol/server-sqlite ./app.db"
 
 ### Complex Reasoning
 ```bash
-agent-cli --model o3 --thinking \
+agent-cli --model gpt-5-thinking --thinking \
           "Plan a distributed system architecture for a social media platform"
 ```
 

Examples/Agent-CLI/Sources/Agent-CLI/Agent.swift

@@ -163,22 +163,21 @@ final class Agent {
         switch model {
         case let .openai(openaiModel):
             switch openaiModel {
-            case .o4Mini,
-                 .gpt5,
+            case .gpt5,
                  .gpt5Pro,
                  .gpt5Mini,
                  .gpt5Nano,
-                 .gpt5Thinking,
-                 .gpt5ThinkingMini,
-                 .gpt5ThinkingNano,
-                 .gpt5ChatLatest:
+                 .gpt55,
+                 .gpt54,
+                 .gpt54Mini,
+                 .gpt54Nano:
                 true
             default:
                 false
             }
         case let .anthropic(anthropicModel):
             switch anthropicModel {
-            case .opus4Thinking, .sonnet4Thinking:
+            case .opus47, .opus45, .opus4, .sonnet46, .sonnet45:
                 true
             default:
                 false

Examples/Agent-CLI/Sources/Agent-CLI/AgentCLI.swift

@@ -27,7 +27,7 @@ struct AgentCLI: AsyncParsableCommand {
     @Argument(help: "Query or task for the agent")
     var query: String?
 
-    @Option(name: .shortAndLong, help: "AI model to use (e.g., gpt-5, claude, o3)")
+    @Option(name: .shortAndLong, help: "AI model to use (e.g., gpt-5.5, gpt-5-thinking, claude)")
     var model: String = "gpt-5"
 
     @Flag(name: .shortAndLong, help: "Interactive conversation mode")
@@ -500,10 +500,13 @@ struct AgentCLI: AsyncParsableCommand {
         case .openai: .openai
         case .anthropic: .anthropic
         case .google: .google
+        case .minimax: .minimax
+        case .minimaxCN: .minimaxCN
         case .mistral: .mistral
         case .groq: .groq
         case .grok: .grok
         case .ollama: .ollama
+        case .lmstudio: .lmstudio
         default: .custom("unknown")
         }
 
@@ -548,7 +551,7 @@ enum CLIError: LocalizedError {
 /// Extension to add provider helpers
 extension Provider {
     static var allStandard: [Provider] {
-        [.openai, .anthropic, .google, .mistral, .groq, .grok, .ollama]
+        [.openai, .anthropic, .google, .minimax, .minimaxCN, .mistral, .groq, .grok, .ollama]
     }
 
     var displayName: String {
@@ -556,6 +559,8 @@ extension Provider {
         case .openai: "OpenAI"
         case .anthropic: "Anthropic"
         case .google: "Google"
+        case .minimax: "MiniMax"
+        case .minimaxCN: "MiniMax China"
         case .mistral: "Mistral"
         case .groq: "Groq"
         case .grok: "Grok"

Examples/Demos/TachikomaDemo.sh

@@ -49,7 +49,7 @@ echo "  • Stream token handling"
 echo ""
 
 echo "🔑 API Integration Examples (require valid API keys):"
-echo "  • OpenAI GPT-4o, GPT-4.1, o3 generation"
+echo "  • OpenAI GPT-5 generation"
 echo "  • Anthropic Claude Opus 4, Sonnet 4 generation"
 echo "  • Grok 4 and Grok 2 Vision models"
 echo "  • Ollama local models (llama3.3, llava)"
@@ -70,7 +70,7 @@ echo "🚀 How to Use Tachikoma:"
 echo "========================"
 echo ""
 echo "1. Basic Generation:"
-echo '   let answer = try await generate("What is 2+2?", using: .openai(.gpt4o))'
+echo '   let answer = try await generate("What is 2+2?", using: .openai(.gpt55))'
 echo ""
 echo "2. With Tools:"
 echo '   @ToolKit'
@@ -88,4 +88,4 @@ echo '   let response = try await conversation.continue(using: .anthropic(.opus4
 echo ""
 
 echo "🕷️  Tachikoma - Intelligent • Adaptable • Reliable"
-echo "   All examples completed successfully!"
+echo "   All examples completed successfully!"

Examples/HarmonyFeatures.swift

@@ -9,7 +9,7 @@ func demonstrateMultiChannelResponse() async throws {
     print("=== Multi-Channel Response Demo ===\n")
 
     let result = try await generateText(
-        model: .openai(.gpt4o),
+        model: .openai(.gpt55),
         messages: [
             .user("Explain how recursion works in programming"),
         ],
@@ -36,7 +36,7 @@ func demonstrateReasoningEffort() async throws {
     // High effort for complex problem
     print("High effort response:")
     let complexResult = try await generateText(
-        model: .openai(.gpt4o),
+        model: .openai(.gpt55),
         messages: [
             .user("Design a distributed system for real-time collaboration"),
         ],
@@ -50,7 +50,7 @@ func demonstrateReasoningEffort() async throws {
     // Low effort for simple query
     print("\nLow effort response:")
     let simpleResult = try await generateText(
-        model: .openai(.gpt4o),
+        model: .openai(.gpt55),
         messages: [
             .user("What is the capital of Japan?"),
         ],
@@ -82,7 +82,7 @@ func demonstrateRetryHandler() async throws {
                 print("Attempting API call...")
                 // Simulate a call that might fail
                 return try await generateText(
-                    model: .openai(.gpt4o),
+                    model: .openai(.gpt55),
                     messages: [.user("Hello")],
                 )
             },
@@ -144,7 +144,7 @@ func demonstrateEnhancedTools() async throws {
     }
 
     let result = try await generateText(
-        model: .openai(.gpt4o),
+        model: .openai(.gpt55),
         messages: [
             .user("What's 25 * 4 and what's the weather in Tokyo?"),
         ],
@@ -278,7 +278,7 @@ func demonstrateIntegratedFeatures() async throws {
     do {
         let result = try await retryHandler.execute {
             try await generateText(
-                model: .openai(.gpt4o),
+                model: .openai(.gpt55),
                 messages: [
                     .system("You are a helpful assistant. Use channels to organize your response."),
                     .user("Analyze the benefits of functional programming"),

Examples/RealtimeAPIDemo.swift

@@ -16,7 +16,7 @@ public struct RealtimeAPIDemo {
         // 1. Create configuration
         print("\n1️⃣ Creating Configuration...")
         let config = SessionConfiguration(
-            model: "gpt-4o-realtime-preview",
+            model: "gpt-realtime",
             voice: .nova,
             instructions: "You are a helpful voice assistant",
             inputAudioFormat: .pcm16,
@@ -108,7 +108,7 @@ public struct RealtimeAPIDemo {
         Example:
         let conversation = try RealtimeConversation(configuration: config)
         try await conversation.start(
-            model: .gpt4oRealtime,
+            model: .custom("gpt-realtime"),
             voice: .nova
         )
         ════════════════════════════════════════════════
@@ -118,7 +118,7 @@ public struct RealtimeAPIDemo {
     /// Create a sample configuration for testing
     public static func createSampleConfiguration() -> SessionConfiguration {
         SessionConfiguration.voiceConversation(
-            model: "gpt-4o-realtime-preview",
+            model: "gpt-realtime",
             voice: .nova,
         )
     }
@@ -127,7 +127,7 @@ public struct RealtimeAPIDemo {
     public static func validateTypes() -> Bool {
         // Test configuration creation
         let config = self.createSampleConfiguration()
-        guard config.model == "gpt-4o-realtime-preview" else { return false }
+        guard config.model == "gpt-realtime" else { return false }
 
         // Test VAD configuration
         let vad = RealtimeTurnDetection.serverVAD

Examples/RealtimeExample.swift

@@ -49,7 +49,7 @@ class RealtimeVoiceAssistant {
 
         // Start the conversation
         self.conversation = try await startRealtimeConversation(
-            model: .gpt4oRealtime,
+            model: .custom("gpt-realtime"),
             voice: .nova,
             instructions: """
             You are a helpful voice assistant. Keep responses concise and natural.
@@ -151,7 +151,7 @@ class RealtimeDemo {
 
         // Server VAD automatically detects when user starts/stops speaking
         let conversation = try await startRealtimeConversation(
-            model: .gpt4oRealtime,
+            model: .custom("gpt-realtime"),
             voice: .alloy,
             instructions: "You are a voice assistant with server-side voice activity detection",
         )
@@ -171,7 +171,7 @@ class RealtimeDemo {
 
         // Configure for both text and audio responses
         let conversation = try await startRealtimeConversation(
-            model: .gpt4oRealtime,
+            model: .custom("gpt-realtime"),
             voice: .shimmer,
         )
 
@@ -222,7 +222,7 @@ class RealtimeDemo {
         }
 
         let conversation = try await startRealtimeConversation(
-            model: .gpt4oRealtime,
+            model: .custom("gpt-realtime"),
             voice: .echo,
             instructions: "You are a smart home assistant. Use the available tools to control devices.",
             tools: [smartHomeTool],
@@ -315,7 +315,7 @@ struct RealtimeVoiceView: View {
     private func setupConversation() async {
         do {
             conversation = try await startRealtimeConversation(
-                model: .gpt4oRealtime,
+                model: .custom("gpt-realtime"),
                 voice: .nova,
             )
 

Examples/RealtimeQuickTest.swift

@@ -13,7 +13,7 @@ func testRealtimeConfiguration() async throws {
     // Test 1: Session Configuration
     print("\n1️⃣ Testing Session Configuration:")
     let voiceConfig = SessionConfiguration.voiceConversation(
-        model: "gpt-4o-realtime-preview",
+        model: "gpt-realtime",
         voice: .nova,
     )
     print("   ✅ Model: \(voiceConfig.model)")

Examples/RealtimeUsageExamples.swift

@@ -24,7 +24,7 @@ class BasicVoiceAssistant: ObservableObject {
 
         // Start conversation with voice
         try await self.conversation?.start(
-            model: .gpt4oRealtime,
+            model: .custom("gpt-realtime"),
             voice: .nova,
             instructions: "You are a helpful voice assistant. Keep responses concise.",
         )
@@ -307,7 +307,7 @@ class AudioStreamingExample {
     func setupAudioStreaming(apiKey: String) async throws {
         // Configure for audio streaming
         var config = EnhancedSessionConfiguration(
-            model: "gpt-4o-realtime-preview",
+            model: "gpt-realtime",
             voice: .echo,
             inputAudioFormat: .pcm16,
             outputAudioFormat: .pcm16,
@@ -371,7 +371,7 @@ class MultiTurnConversation {
     func runConversation(apiKey: String) async throws {
         // Configure for multi-turn dialogue
         let config = EnhancedSessionConfiguration(
-            model: "gpt-4o-realtime-preview",
+            model: "gpt-realtime",
             voice: .fable,
             instructions: """
             You are a knowledgeable assistant engaged in a multi-turn conversation.
@@ -547,7 +547,7 @@ class VoiceAssistantViewController: UIViewController {
                 self.conversation = try RealtimeConversation(configuration: config)
 
                 try await self.conversation?.start(
-                    model: .gpt4oRealtime,
+                    model: .custom("gpt-realtime"),
                     voice: .shimmer,
                 )
 

Examples/RealtimeVoiceAssistant.swift

@@ -29,7 +29,7 @@ class RealtimeVoiceAssistant {
 
         // Start with voice configuration
         try await conversation.start(
-            model: .gpt4oRealtime,
+            model: .custom("gpt-realtime"),
             voice: .nova,
             instructions: "You are a helpful, witty, and friendly AI assistant. Keep responses concise.",
         )
@@ -77,7 +77,7 @@ class RealtimeVoiceAssistant {
 
         // Advanced configuration with all features
         let config = SessionConfiguration(
-            model: "gpt-4o-realtime-preview",
+            model: "gpt-realtime",
             voice: .nova,
             instructions: """
             You are an expert AI assistant with deep knowledge across many domains.
@@ -147,7 +147,7 @@ class RealtimeVoiceAssistant {
 
         // Configuration with tools
         let config = SessionConfiguration.withTools(
-            model: "gpt-4o-realtime-preview",
+            model: "gpt-realtime",
             voice: .nova,
             tools: [
                 // Weather tool

Package.resolved

@@ -1,6 +1,15 @@
 {
-  "originHash" : "243e2c6245528e2e7d881e19f057c59ff495268ec997e5ed087ee704fa9a80d8",
+  "originHash" : "12a454cd38a6ae2519d652cc0872f7f18feb64690ce83d1507bae6db71c1841c",
   "pins" : [
+    {
+      "identity" : "commander",
+      "kind" : "remoteSourceControl",
+      "location" : "https://github.com/steipete/Commander.git",
+      "state" : {
+        "revision" : "ae2ce746b386ff94b26648cfe5625cfa8d02639b",
+        "version" : "0.2.2"
+      }
+    },
     {
       "identity" : "eventsource",
       "kind" : "remoteSourceControl",
@@ -24,8 +33,8 @@
       "kind" : "remoteSourceControl",
       "location" : "https://github.com/apple/swift-asn1.git",
       "state" : {
-        "revision" : "eb50cbd14606a9161cbc5d452f18797c90ef0bab",
-        "version" : "1.7.0"
+        "revision" : "a9a5efd40eaf558a2bcd48d64b1d1646be686008",
+        "version" : "1.7.1"
       }
     },
     {
@@ -33,8 +42,8 @@
       "kind" : "remoteSourceControl",
       "location" : "https://github.com/apple/swift-async-algorithms.git",
       "state" : {
-        "revision" : "9d349bcc328ac3c31ce40e746b5882742a0d1272",
-        "version" : "1.1.3"
+        "revision" : "d0b4a06d0f173a2f3be27d3ea21b3c3aa18db440",
+        "version" : "1.1.4"
       }
     },
     {
@@ -51,8 +60,8 @@
       "kind" : "remoteSourceControl",
       "location" : "https://github.com/apple/swift-collections",
       "state" : {
-        "revision" : "6675bc0ff86e61436e615df6fc5174e043e57924",
-        "version" : "1.4.1"
+        "revision" : "a0cb0954ecb21e4e31b0070e6ed5674e8556685a",
+        "version" : "1.6.0"
       }
     },
     {
@@ -78,8 +87,8 @@
       "kind" : "remoteSourceControl",
       "location" : "https://github.com/apple/swift-log.git",
       "state" : {
-        "revision" : "5073617dac96330a486245e4c0179cb0a6fd2256",
-        "version" : "1.12.0"
+        "revision" : "92448c359f00ebe36ae97d3bd9086f13c7692b5a",
+        "version" : "1.13.2"
       }
     },
     {
@@ -87,8 +96,8 @@
       "kind" : "remoteSourceControl",
       "location" : "https://github.com/apple/swift-nio.git",
       "state" : {
-        "revision" : "f71c8d2a5e74a2c6d11a0fbe324774b5d6084237",
-        "version" : "2.99.0"
+        "revision" : "a8e036cb8628fcc1ff67dfec6ce8168617172c9b",
+        "version" : "2.101.1"
       }
     },
     {
@@ -105,8 +114,8 @@
       "kind" : "remoteSourceControl",
       "location" : "https://github.com/modelcontextprotocol/swift-sdk.git",
       "state" : {
-        "revision" : "6132fd4b5b4217ce4717c4775e4607f5c3120129",
-        "version" : "0.12.0"
+        "revision" : "a0ae212ebf6eab5f754c3129608bc5557637e605",
+        "version" : "0.12.1"
       }
     },
     {
@@ -123,8 +132,8 @@
       "kind" : "remoteSourceControl",
       "location" : "https://github.com/apple/swift-system.git",
       "state" : {
-        "revision" : "7c6ad0fc39d0763e0b699210e4124afd5041c5df",
-        "version" : "1.6.4"
+        "revision" : "7502b711c92a17741fa625d722b0ccbd595d8ed1",
+        "version" : "1.7.2"
       }
     }
   ],

Package.swift

@@ -1,8 +1,18 @@
 // swift-tools-version: 6.2
 // The swift-tools-version declares the minimum version of Swift required to build this package.
 
+import Foundation
 import PackageDescription
 
+let packageDirectory = URL(fileURLWithPath: #filePath).deletingLastPathComponent()
+let localCommanderPath = packageDirectory
+    .deletingLastPathComponent()
+    .appendingPathComponent("Commander")
+    .appendingPathComponent("Package.swift")
+let commanderDependency: Package.Dependency = FileManager.default.fileExists(atPath: localCommanderPath.path)
+    ? .package(path: "../Commander")
+    : .package(url: "https://github.com/steipete/Commander.git", from: "0.2.2")
+
 let package = Package(
     name: "Tachikoma",
     platforms: [
@@ -49,7 +59,7 @@ let package = Package(
             targets: ["TachikomaConfigCLI"]),
     ],
     dependencies: [
-        .package(path: "../Commander"),
+        commanderDependency,
         .package(url: "https://github.com/apple/swift-log.git", from: "1.6.4"),
         .package(url: "https://github.com/modelcontextprotocol/swift-sdk.git", from: "0.12.0"),
         .package(url: "https://github.com/apple/swift-configuration", from: "1.0.0"),

README.md

@@ -10,6 +10,8 @@
   <a href="https://github.com/steipete/Tachikoma/actions/workflows/ci.yml"><img src="https://img.shields.io/github/actions/workflow/status/steipete/Tachikoma/ci.yml?branch=main&style=for-the-badge&label=tests" alt="CI Status"></a>
 </p>
 
+![Tachikoma banner](docs/assets/readme-banner.jpg)
+
 Modern, Swift-native APIs for text, vision, tools, and realtime voice.
 </div>
 
@@ -39,7 +41,7 @@ print(text)
 ```swift
 import Tachikoma
 
-let stream = try await stream("Explain actors in Swift.", using: .openai(.gpt52))
+let stream = try await stream("Explain actors in Swift.", using: .openai(.gpt54))
 for try await delta in stream {
     print(delta.content ?? "", terminator: "")
 }
@@ -64,7 +66,7 @@ import Tachikoma
 
 let pngData: Data = /* ... */
 let image = ImageInput(data: pngData, mimeType: "image/png")
-let answer = try await analyze(image: image, prompt: "What’s in this image?", using: .openai(.gpt4o))
+let answer = try await analyze(image: image, prompt: "What’s in this image?", using: .openai(.gpt55))
 print(answer)
 ```
 
@@ -87,7 +89,7 @@ let tool = createTool(
 }
 
 let result = try await generateText(
-    model: .openai(.gpt52),
+    model: .openai(.gpt54),
     messages: [.user("Compute 123 + 456 using the add tool.")],
     tools: [tool],
     maxSteps: 3
@@ -98,10 +100,10 @@ print(result.text)
 ## Models
 
 Common picks:
-- Anthropic: `claude-opus-4-5` (`LanguageModel.default`)
-- OpenAI: `gpt-5.2` (flagship), `gpt-5` (coding/agents), `o4-mini` (reasoning), `gpt-4o` (vision)
-- Google: `gemini-3-flash`
-- Grok: `grok-4-fast-reasoning`
+- Anthropic: `claude-opus-4-8` (`LanguageModel.default`, non-streaming), `claude-fable-5` (explicit opt-in)
+- OpenAI: `gpt-5.5` (`LanguageModel.defaultStreaming`), `gpt-5.4` / `gpt-5.4-mini` / `gpt-5.4-nano`, `gpt-5`
+- Google: `gemini-3.1-pro-preview`, `gemini-3-flash`
+- Grok: `grok-4.3`
 - Local: `ollama/llama3.3`
 
 Full catalog (including enum case names + provider notes): [`docs/models.md`](docs/models.md).

Sources/Tachikoma/Auth/AuthManager.swift

@@ -25,6 +25,7 @@ public enum TKProviderId: String, CaseIterable, Sendable {
     case anthropic
     case grok
     case gemini
+    case openrouter
 
     public static func normalize(_ value: String) -> TKProviderId? {
         let lower = value.lowercased()
@@ -38,6 +39,7 @@ public enum TKProviderId: String, CaseIterable, Sendable {
         case .anthropic: "Anthropic"
         case .grok: "Grok (xAI)"
         case .gemini: "Gemini"
+        case .openrouter: "OpenRouter"
         }
     }
 
@@ -55,6 +57,7 @@ public enum TKProviderId: String, CaseIterable, Sendable {
         // swiftformat:enable indent
         case .grok: ["X_AI_API_KEY", "XAI_API_KEY", "GROK_API_KEY"]
         case .gemini: ["GEMINI_API_KEY"]
+        case .openrouter: ["OPENROUTER_API_KEY"]
         }
     }
 
@@ -78,8 +81,7 @@ public struct TKCredentialStore {
     public init() {}
 
     private var baseDir: String {
-        let dir = TachikomaConfiguration.profileDirectoryName
-        return NSString(string: "~/" + dir).expandingTildeInPath
+        TachikomaConfiguration.profileDirectoryPath
     }
 
     private var credentialsPath: String {
@@ -137,8 +139,8 @@ public final class TKAuthManager {
 
     private init() {}
 
-    private func environmentValue(for key: String) -> String? {
-        guard !self.ignoreEnv else { return nil }
+    private func environmentValue(for key: String, ignoringEnvironment: Bool) -> String? {
+        guard !ignoringEnvironment else { return nil }
         let value = key.withCString { keyPtr -> String? in
             guard let cValue = getenv(keyPtr) else { return nil }
             let string = String(cString: cValue)
@@ -149,6 +151,14 @@ public final class TKAuthManager {
         return nil
     }
 
+    private func authState() -> (ignoreEnv: Bool, creds: [String: String]) {
+        self.lock.lock()
+        let ignoreEnv = self.ignoreEnv
+        let creds = self.ignoreStore ? [:] : self.store.load()
+        self.lock.unlock()
+        return (ignoreEnv, creds)
+    }
+
     @discardableResult
     public func setIgnoreEnvironment(_ value: Bool) -> Bool {
         self.lock.lock()
@@ -168,20 +178,17 @@ public final class TKAuthManager {
     }
 
     public func credentialValue(for key: String) -> String? {
-        self.lock.lock()
-        let creds = self.ignoreStore ? [:] : self.store.load()
-        self.lock.unlock()
-        if let env = self.environmentValue(for: key) { return env }
-        return creds[key]
+        let state = self.authState()
+        if let env = self.environmentValue(for: key, ignoringEnvironment: state.ignoreEnv) { return env }
+        return state.creds[key]
     }
 
     public func resolveAuth(for provider: TKProviderId) -> TKAuthValue? {
-        self.lock.lock()
-        let creds = self.ignoreStore ? [:] : self.store.load()
-        self.lock.unlock()
+        let state = self.authState()
+        let creds = state.creds
         switch provider {
         case .openai:
-            if let env = self.environmentValue(for: "OPENAI_API_KEY") {
+            if let env = self.environmentValue(for: "OPENAI_API_KEY", ignoringEnvironment: state.ignoreEnv) {
                 return .bearer(env, betaHeader: nil)
             }
             if let access = creds["OPENAI_ACCESS_TOKEN"], !access.isEmpty {
@@ -191,7 +198,7 @@ public final class TKAuthManager {
                 return .apiKey(key)
             }
         case .anthropic:
-            if let env = self.environmentValue(for: "ANTHROPIC_API_KEY") {
+            if let env = self.environmentValue(for: "ANTHROPIC_API_KEY", ignoringEnvironment: state.ignoreEnv) {
                 return .apiKey(env)
             }
             if let access = creds["ANTHROPIC_ACCESS_TOKEN"], !access.isEmpty {
@@ -204,7 +211,7 @@ public final class TKAuthManager {
         case .grok:
             let envOrder = ["X_AI_API_KEY", "XAI_API_KEY", "GROK_API_KEY"]
             for k in envOrder {
-                if let env = self.environmentValue(for: k) {
+                if let env = self.environmentValue(for: k, ignoringEnvironment: state.ignoreEnv) {
                     return .bearer(env, betaHeader: nil)
                 }
             }
@@ -212,10 +219,17 @@ public final class TKAuthManager {
                 if let val = creds[k], !val.isEmpty { return .bearer(val, betaHeader: nil) }
             }
         case .gemini:
-            if let env = self.environmentValue(for: "GEMINI_API_KEY") {
+            if let env = self.environmentValue(for: "GEMINI_API_KEY", ignoringEnvironment: state.ignoreEnv) {
                 return .apiKey(env)
             }
             if let val = creds["GEMINI_API_KEY"], !val.isEmpty { return .apiKey(val) }
+        case .openrouter:
+            if let env = self.environmentValue(for: "OPENROUTER_API_KEY", ignoringEnvironment: state.ignoreEnv) {
+                return .bearer(env, betaHeader: nil)
+            }
+            if let val = creds["OPENROUTER_API_KEY"], !val.isEmpty {
+                return .bearer(val, betaHeader: nil)
+            }
         }
         return nil
     }
@@ -311,7 +325,7 @@ public final class TKAuthManager {
                 requiresStateInTokenExchange: true,
                 pkce: pkce,
             )
-        case .grok, .gemini:
+        case .grok, .gemini, .openrouter:
             OAuthConfig(
                 prefix: "",
                 authorize: "",
@@ -575,7 +589,7 @@ struct TKProviderValidator {
             request.setValue("application/json", forHTTPHeaderField: "Content-Type")
             request.setValue("2023-06-01", forHTTPHeaderField: "anthropic-version")
             request.httpBody = try? JSONSerialization.data(withJSONObject: [
-                "model": "claude-3-haiku-20241022",
+                "model": "claude-haiku-4-5",
                 "max_tokens": 1,
                 "messages": [
                     ["role": "user", "content": "ping"],
@@ -594,6 +608,13 @@ struct TKProviderValidator {
             var request = URLRequest(url: URL(string: url)!)
             request.httpMethod = "GET"
             return await HTTP.perform(request: request, timeoutSeconds: self.timeoutSeconds)
+        case .openrouter:
+            return await self.validateBearer(
+                url: "https://openrouter.ai/api/v1/key",
+                secret: secret,
+                header: "Authorization",
+                valuePrefix: "Bearer ",
+            )
         }
     }
 

Sources/Tachikoma/Core/AnthropicMessageConversion.swift

@@ -1,18 +1,56 @@
 import Foundation
 
+struct AnthropicReasoningReplayTarget {
+    let provider: String
+    let modelId: String
+    let endpointIdentity: String?
+    let allowsLegacyUnknown: Bool
+
+    func matches(_ customData: [String: String]) -> Bool {
+        guard customData["tachikoma.reasoning.provider"] == self.provider else {
+            return false
+        }
+        guard customData["tachikoma.reasoning.model"] == self.modelId else {
+            return false
+        }
+        return customData["tachikoma.reasoning.base_url"] == self.endpointIdentity
+    }
+}
+
 enum AnthropicMessageConversion {
     static func convertMessagesToAnthropic(
         _ messages: [ModelMessage],
         thinkingEnabled: Bool,
+        reasoningTarget: AnthropicReasoningReplayTarget? = nil,
     ) throws
         -> (String?, [AnthropicMessage])
     {
         var systemMessage: String?
         var anthropicMessages: [AnthropicMessage] = []
-        var pendingSignedThinking: (text: String, signature: String, type: String)?
+        var pendingThinkingBlocks: [(text: String, signature: String?, type: String)] = []
         let thinkingSignatureKey = "anthropic.thinking.signature"
         let thinkingTypeKey = "anthropic.thinking.type"
 
+        func appendThinkingBlocks(
+            _ pendingBlocks: [(text: String, signature: String?, type: String)],
+            to content: inout [AnthropicContent],
+        ) {
+            for pending in pendingBlocks {
+                if pending.type == "redacted_thinking" {
+                    content.append(.redactedThinking(.init(
+                        type: "redacted_thinking",
+                        data: pending.text,
+                    )))
+                } else if let signature = pending.signature {
+                    content.append(.thinking(.init(
+                        type: "thinking",
+                        thinking: pending.text,
+                        signature: signature,
+                    )))
+                }
+            }
+        }
+
         for message in messages {
             switch message.role {
             case .system:
@@ -48,29 +86,32 @@ enum AnthropicMessageConversion {
                     }.joined()
                     let signature = message.metadata?.customData?[thinkingSignatureKey]
                     let type = message.metadata?.customData?[thinkingTypeKey] ?? "thinking"
-                    if let signature, !signature.isEmpty {
-                        pendingSignedThinking = (text: text, signature: signature, type: type)
+                    let customData = message.metadata?.customData ?? [:]
+                    if
+                        customData["tachikoma.reasoning.provider"] != nil ||
+                        customData["tachikoma.reasoning.model"] != nil ||
+                        customData["tachikoma.reasoning.base_url"] != nil ||
+                        customData["anthropic.thinking.model"] != nil
+                    {
+                        guard reasoningTarget?.matches(customData) == true else {
+                            continue
+                        }
+                    } else if reasoningTarget?.allowsLegacyUnknown != true {
+                        continue
+                    }
+                    if type == "redacted_thinking" {
+                        pendingThinkingBlocks.append((text: text, signature: nil, type: type))
+                    } else if let signature, !signature.isEmpty {
+                        pendingThinkingBlocks.append((text: text, signature: signature, type: type))
                     }
                     continue
                 }
 
                 var content: [AnthropicContent] = []
 
-                if thinkingEnabled, let pending = pendingSignedThinking {
-                    if pending.type == "redacted_thinking" {
-                        content.append(.redactedThinking(.init(
-                            type: "redacted_thinking",
-                            redactedThinking: pending.text,
-                            signature: pending.signature,
-                        )))
-                    } else {
-                        content.append(.thinking(.init(
-                            type: "thinking",
-                            thinking: pending.text,
-                            signature: pending.signature,
-                        )))
-                    }
-                    pendingSignedThinking = nil
+                if thinkingEnabled, !pendingThinkingBlocks.isEmpty {
+                    appendThinkingBlocks(pendingThinkingBlocks, to: &content)
+                    pendingThinkingBlocks.removeAll()
                 }
 
                 // Process each content part
@@ -139,21 +180,12 @@ enum AnthropicMessageConversion {
             }
         }
 
-        if thinkingEnabled, let pending = pendingSignedThinking {
-            let thinkingContent: AnthropicContent = if pending.type == "redacted_thinking" {
-                .redactedThinking(.init(
-                    type: "redacted_thinking",
-                    redactedThinking: pending.text,
-                    signature: pending.signature,
-                ))
-            } else {
-                .thinking(.init(
-                    type: "thinking",
-                    thinking: pending.text,
-                    signature: pending.signature,
-                ))
+        if thinkingEnabled, !pendingThinkingBlocks.isEmpty {
+            var content: [AnthropicContent] = []
+            appendThinkingBlocks(pendingThinkingBlocks, to: &content)
+            if !content.isEmpty {
+                anthropicMessages.append(AnthropicMessage(role: "assistant", content: content))
             }
-            anthropicMessages.append(AnthropicMessage(role: "assistant", content: [thinkingContent]))
         }
 
         return (systemMessage, anthropicMessages)

Sources/Tachikoma/Core/BaseProviders.swift

@@ -1,3 +1,12 @@
+#if canImport(CryptoKit)
+import CryptoKit
+
+private typealias ReasoningEndpointHasher = CryptoKit.SHA256
+#else
+import Crypto
+
+private typealias ReasoningEndpointHasher = Crypto.SHA256
+#endif
 import Foundation
 #if canImport(FoundationNetworking)
 import FoundationNetworking
@@ -16,18 +25,46 @@ public final class AnthropicProvider: ModelProvider {
     private let model: LanguageModel.Anthropic
     private let auth: TKAuthValue
     private let betaHeader: String
+    private let additionalHeaders: [String: String]
+    private let reasoningProvider: String
+    private let reasoningModelId: String
+    private let reasoningBaseURL: String?
+    private let urlSession: URLSession
 
     private static let requiredBetaFlags: [String] = [
         "interleaved-thinking-2025-05-14",
         "fine-grained-tool-streaming-2025-05-14",
     ]
-
-    public init(model: LanguageModel.Anthropic, configuration: TachikomaConfiguration) throws {
+    public init(
+        model: LanguageModel.Anthropic,
+        configuration: TachikomaConfiguration,
+        additionalHeaders: [String: String] = [:],
+        authOverride: TKAuthValue? = nil,
+        reasoningProvider: String = "anthropic",
+        reasoningModelId: String? = nil,
+        reasoningBaseURL: String? = nil,
+        urlSession: URLSession = .shared,
+    ) throws {
         self.model = model
         self.modelId = model.modelId
         self.baseURL = configuration.getBaseURL(for: .anthropic) ?? "https://api.anthropic.com"
+        self.additionalHeaders = additionalHeaders
+        self.reasoningProvider = reasoningProvider
+        self.reasoningModelId = reasoningModelId ?? model.modelId
+        self.reasoningBaseURL = ReasoningEndpointIdentity.canonical(
+            reasoningBaseURL ?? (reasoningProvider == "anthropic" ? self.baseURL : nil),
+        )
+        self.urlSession = urlSession
 
-        if let key = configuration.getAPIKey(for: .anthropic) {
+        if let authOverride {
+            self.auth = authOverride
+            switch authOverride {
+            case let .apiKey(key):
+                self.apiKey = key
+            case let .bearer(token, _):
+                self.apiKey = token
+            }
+        } else if let key = configuration.getAPIKey(for: .anthropic) {
             self.auth = .apiKey(key)
             self.apiKey = key
         } else if let auth = TKAuthManager.shared.resolveAuth(for: .anthropic) {
@@ -42,16 +79,18 @@ public final class AnthropicProvider: ModelProvider {
             throw TachikomaError.authenticationFailed("ANTHROPIC_API_KEY not found")
         }
 
-        self.betaHeader = Self.mergedBetaHeader(configuration: configuration, auth: self.auth)
+        self.betaHeader = Self.mergedBetaHeader(configuration: configuration, auth: self.auth, model: model)
 
+        let isFable = Self.isFable(model: model)
+        let supportsSafeStreaming = !Self.hasStreamingRefusalRisk(model: model)
         self.capabilities = ModelCapabilities(
             supportsVision: model.supportsVision,
             supportsTools: model.supportsTools,
-            supportsStreaming: true,
+            supportsStreaming: supportsSafeStreaming,
             supportsAudioInput: model.supportsAudioInput,
             supportsAudioOutput: model.supportsAudioOutput,
-            contextLength: model.contextLength,
-            maxOutputTokens: 4096,
+            contextLength: isFable ? 1_000_000 : model.contextLength,
+            maxOutputTokens: isFable ? 128_000 : model.maxOutputTokens,
         )
     }
 
@@ -79,6 +118,16 @@ public final class AnthropicProvider: ModelProvider {
     }
 
     private static func mergedBetaHeader(configuration: TachikomaConfiguration, auth: TKAuthValue) -> String {
+        self.mergedBetaHeader(configuration: configuration, auth: auth, model: nil)
+    }
+
+    private static func mergedBetaHeader(
+        configuration: TachikomaConfiguration,
+        auth: TKAuthValue,
+        model: LanguageModel.Anthropic?,
+    )
+        -> String
+    {
         var existing: String?
         if case let .bearer(_, betaHeader) = auth {
             existing = betaHeader
@@ -88,19 +137,88 @@ public final class AnthropicProvider: ModelProvider {
             existing = configuration.credentialValue(for: "ANTHROPIC_BETA_HEADER")
         }
 
+        if let model, Self.isFable(model: model) {
+            return existing?
+                .split(separator: ",")
+                .map { $0.trimmingCharacters(in: .whitespacesAndNewlines) }
+                .filter { !$0.isEmpty }
+                .joined(separator: ",") ?? ""
+        }
+
         return Self.mergedBetaHeader(existing: existing)
     }
 
-    private func anthropicThinking(from mode: AnthropicOptions.ThinkingMode?) -> AnthropicThinking? {
+    private func anthropicThinking(
+        from mode: AnthropicOptions.ThinkingMode?,
+        model: LanguageModel.Anthropic,
+    )
+        -> AnthropicThinking?
+    {
         guard let mode else { return nil }
         switch mode {
         case .disabled:
             return nil
+        case .adaptive:
+            if Self.isFable(model: model) { return nil }
+            guard self.usesAdaptiveThinking(model: model) else { return nil }
+            return AnthropicThinking(type: "adaptive", budgetTokens: nil)
         case let .enabled(budgetTokens):
+            if Self.isFable(model: model) {
+                return nil
+            }
+            if case .opus48 = model {
+                return AnthropicThinking(type: "adaptive", budgetTokens: nil)
+            }
+            if case .opus47 = model {
+                return AnthropicThinking(type: "adaptive", budgetTokens: nil)
+            }
             return AnthropicThinking(type: "enabled", budgetTokens: budgetTokens)
         }
     }
 
+    private func anthropicOutputConfig(
+        from mode: AnthropicOptions.ThinkingMode?,
+        settings: GenerationSettings,
+        model: LanguageModel.Anthropic,
+    )
+        -> AnthropicOutputConfig?
+    {
+        guard self.supportsEffort(model: model) else { return nil }
+        if let effort = settings.reasoningEffort?.rawValue {
+            return AnthropicOutputConfig(effort: effort)
+        }
+        if case .disabled = mode { return nil }
+
+        let effort = self.usesAdaptiveThinking(model: model) ? self.adaptiveEffort(from: mode) : nil
+        return effort.map { AnthropicOutputConfig(effort: $0) }
+    }
+
+    private func usesAdaptiveThinking(model: LanguageModel.Anthropic) -> Bool {
+        if Self.isFable(model: model) { return true }
+        if case .opus48 = model { return true }
+        if case .opus47 = model { return true }
+        if case .sonnet46 = model { return true }
+        return false
+    }
+
+    private func supportsEffort(model: LanguageModel.Anthropic) -> Bool {
+        if Self.isFable(model: model) { return true }
+        switch model {
+        case .opus48, .opus47, .opus45, .sonnet46:
+            return true
+        default:
+            return false
+        }
+    }
+
+    private func adaptiveEffort(from mode: AnthropicOptions.ThinkingMode?) -> String? {
+        guard case let .enabled(budgetTokens) = mode else { return nil }
+
+        if budgetTokens <= 4096 { return ReasoningEffort.low.rawValue }
+        if budgetTokens <= 12000 { return ReasoningEffort.medium.rawValue }
+        return ReasoningEffort.high.rawValue
+    }
+
     private func messagesEndpointURL() throws -> URL {
         guard let baseURL = self.baseURL, let url = URL(string: baseURL) else {
             throw TachikomaError.invalidConfiguration("Invalid Anthropic base URL: \(self.baseURL ?? "<nil>")")
@@ -127,16 +245,49 @@ public final class AnthropicProvider: ModelProvider {
         self.applyAuth(to: &urlRequest, secret: apiKey)
         urlRequest.setValue("application/json", forHTTPHeaderField: "Content-Type")
         urlRequest.setValue("2023-06-01", forHTTPHeaderField: "anthropic-version")
+        for (key, value) in self.additionalHeaders {
+            urlRequest.setValue(value, forHTTPHeaderField: key)
+        }
 
-        let requestedThinking = self.anthropicThinking(from: request.settings.providerOptions.anthropic?.thinking)
+        let validatedSettings = request.settings.validated(for: .anthropic(self.model))
+        if
+            Self.isFable(model: self.model),
+            case .disabled = validatedSettings.providerOptions.anthropic?.thinking
+        {
+            throw TachikomaError.invalidConfiguration(
+                "Claude Fable 5 always uses adaptive thinking; disabled thinking is not supported",
+            )
+        }
+        if Self.isFable(model: self.model), request.messages.last?.role == .assistant {
+            throw TachikomaError.invalidConfiguration(
+                "Claude Fable 5 does not support assistant prefill requests",
+            )
+        }
+        let requestedThinking = self.anthropicThinking(
+            from: validatedSettings.providerOptions.anthropic?.thinking,
+            model: self.model,
+        )
+        let outputConfig = self.anthropicOutputConfig(
+            from: validatedSettings.providerOptions.anthropic?.thinking,
+            settings: validatedSettings,
+            model: self.model,
+        )
         var thinking: AnthropicThinking?
         let systemMessage: String?
         let messages: [AnthropicMessage]
+        let preserveSignedThinking = requestedThinking != nil || self.requiresSignedThinkingReplay(model: self.model)
+        let reasoningTarget = AnthropicReasoningReplayTarget(
+            provider: self.reasoningProvider,
+            modelId: self.reasoningModelId,
+            endpointIdentity: self.reasoningBaseURL,
+            allowsLegacyUnknown: !Self.isFable(model: self.model),
+        )
         do {
             thinking = requestedThinking
             (systemMessage, messages) = try AnthropicMessageConversion.convertMessagesToAnthropic(
                 request.messages,
-                thinkingEnabled: requestedThinking != nil,
+                thinkingEnabled: preserveSignedThinking,
+                reasoningTarget: reasoningTarget,
             )
         } catch {
             // If we can't provide signed thinking blocks for a cached/history session, fall back to non-thinking mode.
@@ -145,19 +296,26 @@ public final class AnthropicProvider: ModelProvider {
                 (systemMessage, messages) = try AnthropicMessageConversion.convertMessagesToAnthropic(
                     request.messages,
                     thinkingEnabled: false,
+                    reasoningTarget: reasoningTarget,
                 )
             } else {
                 throw error
             }
         }
+        let maxTokens = validatedSettings.maxTokens ?? self.defaultMaxTokens(for: self.model)
+        if !stream, Self.requiresExtendedNonStreamingTimeout(model: self.model, maxTokens: maxTokens) {
+            urlRequest.timeoutInterval = 1800
+        }
+
         let anthropicRequest = try AnthropicMessageRequest(
             model: modelId,
-            maxTokens: request.settings.maxTokens ?? 1024,
-            temperature: request.settings.temperature,
+            maxTokens: maxTokens,
+            temperature: thinking == nil ? validatedSettings.temperature : nil,
             system: systemMessage,
             messages: messages,
             tools: request.tools?.map { try self.convertToolToAnthropic($0) },
             thinking: thinking,
+            outputConfig: outputConfig,
             stream: stream,
         )
 
@@ -187,7 +345,7 @@ public final class AnthropicProvider: ModelProvider {
             }
         }
 
-        let (data, response) = try await URLSession.shared.data(for: urlRequest)
+        let (data, response) = try await self.urlSession.data(for: urlRequest)
 
         guard let httpResponse = response as? HTTPURLResponse else {
             throw TachikomaError.networkError(NSError(domain: "Invalid response", code: 0))
@@ -219,7 +377,7 @@ public final class AnthropicProvider: ModelProvider {
             switch content {
             case let .text(textContent):
                 textContent.text
-            case .toolUse:
+            case .thinking, .redactedThinking, .toolUse:
                 nil
             }
         }.joined()
@@ -229,19 +387,63 @@ public final class AnthropicProvider: ModelProvider {
             outputTokens: anthropicResponse.usage.outputTokens,
         )
 
-        let finishReason: FinishReason? = switch anthropicResponse.stopReason {
-        case "end_turn": .stop
-        case "max_tokens": .length
-        case "tool_use": .toolCalls
-        case "stop_sequence": .stop
-        default: .other
+        let finishReason = Self.mapFinishReason(anthropicResponse.stopReason)
+        if finishReason == .contentFilter {
+            let fallbackRefusalText = if let category = anthropicResponse.stopDetails?.category {
+                "Request refused by Anthropic content filter (\(category))"
+            } else {
+                "Request refused by Anthropic content filter"
+            }
+            let refusalText = anthropicResponse.stopDetails?.explanation ?? fallbackRefusalText
+            return ProviderResponse(
+                text: refusalText,
+                usage: usage,
+                finishReason: finishReason,
+                toolCalls: nil,
+                reasoning: [],
+                assistantMessages: [],
+                isBillable: usage.outputTokens > 0,
+            )
         }
 
-        // Convert tool calls if present
-        let toolCalls = anthropicResponse.content.compactMap { content -> AgentToolCall? in
+        var reasoning: [ProviderReasoningBlock] = []
+        var toolCalls: [AgentToolCall] = []
+        var assistantMessages: [ModelMessage] = []
+
+        for content in anthropicResponse.content {
             switch content {
-            case .text:
-                return nil
+            case let .text(textContent):
+                if !textContent.text.isEmpty {
+                    assistantMessages.append(.assistant(textContent.text))
+                }
+            case let .thinking(thinking):
+                let block = ProviderReasoningBlock(
+                    text: thinking.thinking,
+                    signature: thinking.signature,
+                    type: thinking.type,
+                )
+                reasoning.append(block)
+                assistantMessages.append(ModelMessage(
+                    role: .assistant,
+                    content: [.text(thinking.thinking)],
+                    channel: .thinking,
+                    metadata: .init(customData: self.reasoningMetadata(
+                        type: thinking.type,
+                        signature: thinking.signature,
+                    )),
+                ))
+            case let .redactedThinking(thinking):
+                let block = ProviderReasoningBlock(
+                    text: thinking.data,
+                    type: thinking.type,
+                )
+                reasoning.append(block)
+                assistantMessages.append(ModelMessage(
+                    role: .assistant,
+                    content: [.text(thinking.data)],
+                    channel: .thinking,
+                    metadata: .init(customData: self.reasoningMetadata(type: thinking.type)),
+                ))
             case let .toolUse(toolUse):
                 // Convert input to AnyAgentToolValue dictionary
                 var arguments: [String: AnyAgentToolValue] = [:]
@@ -257,11 +459,13 @@ public final class AnthropicProvider: ModelProvider {
                     }
                 }
 
-                return AgentToolCall(
+                let toolCall = AgentToolCall(
                     id: toolUse.id,
                     name: toolUse.name,
                     arguments: arguments,
                 )
+                toolCalls.append(toolCall)
+                assistantMessages.append(ModelMessage(role: .assistant, content: [.toolCall(toolCall)]))
             }
         }
 
@@ -270,9 +474,61 @@ public final class AnthropicProvider: ModelProvider {
             usage: usage,
             finishReason: finishReason,
             toolCalls: toolCalls.isEmpty ? nil : toolCalls,
+            reasoning: reasoning,
+            assistantMessages: assistantMessages,
         )
     }
 
+    private func reasoningMetadata(type: String, signature: String? = nil) -> [String: String] {
+        var metadata = [
+            "anthropic.thinking.model": self.reasoningModelId,
+            "anthropic.thinking.type": type,
+            "tachikoma.reasoning.provider": self.reasoningProvider,
+            "tachikoma.reasoning.model": self.reasoningModelId,
+        ]
+        if let signature, !signature.isEmpty {
+            metadata["anthropic.thinking.signature"] = signature
+        }
+        if let reasoningBaseURL {
+            metadata["tachikoma.reasoning.base_url"] = reasoningBaseURL
+        }
+        return metadata
+    }
+
+    private func requiresSignedThinkingReplay(model: LanguageModel.Anthropic) -> Bool {
+        Self.isFable(model: model)
+    }
+
+    private func defaultMaxTokens(for model: LanguageModel.Anthropic) -> Int {
+        if Self.isFable(model: model) { return min(128_000, 16384) }
+        return 1024
+    }
+
+    private static func isFable(model: LanguageModel.Anthropic) -> Bool {
+        LanguageModel.Anthropic.isFable(modelId: model.modelId)
+    }
+
+    private static func hasStreamingRefusalRisk(model: LanguageModel.Anthropic) -> Bool {
+        LanguageModel.Anthropic.hasStreamingRefusalRisk(modelId: model.modelId)
+    }
+
+    private static func requiresExtendedNonStreamingTimeout(model: LanguageModel.Anthropic, maxTokens: Int) -> Bool {
+        self.isFable(model: model) || maxTokens >= 64000
+    }
+
+    static func mapFinishReason(_ stopReason: String?) -> FinishReason? {
+        switch stopReason {
+        case "end_turn": .stop
+        case "max_tokens": .length
+        case "tool_use": .toolCalls
+        case "stop_sequence": .stop
+        case "model_context_window_exceeded": .length
+        case "refusal": .contentFilter
+        case nil: nil
+        default: .other
+        }
+    }
+
     private func applyAuth(to request: inout URLRequest, secret: String) {
         switch self.auth {
         case .apiKey:
@@ -280,10 +536,19 @@ public final class AnthropicProvider: ModelProvider {
         case .bearer:
             request.setValue("Bearer " + secret, forHTTPHeaderField: "Authorization")
         }
-        request.setValue(self.betaHeader, forHTTPHeaderField: "anthropic-beta")
+        if !self.betaHeader.isEmpty {
+            request.setValue(self.betaHeader, forHTTPHeaderField: "anthropic-beta")
+        }
     }
 
     public func streamText(request: ProviderRequest) async throws -> AsyncThrowingStream<TextStreamDelta, Error> {
+        guard !Self.hasStreamingRefusalRisk(model: self.model) else {
+            let message = "\(self.model.modelId) streaming is disabled because Anthropic refusals require rollback-aware handling"
+            throw TachikomaError.invalidConfiguration(
+                "\(message); use generateText instead",
+            )
+        }
+
         let urlRequest = try self.makeURLRequest(for: request, stream: true)
 
         // Debug logging only when explicitly enabled
@@ -334,7 +599,7 @@ public final class AnthropicProvider: ModelProvider {
             (Data, URLResponse),
             Error,
         >) in
-            URLSession.shared.dataTask(with: urlRequest) { data, response, error in
+            self.urlSession.dataTask(with: urlRequest) { data, response, error in
                 if let error {
                     continuation.resume(throwing: error)
                 } else if let data, let response {
@@ -362,7 +627,7 @@ public final class AnthropicProvider: ModelProvider {
         let lines = String(data: data, encoding: .utf8)?.components(separatedBy: "\n") ?? []
         #else
         // macOS/iOS: Use streaming API
-        let (bytes, response) = try await URLSession.shared.bytes(for: urlRequest)
+        let (bytes, response) = try await self.urlSession.bytes(for: urlRequest)
 
         guard let httpResponse = response as? HTTPURLResponse else {
             throw TachikomaError.networkError(NSError(domain: "Invalid response", code: 0))
@@ -386,6 +651,7 @@ public final class AnthropicProvider: ModelProvider {
                 var currentReasoningSignature: String?
                 var currentReasoningType: String?
                 var reasoningSignatureEmitted = false
+                var finishReason: FinishReason?
 
                 do {
                     for try await line in bytes.lines {
@@ -419,7 +685,7 @@ public final class AnthropicProvider: ModelProvider {
                                     currentReasoningType = nil
                                     reasoningSignatureEmitted = false
                                 }
-                                continuation.yield(TextStreamDelta.done())
+                                continuation.yield(.done(finishReason: finishReason))
                                 break
                             }
 
@@ -450,6 +716,12 @@ public final class AnthropicProvider: ModelProvider {
                                             currentReasoningSignature = nil
                                             currentReasoningType = block.type
                                             reasoningSignatureEmitted = false
+                                            if block.type == "redacted_thinking", let data = block.data {
+                                                continuation.yield(TextStreamDelta.reasoning(
+                                                    data,
+                                                    type: "redacted_thinking",
+                                                ))
+                                            }
                                             continue
                                         }
                                     }
@@ -554,7 +826,9 @@ public final class AnthropicProvider: ModelProvider {
 
                                 case "message_delta":
                                     // Message-level updates (usage, etc.)
-                                    // Usage is typically included in the done event, not separately
+                                    if let stopReason = event.delta?.stopReason {
+                                        finishReason = Self.mapFinishReason(stopReason)
+                                    }
                                     continue
 
                                 case "message_stop":
@@ -574,7 +848,7 @@ public final class AnthropicProvider: ModelProvider {
                                         currentReasoningType = nil
                                         reasoningSignatureEmitted = false
                                     }
-                                    continuation.yield(TextStreamDelta.done())
+                                    continuation.yield(.done(finishReason: finishReason))
 
                                 default:
                                     // Unknown event type, skip
@@ -611,6 +885,7 @@ public final class AnthropicProvider: ModelProvider {
                 var currentReasoningSignature: String?
                 var currentReasoningType: String?
                 var reasoningSignatureEmitted = false
+                var finishReason: FinishReason?
 
                 do {
                     for line in lines {
@@ -637,7 +912,7 @@ public final class AnthropicProvider: ModelProvider {
                                         type: currentReasoningType,
                                     ))
                                 }
-                                continuation.yield(TextStreamDelta.done())
+                                continuation.yield(.done(finishReason: finishReason))
                                 break
                             }
 
@@ -656,6 +931,12 @@ public final class AnthropicProvider: ModelProvider {
                                         currentReasoningSignature = nil
                                         currentReasoningType = block.type
                                         reasoningSignatureEmitted = false
+                                        if block.type == "redacted_thinking", let data = block.data {
+                                            continuation.yield(TextStreamDelta.reasoning(
+                                                data,
+                                                type: "redacted_thinking",
+                                            ))
+                                        }
                                     }
                                 case "content_block_delta":
                                     if let delta = event.delta {
@@ -678,6 +959,10 @@ public final class AnthropicProvider: ModelProvider {
                                             accumulatedReasoning += thinking
                                         }
                                     }
+                                case "message_delta":
+                                    if let stopReason = event.delta?.stopReason {
+                                        finishReason = Self.mapFinishReason(stopReason)
+                                    }
                                 case "message_stop":
                                     if !accumulatedText.isEmpty {
                                         continuation.yield(TextStreamDelta.text(accumulatedText))
@@ -689,7 +974,7 @@ public final class AnthropicProvider: ModelProvider {
                                             type: currentReasoningType,
                                         ))
                                     }
-                                    continuation.yield(TextStreamDelta.done())
+                                    continuation.yield(.done(finishReason: finishReason))
                                 default:
                                     continue
                                 }
@@ -753,6 +1038,37 @@ public final class AnthropicProvider: ModelProvider {
     }
 }
 
+@available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
+enum ReasoningEndpointIdentity {
+    static func canonical(_ rawValue: String?) -> String? {
+        guard
+            let trimmed = rawValue?.trimmingCharacters(in: .whitespacesAndNewlines),
+            !trimmed.isEmpty,
+            var components = URLComponents(string: trimmed),
+            let scheme = components.scheme?.lowercased(),
+            let host = components.host?.lowercased() else
+        {
+            return nil
+        }
+
+        components.scheme = scheme
+        components.host = host
+        components.user = nil
+        components.password = nil
+        components.fragment = nil
+        while components.path.count > 1, components.path.hasSuffix("/") {
+            components.path.removeLast()
+        }
+
+        guard let value = components.string else { return nil }
+        guard let data = value.data(using: .utf8) else { return nil }
+        let digest = ReasoningEndpointHasher.hash(data: data)
+            .map { String(format: "%02x", $0) }
+            .joined()
+        return "sha256:\(digest)"
+    }
+}
+
 /// Provider for Ollama models
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
 public final class OllamaProvider: ModelProvider {

Sources/Tachikoma/Core/Configuration.swift

@@ -11,12 +11,59 @@ public typealias ProviderFactoryOverride = (LanguageModel, TachikomaConfiguratio
 public final class TachikomaConfiguration: @unchecked Sendable {
     // MARK: - Profile Directory (for config/credentials)
 
-    /// Name of the profile directory under the user's HOME used to store
-    /// configuration and credentials (e.g. \.tachikoma, \.peekaboo).
+    /// Name of the profile directory under the user's HOME, or an absolute path,
+    /// used to store configuration and credentials (e.g. \.tachikoma, \.peekaboo).
     /// Defaults to ".tachikoma". Host applications (like Peekaboo) should set this
-    /// to their own folder name during startup.
+    /// to their own folder/path during startup.
     public nonisolated(unsafe) static var profileDirectoryName: String = ".tachikoma"
 
+    public static var profileDirectoryPath: String {
+        let profile = self.profileDirectoryName
+        if self.isAbsoluteProfilePath(profile) || profile.hasPrefix("~") {
+            return NSString(string: profile).expandingTildeInPath
+        }
+
+        guard let homeDirectory = self.homeDirectoryPath else {
+            return profile
+        }
+
+        return "\(homeDirectory)/\(profile)"
+    }
+
+    private static func isAbsoluteProfilePath(_ profile: String) -> Bool {
+        if profile.hasPrefix("/") {
+            return true
+        }
+
+        #if os(Windows)
+        if profile.hasPrefix("\\\\") {
+            return true
+        }
+
+        if profile.count >= 3 {
+            let driveIndex = profile.index(after: profile.startIndex)
+            let separatorIndex = profile.index(after: driveIndex)
+            let drive = profile[profile.startIndex]
+            let separator = profile[separatorIndex]
+            return drive.isASCII && drive.isLetter && profile[driveIndex] == ":" &&
+                (separator == "\\" || separator == "/")
+        }
+        #endif
+
+        return false
+    }
+
+    private static var homeDirectoryPath: String? {
+        #if os(Windows)
+        let homeDirectory = ProcessInfo.processInfo.environment["USERPROFILE"] ??
+            ((ProcessInfo.processInfo.environment["HOMEDRIVE"] ?? "") +
+                (ProcessInfo.processInfo.environment["HOMEPATH"] ?? ""))
+        return homeDirectory.isEmpty ? nil : homeDirectory
+        #else
+        return ProcessInfo.processInfo.environment["HOME"]
+        #endif
+    }
+
     private let lock = NSLock()
     private var _apiKeys: [String: String] = [:]
     private var _baseURLs: [String: String] = [:]
@@ -95,6 +142,10 @@ public final class TachikomaConfiguration: @unchecked Sendable {
                 return configuredKey
             }
 
+            if provider == .minimaxCN, let sharedMiniMaxKey = self._apiKeys[Provider.minimax.identifier] {
+                return sharedMiniMaxKey
+            }
+
             // Fall back to environment variable only if loadFromEnvironment is true
             if self._loadFromEnvironment {
                 return provider.loadAPIKeyFromEnvironment()
@@ -281,7 +332,17 @@ public final class TachikomaConfiguration: @unchecked Sendable {
         self.loadFromEnvironment()
     }
 
-    /// Load configuration from environment variables
+    /// Load configuration from environment variables.
+    ///
+    /// Marked `@inline(never)` to work around a Swift release-mode optimizer
+    /// issue where inlining this function into the singleton-init path causes
+    /// the `_baseURLs[.anthropic]` write to be incorrectly eliminated, leaving
+    /// `AnthropicProvider.init` reading a stale (default) value via
+    /// `getBaseURL`. OpenAI / Ollama / MiniMax / Azure writes happen to
+    /// survive the optimization; only Anthropic is empirically affected.
+    /// Removing this annotation reintroduces the bug observed in
+    /// openclaw/Peekaboo release builds for `claude-*` models. See #17.
+    @inline(never)
     private func loadFromEnvironment() {
         // Load API keys for all standard providers from environment
         for provider in Provider.standardProviders {
@@ -294,6 +355,8 @@ public final class TachikomaConfiguration: @unchecked Sendable {
         let urlMappings: [Provider: String] = [
             .openai: "OPENAI_BASE_URL",
             .anthropic: "ANTHROPIC_BASE_URL",
+            .minimax: "MINIMAX_BASE_URL",
+            .minimaxCN: "MINIMAX_CN_BASE_URL",
             .ollama: "OLLAMA_BASE_URL",
             .azureOpenAI: "AZURE_OPENAI_ENDPOINT",
         ]
@@ -312,23 +375,13 @@ public final class TachikomaConfiguration: @unchecked Sendable {
     /// Load configuration from credentials file
     private func loadFromCredentials() {
         // Load configuration from credentials file
-        #if os(Windows)
-        let homeDirectory = ProcessInfo.processInfo.environment["USERPROFILE"] ??
-            (ProcessInfo.processInfo.environment["HOMEDRIVE"] ?? "" +
-                (ProcessInfo.processInfo.environment["HOMEPATH"] ?? ""))
-        guard !homeDirectory.isEmpty else { return }
-        #else
-        guard let homeDirectory = ProcessInfo.processInfo.environment["HOME"] else {
-            return
-        }
-        #endif
+        // Primary: configured profile directory/path (e.g. .peekaboo or PEEKABOO_CONFIG_DIR)
+        let primaryCredentialsPath = "\(Self.profileDirectoryPath)/credentials"
 
-        // Primary: configured profile directory (e.g. .peekaboo)
-        let primaryCredentialsPath = "\(homeDirectory)/\(Self.profileDirectoryName)/credentials"
         // Fallback: legacy .tachikoma directory for non-Peekaboo users
-        let fallbackCredentialsPath = "\(homeDirectory)/.tachikoma/credentials"
+        let fallbackCredentialsPath = Self.homeDirectoryPath.map { "\($0)/.tachikoma/credentials" }
 
-        let candidates = [primaryCredentialsPath, fallbackCredentialsPath]
+        let candidates = [primaryCredentialsPath, fallbackCredentialsPath].compactMap(\.self)
         let credentialsPath = candidates.first { FileManager.default.fileExists(atPath: $0) }
         guard let path = credentialsPath else { return }
         let credentialsURL = URL(fileURLWithPath: path)
@@ -356,46 +409,32 @@ public final class TachikomaConfiguration: @unchecked Sendable {
                 let key = components[0].trimmingCharacters(in: .whitespacesAndNewlines)
                 let value = components[1...].joined(separator: "=").trimmingCharacters(in: .whitespacesAndNewlines)
 
-                // Map credential keys to providers
-                let lowercaseKey = key.lowercased()
-                if lowercaseKey.contains("openai") {
-                    self.setAPIKey(value, for: .openai)
-                } else if lowercaseKey.contains("anthropic") || lowercaseKey.contains("claude") {
-                    self.setAPIKey(value, for: .anthropic)
-                } else if lowercaseKey.contains("grok") {
-                    self.setAPIKey(value, for: .grok)
-                } else if lowercaseKey.contains("groq") {
-                    self.setAPIKey(value, for: .groq)
-                } else if lowercaseKey.contains("mistral") {
-                    self.setAPIKey(value, for: .mistral)
-                } else if lowercaseKey.contains("google") || lowercaseKey.contains("gemini") {
-                    self.setAPIKey(value, for: .google)
-                } else if lowercaseKey.contains("ollama") {
-                    self.setAPIKey(value, for: .ollama)
+                if let provider = Self.provider(forCredentialKey: key) {
+                    self.setAPIKey(value, for: provider)
                 }
             }
         }
     }
 
+    private static func provider(forCredentialKey key: String) -> Provider? {
+        let normalizedKey = key.trimmingCharacters(in: .whitespacesAndNewlines).uppercased()
+        return Provider.standardProviders.first { provider in
+            let keys = ([provider.environmentVariable] + provider.alternativeEnvironmentVariables)
+                .filter { !$0.isEmpty }
+                .map { $0.uppercased() }
+            return keys.contains(normalizedKey)
+        }
+    }
+
     // MARK: - Persistence
 
     /// Save current configuration to credentials file
     public func saveCredentials() throws {
         // Save current configuration to credentials file
-        #if os(Windows)
-        let homeDirectory = ProcessInfo.processInfo.environment["USERPROFILE"] ??
-            (ProcessInfo.processInfo.environment["HOMEDRIVE"] ?? "" +
-                (ProcessInfo.processInfo.environment["HOMEPATH"] ?? ""))
-        guard !homeDirectory.isEmpty else {
-            throw TachikomaError.invalidConfiguration("USERPROFILE directory not found")
+        let profileDir = Self.profileDirectoryPath
+        guard !profileDir.isEmpty else {
+            throw TachikomaError.invalidConfiguration("Profile directory not found")
         }
-        #else
-        guard let homeDirectory = ProcessInfo.processInfo.environment["HOME"] else {
-            throw TachikomaError.invalidConfiguration("HOME directory not found")
-        }
-        #endif
-
-        let profileDir = "\(homeDirectory)/\(Self.profileDirectoryName)"
         let credentialsPath = "\(profileDir)/credentials"
 
         // Create directory if needed
@@ -410,7 +449,8 @@ public final class TachikomaConfiguration: @unchecked Sendable {
 
         self.lock.withLock {
             for (provider, key) in self._apiKeys {
-                let envVarName = "\(provider.uppercased())_API_KEY"
+                let standardEnvVar = Provider.from(identifier: provider).environmentVariable
+                let envVarName = standardEnvVar.isEmpty ? "\(provider.uppercased())_API_KEY" : standardEnvVar
                 lines.append("\(envVarName)=\(key)")
             }
         }

Sources/Tachikoma/Core/CustomProviders.swift

@@ -7,6 +7,7 @@ public struct CustomProviderInfo: Sendable {
     public let id: String
     public let kind: Kind
     public let baseURL: String
+    public let apiKey: String?
     public let headers: [String: String]
     public let models: [String: String] // map of alias->modelId (optional usage)
 }
@@ -29,6 +30,7 @@ public final class CustomProviderRegistry: @unchecked Sendable {
             guard
                 let options = dict["options"] as? [String: Any],
                 let baseURL = options["baseURL"] as? String else { continue }
+            let apiKey = options["apiKey"] as? String
             let headers = options["headers"] as? [String: String] ?? [:]
             var models: [String: String] = [:]
             if let modelsDict = dict["models"] as? [String: Any] {
@@ -36,7 +38,14 @@ public final class CustomProviderRegistry: @unchecked Sendable {
                     if let m = (v as? [String: Any])?["name"] as? String { models[k] = m }
                 }
             }
-            out[id] = CustomProviderInfo(id: id, kind: kind, baseURL: baseURL, headers: headers, models: models)
+            out[id] = CustomProviderInfo(
+                id: id,
+                kind: kind,
+                baseURL: baseURL,
+                apiKey: apiKey,
+                headers: headers,
+                models: models,
+            )
         }
         self.providers = out
     }
@@ -52,12 +61,7 @@ public final class CustomProviderRegistry: @unchecked Sendable {
     // MARK: - Helpers
 
     private static func profileDirectoryPath() -> String {
-        #if os(Windows)
-        let home = ProcessInfo.processInfo.environment["USERPROFILE"] ?? ""
-        #else
-        let home = ProcessInfo.processInfo.environment["HOME"] ?? ""
-        #endif
-        return "\(home)/\(TachikomaConfiguration.profileDirectoryName)"
+        TachikomaConfiguration.profileDirectoryPath
     }
 
     private static func profileConfigPath() -> String {

Sources/Tachikoma/Core/Embeddings.swift

@@ -113,10 +113,10 @@ public enum EmbeddingModel: Sendable {
         // Convert to LanguageModel for usage tracking
         switch self {
         case .openai:
-            .openai(.gpt4o) // Placeholder for tracking
+            .openai(.gpt55) // Placeholder for tracking
         case .cohere, .voyage, .custom:
             // Return a dummy model for tracking purposes
-            .openai(.gpt4o)
+            .openai(.gpt55)
         }
     }
 }

Sources/Tachikoma/Core/Generation.swift

@@ -35,10 +35,11 @@ public func generateText(
     var currentMessages = messages
     var allSteps: [GenerationStep] = []
     var totalUsage = Usage(inputTokens: 0, outputTokens: 0)
+    var finalResponseStartIndex = messages.count
 
     for stepIndex in 0..<maxSteps {
         let request = ProviderRequest(
-            messages: currentMessages,
+            messages: currentMessages.sanitizedForProvider(model, configuration: resolvedConfiguration),
             tools: tools,
             settings: settings,
         )
@@ -51,8 +52,23 @@ public func generateText(
             try await provider.generateText(request: request)
         }
 
-        // Track usage with proper session management
-        if let usage = response.usage {
+        let isContentFiltered = response.finishReason == .contentFilter
+        let responseText = isContentFiltered ? "" : response.text
+        let responseToolCalls = isContentFiltered ? [] : (response.toolCalls ?? [])
+        let responseReasoning = isContentFiltered ? [] : response.reasoning
+        let responseAssistantMessages = isContentFiltered ? [] : response.assistantMessages
+        let responseMessageStartIndex = currentMessages.count
+        finalResponseStartIndex = responseMessageStartIndex
+        let responseHistoryMessages = model.responseHistoryMessages(
+            nativeMessages: responseAssistantMessages,
+            text: responseText,
+            reasoning: responseReasoning,
+            toolCalls: responseToolCalls,
+            configuration: resolvedConfiguration,
+        )
+
+        // Track billable usage with proper session management.
+        if response.isBillable, let usage = response.usage {
             let actualSessionId = sessionId ?? "generation-\(UUID().uuidString)"
 
             // Start session if not already started
@@ -86,8 +102,8 @@ public func generateText(
         // Create step record
         let step = GenerationStep(
             stepIndex: stepIndex,
-            text: response.text,
-            toolCalls: response.toolCalls ?? [],
+            text: responseText,
+            toolCalls: responseToolCalls,
             toolResults: [],
             usage: response.usage,
             finishReason: response.finishReason,
@@ -95,18 +111,26 @@ public func generateText(
 
         allSteps.append(step)
 
-        // Add assistant message
-        var assistantContent: [ModelMessage.ContentPart] = [.text(response.text)]
+        if isContentFiltered {
+            break
+        }
+
+        if !responseHistoryMessages.isEmpty {
+            currentMessages.append(contentsOf: responseHistoryMessages)
+            if responseHistoryMessages.allSatisfy({ $0.channel == .thinking }) {
+                currentMessages.append(ModelMessage(
+                    role: .assistant,
+                    content: [.text("")],
+                    metadata: .init(customData: ["tachikoma.internal.boundary": "reasoning_only"]),
+                ))
+            }
+        }
 
         // Handle tool calls
-        if let toolCalls = response.toolCalls, !toolCalls.isEmpty {
-            // Add tool calls to assistant message
-            assistantContent.append(contentsOf: toolCalls.map { .toolCall($0) })
-            currentMessages.append(ModelMessage(role: .assistant, content: assistantContent))
-
+        if !responseToolCalls.isEmpty {
             // Execute tools
             var toolResults: [AgentToolResult] = []
-            for toolCall in toolCalls {
+            for toolCall in responseToolCalls {
                 if let tool = tools?.first(where: { $0.name == toolCall.name }) {
                     do {
                         // Debug: Log tool call details in verbose mode
@@ -124,7 +148,7 @@ public func generateText(
 
                         // Create execution context with full conversation and model info
                         let context = ToolExecutionContext(
-                            messages: currentMessages,
+                            messages: currentMessages.sanitizedForToolContext(),
                             model: model,
                             settings: settings,
                             sessionId: sessionId ?? "generation-\(UUID().uuidString)",
@@ -161,8 +185,8 @@ public func generateText(
             // Update step with tool results
             allSteps[stepIndex] = GenerationStep(
                 stepIndex: stepIndex,
-                text: response.text,
-                toolCalls: toolCalls,
+                text: responseText,
+                toolCalls: responseToolCalls,
                 toolResults: toolResults,
                 usage: response.usage,
                 finishReason: response.finishReason,
@@ -174,17 +198,17 @@ public func generateText(
             }
         } else {
             // No tool calls, we're done
-            currentMessages.append(ModelMessage(role: .assistant, content: assistantContent))
             break
         }
     }
 
     // Extract final text from last step
     var finalText = allSteps.last?.text ?? ""
+    let originalFinalText = finalText
     var finalFinishReason = allSteps.last?.finishReason ?? .other
 
     // Apply stop conditions if configured
-    if let stopCondition = settings.stopConditions {
+    if finalFinishReason != .contentFilter, let stopCondition = settings.stopConditions {
         // Check if we should stop and truncate the text
         if await stopCondition.shouldStop(text: finalText, delta: nil) {
             // Truncate text based on the type of stop condition
@@ -217,13 +241,16 @@ public func generateText(
             }
         }
     }
+    let finalMessages = finalText == originalFinalText
+        ? currentMessages
+        : currentMessages.replacingGeneratedAssistantText(after: finalResponseStartIndex, with: finalText)
 
     return GenerateTextResult(
         text: finalText,
         usage: totalUsage,
         finishReason: finalFinishReason,
         steps: allSteps,
-        messages: currentMessages,
+        messages: finalMessages,
     )
 }
 
@@ -254,6 +281,9 @@ public func streamText(
 {
     // Debug logging only when explicitly enabled via environment variable or verbose flag
     let resolvedConfiguration = TachikomaConfiguration.resolve(configuration)
+    guard model.supportsStreaming else {
+        throw TachikomaError.invalidConfiguration("\(model.modelId) does not support streaming")
+    }
     let debugEnabled = ProcessInfo.processInfo.environment["DEBUG_TACHIKOMA"] != nil ||
         resolvedConfiguration.verbose
     if debugEnabled {
@@ -275,7 +305,7 @@ public func streamText(
     }
 
     let request = ProviderRequest(
-        messages: messages,
+        messages: messages.sanitizedForProvider(model, configuration: resolvedConfiguration),
         tools: tools,
         settings: settings,
     )
@@ -296,12 +326,6 @@ public func streamText(
         stream = try await provider.streamText(request: request)
     }
 
-    // Apply stop conditions if configured
-    if let stopCondition = settings.stopConditions {
-        // Wrap the stream with stop condition checking
-        stream = stream.stopWhen(stopCondition)
-    }
-
     // Use provided session or create a new one for tracking streaming usage
     let actualSessionId = sessionId ?? "streaming-\(UUID().uuidString)"
     if sessionId == nil {
@@ -311,23 +335,84 @@ public func streamText(
     // Wrap the stream to track usage when it completes
     let capturedModel = model
     let capturedSessionId = actualSessionId
-    let capturedStream = stream
     let shouldEndSession = sessionId == nil
+    let buffersUntilDone = model.buffersTextStreamUntilDone(settings: settings)
+    if !buffersUntilDone, let stopCondition = settings.stopConditions {
+        stream = stream.stopWhen(stopCondition)
+    }
+    let capturedStream = stream
+    let capturedStopCondition = buffersUntilDone ? settings.stopConditions : nil
 
     let trackedStream = AsyncThrowingStream<TextStreamDelta, Error> { continuation in
         Task {
             do {
                 let totalInputTokens = 0
                 var totalOutputTokens = 0
+                var bufferedDeltas: [TextStreamDelta] = []
+                var bufferedVisibleText = ""
+                var didReceiveTerminal = false
+                var didTriggerLocalStop = false
 
-                for try await delta in capturedStream {
-                    continuation.yield(delta)
-
+                func track(_ delta: TextStreamDelta) {
                     // Track tokens as they come in (approximate)
                     if case .textDelta = delta.type, let content = delta.content {
                         // Rough approximation: ~4 characters per token
                         totalOutputTokens += max(1, content.count / 4)
                     }
+                }
+
+                func yieldAndTrack(_ delta: TextStreamDelta) {
+                    track(delta)
+                    continuation.yield(delta)
+                }
+
+                if let capturedStopCondition {
+                    await capturedStopCondition.reset()
+                }
+
+                for try await delta in capturedStream {
+                    if buffersUntilDone, delta.type != .done {
+                        if !didTriggerLocalStop {
+                            bufferedDeltas.append(delta)
+                            track(delta)
+                            if
+                                let capturedStopCondition,
+                                case .textDelta = delta.type,
+                                let content = delta.content
+                            {
+                                bufferedVisibleText += content
+                                didTriggerLocalStop = await capturedStopCondition.shouldStop(
+                                    text: bufferedVisibleText,
+                                    delta: content,
+                                )
+                            }
+                        }
+                        continue
+                    }
+
+                    if case .done = delta.type {
+                        didReceiveTerminal = true
+                        if buffersUntilDone {
+                            if delta.finishReason == .contentFilter {
+                                bufferedDeltas.removeAll()
+                                yieldAndTrack(delta)
+                            } else {
+                                for bufferedDelta in bufferedDeltas {
+                                    continuation.yield(bufferedDelta)
+                                }
+                                bufferedDeltas.removeAll()
+                                if didTriggerLocalStop {
+                                    yieldAndTrack(TextStreamDelta.done(usage: delta.usage, finishReason: .stop))
+                                } else {
+                                    yieldAndTrack(delta)
+                                }
+                            }
+                        } else {
+                            yieldAndTrack(delta)
+                        }
+                    } else {
+                        yieldAndTrack(delta)
+                    }
 
                     if case .done = delta.type {
                         // Record final usage (this is approximate for streaming)
@@ -348,6 +433,10 @@ public func streamText(
                     }
                 }
 
+                if buffersUntilDone, !didReceiveTerminal {
+                    throw TachikomaError.apiError("Stream ended before provider completion status was received")
+                }
+
                 continuation.finish()
             } catch {
                 if shouldEndSession {
@@ -392,7 +481,7 @@ public func generateObject<T: Codable & Sendable>(
     let provider = try resolvedConfiguration.makeProvider(for: model)
 
     let request = ProviderRequest(
-        messages: messages,
+        messages: messages.sanitizedForProvider(model, configuration: resolvedConfiguration),
         tools: nil,
         settings: settings,
         outputFormat: .json,
@@ -406,6 +495,10 @@ public func generateObject<T: Codable & Sendable>(
         try await provider.generateText(request: request)
     }
 
+    if response.finishReason == .contentFilter {
+        throw TachikomaError.apiError("Response was blocked by the provider content filter")
+    }
+
     // Parse the JSON response into the expected type
     guard let jsonData = response.text.data(using: .utf8) else {
         throw TachikomaError.invalidInput("Response text is not valid UTF-8")
@@ -446,11 +539,14 @@ public func streamObject<T: Codable & Sendable>(
     -> StreamObjectResult<T>
 {
     let resolvedConfiguration = TachikomaConfiguration.resolve(configuration)
+    guard model.supportsStreaming else {
+        throw TachikomaError.invalidConfiguration("\(model.modelId) does not support streaming")
+    }
     let provider = try resolvedConfiguration.makeProvider(for: model)
 
     // Create request with JSON output format
     let request = ProviderRequest(
-        messages: messages,
+        messages: messages.sanitizedForProvider(model, configuration: resolvedConfiguration),
         tools: nil,
         settings: settings,
         outputFormat: .json,
@@ -458,6 +554,7 @@ public func streamObject<T: Codable & Sendable>(
 
     // Get the text stream from the provider
     let stream = try await provider.streamText(request: request)
+    let buffersUntilDone = model.buffersObjectStreamUntilDone(settings: settings)
 
     // Create a new stream that attempts to parse partial JSON objects
     let objectStream = AsyncThrowingStream<ObjectStreamDelta<T>, Error> { continuation in
@@ -466,6 +563,37 @@ public func streamObject<T: Codable & Sendable>(
                 var accumulatedText = ""
                 var lastValidObject: T?
                 var hasStarted = false
+                var bufferedStartDelta: ObjectStreamDelta<T>?
+                var didFinishObject = false
+
+                func publishCompleteObject(allowLastValidObjectFallback: Bool) throws {
+                    if buffersUntilDone, let bufferedStartDelta {
+                        continuation.yield(bufferedStartDelta)
+                    }
+                    if
+                        let jsonData = accumulatedText.data(using: .utf8),
+                        let finalObject = try? JSONDecoder().decode(T.self, from: jsonData)
+                    {
+                        continuation.yield(ObjectStreamDelta(
+                            type: .complete,
+                            object: finalObject,
+                            rawText: accumulatedText,
+                        ))
+                    } else if allowLastValidObjectFallback, let lastValidObject {
+                        // If we have a last valid object, use it as complete
+                        continuation.yield(ObjectStreamDelta(
+                            type: .complete,
+                            object: lastValidObject,
+                            rawText: accumulatedText,
+                        ))
+                    } else {
+                        throw TachikomaError.invalidInput(
+                            "Failed to parse complete object from stream",
+                        )
+                    }
+                    continuation.yield(ObjectStreamDelta(type: .done))
+                    didFinishObject = true
+                }
 
                 for try await delta in stream {
                     if case .textDelta = delta.type, let content = delta.content {
@@ -474,7 +602,16 @@ public func streamObject<T: Codable & Sendable>(
                         // Signal stream start
                         if !hasStarted {
                             hasStarted = true
-                            continuation.yield(ObjectStreamDelta(type: .start))
+                            let startDelta = ObjectStreamDelta<T>(type: .start)
+                            if buffersUntilDone {
+                                bufferedStartDelta = startDelta
+                            } else {
+                                continuation.yield(startDelta)
+                            }
+                        }
+
+                        if buffersUntilDone {
+                            continue
                         }
 
                         // Attempt to parse the accumulated JSON
@@ -482,44 +619,51 @@ public func streamObject<T: Codable & Sendable>(
                             // Try to parse as complete object
                             if let object = try? JSONDecoder().decode(T.self, from: jsonData) {
                                 lastValidObject = object
-                                continuation.yield(ObjectStreamDelta(
+                                let objectDelta = ObjectStreamDelta(
                                     type: .partial,
                                     object: object,
                                     rawText: accumulatedText,
-                                ))
+                                )
+                                continuation.yield(objectDelta)
                             } else if let partialObject = attemptPartialParse(T.self, from: accumulatedText) {
                                 // Attempt to parse as partial object
                                 lastValidObject = partialObject
-                                continuation.yield(ObjectStreamDelta(
+                                let objectDelta = ObjectStreamDelta(
                                     type: .partial,
                                     object: partialObject,
                                     rawText: accumulatedText,
-                                ))
+                                )
+                                continuation.yield(objectDelta)
                             }
                         }
                     } else if case .done = delta.type {
-                        // Final parse attempt
-                        if
-                            let jsonData = accumulatedText.data(using: .utf8),
-                            let finalObject = try? JSONDecoder().decode(T.self, from: jsonData)
-                        {
-                            continuation.yield(ObjectStreamDelta(
-                                type: .complete,
-                                object: finalObject,
-                                rawText: accumulatedText,
-                            ))
-                        } else if let lastValidObject {
-                            // If we have a last valid object, use it as complete
-                            continuation.yield(ObjectStreamDelta(
-                                type: .complete,
-                                object: lastValidObject,
-                                rawText: accumulatedText,
-                            ))
-                        } else {
-                            throw TachikomaError.invalidInput(
-                                "Failed to parse complete object from stream",
-                            )
+                        if delta.finishReason == .contentFilter {
+                            throw TachikomaError.apiError("Response was blocked by the provider content filter")
                         }
+                        try publishCompleteObject(allowLastValidObjectFallback: delta.finishReason == .stop || delta
+                            .finishReason == nil)
+                    }
+                }
+
+                if !didFinishObject, hasStarted {
+                    if buffersUntilDone {
+                        throw TachikomaError.apiError("Stream ended before provider completion status was received")
+                    } else if
+                        let jsonData = accumulatedText.data(using: .utf8),
+                        let finalObject = try? JSONDecoder().decode(T.self, from: jsonData)
+                    {
+                        continuation.yield(ObjectStreamDelta(
+                            type: .complete,
+                            object: finalObject,
+                            rawText: accumulatedText,
+                        ))
+                        continuation.yield(ObjectStreamDelta(type: .done))
+                    } else if let lastValidObject {
+                        continuation.yield(ObjectStreamDelta(
+                            type: .complete,
+                            object: lastValidObject,
+                            rawText: accumulatedText,
+                        ))
                         continuation.yield(ObjectStreamDelta(type: .done))
                     }
                 }
@@ -599,6 +743,466 @@ private func fixPartialJSON(_ json: String) -> String {
     return fixed
 }
 
+extension LanguageModel {
+    fileprivate func buffersTextStreamUntilDone(settings: GenerationSettings) -> Bool {
+        self.hasAnthropicStreamingRefusalRisk ||
+            settings.streamBuffering == .untilTerminal ||
+            (settings.stopConditions != nil && self.canEmitTerminalContentFilterAfterText)
+    }
+
+    fileprivate func buffersObjectStreamUntilDone(settings: GenerationSettings) -> Bool {
+        settings.streamBuffering == .untilTerminal ||
+            self.hasAnthropicStreamingRefusalRisk
+    }
+
+    private var hasAnthropicStreamingRefusalRisk: Bool {
+        switch self {
+        case let .anthropic(model):
+            return LanguageModel.Anthropic.hasStreamingRefusalRisk(modelId: model.modelId)
+        case let .anthropicCompatible(modelId, _):
+            return LanguageModel.Anthropic.hasStreamingRefusalRisk(modelId: modelId)
+        case let .openRouter(modelId), let .together(modelId):
+            return LanguageModel.Anthropic.hasStreamingRefusalRisk(modelId: modelId)
+        case let .openaiCompatible(modelId, _):
+            return LanguageModel.Anthropic.hasStreamingRefusalRisk(modelId: modelId)
+        case let .custom(provider):
+            if
+                let parsed = ProviderParser.parse(provider.modelId),
+                CustomProviderRegistry.shared.get(parsed.provider)?.kind == .anthropic
+            {
+                return LanguageModel.Anthropic.hasStreamingRefusalRisk(modelId: parsed.model)
+            }
+            return LanguageModel.Anthropic.hasStreamingRefusalRisk(modelId: provider.modelId)
+        default:
+            return false
+        }
+    }
+
+    private var canEmitTerminalContentFilterAfterText: Bool {
+        switch self {
+        case .openai,
+             .openaiCompatible,
+             .openRouter,
+             .together,
+             .replicate,
+             .google,
+             .mistral,
+             .groq,
+             .grok,
+             .azureOpenAI:
+            return true
+        case let .custom(provider):
+            guard
+                let parsed = ProviderParser.parse(provider.modelId),
+                let registeredProvider = CustomProviderRegistry.shared.get(parsed.provider) else
+            {
+                return false
+            }
+            switch registeredProvider.kind {
+            case .openai:
+                return true
+            case .anthropic:
+                return false
+            }
+        default:
+            return false
+        }
+    }
+}
+
+private struct ReasoningReplayTarget {
+    let provider: String
+    let modelId: String
+    let baseURL: String?
+    let allowsLegacyUnknown: Bool
+
+    func matches(_ customData: [String: String]) -> Bool {
+        guard customData["tachikoma.reasoning.provider"] == self.provider else {
+            return false
+        }
+        guard customData["tachikoma.reasoning.model"] == self.modelId else {
+            return false
+        }
+        return customData["tachikoma.reasoning.base_url"] == self.endpointIdentity
+    }
+
+    var endpointIdentity: String? {
+        ReasoningEndpointIdentity.canonical(self.baseURL)
+    }
+}
+
+extension [ModelMessage] {
+    fileprivate func replacingGeneratedAssistantText(after prefixCount: Int, with text: String) -> [ModelMessage] {
+        guard self.indices.contains(prefixCount) else {
+            return self
+        }
+
+        var messages = self
+        var cursor = text.startIndex
+        for messageIndex in prefixCount..<messages.count {
+            let message = messages[messageIndex]
+            guard message.role == .assistant, message.channel != .thinking else {
+                continue
+            }
+
+            var content: [ModelMessage.ContentPart] = []
+            for part in message.content {
+                guard case let .text(originalText) = part else {
+                    content.append(part)
+                    continue
+                }
+
+                guard cursor < text.endIndex else {
+                    continue
+                }
+
+                let remainingCount = text.distance(from: cursor, to: text.endIndex)
+                let takeCount = Swift.min(originalText.count, remainingCount)
+                let endIndex = text.index(cursor, offsetBy: takeCount)
+                content.append(.text(String(text[cursor..<endIndex])))
+                cursor = endIndex
+            }
+
+            messages[messageIndex] = ModelMessage(
+                id: message.id,
+                role: message.role,
+                content: content,
+                timestamp: message.timestamp,
+                channel: message.channel,
+                metadata: message.metadata,
+            )
+        }
+        return messages
+    }
+}
+
+extension [ModelMessage] {
+    fileprivate func sanitizedForProvider(
+        _ model: LanguageModel,
+        configuration: TachikomaConfiguration,
+    )
+        -> [ModelMessage]
+    {
+        if let target = model.anthropicThinkingReplayTarget(configuration: configuration) {
+            var sanitized: [ModelMessage] = []
+            for message in self {
+                if message.isSyntheticReasoningBoundary {
+                    if sanitized.last?.channel == .thinking {
+                        sanitized.append(message)
+                    }
+                    continue
+                }
+                guard message.channel == .thinking else {
+                    sanitized.append(message)
+                    continue
+                }
+                guard !message.hasOpenRouterReasoningReplayMetadata else {
+                    continue
+                }
+                guard let producerModel = message.metadata?.customData?["anthropic.thinking.model"] else {
+                    if
+                        target.allowsLegacyUnknown,
+                        message.metadata?.customData?["anthropic.thinking.type"] != nil
+                    {
+                        sanitized.append(message)
+                    }
+                    continue
+                }
+                let customData = message.metadata?.customData ?? [:]
+                if producerModel == target.modelId, target.matches(customData) {
+                    sanitized.append(message)
+                }
+            }
+            return sanitized
+        }
+
+        if let target = model.openRouterReasoningReplayTarget(configuration: configuration) {
+            var sanitized: [ModelMessage] = []
+            for message in self {
+                if message.isSyntheticReasoningBoundary {
+                    if sanitized.last?.channel == .thinking {
+                        sanitized.append(message)
+                    }
+                    continue
+                }
+                guard message.channel == .thinking else {
+                    sanitized.append(message)
+                    continue
+                }
+                guard message.hasOpenRouterReasoningReplayMetadata else {
+                    continue
+                }
+                if target.matches(message.metadata?.customData ?? [:]) {
+                    sanitized.append(message)
+                }
+            }
+            return sanitized
+        }
+
+        return self.filter { !$0.isSyntheticReasoningBoundary && $0.channel != .thinking }
+    }
+}
+
+extension ModelMessage {
+    private var hasAnthropicThinkingReplayMetadata: Bool {
+        guard let customData = metadata?.customData else { return false }
+        return customData["anthropic.thinking.model"] != nil ||
+            customData["anthropic.thinking.type"] != nil ||
+            customData["anthropic.thinking.signature"] != nil
+    }
+
+    fileprivate var hasOpenRouterReasoningReplayMetadata: Bool {
+        guard let customData = metadata?.customData else { return false }
+        return customData["openrouter.reasoning_details"] != nil ||
+            customData["openrouter.reasoning"] != nil
+    }
+
+    private var hasProviderReasoningReplayMetadata: Bool {
+        self.hasAnthropicThinkingReplayMetadata || self.hasOpenRouterReasoningReplayMetadata
+    }
+
+    fileprivate var isSyntheticReasoningBoundary: Bool {
+        metadata?.customData?["tachikoma.internal.boundary"] == "reasoning_only"
+    }
+}
+
+extension [ModelMessage] {
+    fileprivate func sanitizedForToolContext() -> [ModelMessage] {
+        self.filter { $0.channel != .thinking && !$0.isSyntheticReasoningBoundary }
+    }
+
+    fileprivate func containsAssistantText(_ text: String) -> Bool {
+        guard !text.isEmpty else { return true }
+        let assistantTexts = self.flatMap { message -> [String] in
+            guard message.role == .assistant, message.channel != .thinking else {
+                return []
+            }
+            return message.content.compactMap { part in
+                if case let .text(value) = part {
+                    return value
+                }
+                return nil
+            }
+        }
+        return assistantTexts.contains(text) || assistantTexts.joined() == text
+    }
+
+    fileprivate func containsReasoningBlock(_ reasoning: ProviderReasoningBlock) -> Bool {
+        self.contains { message in
+            message.role == .assistant && message.channel == .thinking && message.content.contains { part in
+                guard case let .text(value) = part else { return false }
+                if let signature = reasoning.signature, !signature.isEmpty {
+                    return message.metadata?.customData?["anthropic.thinking.signature"] == signature ||
+                        message.metadata?.customData?["tachikoma.reasoning.signature"] == signature
+                }
+                return value == reasoning.text
+            }
+        }
+    }
+
+    fileprivate func containsToolCall(id: String) -> Bool {
+        self.contains { message in
+            message.role == .assistant && message.content.contains { part in
+                if case let .toolCall(toolCall) = part {
+                    return toolCall.id == id
+                }
+                return false
+            }
+        }
+    }
+}
+
+extension LanguageModel {
+    fileprivate func responseHistoryMessages(
+        nativeMessages: [ModelMessage],
+        text: String,
+        reasoning: [ProviderReasoningBlock],
+        toolCalls: [AgentToolCall],
+        configuration: TachikomaConfiguration,
+    )
+        -> [ModelMessage]
+    {
+        var history = nativeMessages
+
+        for reasoningBlock in reasoning where !history.containsReasoningBlock(reasoningBlock) {
+            history.append(ModelMessage(
+                role: .assistant,
+                content: [.text(reasoningBlock.text)],
+                channel: .thinking,
+                metadata: .init(customData: self.anthropicThinkingMetadata(
+                    for: reasoningBlock,
+                    configuration: configuration,
+                )),
+            ))
+        }
+
+        let missingToolCalls = toolCalls.filter { !history.containsToolCall(id: $0.id) }
+        let isMissingText = !history.containsAssistantText(text)
+        let needsFallbackBoundary = nativeMessages.isEmpty && text.isEmpty && missingToolCalls.isEmpty
+
+        guard isMissingText || !missingToolCalls.isEmpty || needsFallbackBoundary else {
+            return history
+        }
+
+        var fallbackContent: [ModelMessage.ContentPart] = []
+        if isMissingText || needsFallbackBoundary {
+            fallbackContent.append(.text(text))
+        }
+        fallbackContent.append(contentsOf: missingToolCalls.map { .toolCall($0) })
+        let fallbackMetadata = needsFallbackBoundary
+            ? MessageMetadata(customData: ["tachikoma.internal.boundary": "reasoning_only"])
+            : nil
+        history.append(ModelMessage(role: .assistant, content: fallbackContent, metadata: fallbackMetadata))
+        return history
+    }
+
+    fileprivate func anthropicThinkingReplayTarget(configuration: TachikomaConfiguration) -> ReasoningReplayTarget? {
+        switch self {
+        case let .anthropic(model):
+            return ReasoningReplayTarget(
+                provider: "anthropic",
+                modelId: model.modelId,
+                baseURL: configuration.getBaseURL(for: .anthropic) ?? Provider.anthropic.defaultBaseURL,
+                allowsLegacyUnknown: !LanguageModel.Anthropic.isFable(modelId: model.modelId),
+            )
+        case let .anthropicCompatible(modelId, baseURL):
+            return ReasoningReplayTarget(
+                provider: "anthropic-compatible",
+                modelId: modelId,
+                baseURL: baseURL,
+                allowsLegacyUnknown: !LanguageModel.Anthropic.isFable(modelId: modelId),
+            )
+        case let .minimax(model):
+            return ReasoningReplayTarget(
+                provider: "minimax",
+                modelId: model.modelId,
+                baseURL: configuration.getBaseURL(for: .minimax) ?? Provider.minimax.defaultBaseURL,
+                allowsLegacyUnknown: true,
+            )
+        case let .minimaxCN(model):
+            return ReasoningReplayTarget(
+                provider: "minimax-cn",
+                modelId: model.modelId,
+                baseURL: configuration.getBaseURL(for: .minimaxCN) ?? Provider.minimaxCN.defaultBaseURL,
+                allowsLegacyUnknown: true,
+            )
+        case let .custom(provider):
+            if let directAnthropicProvider = provider as? AnthropicProvider {
+                return ReasoningReplayTarget(
+                    provider: "anthropic",
+                    modelId: directAnthropicProvider.modelId,
+                    baseURL: directAnthropicProvider.baseURL ?? Provider.anthropic.defaultBaseURL,
+                    allowsLegacyUnknown: !LanguageModel.Anthropic.isFable(modelId: directAnthropicProvider.modelId),
+                )
+            }
+            if let compatibleProvider = provider as? AnthropicCompatibleProvider {
+                return ReasoningReplayTarget(
+                    provider: "anthropic-compatible",
+                    modelId: compatibleProvider.modelId,
+                    baseURL: compatibleProvider.baseURL,
+                    allowsLegacyUnknown: !LanguageModel.Anthropic.isFable(modelId: compatibleProvider.modelId),
+                )
+            }
+            guard
+                let parsed = ProviderParser.parse(provider.modelId),
+                let registeredProvider = CustomProviderRegistry.shared.get(parsed.provider),
+                registeredProvider.kind == .anthropic else
+            {
+                return provider.modelId.contains("claude") || provider.modelId.contains("anthropic")
+                    ? ReasoningReplayTarget(
+                        provider: "custom-anthropic",
+                        modelId: provider.modelId,
+                        baseURL: provider.baseURL,
+                        allowsLegacyUnknown: !LanguageModel.Anthropic.isFable(modelId: provider.modelId),
+                    )
+                    : nil
+            }
+            return ReasoningReplayTarget(
+                provider: "custom-anthropic",
+                modelId: parsed.model,
+                baseURL: registeredProvider.baseURL,
+                allowsLegacyUnknown: !LanguageModel.Anthropic.isFable(modelId: parsed.model),
+            )
+        default:
+            return nil
+        }
+    }
+
+    fileprivate func openRouterReasoningReplayTarget(configuration: TachikomaConfiguration) -> ReasoningReplayTarget? {
+        switch self {
+        case let .openRouter(modelId):
+            ReasoningReplayTarget(
+                provider: "openrouter",
+                modelId: modelId,
+                baseURL: configuration.getBaseURL(for: .custom("openrouter")) ?? "https://openrouter.ai/api/v1",
+                allowsLegacyUnknown: false,
+            )
+        default:
+            nil
+        }
+    }
+
+    private func anthropicThinkingMetadata(
+        for reasoning: ProviderReasoningBlock,
+        configuration: TachikomaConfiguration,
+    )
+        -> [String: String]
+    {
+        if
+            let rawJSON = reasoning.rawJSON,
+            let target = self.openRouterReasoningReplayTarget(configuration: configuration)
+        {
+            var metadata = [
+                "openrouter.reasoning_details": rawJSON,
+                "tachikoma.reasoning.type": reasoning.type,
+                "tachikoma.reasoning.provider": target.provider,
+                "tachikoma.reasoning.model": target.modelId,
+            ]
+            if let endpointIdentity = target.endpointIdentity {
+                metadata["tachikoma.reasoning.base_url"] = endpointIdentity
+            }
+            return metadata
+        }
+        if
+            reasoning.type == "openrouter_reasoning",
+            let target = self.openRouterReasoningReplayTarget(configuration: configuration)
+        {
+            var metadata = [
+                "openrouter.reasoning": reasoning.text,
+                "tachikoma.reasoning.type": reasoning.type,
+                "tachikoma.reasoning.provider": target.provider,
+                "tachikoma.reasoning.model": target.modelId,
+            ]
+            if let endpointIdentity = target.endpointIdentity {
+                metadata["tachikoma.reasoning.base_url"] = endpointIdentity
+            }
+            return metadata
+        }
+
+        guard let target = self.anthropicThinkingReplayTarget(configuration: configuration) else {
+            var customData = ["tachikoma.reasoning.type": reasoning.type]
+            if let signature = reasoning.signature, !signature.isEmpty {
+                customData["tachikoma.reasoning.signature"] = signature
+            }
+            return customData
+        }
+
+        var customData = [
+            "anthropic.thinking.type": reasoning.type,
+            "anthropic.thinking.model": target.modelId,
+            "tachikoma.reasoning.provider": target.provider,
+            "tachikoma.reasoning.model": target.modelId,
+        ]
+        if let endpointIdentity = target.endpointIdentity {
+            customData["tachikoma.reasoning.base_url"] = endpointIdentity
+        }
+        if let signature = reasoning.signature, !signature.isEmpty {
+            customData["anthropic.thinking.signature"] = signature
+        }
+        return customData
+    }
+}
+
 // MARK: - Convenience Functions
 
 /// Simple text generation from a prompt (convenience wrapper) - with Model enum
@@ -670,7 +1274,7 @@ public func analyze(
         model
     } else {
         // Use a vision-capable model by default
-        .openai(.gpt4o)
+        .openai(.gpt55)
     }
 
     // Ensure the model supports vision
@@ -746,7 +1350,7 @@ public func analyze(
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
 public func stream(
     _ prompt: String,
-    using model: LanguageModel = .default,
+    using model: LanguageModel = .defaultStreaming,
     system: String? = nil,
     maxTokens: Int? = nil,
     temperature: Double? = nil,

Sources/Tachikoma/Core/ModelCapabilities.swift

@@ -206,6 +206,10 @@ public final class ModelCapabilityRegistry: @unchecked Sendable {
             "ollama:\(submodel.modelId)"
         case let .lmstudio(submodel):
             "lmstudio:\(submodel.modelId)"
+        case let .minimax(submodel):
+            "minimax:\(submodel.modelId)"
+        case let .minimaxCN(submodel):
+            "minimax-cn:\(submodel.modelId)"
         case let .openRouter(modelId):
             "openrouter:\(modelId)"
         case let .together(modelId):
@@ -237,46 +241,16 @@ public final class ModelCapabilityRegistry: @unchecked Sendable {
             excludedParameters: ["temperature", "topP", "frequencyPenalty", "presencePenalty"],
         )
 
-        self.capabilities["openai:gpt-5.1"] = gpt5Capabilities
-        self.capabilities["openai:gpt-5.2"] = gpt5Capabilities
+        self.capabilities["openai:chat-latest"] = gpt5Capabilities
+        self.capabilities["openai:gpt-5-chat-latest"] = gpt5Capabilities
+        self.capabilities["openai:gpt-5.5"] = gpt5Capabilities
+        self.capabilities["openai:gpt-5.4"] = gpt5Capabilities
+        self.capabilities["openai:gpt-5.4-mini"] = gpt5Capabilities
+        self.capabilities["openai:gpt-5.4-nano"] = gpt5Capabilities
         self.capabilities["openai:gpt-5"] = gpt5Capabilities
         self.capabilities["openai:gpt-5-pro"] = gpt5Capabilities
         self.capabilities["openai:gpt-5-mini"] = gpt5Capabilities
         self.capabilities["openai:gpt-5-nano"] = gpt5Capabilities
-        self.capabilities["openai:gpt-5-thinking"] = gpt5Capabilities
-        self.capabilities["openai:gpt-5-thinking-mini"] = gpt5Capabilities
-        self.capabilities["openai:gpt-5-thinking-nano"] = gpt5Capabilities
-        self.capabilities["openai:gpt-5-chat-latest"] = gpt5Capabilities
-
-        // O4/GPT-5 reasoning models (fixed temperature, reasoning effort)
-        let reasoningCapabilities = ModelParameterCapabilities(
-            supportsTemperature: false,
-            supportsTopP: false,
-            supportedProviderOptions: .init(
-                supportsReasoningEffort: true,
-                supportsPreviousResponseId: true,
-            ),
-            forcedTemperature: 1.0,
-            excludedParameters: ["temperature", "topP"],
-        )
-
-        self.capabilities["openai:o4"] = reasoningCapabilities
-        self.capabilities["openai:o4-mini"] = reasoningCapabilities
-
-        // Standard GPT-4 models
-        let gpt4Capabilities = ModelParameterCapabilities(
-            supportedProviderOptions: .init(
-                supportsParallelToolCalls: true,
-                supportsResponseFormat: true,
-                supportsLogprobs: true,
-            ),
-        )
-
-        self.capabilities["openai:gpt-4o"] = gpt4Capabilities
-        self.capabilities["openai:gpt-4o-mini"] = gpt4Capabilities
-        self.capabilities["openai:gpt-4.1"] = gpt4Capabilities
-        self.capabilities["openai:gpt-4.1-mini"] = gpt4Capabilities
-        self.capabilities["openai:gpt-4-turbo"] = gpt4Capabilities
 
         // Claude 4 models with thinking
         let claude4Capabilities = ModelParameterCapabilities(
@@ -286,11 +260,25 @@ public final class ModelCapabilityRegistry: @unchecked Sendable {
             ),
         )
 
+        // Fable 5 and Opus 4.7+ map requested thinking to Anthropic's adaptive thinking request shape.
+        let claudeAdaptiveThinkingCapabilities = ModelParameterCapabilities(
+            supportsTemperature: false,
+            supportsTopP: false,
+            supportsTopK: false,
+            supportedProviderOptions: .init(
+                supportsThinking: true,
+                supportsCacheControl: true,
+            ),
+            excludedParameters: ["temperature", "topP", "topK"],
+        )
+        self.capabilities["anthropic:claude-fable-5"] = claudeAdaptiveThinkingCapabilities
+        self.capabilities["anthropic:claude-opus-4-8"] = claudeAdaptiveThinkingCapabilities
+        self.capabilities["anthropic:claude-opus-4-7"] = claudeAdaptiveThinkingCapabilities
         self.capabilities["anthropic:claude-opus-4-5"] = claude4Capabilities
         self.capabilities["anthropic:claude-opus-4-1-20250805"] = claude4Capabilities
-        self.capabilities["anthropic:claude-sonnet-4-20250514"] = claude4Capabilities
+        self.capabilities["anthropic:claude-sonnet-4-6"] = claude4Capabilities
         self.capabilities["anthropic:claude-sonnet-4-5-20250929"] = claude4Capabilities
-        self.capabilities["anthropic:claude-haiku-4.5"] = claude4Capabilities
+        self.capabilities["anthropic:claude-haiku-4-5"] = claude4Capabilities
 
         // Google Gemini with thinking
         let geminiCapabilities = ModelParameterCapabilities(
@@ -301,6 +289,9 @@ public final class ModelCapabilityRegistry: @unchecked Sendable {
             ),
         )
 
+        self.capabilities["google:gemini-3.5-flash"] = geminiCapabilities
+        self.capabilities["google:gemini-3.1-pro-preview"] = geminiCapabilities
+        self.capabilities["google:gemini-3.1-flash-lite"] = geminiCapabilities
         self.capabilities["google:gemini-2.5-pro"] = geminiCapabilities
         self.capabilities["google:gemini-2.5-flash"] = geminiCapabilities
         self.capabilities["google:gemini-2.5-flash-lite"] = geminiCapabilities
@@ -314,8 +305,12 @@ public final class ModelCapabilityRegistry: @unchecked Sendable {
             ),
         )
 
-        self.capabilities["mistral:mistral-large-2"] = mistralCapabilities
-        self.capabilities["mistral:codestral"] = mistralCapabilities
+        self.capabilities["mistral:mistral-large-latest"] = mistralCapabilities
+        self.capabilities["mistral:mistral-medium-latest"] = mistralCapabilities
+        self.capabilities["mistral:mistral-medium-3-5"] = mistralCapabilities
+        self.capabilities["mistral:mistral-small-latest"] = mistralCapabilities
+        self.capabilities["mistral:open-mistral-nemo-2407"] = mistralCapabilities
+        self.capabilities["mistral:codestral-latest"] = mistralCapabilities
 
         // Groq models (ultra-fast inference)
         let groqCapabilities = ModelParameterCapabilities(
@@ -324,12 +319,12 @@ public final class ModelCapabilityRegistry: @unchecked Sendable {
             ),
         )
 
-        self.capabilities["groq:llama-3.1-70b"] = groqCapabilities
-        self.capabilities["groq:llama-3.1-8b"] = groqCapabilities
-        self.capabilities["groq:llama-3-70b"] = groqCapabilities
-        self.capabilities["groq:llama-3-8b"] = groqCapabilities
-        self.capabilities["groq:mixtral-8x7b"] = groqCapabilities
-        self.capabilities["groq:gemma2-9b"] = groqCapabilities
+        self.capabilities["groq:openai/gpt-oss-120b"] = groqCapabilities
+        self.capabilities["groq:openai/gpt-oss-20b"] = groqCapabilities
+        self.capabilities["groq:llama-3.3-70b-versatile"] = groqCapabilities
+        self.capabilities["groq:llama-3.1-8b-instant"] = groqCapabilities
+        self.capabilities["groq:meta-llama/llama-4-maverick-17b-128e-instruct"] = groqCapabilities
+        self.capabilities["groq:meta-llama/llama-4-scout-17b-16e-instruct"] = groqCapabilities
 
         // Grok models
         let grokCapabilities = ModelParameterCapabilities(
@@ -339,17 +334,9 @@ public final class ModelCapabilityRegistry: @unchecked Sendable {
             ),
         )
 
-        self.capabilities["grok:grok-4-0709"] = grokCapabilities
-        self.capabilities["grok:grok-4-fast-reasoning"] = grokCapabilities
-        self.capabilities["grok:grok-4-fast-non-reasoning"] = grokCapabilities
-        self.capabilities["grok:grok-code-fast-1"] = grokCapabilities
-        self.capabilities["grok:grok-3"] = grokCapabilities
-        self.capabilities["grok:grok-3-mini"] = grokCapabilities
-        self.capabilities["grok:grok-2-1212"] = grokCapabilities
-        self.capabilities["grok:grok-2-vision-1212"] = grokCapabilities
-        self.capabilities["grok:grok-2-image-1212"] = grokCapabilities
-        self.capabilities["grok:grok-vision-beta"] = grokCapabilities
-        self.capabilities["grok:grok-beta"] = grokCapabilities
+        self.capabilities["grok:grok-4.3"] = grokCapabilities
+        self.capabilities["grok:grok-4.20-0309-reasoning"] = grokCapabilities
+        self.capabilities["grok:grok-4.20-0309-non-reasoning"] = grokCapabilities
     }
 
     private func defaultCapabilities(for model: LanguageModel) -> ModelParameterCapabilities {
@@ -359,6 +346,19 @@ public final class ModelCapabilityRegistry: @unchecked Sendable {
             return registered
         }
 
+        if self.isAnthropicFableCompatible(model) {
+            return ModelParameterCapabilities(
+                supportsTemperature: false,
+                supportsTopP: false,
+                supportsTopK: false,
+                supportedProviderOptions: .init(
+                    supportsThinking: true,
+                    supportsCacheControl: true,
+                ),
+                excludedParameters: ["temperature", "topP", "topK"],
+            )
+        }
+
         // Return provider-based defaults
         switch model {
         case .openai:
@@ -369,6 +369,7 @@ public final class ModelCapabilityRegistry: @unchecked Sendable {
             // Default Anthropic capabilities
             return ModelParameterCapabilities(
                 supportedProviderOptions: .init(
+                    supportsThinking: true,
                     supportsCacheControl: true,
                 ),
             )
@@ -388,11 +389,38 @@ public final class ModelCapabilityRegistry: @unchecked Sendable {
                 supportsSeed: true,
             )
 
+        case .minimax, .minimaxCN:
+            return ModelParameterCapabilities()
+
         default:
             // Default capabilities for unknown models
             return ModelParameterCapabilities()
         }
     }
+
+    private func isAnthropicFableCompatible(_ model: LanguageModel) -> Bool {
+        switch model {
+        case let .anthropic(anthropic):
+            return LanguageModel.Anthropic.isFable(modelId: anthropic.modelId)
+        case let .anthropicCompatible(modelId, _):
+            return LanguageModel.Anthropic.isFable(modelId: modelId)
+        case let .openRouter(modelId),
+             let .openaiCompatible(modelId, _),
+             let .together(modelId):
+            return LanguageModel.Anthropic.isFable(modelId: modelId)
+        case let .custom(provider):
+            guard
+                let parsed = ProviderParser.parse(provider.modelId),
+                LanguageModel.Anthropic.isFable(modelId: parsed.model),
+                CustomProviderRegistry.shared.get(parsed.provider)?.kind == .anthropic else
+            {
+                return false
+            }
+            return true
+        default:
+            return false
+        }
+    }
 }
 
 // MARK: - GenerationSettings Extension
@@ -406,6 +434,7 @@ extension GenerationSettings {
 
         var adjustedTemperature = temperature
         var adjustedTopP = topP
+        var adjustedTopK = topK
         var adjustedFrequencyPenalty = frequencyPenalty
         var adjustedPresencePenalty = presencePenalty
         var adjustedProviderOptions = providerOptions
@@ -417,6 +446,9 @@ extension GenerationSettings {
         if capabilities.excludedParameters.contains("topP") {
             adjustedTopP = nil
         }
+        if capabilities.excludedParameters.contains("topK") {
+            adjustedTopK = nil
+        }
         if capabilities.excludedParameters.contains("frequencyPenalty") {
             adjustedFrequencyPenalty = nil
         }
@@ -440,7 +472,7 @@ extension GenerationSettings {
             maxTokens: maxTokens,
             temperature: adjustedTemperature,
             topP: adjustedTopP,
-            topK: topK,
+            topK: adjustedTopK,
             frequencyPenalty: adjustedFrequencyPenalty,
             presencePenalty: adjustedPresencePenalty,
             stopSequences: stopSequences,
@@ -448,6 +480,7 @@ extension GenerationSettings {
             stopConditions: stopConditions,
             seed: seed,
             providerOptions: adjustedProviderOptions,
+            streamBuffering: self.streamBuffering,
         )
     }
 

Sources/Tachikoma/Core/OpenAICompatibleHelper.swift

@@ -36,14 +36,24 @@ struct OpenAICompatibleHelper {
         }
 
         // Extract stop sequences from stop conditions
-        let stopSequences = Self.extractStopSequences(from: request.settings.stopConditions)
+        let settings = Self.validatedSettings(
+            request.settings,
+            providerName: providerName,
+            modelId: modelId,
+            baseURL: baseURL,
+        )
+        let stopSequences = Self.extractStopSequences(from: settings.stopConditions)
 
         // Convert request to OpenAI-compatible format
         let openAIRequest = try OpenAIChatRequest(
             model: modelId,
-            messages: convertMessages(request.messages),
-            temperature: request.settings.temperature,
-            maxTokens: request.settings.maxTokens,
+            messages: convertMessages(
+                request.messages,
+                replayOpenRouterReasoningForModel: providerName == "OpenRouter" ? modelId : nil,
+                replayOpenRouterReasoningForBaseURL: providerName == "OpenRouter" ? baseURL : nil,
+            ),
+            temperature: settings.temperature,
+            maxTokens: settings.maxTokens,
             tools: request.tools?.compactMap { try self.convertTool($0) },
             stream: false,
             stop: stopSequences.isEmpty ? nil : stopSequences,
@@ -100,14 +110,9 @@ struct OpenAICompatibleHelper {
         let usage = openAIResponse.usage.map {
             Usage(inputTokens: $0.promptTokens ?? 0, outputTokens: $0.completionTokens ?? 0)
         }
+        let reasoning = Self.reasoningBlocks(from: choice.message)
 
-        let finishReason: FinishReason? = switch choice.finishReason {
-        case "stop": .stop
-        case "length": .length
-        case "tool_calls": .toolCalls
-        case "content_filter": .contentFilter
-        default: .other
-        }
+        let finishReason = Self.mapFinishReason(choice.finishReason)
 
         // Convert tool calls if present
         let toolCalls = choice.message.toolCalls?.compactMap { openAIToolCall -> AgentToolCall? in
@@ -142,6 +147,7 @@ struct OpenAICompatibleHelper {
             usage: usage,
             finishReason: finishReason,
             toolCalls: toolCalls,
+            reasoning: reasoning,
         )
     }
 
@@ -173,14 +179,27 @@ struct OpenAICompatibleHelper {
         }
 
         // Extract stop sequences from stop conditions
-        let stopSequences = Self.extractStopSequences(from: request.settings.stopConditions)
+        guard !LanguageModel.Anthropic.hasStreamingRefusalRisk(modelId: modelId) else {
+            throw TachikomaError.invalidConfiguration("\(modelId) does not support streaming")
+        }
+        let settings = Self.validatedSettings(
+            request.settings,
+            providerName: providerName,
+            modelId: modelId,
+            baseURL: baseURL,
+        )
+        let stopSequences = Self.extractStopSequences(from: settings.stopConditions)
 
         // Convert request to OpenAI-compatible format
         let openAIRequest = try OpenAIChatRequest(
             model: modelId,
-            messages: convertMessages(request.messages),
-            temperature: request.settings.temperature,
-            maxTokens: request.settings.maxTokens,
+            messages: convertMessages(
+                request.messages,
+                replayOpenRouterReasoningForModel: providerName == "OpenRouter" ? modelId : nil,
+                replayOpenRouterReasoningForBaseURL: providerName == "OpenRouter" ? baseURL : nil,
+            ),
+            temperature: settings.temperature,
+            maxTokens: settings.maxTokens,
             tools: request.tools?.compactMap { try self.convertTool($0) },
             stream: true,
             stop: stopSequences.isEmpty ? nil : stopSequences,
@@ -346,8 +365,10 @@ struct OpenAICompatibleHelper {
                                         }
                                     }
 
-                                    if choice.finishReason != nil {
-                                        continuation.yield(TextStreamDelta.done())
+                                    if let finishReason = choice.finishReason {
+                                        continuation.yield(TextStreamDelta.done(
+                                            finishReason: Self.mapFinishReason(finishReason),
+                                        ))
                                         break
                                     }
                                 }
@@ -427,9 +448,10 @@ struct OpenAICompatibleHelper {
                                     }
 
                                     if let finishReason = choice.finishReason {
-                                        if finishReason == "stop" || finishReason == "tool_calls" {
-                                            continuation.yield(TextStreamDelta.done())
-                                        }
+                                        continuation.yield(TextStreamDelta.done(
+                                            finishReason: Self.mapFinishReason(finishReason),
+                                        ))
+                                        break
                                     }
                                 }
                             } catch {
@@ -456,6 +478,39 @@ struct OpenAICompatibleHelper {
 
     // MARK: - Helper Methods
 
+    private static func mapFinishReason(_ reason: String?) -> FinishReason? {
+        switch reason {
+        case "stop": .stop
+        case "length": .length
+        case "tool_calls": .toolCalls
+        case "content_filter": .contentFilter
+        case nil: nil
+        default: .other
+        }
+    }
+
+    private static func validatedSettings(
+        _ settings: GenerationSettings,
+        providerName: String,
+        modelId: String,
+        baseURL: String,
+    )
+        -> GenerationSettings
+    {
+        settings.validated(for: self.languageModel(providerName: providerName, modelId: modelId, baseURL: baseURL))
+    }
+
+    private static func languageModel(providerName: String, modelId: String, baseURL: String) -> LanguageModel {
+        switch providerName.lowercased() {
+        case "openrouter":
+            .openRouter(modelId: modelId)
+        case "together":
+            .together(modelId: modelId)
+        default:
+            .openaiCompatible(modelId: modelId, baseURL: baseURL)
+        }
+    }
+
     /// Extract native stop sequences from stop conditions
     private static func extractStopSequences(from stopCondition: (any StopCondition)?) -> [String] {
         // Extract native stop sequences from stop conditions
@@ -512,18 +567,55 @@ struct OpenAICompatibleHelper {
         }
     }
 
-    private static func convertMessages(_ messages: [ModelMessage]) throws -> [OpenAIChatMessage] {
-        messages.map { message in
+    private static func convertMessages(
+        _ messages: [ModelMessage],
+        replayOpenRouterReasoningForModel modelId: String?,
+        replayOpenRouterReasoningForBaseURL baseURL: String?,
+    ) throws
+        -> [OpenAIChatMessage]
+    {
+        var converted: [OpenAIChatMessage] = []
+        var pendingReasoningDetails: [JSONValue] = []
+        var pendingReasoningText: [String] = []
+        let endpointIdentity = ReasoningEndpointIdentity.canonical(baseURL)
+
+        for message in messages {
+            if
+                message.channel == .thinking,
+                let customData = message.metadata?.customData,
+                customData["tachikoma.reasoning.provider"] == "openrouter",
+                customData["tachikoma.reasoning.model"] == modelId,
+                customData["tachikoma.reasoning.base_url"] == endpointIdentity,
+                let rawReasoningDetails = customData["openrouter.reasoning_details"]
+            {
+                pendingReasoningDetails.append(contentsOf: Self.decodeReasoningDetails(rawReasoningDetails))
+                continue
+            }
+            if
+                message.channel == .thinking,
+                let customData = message.metadata?.customData,
+                customData["tachikoma.reasoning.provider"] == "openrouter",
+                customData["tachikoma.reasoning.model"] == modelId,
+                customData["tachikoma.reasoning.base_url"] == endpointIdentity,
+                let reasoning = customData["openrouter.reasoning"]
+            {
+                pendingReasoningText.append(reasoning)
+                continue
+            }
+            if message.channel == .thinking {
+                continue
+            }
+
             switch message.role {
             case .system:
-                return OpenAIChatMessage(role: "system", content: message.content.compactMap { part in
+                converted.append(OpenAIChatMessage(role: "system", content: message.content.compactMap { part in
                     if case let .text(text) = part { return text }
                     return nil
-                }.joined())
+                }.joined()))
             case .user:
                 if message.content.count == 1, case let .text(text) = message.content.first! {
                     // Simple text message
-                    return OpenAIChatMessage(role: "user", content: text)
+                    converted.append(OpenAIChatMessage(role: "user", content: text))
                 } else {
                     // Multi-modal message
                     let content = message.content.compactMap { contentPart -> OpenAIChatMessageContent? in
@@ -540,7 +632,7 @@ struct OpenAICompatibleHelper {
                             return nil // Skip tool calls and results in user messages
                         }
                     }
-                    return OpenAIChatMessage(role: "user", content: content)
+                    converted.append(OpenAIChatMessage(role: "user", content: content))
                 }
             case .assistant:
                 // Check if this assistant message contains tool calls
@@ -571,15 +663,25 @@ struct OpenAICompatibleHelper {
 
                 // If we have tool calls, create a message with tool calls
                 if !toolCalls.isEmpty {
-                    return OpenAIChatMessage(
+                    converted.append(OpenAIChatMessage(
                         role: "assistant",
                         content: textContent.isEmpty ? nil : textContent,
                         toolCalls: toolCalls,
-                    )
+                        reasoning: pendingReasoningText.isEmpty ? nil : pendingReasoningText.joined(separator: "\n"),
+                        reasoningDetails: pendingReasoningDetails.isEmpty ? nil : pendingReasoningDetails,
+                    ))
                 } else {
                     // Regular text message
-                    return OpenAIChatMessage(role: "assistant", content: textContent)
+                    converted.append(OpenAIChatMessage(
+                        role: "assistant",
+                        content: textContent,
+                        toolCalls: nil,
+                        reasoning: pendingReasoningText.isEmpty ? nil : pendingReasoningText.joined(separator: "\n"),
+                        reasoningDetails: pendingReasoningDetails.isEmpty ? nil : pendingReasoningDetails,
+                    ))
                 }
+                pendingReasoningText.removeAll()
+                pendingReasoningDetails.removeAll()
             case .tool:
                 // Extract tool call ID and result content from tool result
                 var toolCallId: String?
@@ -598,9 +700,44 @@ struct OpenAICompatibleHelper {
                     }
                 }
 
-                return OpenAIChatMessage(role: "tool", content: resultContent, toolCallId: toolCallId)
+                converted.append(OpenAIChatMessage(role: "tool", content: resultContent, toolCallId: toolCallId))
             }
         }
+
+        return converted
+    }
+
+    private static func reasoningBlocks(from message: OpenAIChatResponse.Message) -> [ProviderReasoningBlock] {
+        var blocks: [ProviderReasoningBlock] = []
+        if let details = message.reasoningDetails, !details.isEmpty {
+            blocks.append(ProviderReasoningBlock(
+                text: message.reasoning ?? "",
+                type: "openrouter_reasoning_details",
+                rawJSON: Self.encodeReasoningDetails(details),
+            ))
+        } else if let reasoning = message.reasoning, !reasoning.isEmpty {
+            blocks.append(ProviderReasoningBlock(
+                text: reasoning,
+                type: "openrouter_reasoning",
+                rawJSON: nil,
+            ))
+        }
+        return blocks
+    }
+
+    private static func encodeReasoningDetails(_ details: [JSONValue]) -> String? {
+        guard let data = try? JSONEncoder().encode(details) else { return nil }
+        return String(data: data, encoding: .utf8)
+    }
+
+    private static func decodeReasoningDetails(_ rawJSON: String) -> [JSONValue] {
+        guard
+            let data = rawJSON.data(using: .utf8),
+            let details = try? JSONDecoder().decode([JSONValue].self, from: data) else
+        {
+            return []
+        }
+        return details
     }
 
     private static func convertTool(_ tool: AgentTool) throws -> OpenAITool {

Sources/Tachikoma/Core/Provider.swift

@@ -41,6 +41,12 @@ public enum Provider: Sendable, Hashable, Codable {
     /// Google provider (Gemini models)
     case google
 
+    /// MiniMax provider (Anthropic-compatible hosted models)
+    case minimax
+
+    /// MiniMax China provider (Anthropic-compatible hosted models)
+    case minimaxCN
+
     /// Ollama provider (local model hosting)
     case ollama
 
@@ -62,6 +68,8 @@ public enum Provider: Sendable, Hashable, Codable {
         case .groq: "groq"
         case .mistral: "mistral"
         case .google: "google"
+        case .minimax: "minimax"
+        case .minimaxCN: "minimax-cn"
         case .ollama: "ollama"
         case .lmstudio: "lmstudio"
         case .azureOpenAI: "azure-openai"
@@ -78,6 +86,8 @@ public enum Provider: Sendable, Hashable, Codable {
         case .groq: "Groq"
         case .mistral: "Mistral"
         case .google: "Google"
+        case .minimax: "MiniMax"
+        case .minimaxCN: "MiniMax China"
         case .ollama: "Ollama"
         case .lmstudio: "LMStudio"
         case .azureOpenAI: "Azure OpenAI"
@@ -94,6 +104,8 @@ public enum Provider: Sendable, Hashable, Codable {
         case .groq: "GROQ_API_KEY"
         case .mistral: "MISTRAL_API_KEY"
         case .google: "GEMINI_API_KEY"
+        case .minimax: "MINIMAX_API_KEY"
+        case .minimaxCN: "MINIMAX_CN_API_KEY"
         case .ollama: "OLLAMA_API_KEY"
         case .lmstudio: "" // LMStudio doesn't need API keys
         case .azureOpenAI: "AZURE_OPENAI_API_KEY"
@@ -106,6 +118,7 @@ public enum Provider: Sendable, Hashable, Codable {
         switch self {
         case .grok: ["XAI_API_KEY", "GROK_API_KEY"] // Additional Grok aliases
         case .google: ["GOOGLE_API_KEY"] // Backwards compatibility
+        case .minimaxCN: ["MINIMAX_API_KEY"]
         case .azureOpenAI: ["AZURE_OPENAI_TOKEN", "AZURE_OPENAI_BEARER_TOKEN"]
         default: []
         }
@@ -120,6 +133,8 @@ public enum Provider: Sendable, Hashable, Codable {
         case .groq: "https://api.groq.com/openai/v1"
         case .mistral: "https://api.mistral.ai/v1"
         case .google: "https://generativelanguage.googleapis.com/v1beta"
+        case .minimax: "https://api.minimax.io/anthropic"
+        case .minimaxCN: "https://api.minimaxi.com/anthropic"
         case .ollama: "http://localhost:11434"
         case .lmstudio: "http://localhost:1234/v1"
         case .azureOpenAI: nil // Requires resource or endpoint
@@ -139,7 +154,7 @@ public enum Provider: Sendable, Hashable, Codable {
 
     /// All standard providers (excludes custom)
     public static var standardProviders: [Provider] {
-        [.openai, .anthropic, .grok, .groq, .mistral, .google, .ollama, .azureOpenAI]
+        [.openai, .anthropic, .grok, .groq, .mistral, .google, .minimax, .minimaxCN, .ollama, .azureOpenAI]
     }
 
     /// Create provider from string identifier
@@ -151,7 +166,9 @@ public enum Provider: Sendable, Hashable, Codable {
         case "grok", "xai": .grok
         case "groq": .groq
         case "mistral": .mistral
-        case "google": .google
+        case "google", "gemini": .google
+        case "minimax": .minimax
+        case "minimax-cn", "minimax_cn", "minimaxi": .minimaxCN
         case "ollama": .ollama
         case "azure-openai", "azure_openai", "azureopenai": .azureOpenAI
         default: .custom(identifier)

Sources/Tachikoma/Core/ProviderOptions.swift

@@ -56,7 +56,7 @@ public struct OpenAIOptions: Sendable, Codable {
     /// Verbosity level for GPT-5 models
     public var verbosity: Verbosity?
 
-    /// Reasoning effort for O3/O4 models
+    /// Reasoning effort for GPT-5 models
     public var reasoningEffort: ReasoningEffort?
 
     /// Previous response ID for Responses API chaining
@@ -159,6 +159,7 @@ public struct AnthropicOptions: Sendable, Codable {
     public enum ThinkingMode: Sendable, Codable {
         case disabled
         case enabled(budgetTokens: Int)
+        case adaptive
 
         private enum CodingKeys: String, CodingKey {
             case type
@@ -174,6 +175,8 @@ public struct AnthropicOptions: Sendable, Codable {
             case "enabled":
                 let budget = try container.decode(Int.self, forKey: .budgetTokens)
                 self = .enabled(budgetTokens: budget)
+            case "adaptive":
+                self = .adaptive
             default:
                 throw DecodingError.dataCorruptedError(
                     forKey: .type,
@@ -191,6 +194,8 @@ public struct AnthropicOptions: Sendable, Codable {
             case let .enabled(budgetTokens):
                 try container.encode("enabled", forKey: .type)
                 try container.encode(budgetTokens, forKey: .budgetTokens)
+            case .adaptive:
+                try container.encode("adaptive", forKey: .type)
             }
         }
     }

Sources/Tachikoma/Core/StopConditions.swift

@@ -13,6 +13,17 @@ public protocol StopCondition: Sendable {
     func reset() async
 }
 
+@available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
+protocol StableCacheKeyStopCondition {
+    var stableCacheKey: String? { get }
+}
+
+@available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
+private func compositeStableCacheKey(kind: String, children: [String]) -> String {
+    let encodedChildren = children.map { "\($0.utf8.count):\($0)" }.joined()
+    return "\(kind):[\(encodedChildren)]"
+}
+
 // MARK: - Built-in Stop Conditions
 
 /// Stop when a specific string is encountered
@@ -39,6 +50,13 @@ public struct StringStopCondition: StopCondition {
     public func reset() async {}
 }
 
+@available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
+extension StringStopCondition: StableCacheKeyStopCondition {
+    var stableCacheKey: String? {
+        "string:\(self.caseSensitive):\(self.stopString)"
+    }
+}
+
 /// Stop when a regex pattern is matched
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
 public struct RegexStopCondition: StopCondition {
@@ -82,6 +100,13 @@ public struct RegexStopCondition: StopCondition {
     }
 }
 
+@available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
+extension RegexStopCondition: StableCacheKeyStopCondition {
+    var stableCacheKey: String? {
+        "regex:\(self.pattern)"
+    }
+}
+
 /// Stop after a certain number of tokens
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
 public actor TokenCountStopCondition: StopCondition {
@@ -182,6 +207,15 @@ public struct AnyStopCondition: StopCondition {
     }
 }
 
+@available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
+extension AnyStopCondition: StableCacheKeyStopCondition {
+    var stableCacheKey: String? {
+        let keys = self.conditions.compactMap { ($0 as? StableCacheKeyStopCondition)?.stableCacheKey }
+        guard keys.count == self.conditions.count else { return nil }
+        return compositeStableCacheKey(kind: "any", children: keys)
+    }
+}
+
 /// Stop when all conditions are met
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
 public struct AllStopCondition: StopCondition {
@@ -211,6 +245,15 @@ public struct AllStopCondition: StopCondition {
     }
 }
 
+@available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
+extension AllStopCondition: StableCacheKeyStopCondition {
+    var stableCacheKey: String? {
+        let keys = self.conditions.compactMap { ($0 as? StableCacheKeyStopCondition)?.stableCacheKey }
+        guard keys.count == self.conditions.count else { return nil }
+        return compositeStableCacheKey(kind: "all", children: keys)
+    }
+}
+
 // MARK: - Stateful Stop Conditions
 
 /// Stop when a pattern appears consecutively N times
@@ -386,6 +429,13 @@ public struct NeverStopCondition: StopCondition {
     public func reset() async {}
 }
 
+@available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
+extension NeverStopCondition: StableCacheKeyStopCondition {
+    var stableCacheKey: String? {
+        "never"
+    }
+}
+
 // MARK: - Integration with Generation Functions
 
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
@@ -431,9 +481,8 @@ extension AsyncThrowingStream where Element == TextStreamDelta {
 
                             // Check stop condition
                             if await condition.shouldStop(text: accumulatedText, delta: content) {
-                                // Yield the current delta then stop
                                 continuation.yield(delta)
-                                continuation.yield(TextStreamDelta.done())
+                                continuation.yield(TextStreamDelta.done(finishReason: .stop))
                                 continuation.finish()
                                 return
                             }

Sources/Tachikoma/Core/Tachikoma.swift

@@ -27,7 +27,7 @@ import Foundation
 //
 // ```swift
 // // Simple generation
-// let answer = try await generate("What is 2+2?", using: .openai(.gpt4o))
+// let answer = try await generate("What is 2+2?", using: .openai(.gpt55))
 //
 // // Conversation management
 // let conversation = Conversation()
@@ -114,7 +114,7 @@ public enum API {
     /// Model selection system
     public enum Models {
         /// Type-safe model selection
-        /// - `.openai(.gpt4o)`, `.anthropic(.opus4)`, `.grok(.grok4)`, `.ollama(.llama3_3)`
+        /// - `.openai(.gpt55)`, `.anthropic(.opus47)`, `.grok(.grok43)`, `.ollama(.llama3_3)`
         public static let typed = "Provider-specific model enums"
 
         /// Custom endpoints
@@ -196,16 +196,16 @@ public enum API {
 /// Migration guide from legacy API to modern API
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
 public enum MigrationGuide {
-    /// Legacy: `Tachikoma.shared.getModel("gpt-4").getResponse(request)`
-    /// Modern: `generate("Hello", using: .openai(.gpt4o))`
+    /// Legacy: `Tachikoma.shared.getModel("gpt-5.5").getResponse(request)`
+    /// Modern: `generate("Hello", using: .openai(.gpt55))`
     public static let simpleGeneration = """
     // OLD (deprecated)
-    let model = try await Tachikoma.shared.getModel("gpt-4")
+    let model = try await Tachikoma.shared.getModel("gpt-5.5")
     let request = ModelRequest(messages: [.user(content: .text("Hello"))], settings: .default)
     let response = try await model.getResponse(request: request)
 
     // NEW (modern)
-    let response = try await generate("Hello", using: .openai(.gpt4o))
+    let response = try await generate("Hello", using: .openai(.gpt55))
     """
 
     /// Legacy: Complex ModelRequest/ModelResponse handling

Sources/Tachikoma/Core/Types.swift

@@ -186,8 +186,9 @@ public enum OpenAIAPIMode: String, Sendable, CaseIterable {
     public static func defaultMode(for model: LanguageModel.OpenAI) -> OpenAIAPIMode {
         // Determine default API mode for a given model
         switch model {
-        case .o4Mini, .gpt5, .gpt5Pro, .gpt5Mini, .gpt5Nano, .gpt51, .gpt52:
-            .responses // Reasoning models and GPT-5 default to Responses API
+        case .chatLatest, .gpt5ChatLatest,
+             .gpt5, .gpt5Pro, .gpt5Mini, .gpt5Nano, .gpt54, .gpt54Mini, .gpt54Nano, .gpt55:
+            .responses // GPT-5 defaults to Responses API
         default:
             .chat // All other models use Chat Completions API
         }
@@ -325,6 +326,11 @@ public enum ImageInput: Sendable {
 /// Settings for text generation
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
 public struct GenerationSettings: Sendable {
+    public enum StreamBufferingMode: String, Sendable, Codable {
+        case incremental
+        case untilTerminal
+    }
+
     public let maxTokens: Int?
     public let temperature: Double?
     public let topP: Double?
@@ -336,6 +342,7 @@ public struct GenerationSettings: Sendable {
     public let stopConditions: (any StopCondition)?
     public let seed: Int?
     public let providerOptions: ProviderOptions
+    public let streamBuffering: StreamBufferingMode
 
     public init(
         maxTokens: Int? = nil,
@@ -349,6 +356,7 @@ public struct GenerationSettings: Sendable {
         stopConditions: (any StopCondition)? = nil,
         seed: Int? = nil,
         providerOptions: ProviderOptions = .init(),
+        streamBuffering: StreamBufferingMode = .incremental,
     ) {
         self.maxTokens = maxTokens
         self.temperature = temperature
@@ -361,9 +369,27 @@ public struct GenerationSettings: Sendable {
         self.stopConditions = stopConditions
         self.seed = seed
         self.providerOptions = providerOptions
+        self.streamBuffering = streamBuffering
     }
 
     public static let `default` = GenerationSettings()
+
+    public func withStreamBuffering(_ mode: StreamBufferingMode) -> GenerationSettings {
+        GenerationSettings(
+            maxTokens: self.maxTokens,
+            temperature: self.temperature,
+            topP: self.topP,
+            topK: self.topK,
+            frequencyPenalty: self.frequencyPenalty,
+            presencePenalty: self.presencePenalty,
+            stopSequences: self.stopSequences,
+            reasoningEffort: self.reasoningEffort,
+            stopConditions: self.stopConditions,
+            seed: self.seed,
+            providerOptions: self.providerOptions,
+            streamBuffering: mode,
+        )
+    }
 }
 
 /// Manual Codable conformance excluding non-codable stopConditions
@@ -379,6 +405,7 @@ extension GenerationSettings: Codable {
         case reasoningEffort
         case seed
         case providerOptions
+        case streamBuffering
     }
 
     public init(from decoder: Decoder) throws {
@@ -393,6 +420,8 @@ extension GenerationSettings: Codable {
         self.reasoningEffort = try container.decodeIfPresent(ReasoningEffort.self, forKey: .reasoningEffort)
         self.seed = try container.decodeIfPresent(Int.self, forKey: .seed)
         self.providerOptions = try container.decodeIfPresent(ProviderOptions.self, forKey: .providerOptions) ?? .init()
+        self.streamBuffering = try container
+            .decodeIfPresent(StreamBufferingMode.self, forKey: .streamBuffering) ?? .incremental
         self.stopConditions = nil // Can't decode function types
     }
 
@@ -408,6 +437,7 @@ extension GenerationSettings: Codable {
         try container.encodeIfPresent(self.reasoningEffort, forKey: .reasoningEffort)
         try container.encodeIfPresent(self.seed, forKey: .seed)
         try container.encode(self.providerOptions, forKey: .providerOptions)
+        try container.encode(self.streamBuffering, forKey: .streamBuffering)
         // Don't encode stopConditions since it can't be serialized
     }
 }
@@ -614,7 +644,7 @@ public enum ResponseChannel: String, Sendable, Codable, CaseIterable {
     case final // Final answer to the user
 }
 
-/// Reasoning effort level for models that support it (o3, opus-4, etc.)
+/// Reasoning effort level for models that support it (GPT-5 thinking, opus-4, etc.)
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
 public enum ReasoningEffort: String, Sendable, Codable, CaseIterable {
     case low

Sources/Tachikoma/Core/UIIntegration.swift

@@ -132,7 +132,9 @@ extension [ModelMessage] {
     /// Convert model messages to UI messages for display
     public func toUIMessages() -> [UIMessage] {
         // Convert model messages to UI messages for display
-        map { modelMessage in
+        compactMap { modelMessage in
+            guard !modelMessage.isProviderNativeReasoningBlock else { return nil }
+            guard !modelMessage.isSyntheticReasoningBoundary else { return nil }
             var content = ""
             var attachments: [UIAttachment] = []
             var toolCalls: [AgentToolCall] = []
@@ -179,6 +181,20 @@ extension [ModelMessage] {
     }
 }
 
+extension ModelMessage {
+    fileprivate var isProviderNativeReasoningBlock: Bool {
+        guard channel == .thinking, let customData = metadata?.customData else { return false }
+        return customData["anthropic.thinking.model"] != nil ||
+            customData["anthropic.thinking.type"] != nil ||
+            customData["anthropic.thinking.signature"] != nil ||
+            customData["tachikoma.reasoning.provider"] != nil
+    }
+
+    fileprivate var isSyntheticReasoningBoundary: Bool {
+        metadata?.customData?["tachikoma.internal.boundary"] == "reasoning_only"
+    }
+}
+
 // MARK: - Streaming Extensions
 
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)

Sources/Tachikoma/Models/ModelInterface.swift

@@ -88,7 +88,7 @@ public struct ModelSettings: Sendable, Codable {
 
     /// Legacy initializer for backward compatibility
     public init(
-        modelName: String = "gpt-4",
+        modelName: String = "gpt-5.5",
         maxTokens: Int? = nil,
         temperature: Double? = nil,
         topP: Double? = nil,

Sources/Tachikoma/Models/ModelProvider.swift

@@ -84,6 +84,27 @@ public struct ProviderResponse: Sendable {
     public let usage: Usage?
     public let finishReason: FinishReason?
     public let toolCalls: [AgentToolCall]?
+    public let reasoning: [ProviderReasoningBlock]
+    public let assistantMessages: [ModelMessage]
+    public let isBillable: Bool
+
+    public init(
+        text: String,
+        usage: Usage? = nil,
+        finishReason: FinishReason? = nil,
+        toolCalls: [AgentToolCall]? = nil,
+        reasoning: [ProviderReasoningBlock] = [],
+        assistantMessages: [ModelMessage] = [],
+        isBillable: Bool = true,
+    ) {
+        self.text = text
+        self.usage = usage
+        self.finishReason = finishReason
+        self.toolCalls = toolCalls
+        self.reasoning = reasoning
+        self.assistantMessages = assistantMessages
+        self.isBillable = isBillable
+    }
 
     public init(
         text: String,
@@ -91,9 +112,30 @@ public struct ProviderResponse: Sendable {
         finishReason: FinishReason? = nil,
         toolCalls: [AgentToolCall]? = nil,
     ) {
-        self.text = text
-        self.usage = usage
-        self.finishReason = finishReason
-        self.toolCalls = toolCalls
+        self.init(
+            text: text,
+            usage: usage,
+            finishReason: finishReason,
+            toolCalls: toolCalls,
+            reasoning: [],
+            assistantMessages: [],
+            isBillable: true,
+        )
+    }
+}
+
+/// Provider-native signed reasoning block that must be replayed in later requests.
+@available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
+public struct ProviderReasoningBlock: Sendable, Equatable {
+    public let text: String
+    public let signature: String?
+    public let type: String
+    public let rawJSON: String?
+
+    public init(text: String, signature: String? = nil, type: String = "thinking", rawJSON: String? = nil) {
+        self.text = text
+        self.signature = signature
+        self.type = type
+        self.rawJSON = rawJSON
     }
 }

Sources/Tachikoma/Models/ModelSelection.swift

@@ -30,12 +30,61 @@ public struct ModelSelector {
             return .google(googleModel)
         }
 
+        // MiniMax shortcuts and models
+        if let miniMaxModel = parseMiniMaxCNModel(normalized) {
+            return .minimaxCN(miniMaxModel)
+        }
+
+        if let miniMaxModel = parseMiniMaxModel(normalized) {
+            return .minimax(miniMaxModel)
+        }
+
         // Grok shortcuts and models
         if let grokModel = parseGrokModel(normalized) {
             return .grok(grokModel)
         }
 
-        // Ollama shortcuts and models
+        if
+            let qualified = ProviderParser.parse(normalized),
+            ["grok", "xai", "x-ai"].contains(qualified.provider.lowercased())
+        {
+            let model = qualified.model.lowercased()
+            guard
+                !Self.isUnsupportedLegacyGrokModel(model),
+                let grokModel = self.parseGrokModel(model) else
+            {
+                throw ModelValidationError.unsupportedModel(modelString)
+            }
+            return .grok(grokModel)
+        }
+
+        if
+            Self.isUnsupportedLegacyOpenAIModel(normalized) ||
+            Self.isUnsupportedLegacyAnthropicModel(normalized) ||
+            Self.isUnsupportedLegacyGrokModel(normalized)
+        {
+            throw ModelValidationError.unsupportedModel(modelString)
+        }
+
+        // LM Studio shortcuts and models
+        if let lmStudioModel = parseLMStudioModel(normalized) {
+            return .lmstudio(lmStudioModel)
+        }
+
+        if let qualified = ProviderParser.parse(normalized) {
+            let provider = qualified.provider.lowercased()
+            if provider == "openai", let openaiModel = parseOpenAIModel(qualified.model.lowercased()) {
+                return .openai(openaiModel)
+            }
+            if provider == "openrouter" {
+                return .openRouter(modelId: qualified.model)
+            }
+            if !["ollama", "lmstudio", "lm-studio"].contains(provider) {
+                return .openRouter(modelId: normalized)
+            }
+        }
+
+        // Ollama shortcuts and models. Keep after explicit local providers because it accepts custom IDs.
         if let ollamaModel = parseOllamaModel(normalized) {
             return .ollama(ollamaModel)
         }
@@ -46,7 +95,7 @@ public struct ModelSelector {
         }
 
         // Custom model ID - try to infer provider
-        if normalized.contains("gpt") || normalized.contains("o3") || normalized.contains("o4") {
+        if normalized.contains("gpt-5") || normalized.contains("gpt5") {
             return .openai(.custom(normalized))
         }
 
@@ -66,20 +115,24 @@ public struct ModelSelector {
 
     private static func parseOpenAIModel(_ input: String) -> Model.OpenAI? {
         switch input {
-        // GPT-5.2 models
-        case "gpt-5.2", "gpt5.2", "gpt-5-2", "gpt5-2", "gpt52":
-            return .gpt52
-        case "gpt-5.2-mini", "gpt5.2-mini", "gpt52-mini", "gpt52mini", "gpt-5-2-mini", "gpt5-2-mini":
+        case "chat-latest", "chatlatest":
+            return .chatLatest
+        case "gpt-5-chat-latest", "gpt5-chat-latest", "gpt5chatlatest":
+            return .gpt5ChatLatest
+        // GPT-5.5 models
+        case "gpt-5.5", "gpt5.5", "gpt-5-5", "gpt5-5", "gpt55":
+            return .gpt55
+        case "gpt-5.5-mini", "gpt5.5-mini", "gpt55-mini", "gpt55mini", "gpt-5-5-mini", "gpt5-5-mini":
             return .gpt5Mini
-        case "gpt-5.2-nano", "gpt5.2-nano", "gpt52-nano", "gpt52nano", "gpt-5-2-nano", "gpt5-2-nano":
-            return .gpt5Nano
-        // GPT-5.1 models (latest)
-        case "gpt-5.1", "gpt5.1", "gpt-5-1", "gpt5-1", "gpt51":
-            return .gpt51
-        case "gpt-5.1-mini", "gpt5.1-mini", "gpt51-mini", "gpt51mini", "gpt-5-1-mini", "gpt5-1-mini":
-            return .gpt5Mini
-        case "gpt-5.1-nano", "gpt5.1-nano", "gpt51-nano", "gpt51nano", "gpt-5-1-nano", "gpt5-1-nano":
+        case "gpt-5.5-nano", "gpt5.5-nano", "gpt55-nano", "gpt55nano", "gpt-5-5-nano", "gpt5-5-nano":
             return .gpt5Nano
+        // GPT-5.4 models
+        case "gpt-5.4", "gpt5.4", "gpt-5-4", "gpt5-4", "gpt54":
+            return .gpt54
+        case "gpt-5.4-mini", "gpt5.4-mini", "gpt54-mini", "gpt54mini", "gpt-5-4-mini", "gpt5-4-mini":
+            return .gpt54Mini
+        case "gpt-5.4-nano", "gpt5.4-nano", "gpt54-nano", "gpt54nano", "gpt-5-4-nano", "gpt5-4-nano":
+            return .gpt54Nano
         // GPT-5 models
         case "gpt-5", "gpt5":
             return .gpt5
@@ -89,35 +142,17 @@ public struct ModelSelector {
             return .gpt5Mini
         case "gpt-5-nano", "gpt5-nano", "gpt5nano":
             return .gpt5Nano
-        case "gpt-5-thinking", "gpt5-thinking", "gpt5thinking":
-            return .gpt5Thinking
-        case "gpt-5-thinking-mini", "gpt5-thinking-mini", "gpt5thinkingmini":
-            return .gpt5ThinkingMini
-        case "gpt-5-thinking-nano", "gpt5-thinking-nano", "gpt5thinkingnano":
-            return .gpt5ThinkingNano
-        case "gpt-5-chat-latest", "gpt5-chat-latest":
-            return .gpt5ChatLatest
-        // Direct matches
-        case "gpt-4o", "gpt4o":
-            return .gpt4o
-        case "gpt-4o-mini", "gpt4o-mini", "gpt4omini":
-            return .gpt4oMini
-        case "gpt-4.1", "gpt4.1", "gpt41":
-            return .gpt41
-        case "gpt-4.1-mini", "gpt4.1-mini", "gpt41mini":
-            return .gpt41Mini
-        case "o4-mini", "o4mini":
-            return .o4Mini
         // Shortcuts
         case "gpt":
-            return .gpt51 // Default to flagship GPT-5.1
-        case "gpt4", "gpt-4":
-            return .gpt4o // Default to latest GPT-4 variant
+            return .gpt55 // Default to flagship GPT-5.5
         case "openai":
-            return .gpt51 // Default to GPT-5.1
+            return .gpt55 // Default to GPT-5.5
         default:
             // Check if it's an OpenAI model ID
-            if input.hasPrefix("gpt") || input.hasPrefix("o4") {
+            if self.isUnsupportedLegacyOpenAIModel(input) {
+                return nil
+            }
+            if input.hasPrefix("gpt-5") || input.hasPrefix("gpt5") {
                 return .custom(input)
             }
             return nil
@@ -127,31 +162,35 @@ public struct ModelSelector {
     private static func parseAnthropicModel(_ input: String) -> Model.Anthropic? {
         switch input {
         // Direct matches
-        case "claude-opus-4-20250514":
-            return .opus4
-        case "claude-opus-4-20250514-thinking":
-            return .opus4Thinking
+        case "claude-fable-5", "claude-fable-5-latest", "fable-5", "fable.5", "fable5", "fable":
+            return .fable5
+        case "claude-opus-4-8", "claude-opus-4.8", "opus-4-8", "opus-4.8", "opus48",
+             "claude-opus-4-8-latest":
+            return .opus48
+        case "claude-opus-4-7", "claude-opus-4.7", "opus-4-7", "opus-4.7", "opus47":
+            return .opus47
         case "claude-opus-4-5", "claude-opus-4.5", "opus-4-5", "opus-4.5", "opus45":
             return .opus45
-        case "claude-sonnet-4-20250514":
-            return .sonnet4
-        case "claude-sonnet-4-20250514-thinking":
-            return .sonnet4Thinking
+        case "claude-sonnet-4-6", "claude-sonnet-4.6", "sonnet-4-6", "sonnet-4.6", "sonnet46":
+            return .sonnet46
         case "claude-sonnet-4-5-20250929", "claude-sonnet-4.5":
             return .sonnet45
         // Shortcuts
         case "claude":
-            return .sonnet45 // Default plain Claude alias to latest Sonnet
+            return .opus48
         case "claude-opus", "opus":
-            return .opus45
+            return .opus48
         case "claude-sonnet", "sonnet":
-            return .sonnet4
+            return .sonnet46
         case "claude-haiku", "haiku":
             return .haiku45
         case "anthropic":
-            return .opus45 // Default Anthropic model
+            return .opus48
         default:
             // Check if it's a Claude model ID
+            if self.isUnsupportedLegacyAnthropicModel(input) {
+                return nil
+            }
             if input.hasPrefix("claude") {
                 return .custom(input)
             }
@@ -159,8 +198,34 @@ public struct ModelSelector {
         }
     }
 
+    private static func isUnsupportedLegacyOpenAIModel(_ input: String) -> Bool {
+        let normalized = input.lowercased()
+        let compact = normalized.replacingOccurrences(of: "-", with: "").replacingOccurrences(of: ".", with: "")
+        return normalized.hasPrefix("gpt-4") || compact.hasPrefix("gpt4") ||
+            normalized.hasPrefix("gpt-3") || compact.hasPrefix("gpt3") ||
+            normalized.hasPrefix("o3") || normalized.hasPrefix("o4") ||
+            normalized.hasPrefix("gpt-5.1") || normalized.hasPrefix("gpt-5.2") ||
+            compact.hasPrefix("gpt51") || compact.hasPrefix("gpt52") ||
+            normalized.contains("gpt-5-thinking") || compact.contains("gpt5thinking")
+    }
+
+    private static func isUnsupportedLegacyAnthropicModel(_ input: String) -> Bool {
+        let normalized = input.lowercased()
+        let compact = normalized.replacingOccurrences(of: "-", with: "").replacingOccurrences(of: ".", with: "")
+        return normalized.hasPrefix("claude-3") || compact.hasPrefix("claude3") ||
+            normalized == "claude-opus-4-20250514" ||
+            normalized == "claude-sonnet-4-20250514" ||
+            normalized.contains("-thinking")
+    }
+
     private static func parseGoogleModel(_ input: String) -> Model.Google? {
         switch input {
+        case "gemini-3.5-flash", "gemini35flash", "gemini-3-5-flash":
+            .gemini35Flash
+        case "gemini-3.1-pro-preview", "gemini-3.1-pro", "gemini31pro", "gemini31propreview":
+            .gemini31ProPreview
+        case "gemini-3.1-flash-lite", "gemini31flashlite", "gemini-3.1-flashlite":
+            .gemini31FlashLite
         case "gemini-3-flash", "gemini-3-flash-preview", "gemini3flash", "gemini-3flash":
             .gemini3Flash
         case "gemini-2.5-pro", "gemini25pro", "gemini2.5pro":
@@ -170,9 +235,9 @@ public struct ModelSelector {
         case "gemini-2.5-flash-lite", "gemini25flashlite", "gemini-2.5-flashlite":
             .gemini25FlashLite
         case "gemini":
-            .gemini3Flash
+            .gemini35Flash
         case "google":
-            .gemini25Pro
+            .gemini35Flash
         default:
             nil
         }
@@ -181,35 +246,22 @@ public struct ModelSelector {
     private static func parseGrokModel(_ input: String) -> Model.Grok? {
         switch input {
         // Direct matches for available models only
-        case "grok-4-0709":
-            return .grok4
-        case "grok-4-fast-reasoning":
-            return .grok4FastReasoning
-        case "grok-4-fast-non-reasoning":
-            return .grok4FastNonReasoning
-        case "grok-code-fast-1":
-            return .grokCodeFast1
-        case "grok-3", "grok3":
-            return .grok3
-        case "grok-3-mini":
-            return .grok3Mini
-        case "grok-2-1212", "grok-2":
-            return .grok2
-        case "grok-2-vision-1212":
-            return .grok2Vision
-        case "grok-2-image-1212":
-            return .grok2Image
-        case "grok-vision-beta":
-            return .grokVisionBeta
-        case "grok-beta":
-            return .grokBeta
+        case "grok-4.3", "grok-4-3", "grok43", "grok-4.3-latest", "grok-4-latest", "grok-4", "grok-latest":
+            return .grok43
+        case "grok-4.20-0309-reasoning", "grok-4-20-0309-reasoning":
+            return .grok420Reasoning
+        case "grok-4.20-0309-non-reasoning", "grok-4-20-0309-non-reasoning":
+            return .grok420NonReasoning
         // Shortcuts
         case "grok":
-            return .grok4FastReasoning // Default to the latest fast Grok model
+            return .grok43
         case "xai":
-            return .grok3 // Default xAI model
+            return .grok43
         default:
             // Check if it's a Grok model ID
+            if self.isUnsupportedLegacyGrokModel(input) {
+                return nil
+            }
             if input.hasPrefix("grok") {
                 return .custom(input)
             }
@@ -217,6 +269,51 @@ public struct ModelSelector {
         }
     }
 
+    private static func parseMiniMaxModel(_ input: String) -> Model.MiniMax? {
+        switch input {
+        case "minimax-m2.7", "minimax-m2-7", "m2.7", "m2-7", "minimax/m2.7", "minimax/m2-7",
+             "minimax/minimax-m2.7", "minimax/minimax-m2-7":
+            .m27
+        case "minimax-m2.7-highspeed", "minimax-m2-7-highspeed", "m2.7-highspeed", "m2-7-highspeed",
+             "minimax/m2.7-highspeed", "minimax/m2-7-highspeed", "minimax/minimax-m2.7-highspeed",
+             "minimax/minimax-m2-7-highspeed":
+            .m27Highspeed
+        case "minimax":
+            .m27
+        default:
+            nil
+        }
+    }
+
+    private static func parseMiniMaxCNModel(_ input: String) -> Model.MiniMax? {
+        switch input {
+        case "minimax-cn-m2.7", "minimax-cn-m2-7", "minimaxi-m2.7", "minimaxi-m2-7",
+             "minimax_cn/m2.7", "minimax_cn/m2-7", "minimax_cn/minimax-m2.7",
+             "minimax_cn/minimax-m2-7",
+             "minimax-cn/m2.7", "minimax-cn/m2-7", "minimax-cn/minimax-m2.7",
+             "minimax-cn/minimax-m2-7", "minimaxi/m2.7", "minimaxi/m2-7":
+            .m27
+        case "minimax-cn-m2.7-highspeed", "minimax-cn-m2-7-highspeed", "minimaxi-m2.7-highspeed",
+             "minimaxi-m2-7-highspeed", "minimax-cn/m2.7-highspeed", "minimax-cn/m2-7-highspeed",
+             "minimax_cn/m2.7-highspeed", "minimax_cn/m2-7-highspeed",
+             "minimax_cn/minimax-m2.7-highspeed", "minimax_cn/minimax-m2-7-highspeed",
+             "minimax-cn/minimax-m2.7-highspeed", "minimax-cn/minimax-m2-7-highspeed",
+             "minimaxi/m2.7-highspeed", "minimaxi/m2-7-highspeed":
+            .m27Highspeed
+        case "minimax-cn", "minimax_cn", "minimaxi":
+            .m27
+        default:
+            nil
+        }
+    }
+
+    private static func isUnsupportedLegacyGrokModel(_ input: String) -> Bool {
+        let normalized = input.lowercased()
+        return normalized.contains("grok-4.20-multi-agent") ||
+            normalized.contains("grok-4-20-multi-agent") ||
+            normalized.contains("grok420multiagent")
+    }
+
     private static func parseOllamaModel(_ input: String) -> Model.Ollama? {
         switch input {
         // Direct matches
@@ -238,22 +335,12 @@ public struct ModelSelector {
             .qwen25vl7b
         case "qwen2.5vl:32b":
             .qwen25vl32b
-        case "llama2", "llama2:latest":
-            .llama2
         case "llama4", "llama4:latest":
             .llama4
-        case "codellama", "codellama:latest":
-            .codellama
         case "mistral", "mistral:latest":
             .mistral
         case "mistral-nemo", "mistral-nemo:latest":
             .mistralNemo
-        case "mixtral", "mixtral:latest":
-            .mixtral
-        case "neural-chat", "neural-chat:latest":
-            .neuralChat
-        case "gemma", "gemma:latest":
-            .gemma
         case "devstral", "devstral:latest":
             .devstral
         case "deepseek-r1:8b":
@@ -277,6 +364,33 @@ public struct ModelSelector {
         }
     }
 
+    private static func parseLMStudioModel(_ input: String) -> Model.LMStudio? {
+        if input == "lmstudio" || input == "lm-studio" {
+            return .gptOSS120B
+        }
+
+        for prefix in ["lmstudio/", "lm-studio/"] where input.hasPrefix(prefix) {
+            let modelId = String(input.dropFirst(prefix.count))
+            guard !modelId.isEmpty else { return nil }
+            return Self.parseLMStudioModelIdentifier(modelId)
+        }
+
+        return nil
+    }
+
+    private static func parseLMStudioModelIdentifier(_ input: String) -> Model.LMStudio {
+        switch input {
+        case "openai/gpt-oss-120b", "gpt-oss-120b", "gpt-oss:120b":
+            .gptOSS120B
+        case "openai/gpt-oss-20b", "gpt-oss-20b", "gpt-oss:20b":
+            .gptOSS20B
+        case "meta/llama-3.3-70b", "llama-3.3-70b", "llama3.3-70b":
+            .llama3370B
+        default:
+            .custom(input)
+        }
+    }
+
     // MARK: - Model Information
 
     /// Get available models for a specific provider
@@ -302,11 +416,18 @@ public struct ModelSelector {
             }
         case "google", "gemini":
             return Model.Google.allCases.map(\.userFacingModelId)
+        case "minimax", "minimax-cn", "minimaxi":
+            return Model.MiniMax.allCases.map(\.modelId)
         case "ollama":
             return Model.Ollama.allCases.compactMap {
                 if case .custom = $0 { return nil }
                 return $0.modelId
             }
+        case "lmstudio", "lm-studio":
+            return Model.LMStudio.allCases.compactMap {
+                if case .custom = $0 { return nil }
+                return $0.modelId
+            }
         default:
             return []
         }
@@ -377,6 +498,16 @@ public func getAllAvailableModels() -> String {
         models: ModelSelector.availableModels(for: "google"),
     )
 
+    output += formatModelList(
+        title: "MiniMax",
+        models: ModelSelector.availableModels(for: "minimax"),
+    )
+
+    output += formatModelList(
+        title: "MiniMax China",
+        models: ModelSelector.availableModels(for: "minimax-cn"),
+    )
+
     output += formatModelList(
         title: "Grok (xAI)",
         models: ModelSelector.availableModels(for: "grok"),
@@ -388,14 +519,17 @@ public func getAllAvailableModels() -> String {
     )
 
     output += "\nShortcuts:\n"
-    output += "  • claude, claude-opus, opus → claude-opus-4-20250514\n"
-    output += "  • gpt, gpt4 → gpt-4.1\n"
-    output += "  • gemini → gemini-3-flash\n"
-    output += "  • grok → grok-4-fast-reasoning\n"
+    output += "  • claude, claude-opus, opus → claude-opus-4-8\n"
+    output += "  • fable → claude-fable-5\n"
+    output += "  • gpt → gpt-5.5\n"
+    output += "  • gemini → gemini-3.5-flash\n"
+    output += "  • minimax → MiniMax-M2.7\n"
+    output += "  • minimax-cn → MiniMax-M2.7 via api.minimaxi.com\n"
+    output += "  • grok → grok-4.3\n"
     output += "  • llama, llama3 → llama3.3\n"
 
     output += "\nCustom Models:\n"
-    output += "  • OpenRouter: anthropic/claude-3.5-sonnet\n"
+    output += "  • OpenRouter: anthropic/claude-fable-5\n"
     output += "  • Custom OpenAI: custom-gpt-model\n"
     output += "  • Local Ollama: any-model:tag\n"
 
@@ -423,15 +557,15 @@ extension ModelSelector {
         // Get recommended models for specific use cases
         switch useCase {
         case .coding:
-            [.claude, .gpt4o, .google(.gemini25Pro)]
+            [.claude, .openai(.gpt55), .google(.gemini35Flash)]
         case .vision:
-            [.claude, .gpt4o, .google(.gemini3Flash)]
+            [.claude, .openai(.gpt55), .google(.gemini35Flash)]
         case .reasoning:
-            [.openai(.gpt5Mini), .claude, .google(.gemini25Pro)]
+            [.openai(.gpt54), .claude, .google(.gemini35Flash)]
         case .local:
             [.llama, .ollama(.mistralNemo), .ollama(.commandRPlus)]
         case .general:
-            [.claude, .gpt4o, .google(.gemini3Flash), .grok(.grok4FastReasoning), .llama]
+            [.claude, .openai(.gpt55), .google(.gemini35Flash), .grok(.grok43), .llama]
         }
     }
 }
@@ -449,6 +583,7 @@ public enum UseCase {
 public enum ModelValidationError: Error, LocalizedError {
     case visionNotSupported(String)
     case toolsNotSupported(String)
+    case unsupportedModel(String)
 
     public var errorDescription: String? {
         switch self {
@@ -456,6 +591,8 @@ public enum ModelValidationError: Error, LocalizedError {
             "Model '\(modelId)' does not support vision inputs"
         case let .toolsNotSupported(modelId):
             "Model '\(modelId)' does not support tool calling"
+        case let .unsupportedModel(modelId):
+            "Model '\(modelId)' is no longer supported"
         }
     }
 }

Sources/Tachikoma/Providers/Anthropic/AnthropicTypes.swift

@@ -89,11 +89,13 @@ struct AnthropicMessageRequest: Codable {
     let messages: [AnthropicMessage]
     let tools: [AnthropicTool]?
     let thinking: AnthropicThinking?
+    let outputConfig: AnthropicOutputConfig?
     let stream: Bool?
 
     enum CodingKeys: String, CodingKey {
         case model, temperature, system, messages, tools, thinking, stream
         case maxTokens = "max_tokens"
+        case outputConfig = "output_config"
     }
 }
 
@@ -107,6 +109,10 @@ struct AnthropicThinking: Codable {
     }
 }
 
+struct AnthropicOutputConfig: Codable {
+    let effort: String?
+}
+
 struct AnthropicMessage: Codable {
     let role: String
     let content: [AnthropicContent]
@@ -133,13 +139,11 @@ enum AnthropicContent: Codable {
 
     struct RedactedThinkingContent: Codable {
         let type: String
-        let redactedThinking: String
-        let signature: String
+        let data: String
 
         enum CodingKeys: String, CodingKey {
             case type
-            case redactedThinking = "redacted_thinking"
-            case signature
+            case data
         }
     }
 
@@ -396,12 +400,14 @@ struct AnthropicMessageResponse: Codable {
     let model: String
     let stopReason: String?
     let stopSequence: String?
+    let stopDetails: StopDetails?
     let usage: AnthropicUsage
 
     enum CodingKeys: String, CodingKey {
         case id, type, role, content, model, usage
         case stopReason = "stop_reason"
         case stopSequence = "stop_sequence"
+        case stopDetails = "stop_details"
     }
 
     init(from decoder: Decoder) throws {
@@ -413,12 +419,20 @@ struct AnthropicMessageResponse: Codable {
         self.model = try container.decode(String.self, forKey: .model)
         self.stopReason = try container.decodeIfPresent(String.self, forKey: .stopReason)
         self.stopSequence = try container.decodeIfPresent(String.self, forKey: .stopSequence)
+        self.stopDetails = try container.decodeIfPresent(StopDetails.self, forKey: .stopDetails)
         self.usage = try container.decode(AnthropicUsage.self, forKey: .usage)
     }
+
+    struct StopDetails: Codable {
+        let category: String?
+        let explanation: String?
+    }
 }
 
 enum AnthropicResponseContent: Codable {
     case text(TextContent)
+    case thinking(ThinkingContent)
+    case redactedThinking(RedactedThinkingContent)
     case toolUse(ToolUseContent)
 
     struct TextContent: Codable {
@@ -455,6 +469,22 @@ enum AnthropicResponseContent: Codable {
         }
     }
 
+    struct ThinkingContent: Codable {
+        let type: String
+        let thinking: String
+        let signature: String
+    }
+
+    struct RedactedThinkingContent: Codable {
+        let type: String
+        let data: String
+
+        enum CodingKeys: String, CodingKey {
+            case type
+            case data
+        }
+    }
+
     struct ToolUseContent: Codable {
         let type: String
         let id: String
@@ -534,6 +564,10 @@ enum AnthropicResponseContent: Codable {
         switch type {
         case "text":
             self = try .text(TextContent(from: decoder))
+        case "thinking":
+            self = try .thinking(ThinkingContent(from: decoder))
+        case "redacted_thinking":
+            self = try .redactedThinking(RedactedThinkingContent(from: decoder))
         case "tool_use":
             self = try .toolUse(ToolUseContent(from: decoder))
         default:
@@ -552,6 +586,10 @@ enum AnthropicResponseContent: Codable {
         switch self {
         case let .text(content):
             try content.encode(to: encoder)
+        case let .thinking(content):
+            try content.encode(to: encoder)
+        case let .redactedThinking(content):
+            try content.encode(to: encoder)
         case let .toolUse(content):
             try content.encode(to: encoder)
         }
@@ -561,6 +599,8 @@ enum AnthropicResponseContent: Codable {
         case type
         case text
         case thinking
+        case redactedThinking = "redacted_thinking"
+        case signature
     }
 }
 
@@ -572,6 +612,17 @@ struct AnthropicUsage: Codable {
         case inputTokens = "input_tokens"
         case outputTokens = "output_tokens"
     }
+
+    init(inputTokens: Int, outputTokens: Int) {
+        self.inputTokens = inputTokens
+        self.outputTokens = outputTokens
+    }
+
+    init(from decoder: Decoder) throws {
+        let container = try decoder.container(keyedBy: CodingKeys.self)
+        self.inputTokens = try container.decodeIfPresent(Int.self, forKey: .inputTokens) ?? 0
+        self.outputTokens = try container.decodeIfPresent(Int.self, forKey: .outputTokens) ?? 0
+    }
 }
 
 // MARK: - Streaming Types
@@ -606,12 +657,11 @@ struct AnthropicStreamContentBlock: Codable {
     let text: String?
     let input: Any?
     let thinking: String?
-    let redactedThinking: String?
+    let data: String?
     let signature: String?
 
     enum CodingKeys: String, CodingKey {
-        case type, id, name, text, input, thinking, signature
-        case redactedThinking = "redacted_thinking"
+        case type, id, name, text, input, thinking, data, signature
     }
 
     init(from decoder: Decoder) throws {
@@ -621,7 +671,7 @@ struct AnthropicStreamContentBlock: Codable {
         self.name = try? container.decode(String.self, forKey: .name)
         self.text = try? container.decode(String.self, forKey: .text)
         self.thinking = try? container.decode(String.self, forKey: .thinking)
-        self.redactedThinking = try? container.decode(String.self, forKey: .redactedThinking)
+        self.data = try? container.decode(String.self, forKey: .data)
         self.signature = try? container.decode(String.self, forKey: .signature)
 
         // Decode input as generic JSON if present
@@ -647,7 +697,7 @@ struct AnthropicStreamContentBlock: Codable {
         try container.encodeIfPresent(self.name, forKey: .name)
         try container.encodeIfPresent(self.text, forKey: .text)
         try container.encodeIfPresent(self.thinking, forKey: .thinking)
-        try container.encodeIfPresent(self.redactedThinking, forKey: .redactedThinking)
+        try container.encodeIfPresent(self.data, forKey: .data)
         try container.encodeIfPresent(self.signature, forKey: .signature)
         if let input {
             let data = try JSONSerialization.data(withJSONObject: input)
@@ -671,6 +721,17 @@ struct AnthropicStreamDelta: Codable {
         case stopReason = "stop_reason"
         case stopSequence = "stop_sequence"
     }
+
+    init(from decoder: Decoder) throws {
+        let container = try decoder.container(keyedBy: CodingKeys.self)
+        self.type = try container.decodeIfPresent(String.self, forKey: .type) ?? ""
+        self.text = try container.decodeIfPresent(String.self, forKey: .text)
+        self.thinking = try container.decodeIfPresent(String.self, forKey: .thinking)
+        self.signature = try container.decodeIfPresent(String.self, forKey: .signature)
+        self.partialJson = try container.decodeIfPresent(String.self, forKey: .partialJson)
+        self.stopReason = try container.decodeIfPresent(String.self, forKey: .stopReason)
+        self.stopSequence = try container.decodeIfPresent(String.self, forKey: .stopSequence)
+    }
 }
 
 struct AnthropicErrorResponse: Codable {

Sources/Tachikoma/Providers/Compatible/AnthropicCompatibleProvider.swift

@@ -6,32 +6,79 @@ public final class AnthropicCompatibleProvider: ModelProvider {
     public let modelId: String
     public let baseURL: String?
     public let apiKey: String?
+    public let additionalHeaders: [String: String]
     public let capabilities: ModelCapabilities
     private let configuration: TachikomaConfiguration
+    private let auth: TKAuthValue?
+    private let reasoningProvider: String
+    private let reasoningModelId: String
+    private let reasoningBaseURL: String?
 
-    public init(modelId: String, baseURL: String, configuration: TachikomaConfiguration) throws {
+    public init(
+        modelId: String,
+        baseURL: String,
+        configuration: TachikomaConfiguration,
+        apiKey: String? = nil,
+        additionalHeaders: [String: String] = [:],
+        auth: TKAuthValue? = nil,
+        capabilities: ModelCapabilities? = nil,
+        reasoningProvider: String = "anthropic-compatible",
+        reasoningModelId: String? = nil,
+        reasoningBaseURL: String? = nil,
+        includeReasoningBaseURL: Bool = true,
+    ) throws {
         self.modelId = modelId
         self.baseURL = baseURL
         self.configuration = configuration
+        self.additionalHeaders = additionalHeaders
+        self.reasoningProvider = reasoningProvider
+        self.reasoningModelId = reasoningModelId ?? modelId
+        self.reasoningBaseURL = includeReasoningBaseURL ? (reasoningBaseURL ?? baseURL) : nil
 
-        // Try to get API key from configuration, otherwise try common environment variable patterns
-        if let key = configuration.getAPIKey(for: .custom("anthropic_compatible")) {
+        // Try explicit provider key, then configuration, then common environment variable patterns.
+        if let key = apiKey {
             self.apiKey = key
+            self.auth = auth ?? .apiKey(key)
+        } else if let key = configuration.getAPIKey(for: .custom("anthropic_compatible")) {
+            self.apiKey = key
+            self.auth = auth ?? .apiKey(key)
         } else if
             let key = ProcessInfo.processInfo.environment["ANTHROPIC_COMPATIBLE_API_KEY"] ??
             ProcessInfo.processInfo.environment["API_KEY"]
         {
             self.apiKey = key
+            self.auth = auth ?? .apiKey(key)
+        } else if let auth {
+            self.auth = auth
+            switch auth {
+            case let .apiKey(key):
+                self.apiKey = key
+            case let .bearer(token, _):
+                self.apiKey = token
+            }
         } else {
             self.apiKey = nil
+            self.auth = nil
         }
 
-        self.capabilities = ModelCapabilities(
+        let isFable = LanguageModel.Anthropic.isFable(modelId: modelId)
+        let supportsSafeStreaming = !LanguageModel.Anthropic.hasStreamingRefusalRisk(modelId: modelId)
+        let baseCapabilities = capabilities ?? ModelCapabilities(
             supportsVision: true,
             supportsTools: true,
-            supportsStreaming: true,
-            contextLength: 200_000,
-            maxOutputTokens: 8192,
+            supportsStreaming: supportsSafeStreaming,
+            contextLength: isFable ? 1_000_000 : 200_000,
+            maxOutputTokens: isFable ? 128_000 : 8192,
+        )
+        self.capabilities = supportsSafeStreaming ? baseCapabilities : ModelCapabilities(
+            supportsVision: baseCapabilities.supportsVision,
+            supportsTools: baseCapabilities.supportsTools,
+            supportsStreaming: false,
+            supportsAudioInput: baseCapabilities.supportsAudioInput,
+            supportsAudioOutput: baseCapabilities.supportsAudioOutput,
+            contextLength: baseCapabilities.contextLength,
+            maxOutputTokens: baseCapabilities.maxOutputTokens,
+            costPerToken: baseCapabilities.costPerToken,
         )
     }
 
@@ -62,6 +109,11 @@ public final class AnthropicCompatibleProvider: ModelProvider {
         return try AnthropicProvider(
             model: .custom(self.modelId),
             configuration: compatConfig,
+            additionalHeaders: self.additionalHeaders,
+            authOverride: self.auth,
+            reasoningProvider: self.reasoningProvider,
+            reasoningModelId: self.reasoningModelId,
+            reasoningBaseURL: self.reasoningBaseURL,
         )
     }
 }

Sources/Tachikoma/Providers/Compatible/OpenAICompatibleProvider.swift

@@ -1,4 +1,7 @@
 import Foundation
+#if canImport(FoundationNetworking)
+import FoundationNetworking
+#endif
 
 /// Provider for OpenAI-compatible APIs
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
@@ -6,14 +9,27 @@ public final class OpenAICompatibleProvider: ModelProvider {
     public let modelId: String
     public let baseURL: String?
     public let apiKey: String?
+    public let additionalHeaders: [String: String]
     public let capabilities: ModelCapabilities
+    private let session: URLSession
 
-    public init(modelId: String, baseURL: String, configuration: TachikomaConfiguration) throws {
+    public init(
+        modelId: String,
+        baseURL: String,
+        configuration: TachikomaConfiguration,
+        apiKey: String? = nil,
+        additionalHeaders: [String: String] = [:],
+        session: URLSession = .shared,
+    ) throws {
         self.modelId = modelId
         self.baseURL = baseURL
+        self.additionalHeaders = additionalHeaders
+        self.session = session
 
-        // Try to get API key from configuration, otherwise try common environment variable patterns
-        if let key = configuration.getAPIKey(for: .custom("openai_compatible")) {
+        // Try explicit provider key, then configuration, then common environment variable patterns.
+        if let key = apiKey {
+            self.apiKey = key
+        } else if let key = configuration.getAPIKey(for: .custom("openai_compatible")) {
             self.apiKey = key
         } else if
             let key = ProcessInfo.processInfo.environment["OPENAI_COMPATIBLE_API_KEY"] ??
@@ -24,12 +40,13 @@ public final class OpenAICompatibleProvider: ModelProvider {
             self.apiKey = nil // Some compatible APIs don't require keys
         }
 
+        let isFable = LanguageModel.Anthropic.isFable(modelId: modelId)
         self.capabilities = ModelCapabilities(
             supportsVision: false,
             supportsTools: true,
-            supportsStreaming: true,
-            contextLength: 128_000,
-            maxOutputTokens: 4096,
+            supportsStreaming: !LanguageModel.Anthropic.hasStreamingRefusalRisk(modelId: modelId),
+            contextLength: isFable ? 1_000_000 : 128_000,
+            maxOutputTokens: isFable ? 128_000 : 4096,
         )
     }
 
@@ -41,17 +58,25 @@ public final class OpenAICompatibleProvider: ModelProvider {
             baseURL: self.baseURL!,
             apiKey: self.apiKey ?? "",
             providerName: "OpenAICompatible",
+            additionalHeaders: self.additionalHeaders,
+            session: self.session,
         )
     }
 
     public func streamText(request: ProviderRequest) async throws -> AsyncThrowingStream<TextStreamDelta, Error> {
+        guard !LanguageModel.Anthropic.hasStreamingRefusalRisk(modelId: self.modelId) else {
+            throw TachikomaError.invalidConfiguration("\(self.modelId) does not support streaming")
+        }
+
         // Use OpenAI-compatible streaming implementation
-        try await OpenAICompatibleHelper.streamText(
+        return try await OpenAICompatibleHelper.streamText(
             request: request,
             modelId: self.modelId,
             baseURL: self.baseURL!,
             apiKey: self.apiKey ?? "",
             providerName: "OpenAICompatible",
+            additionalHeaders: self.additionalHeaders,
+            session: self.session,
         )
     }
 }

Sources/Tachikoma/Providers/Google/GoogleProvider.swift

@@ -214,16 +214,15 @@ public final class GoogleProvider: ModelProvider {
                 }
             }
 
-            guard !parts.isEmpty else { continue }
             let role = switch message.role {
             case .assistant:
                 "model"
             case .tool:
-                "function"
+                "user"
             default:
                 "user"
             }
-            contents.append(.init(role: role, parts: parts))
+            Self.appendGoogleContent(role: role, parts: parts, to: &contents)
         }
 
         let config = GoogleGenerateRequest.GenerationConfig(
@@ -253,6 +252,20 @@ public final class GoogleProvider: ModelProvider {
 // MARK: - Streaming Helpers
 
 extension GoogleProvider {
+    private static func appendGoogleContent(
+        role: String,
+        parts: [GoogleGenerateRequest.Content.Part],
+        to contents: inout [GoogleGenerateRequest.Content],
+    ) {
+        guard !parts.isEmpty else { return }
+
+        if let last = contents.last, last.role == role {
+            contents[contents.count - 1] = .init(role: role, parts: last.parts + parts)
+        } else {
+            contents.append(.init(role: role, parts: parts))
+        }
+    }
+
     private static func buildToolCallNameMap(from messages: [ModelMessage]) -> [String: String] {
         var map: [String: String] = [:]
         for message in messages {
@@ -271,12 +284,6 @@ extension GoogleProvider {
     }
 
     private static func convertTool(_ tool: AgentTool) throws -> GoogleGenerateRequest.Tool.FunctionDeclaration {
-        var parameters: [String: Any] = [
-            "type": "object",
-            "properties": [:],
-            "required": tool.parameters.required,
-        ]
-
         var properties: [String: Any] = [:]
         for (key, prop) in tool.parameters.properties {
             var propDict: [String: Any] = [
@@ -295,7 +302,21 @@ extension GoogleProvider {
             }
             properties[key] = propDict
         }
-        parameters["properties"] = properties
+
+        // Gemini validates that every name in `required` exists in `properties`. Tools occasionally
+        // ship with a `required` entry whose property got filtered out during MCP→Agent conversion
+        // (e.g. anyOf/oneOf schemas that don't survive the simplified MCP→Agent translation). Drop
+        // any orphan `required` names so we never send Gemini an invalid schema.
+        let declaredKeys = Set(properties.keys)
+        let sanitizedRequired = tool.parameters.required.filter { declaredKeys.contains($0) }
+
+        var parameters: [String: Any] = [
+            "type": "object",
+            "properties": properties,
+        ]
+        if !sanitizedRequired.isEmpty {
+            parameters["required"] = sanitizedRequired
+        }
 
         guard let schema = JSONValue(value: parameters) else {
             throw TachikomaError.invalidInput("Failed to encode tool parameters for '\(tool.name)'")

Sources/Tachikoma/Providers/Grok/GrokProvider.swift

@@ -11,8 +11,15 @@ public final class GrokProvider: ModelProvider {
     private let model: LanguageModel.Grok
 
     public init(model: LanguageModel.Grok, configuration: TachikomaConfiguration) throws {
+        let modelId = model.modelId
+        guard !Self.requiresResponsesAPIRouting(modelId) else {
+            throw TachikomaError.unsupportedOperation(
+                "\(modelId) requires xAI Responses API routing",
+            )
+        }
+
         self.model = model
-        self.modelId = model.modelId
+        self.modelId = modelId
         self.baseURL = configuration.getBaseURL(for: .grok) ?? "https://api.x.ai/v1"
 
         // Get API key from configuration system (environment or credentials)
@@ -31,6 +38,16 @@ public final class GrokProvider: ModelProvider {
         )
     }
 
+    private static func requiresResponsesAPIRouting(_ modelId: String) -> Bool {
+        let normalized = modelId.lowercased()
+        let compact = normalized
+            .replacingOccurrences(of: "-", with: "")
+            .replacingOccurrences(of: ".", with: "")
+        return normalized.contains("grok-4.20-multi-agent") ||
+            normalized.contains("grok-4-20-multi-agent") ||
+            compact.contains("grok420multiagent")
+    }
+
     public func generateText(request: ProviderRequest) async throws -> ProviderResponse {
         // Grok uses OpenAI-compatible API format - delegate to shared implementation
         try await OpenAICompatibleHelper.generateText(

Sources/Tachikoma/Providers/LMStudioProvider.swift

@@ -41,7 +41,7 @@ public actor LMStudioProvider: ModelProvider {
 
     public init(
         baseURL: String = "http://localhost:1234/v1",
-        modelId: String = "current",
+        modelId: String = "openai/gpt-oss-20b",
         apiKey: String? = nil,
         sessionConfiguration: URLSessionConfiguration = .default,
     ) {
@@ -74,8 +74,8 @@ public actor LMStudioProvider: ModelProvider {
 
         for url in commonURLs {
             let provider = LMStudioProvider(baseURL: url)
-            if await (try? provider.healthCheck()) != nil {
-                return provider
+            if let status = await (try? provider.healthCheck()) {
+                return LMStudioProvider(baseURL: url, modelId: status.model ?? "openai/gpt-oss-20b")
             }
         }
 

Sources/Tachikoma/Providers/OpenAI/OpenAIProvider.swift

@@ -20,6 +20,10 @@ public final class OpenAIProvider: ModelProvider {
         configuration: TachikomaConfiguration,
         session: URLSession = .shared,
     ) throws {
+        guard !model.isUnsupportedLegacyFamily else {
+            throw TachikomaError.unsupportedOperation("OpenAI model '\(model.modelId)' is no longer supported")
+        }
+
         self.model = model
         self.modelId = model.modelId
         self.baseURL = configuration.getBaseURL(for: .openai) ?? "https://api.openai.com/v1"

Sources/Tachikoma/Providers/OpenAI/OpenAIResponsesProvider.swift

@@ -2,11 +2,8 @@ import Foundation
 #if canImport(FoundationNetworking)
 import FoundationNetworking
 #endif
-#if canImport(FoundationNetworking)
-import FoundationNetworking
-#endif
 
-/// Provider for OpenAI Responses API (GPT-5, o3, o4)
+/// Provider for OpenAI Responses API (GPT-5)
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
 public final class OpenAIResponsesProvider: ModelProvider {
     public let modelId: String
@@ -17,6 +14,7 @@ public final class OpenAIResponsesProvider: ModelProvider {
     private let model: LanguageModel.OpenAI
     private let configuration: TachikomaConfiguration
     private let session: URLSession
+    private let auth: TKAuthValue
 
     private static let debugLogURL = URL(fileURLWithPath: "/tmp/tachikoma-gpt5.log")
 
@@ -31,15 +29,28 @@ public final class OpenAIResponsesProvider: ModelProvider {
         configuration: TachikomaConfiguration,
         session: URLSession = .shared,
     ) throws {
+        guard !model.isUnsupportedLegacyFamily else {
+            throw TachikomaError.unsupportedOperation("OpenAI model '\(model.modelId)' is no longer supported")
+        }
+
         self.model = model
         self.modelId = model.modelId
         self.configuration = configuration
         self.session = session
         self.baseURL = configuration.getBaseURL(for: .openai) ?? "https://api.openai.com/v1"
 
-        // Get API key from configuration
+        // Prefer configuration-provided key first (test configs use this), then shared OAuth/API-key auth.
         if let key = configuration.getAPIKey(for: .openai) {
+            self.auth = .bearer(key, betaHeader: nil)
             self.apiKey = key
+        } else if let auth = TKAuthManager.shared.resolveAuth(for: .openai) {
+            self.auth = auth
+            switch auth {
+            case let .apiKey(key):
+                self.apiKey = key
+            case let .bearer(token, _):
+                self.apiKey = token
+            }
         } else {
             throw TachikomaError.authenticationFailed("OPENAI_API_KEY not found")
         }
@@ -47,13 +58,15 @@ public final class OpenAIResponsesProvider: ModelProvider {
         // Set capabilities based on model
         let isReasoningModel = Self.isReasoningModel(model)
         let isGPT5 = Self.isGPT5Model(model)
+        let hasLargeOutputWindow = isReasoningModel || isGPT5 || model == .chatLatest
+        let maxOutputTokens = model == .gpt5ChatLatest ? 16384 : (hasLargeOutputWindow ? 128_000 : 4096)
 
         self.capabilities = ModelCapabilities(
             supportsVision: model.supportsVision,
             supportsTools: model.supportsTools,
             supportsStreaming: true,
             contextLength: model.contextLength,
-            maxOutputTokens: isReasoningModel || isGPT5 ? 128_000 : 4096,
+            maxOutputTokens: maxOutputTokens,
         )
     }
 
@@ -65,7 +78,8 @@ public final class OpenAIResponsesProvider: ModelProvider {
         let url = URL(string: "\(baseURL!)/responses")!
         var urlRequest = URLRequest(url: url)
         urlRequest.httpMethod = "POST"
-        urlRequest.setValue("Bearer \(self.apiKey!)", forHTTPHeaderField: "Authorization")
+        let (authHeaderName, prefix, secret) = self.authHeader()
+        urlRequest.setValue("\(prefix)\(secret)", forHTTPHeaderField: authHeaderName)
         urlRequest.setValue("application/json", forHTTPHeaderField: "Content-Type")
 
         // Add OpenAI-specific headers
@@ -134,7 +148,8 @@ public final class OpenAIResponsesProvider: ModelProvider {
         let finalURLRequest: URLRequest = {
             var req = URLRequest(url: url)
             req.httpMethod = "POST"
-            req.setValue("Bearer \(self.apiKey!)", forHTTPHeaderField: "Authorization")
+            let (authHeaderName, prefix, secret) = self.authHeader()
+            req.setValue("\(prefix)\(secret)", forHTTPHeaderField: authHeaderName)
             req.setValue("application/json", forHTTPHeaderField: "Content-Type")
             req.setValue("text/event-stream", forHTTPHeaderField: "Accept")
 
@@ -198,6 +213,21 @@ public final class OpenAIResponsesProvider: ModelProvider {
                     // Parse the entire response for Linux
                     let responseText = String(data: data, encoding: .utf8) ?? ""
                     let lines = responseText.components(separatedBy: "\n")
+                    var streamState = ResponsesStreamState()
+                    for line in lines {
+                        if
+                            try Self.processResponsesStreamLine(
+                                line,
+                                model: self.model,
+                                state: &streamState,
+                                continuation: continuation,
+                            )
+                        {
+                            return
+                        }
+                    }
+                    continuation.finish()
+                    return
                     #else
                     // macOS/iOS: Use streaming API
                     let (bytes, response) = try await self.session.bytes(for: finalURLRequest)
@@ -227,157 +257,18 @@ public final class OpenAIResponsesProvider: ModelProvider {
                         throw TachikomaError.apiError("Failed to start streaming: \(errorMessage)")
                     }
 
-                    var previousContent = "" // Track previously sent content for GPT-5 preambles
-                    struct PartialToolCall {
-                        var id: String
-                        var name: String?
-                        var arguments: String
-                    }
-                    var pendingToolCalls: [String: PartialToolCall] = [:]
+                    var streamState = ResponsesStreamState()
 
                     for try await line in bytes.lines {
-                        // Handle SSE format
-                        if line.hasPrefix("data: ") {
-                            let jsonString = String(line.dropFirst(6))
-
-                            if ProcessInfo.processInfo.environment["DEBUG_TACHIKOMA_STREAM"] != nil {
-                                Self.debugLog("raw stream: \(jsonString)")
-                            }
-
-                            if jsonString == "[DONE]" {
-                                continuation.finish()
-                                return
-                            }
-
-                            if let data = jsonString.data(using: .utf8) {
-                                // Try GPT-5 format first
-                                if Self.isGPT5Model(self.model) {
-                                    if
-                                        let event = try? JSONSerialization.jsonObject(with: data) as? [String: Any],
-                                        let eventType = event["type"] as? String
-                                    {
-                                        if ProcessInfo.processInfo.environment["DEBUG_TACHIKOMA"] != nil {
-                                            Self.debugLog("event: \(eventType) payload: \(event)")
-                                        }
-
-                                        switch eventType {
-                                        case "response.output_text.delta":
-                                            if let delta = event["delta"] as? String, !delta.isEmpty {
-                                                continuation.yield(TextStreamDelta.text(delta))
-                                            }
-
-                                        case "response.output_item.added":
-                                            if
-                                                let item = event["item"] as? [String: Any],
-                                                let itemType = item["type"] as? String,
-                                                itemType == "function_call"
-                                            {
-                                                let identifier = (item["id"] as? String) ??
-                                                    (item["call_id"] as? String) ?? UUID().uuidString
-                                                var partial = pendingToolCalls[identifier] ?? PartialToolCall(
-                                                    id: identifier,
-                                                    name: nil,
-                                                    arguments: "",
-                                                )
-                                                if let name = item["name"] as? String {
-                                                    partial.name = name
-                                                }
-                                                pendingToolCalls[identifier] = partial
-                                            }
-
-                                        case "response.function_call_arguments.delta":
-                                            if
-                                                let itemId = event["item_id"] as? String,
-                                                let delta = event["delta"] as? String
-                                            {
-                                                var partial = pendingToolCalls[itemId] ?? PartialToolCall(
-                                                    id: itemId,
-                                                    name: nil,
-                                                    arguments: "",
-                                                )
-                                                partial.arguments.append(delta)
-                                                pendingToolCalls[itemId] = partial
-                                            }
-
-                                        case "response.function_call_arguments.done":
-                                            if
-                                                let itemId = event["item_id"] as? String,
-                                                let arguments = event["arguments"] as? String
-                                            {
-                                                var partial = pendingToolCalls[itemId] ?? PartialToolCall(
-                                                    id: itemId,
-                                                    name: nil,
-                                                    arguments: "",
-                                                )
-                                                partial.arguments = arguments
-                                                pendingToolCalls[itemId] = partial
-
-                                                if
-                                                    let name = partial.name,
-                                                    let toolCall = Self.makeToolCall(
-                                                        id: itemId,
-                                                        name: name,
-                                                        argumentsJSON: arguments,
-                                                    )
-                                                {
-                                                    continuation.yield(.tool(toolCall))
-                                                    pendingToolCalls.removeValue(forKey: itemId)
-                                                }
-                                            }
-
-                                        case "response.completed":
-                                            continuation.finish()
-                                            return
-
-                                        default:
-                                            break
-                                        }
-                                    }
-                                } else {
-                                    // Try standard Responses API format (O3, etc.)
-                                    do {
-                                        let chunk = try JSONDecoder().decode(
-                                            OpenAIResponsesStreamChunk.self,
-                                            from: data,
-                                        )
-
-                                        // Convert to TextStreamDelta
-                                        if
-                                            let choice = chunk.choices.first,
-                                            let content = choice.delta.content,
-                                            !content.isEmpty
-                                        {
-                                            // Handle accumulated content for models with preambles
-                                            if content.hasPrefix(previousContent), !previousContent.isEmpty {
-                                                // This is accumulated content, extract just the delta
-                                                let delta = String(content.dropFirst(previousContent.count))
-                                                if !delta.isEmpty {
-                                                    continuation.yield(TextStreamDelta.text(delta))
-                                                    previousContent = content // Update the accumulated content
-                                                }
-                                            } else {
-                                                // This is a true delta or the first chunk
-                                                continuation.yield(TextStreamDelta.text(content))
-                                                previousContent += content // Accumulate for comparison
-                                            }
-                                        }
-
-                                        // Check for finish
-                                        if
-                                            let choice = chunk.choices.first,
-                                            choice.finishReason != nil
-                                        {
-                                            continuation.finish()
-                                            return
-                                        }
-                                    } catch {
-                                        // Ignore parsing errors for incomplete chunks
-                                    }
-                                }
-                            }
-                        } else if line.hasPrefix("event: ") {
-                            // Track event types for GPT-5 streaming (but we handle them in data lines)
-                            // This helps us understand the stream structure
+                        if
+                            try Self.processResponsesStreamLine(
+                                line,
+                                model: self.model,
+                                state: &streamState,
+                                continuation: continuation,
+                            )
+                        {
+                            return
                         }
                     }
 
@@ -390,6 +281,210 @@ public final class OpenAIResponsesProvider: ModelProvider {
         }
     }
 
+    private struct ResponsesStreamState {
+        struct PartialToolCall {
+            var id: String
+            var name: String?
+            var arguments: String
+        }
+
+        var previousContent = ""
+        var pendingToolCalls: [String: PartialToolCall] = [:]
+        var didYieldToolCall = false
+        var didReceiveRefusal = false
+    }
+
+    private static func processResponsesStreamLine(
+        _ line: String,
+        model: LanguageModel.OpenAI,
+        state: inout ResponsesStreamState,
+        continuation: AsyncThrowingStream<TextStreamDelta, Error>.Continuation,
+    ) throws
+        -> Bool
+    {
+        guard line.hasPrefix("data: ") else {
+            return false
+        }
+
+        let jsonString = String(line.dropFirst(6))
+        if ProcessInfo.processInfo.environment["DEBUG_TACHIKOMA_STREAM"] != nil {
+            Self.debugLog("raw stream: \(jsonString)")
+        }
+
+        if jsonString == "[DONE]" {
+            continuation.finish()
+            return true
+        }
+
+        guard let data = jsonString.data(using: .utf8) else {
+            return false
+        }
+
+        if Self.usesResponsesEventStream(model) {
+            guard
+                let event = try? JSONSerialization.jsonObject(with: data) as? [String: Any],
+                let eventType = event["type"] as? String else
+            {
+                return false
+            }
+
+            if ProcessInfo.processInfo.environment["DEBUG_TACHIKOMA"] != nil {
+                Self.debugLog("event: \(eventType) payload: \(event)")
+            }
+
+            switch eventType {
+            case "response.output_text.delta":
+                if let delta = event["delta"] as? String, !delta.isEmpty {
+                    continuation.yield(TextStreamDelta.text(delta))
+                }
+
+            case "response.output_item.added":
+                if
+                    let item = event["item"] as? [String: Any],
+                    let itemType = item["type"] as? String,
+                    itemType == "function_call"
+                {
+                    let identifier = (item["id"] as? String) ??
+                        (item["call_id"] as? String) ?? UUID().uuidString
+                    var partial = state.pendingToolCalls[identifier] ?? ResponsesStreamState.PartialToolCall(
+                        id: identifier,
+                        name: nil,
+                        arguments: "",
+                    )
+                    if let name = item["name"] as? String {
+                        partial.name = name
+                    }
+                    state.pendingToolCalls[identifier] = partial
+                }
+
+            case "response.function_call_arguments.delta":
+                if
+                    let itemId = event["item_id"] as? String,
+                    let delta = event["delta"] as? String
+                {
+                    var partial = state.pendingToolCalls[itemId] ?? ResponsesStreamState.PartialToolCall(
+                        id: itemId,
+                        name: nil,
+                        arguments: "",
+                    )
+                    partial.arguments.append(delta)
+                    state.pendingToolCalls[itemId] = partial
+                }
+
+            case "response.function_call_arguments.done":
+                if
+                    let itemId = event["item_id"] as? String,
+                    let arguments = event["arguments"] as? String
+                {
+                    var partial = state.pendingToolCalls[itemId] ?? ResponsesStreamState.PartialToolCall(
+                        id: itemId,
+                        name: nil,
+                        arguments: "",
+                    )
+                    partial.arguments = arguments
+                    state.pendingToolCalls[itemId] = partial
+
+                    if
+                        let name = partial.name,
+                        let toolCall = Self.makeToolCall(id: itemId, name: name, argumentsJSON: arguments)
+                    {
+                        continuation.yield(.tool(toolCall))
+                        state.didYieldToolCall = true
+                        state.pendingToolCalls.removeValue(forKey: itemId)
+                    }
+                }
+
+            case "response.refusal.delta",
+                 "response.refusal.done":
+                state.didReceiveRefusal = true
+
+            case "response.completed":
+                let finishReason: FinishReason = state.didReceiveRefusal
+                    ? .contentFilter
+                    : (state.didYieldToolCall ? .toolCalls : .stop)
+                continuation.yield(.done(finishReason: finishReason))
+                continuation.finish()
+                return true
+
+            case "response.incomplete":
+                let finishReason = Self.finishReasonForIncompleteResponseEvent(event)
+                continuation.yield(.done(finishReason: finishReason))
+                continuation.finish()
+                return true
+
+            case "response.failed",
+                 "error":
+                throw TachikomaError.apiError(Self.errorMessageForResponseStreamEvent(event))
+
+            default:
+                break
+            }
+            return false
+        }
+
+        do {
+            let chunk = try JSONDecoder().decode(OpenAIResponsesStreamChunk.self, from: data)
+            if
+                let choice = chunk.choices.first,
+                let content = choice.delta.content,
+                !content.isEmpty
+            {
+                if content.hasPrefix(state.previousContent), !state.previousContent.isEmpty {
+                    let delta = String(content.dropFirst(state.previousContent.count))
+                    if !delta.isEmpty {
+                        continuation.yield(TextStreamDelta.text(delta))
+                        state.previousContent = content
+                    }
+                } else {
+                    continuation.yield(TextStreamDelta.text(content))
+                    state.previousContent += content
+                }
+            }
+
+            if let choice = chunk.choices.first, let finishReason = choice.finishReason {
+                continuation.yield(.done(finishReason: Self.finishReasonForChatStream(finishReason)))
+                continuation.finish()
+                return true
+            }
+        } catch {
+            // Ignore parsing errors for incomplete chunks.
+        }
+
+        return false
+    }
+
+    private static func finishReasonForChatStream(_ reason: String) -> FinishReason {
+        switch reason {
+        case "stop": .stop
+        case "length": .length
+        case "tool_calls": .toolCalls
+        case "content_filter": .contentFilter
+        default: .other
+        }
+    }
+
+    private static func errorMessageForResponseStreamEvent(_ event: [String: Any]) -> String {
+        let eventType = event["type"] as? String ?? "error"
+        let errorPayload = (event["error"] as? [String: Any]) ??
+            (event["response"] as? [String: Any]).flatMap { $0["error"] as? [String: Any] }
+        if let message = errorPayload?["message"] as? String, !message.isEmpty {
+            return "OpenAI Responses API stream \(eventType): \(message)"
+        }
+        if let message = event["message"] as? String, !message.isEmpty {
+            return "OpenAI Responses API stream \(eventType): \(message)"
+        }
+        return "OpenAI Responses API stream \(eventType)"
+    }
+
+    private func authHeader() -> (String, String, String) {
+        switch self.auth {
+        case let .apiKey(key):
+            ("Authorization", "Bearer ", key)
+        case let .bearer(token, _):
+            ("Authorization", "Bearer ", token)
+        }
+    }
+
     // MARK: - Private Helpers
 
     private func buildResponsesRequest(
@@ -534,6 +629,29 @@ public final class OpenAIResponsesProvider: ModelProvider {
         }
     }
 
+    private static func finishReasonForIncompleteResponseEvent(_ event: [String: Any]) -> FinishReason {
+        guard
+            let response = event["response"] as? [String: Any],
+            let incompleteDetails = response["incomplete_details"] as? [String: Any],
+            let reason = incompleteDetails["reason"] as? String else
+        {
+            return .other
+        }
+
+        return Self.finishReasonForIncompleteReason(reason)
+    }
+
+    private static func finishReasonForIncompleteReason(_ reason: String?) -> FinishReason {
+        switch reason {
+        case "content_filter":
+            .contentFilter
+        case "max_output_tokens":
+            .length
+        default:
+            .other
+        }
+    }
+
     private func makeMessageEntry(role: String, message: ModelMessage) -> ResponsesMessage? {
         let parts = self.convertContentParts(for: message)
         guard !parts.isEmpty else { return nil }
@@ -731,15 +849,16 @@ public final class OpenAIResponsesProvider: ModelProvider {
     }
 
     static func convertToProviderResponse(_ response: OpenAIResponsesResponse) throws -> ProviderResponse {
-        // Handle GPT-5 format (output array) vs O3 format (choices array)
-        let text: String
-        let toolCalls: [AgentToolCall]?
-        let finishReason: FinishReason?
+        // Handle GPT-5 output arrays and alternate choices arrays.
+        var text: String
+        var toolCalls: [AgentToolCall]?
+        var finishReason: FinishReason?
 
         if let outputs = response.output {
             // GPT-5 format with output array
             var collectedText = ""
             var collectedToolCalls: [AgentToolCall] = []
+            var didCollectRefusal = false
 
             for output in outputs {
                 if output.type == "message" {
@@ -750,6 +869,10 @@ public final class OpenAIResponsesProvider: ModelProvider {
                                 if let textSegment = chunk.text {
                                     collectedText.append(textSegment)
                                 }
+                            case "refusal":
+                                if chunk.refusal != nil || chunk.text != nil {
+                                    didCollectRefusal = true
+                                }
                             case "tool_call":
                                 if
                                     let toolCall = chunk.toolCall,
@@ -771,14 +894,24 @@ public final class OpenAIResponsesProvider: ModelProvider {
             }
 
             text = collectedText
-            toolCalls = collectedToolCalls.isEmpty ? nil : collectedToolCalls
-            if let toolCalls, !toolCalls.isEmpty {
-                finishReason = .toolCalls
+            let incompleteFinishReason = response.status == "incomplete"
+                ? Self.finishReasonForIncompleteReason(response.incompleteDetails?.reason)
+                : nil
+            if incompleteFinishReason == .contentFilter || didCollectRefusal {
+                text = ""
+                toolCalls = nil
+                finishReason = .contentFilter
             } else {
-                finishReason = .stop
+                toolCalls = collectedToolCalls.isEmpty ? nil : collectedToolCalls
+            }
+            if finishReason == nil, let toolCalls, !toolCalls.isEmpty {
+                finishReason = .toolCalls
+            }
+            if finishReason == nil {
+                finishReason = incompleteFinishReason ?? .stop
             }
         } else if let choices = response.choices, let choice = choices.first {
-            // O3 format with choices array
+            // Alternate format with choices array.
             text = choice.message.content ?? ""
 
             // Convert tool calls
@@ -790,20 +923,26 @@ public final class OpenAIResponsesProvider: ModelProvider {
                 case "stop": finishReason = .stop
                 case "length": finishReason = .length
                 case "tool_calls": finishReason = .toolCalls
+                case "content_filter": finishReason = .contentFilter
                 default: finishReason = .stop
                 }
             } else {
                 finishReason = nil
             }
+            if finishReason == .contentFilter || choice.message.refusal != nil {
+                text = ""
+                toolCalls = nil
+                finishReason = .contentFilter
+            }
         } else {
             throw TachikomaError.apiError("No output or choices in response")
         }
 
-        // Convert usage (handle both GPT-5 and O3 formats)
+        // Convert usage across Responses API token field variants.
         let usage: Usage?
         if let apiUsage = response.usage {
             // GPT-5 uses input_tokens/output_tokens
-            // O3 uses prompt_tokens/completion_tokens
+            // Alternate responses can use prompt_tokens/completion_tokens.
             let inputTokens = apiUsage.inputTokens ?? apiUsage.promptTokens ?? 0
             let outputTokens = apiUsage.outputTokens ?? apiUsage.completionTokens ?? 0
 
@@ -898,31 +1037,28 @@ public final class OpenAIResponsesProvider: ModelProvider {
         }
     }
 
-    private static func isReasoningModel(_ model: LanguageModel.OpenAI) -> Bool {
+    private static func isReasoningModel(_: LanguageModel.OpenAI) -> Bool {
+        false
+    }
+
+    private static func isGPT5Model(_ model: LanguageModel.OpenAI) -> Bool {
         switch model {
-        case .o4Mini:
+        case .gpt55,
+             .gpt54,
+             .gpt54Mini,
+             .gpt54Nano,
+             .gpt5,
+             .gpt5Pro,
+             .gpt5Mini,
+             .gpt5Nano:
             true
         default:
             false
         }
     }
 
-    private static func isGPT5Model(_ model: LanguageModel.OpenAI) -> Bool {
-        switch model {
-        case .gpt52,
-             .gpt51,
-             .gpt5,
-             .gpt5Pro,
-             .gpt5Mini,
-             .gpt5Nano,
-             .gpt5Thinking,
-             .gpt5ThinkingMini,
-             .gpt5ThinkingNano,
-             .gpt5ChatLatest:
-            true
-        default:
-            false
-        }
+    private static func usesResponsesEventStream(_ model: LanguageModel.OpenAI) -> Bool {
+        model == .chatLatest || model == .gpt5ChatLatest || self.isGPT5Model(model)
     }
 }
 

Sources/Tachikoma/Providers/OpenAI/OpenAIResponsesTypes.swift

@@ -28,7 +28,7 @@ struct OpenAIResponsesRequest: Encodable {
     let serviceTier: String?
     let include: [String]?
 
-    /// Reasoning configuration (for o3/o4/GPT-5)
+    /// Reasoning configuration (for GPT-5 thinking models)
     let reasoning: ReasoningConfig?
 
     /// Truncation for long inputs
@@ -477,18 +477,24 @@ struct OpenAIResponsesResponse: Codable {
     let id: String
     let object: String?
     let createdAt: Int? // GPT-5 uses created_at
-    let created: Int? // O3 uses created
+    let created: Int? // Alternate responses can use created
     let status: String?
     let model: String
     let output: [ResponsesOutput]? // GPT-5 uses output array
-    let choices: [ResponsesChoice]? // O3 uses choices array
+    let choices: [ResponsesChoice]? // Alternate responses can use choices array
     let usage: ResponsesUsage?
     let metadata: ResponsesMetadata?
+    let incompleteDetails: IncompleteDetails?
 
     enum CodingKeys: String, CodingKey {
         case id, object, status, model, output, choices, usage, metadata
         case createdAt = "created_at"
         case created
+        case incompleteDetails = "incomplete_details"
+    }
+
+    struct IncompleteDetails: Codable {
+        let reason: String?
     }
 
     /// GPT-5 output format
@@ -513,17 +519,31 @@ struct OpenAIResponsesResponse: Codable {
         struct OutputContent: Codable {
             let type: String
             let text: String?
+            let refusal: String?
             let toolCall: ResponsesToolCall?
 
+            init(
+                type: String,
+                text: String? = nil,
+                refusal: String? = nil,
+                toolCall: ResponsesToolCall? = nil,
+            ) {
+                self.type = type
+                self.text = text
+                self.refusal = refusal
+                self.toolCall = toolCall
+            }
+
             enum CodingKeys: String, CodingKey {
                 case type
                 case text
+                case refusal
                 case toolCall = "tool_call"
             }
         }
     }
 
-    /// O3 choices format (kept for compatibility)
+    /// Alternate choices format.
     struct ResponsesChoice: Codable {
         let index: Int
         let message: ResponsesOutputMessage
@@ -614,7 +634,7 @@ struct OpenAIResponsesResponse: Codable {
 
 // MARK: - Streaming Response Types
 
-/// Server-sent event for streaming responses (O3 and older models)
+/// Server-sent event for alternate streaming responses.
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
 struct OpenAIResponsesStreamChunk: Codable {
     let id: String

Sources/Tachikoma/Providers/OpenAI/OpenAITypes.swift

@@ -77,11 +77,14 @@ struct OpenAIChatMessage: Codable {
     let content: Either<String, [OpenAIChatMessageContent]>?
     let toolCallId: String?
     let toolCalls: [AgentToolCall]?
+    let reasoning: String?
+    let reasoningDetails: [JSONValue]?
 
     enum CodingKeys: String, CodingKey {
-        case role, content
+        case role, content, reasoning
         case toolCallId = "tool_call_id"
         case toolCalls = "tool_calls"
+        case reasoningDetails = "reasoning_details"
     }
 
     struct AgentToolCall: Codable {
@@ -100,6 +103,8 @@ struct OpenAIChatMessage: Codable {
         self.content = .left(content)
         self.toolCallId = toolCallId
         self.toolCalls = nil
+        self.reasoning = nil
+        self.reasoningDetails = nil
     }
 
     init(role: String, content: [OpenAIChatMessageContent], toolCallId: String? = nil) {
@@ -107,13 +112,23 @@ struct OpenAIChatMessage: Codable {
         self.content = .right(content)
         self.toolCallId = toolCallId
         self.toolCalls = nil
+        self.reasoning = nil
+        self.reasoningDetails = nil
     }
 
-    init(role: String, content: String? = nil, toolCalls: [AgentToolCall]?) {
+    init(
+        role: String,
+        content: String? = nil,
+        toolCalls: [AgentToolCall]?,
+        reasoning: String? = nil,
+        reasoningDetails: [JSONValue]? = nil,
+    ) {
         self.role = role
         self.content = content.map { .left($0) }
         self.toolCallId = nil
         self.toolCalls = toolCalls
+        self.reasoning = reasoning
+        self.reasoningDetails = reasoningDetails
     }
 }
 
@@ -246,10 +261,13 @@ struct OpenAIChatResponse: Codable {
         let role: String
         let content: String?
         let toolCalls: [AgentToolCall]?
+        let reasoning: String?
+        let reasoningDetails: [JSONValue]?
 
         enum CodingKeys: String, CodingKey {
-            case role, content
+            case role, content, reasoning
             case toolCalls = "tool_calls"
+            case reasoningDetails = "reasoning_details"
         }
     }
 

Sources/Tachikoma/Providers/OpenRouter/OpenRouterProvider.swift

@@ -24,21 +24,27 @@ public final class OpenRouterProvider: ModelProvider {
 
         if let key = configuration.getAPIKey(for: .custom("openrouter")) {
             self.apiKey = key
+        } else if let auth = TKAuthManager.shared.resolveAuth(for: .openrouter) {
+            switch auth {
+            case let .apiKey(key), let .bearer(key, _):
+                self.apiKey = key
+            }
         } else {
             throw TachikomaError.authenticationFailed("OPENROUTER_API_KEY not found")
         }
 
+        let isFable = LanguageModel.Anthropic.isFable(modelId: modelId)
         self.capabilities = ModelCapabilities(
             supportsVision: true,
             supportsTools: true,
-            supportsStreaming: true,
-            contextLength: 128_000,
-            maxOutputTokens: 4096,
+            supportsStreaming: !LanguageModel.Anthropic.hasStreamingRefusalRisk(modelId: modelId),
+            contextLength: isFable ? 1_000_000 : 128_000,
+            maxOutputTokens: isFable ? 128_000 : 4096,
         )
 
         self.defaultHeaders = [
             "HTTP-Referer": ProcessInfo.processInfo.environment["OPENROUTER_REFERER"] ?? "https://peekaboo.app",
-            "X-Title": ProcessInfo.processInfo.environment["OPENROUTER_TITLE"] ?? "Peekaboo",
+            "X-OpenRouter-Title": ProcessInfo.processInfo.environment["OPENROUTER_TITLE"] ?? "Peekaboo",
         ]
     }
 
@@ -62,6 +68,9 @@ public final class OpenRouterProvider: ModelProvider {
         guard let baseURL, let apiKey else {
             throw TachikomaError.invalidConfiguration("OpenRouter provider missing base URL or API key")
         }
+        guard !LanguageModel.Anthropic.hasStreamingRefusalRisk(modelId: self.modelId) else {
+            throw TachikomaError.invalidConfiguration("\(self.modelId) does not support streaming")
+        }
 
         return try await OpenAICompatibleHelper.streamText(
             request: request,

Sources/Tachikoma/Providers/ProviderFactory.swift

@@ -15,19 +15,18 @@ public struct ProviderFactory {
         // Create a provider for the specified language model
         switch model {
         case let .openai(openaiModel):
-            // Use Responses API for reasoning models (o4) and GPT-5 family
+            // Use Responses API for the GPT-5 family
             switch openaiModel {
-            case .o4Mini,
-                 .gpt52,
-                 .gpt51,
+            case .chatLatest,
+                 .gpt5ChatLatest,
+                 .gpt55,
+                 .gpt54,
+                 .gpt54Mini,
+                 .gpt54Nano,
                  .gpt5,
                  .gpt5Pro,
                  .gpt5Mini,
-                 .gpt5Nano,
-                 .gpt5Thinking,
-                 .gpt5ThinkingMini,
-                 .gpt5ThinkingNano,
-                 .gpt5ChatLatest:
+                 .gpt5Nano:
                 return try OpenAIResponsesProvider(model: openaiModel, configuration: configuration)
             default:
                 return try OpenAIProvider(model: openaiModel, configuration: configuration)
@@ -59,6 +58,28 @@ public struct ProviderFactory {
                 modelId: lmstudioModel.modelId,
             )
 
+        case let .minimax(minimaxModel):
+            guard let apiKey = configuration.getAPIKey(for: .minimax) else {
+                throw TachikomaError.authenticationFailed("MINIMAX_API_KEY not found")
+            }
+            return try Self.makeMiniMaxProvider(
+                model: minimaxModel,
+                provider: .minimax,
+                apiKey: apiKey,
+                configuration: configuration,
+            )
+
+        case let .minimaxCN(minimaxModel):
+            guard let apiKey = configuration.getAPIKey(for: .minimaxCN) ?? configuration.getAPIKey(for: .minimax) else {
+                throw TachikomaError.authenticationFailed("MINIMAX_CN_API_KEY or MINIMAX_API_KEY not found")
+            }
+            return try Self.makeMiniMaxProvider(
+                model: minimaxModel,
+                provider: .minimaxCN,
+                apiKey: apiKey,
+                configuration: configuration,
+            )
+
         case let .openRouter(modelId):
             return try OpenRouterProvider(modelId: modelId, configuration: configuration)
 
@@ -94,12 +115,18 @@ public struct ProviderFactory {
                             modelId: parsed.model,
                             baseURL: custom.baseURL,
                             configuration: configuration,
+                            apiKey: custom.apiKey,
+                            additionalHeaders: custom.headers,
                         )
                     case .anthropic:
                         return try AnthropicCompatibleProvider(
                             modelId: parsed.model,
                             baseURL: custom.baseURL,
                             configuration: configuration,
+                            apiKey: custom.apiKey,
+                            additionalHeaders: custom.headers,
+                            reasoningProvider: "custom-anthropic",
+                            reasoningBaseURL: custom.baseURL,
                         )
                     }
                 }
@@ -107,6 +134,34 @@ public struct ProviderFactory {
             return provider
         }
     }
+
+    private static func makeMiniMaxProvider(
+        model: LanguageModel.MiniMax,
+        provider: Provider,
+        apiKey: String,
+        configuration: TachikomaConfiguration,
+    ) throws
+        -> any ModelProvider
+    {
+        let baseURL = configuration.getBaseURL(for: provider) ?? provider.defaultBaseURL ?? "https://api.minimax.io/anthropic"
+        return try AnthropicCompatibleProvider(
+            modelId: model.modelId,
+            baseURL: baseURL,
+            configuration: configuration,
+            apiKey: apiKey,
+            // MiniMax's Anthropic-compatible setup uses Claude Code-style Authorization auth, not Anthropic x-api-key.
+            auth: .bearer(apiKey, betaHeader: nil),
+            capabilities: ModelCapabilities(
+                supportsVision: model.supportsVision,
+                supportsTools: model.supportsTools,
+                supportsStreaming: true,
+                contextLength: model.contextLength,
+                maxOutputTokens: 8192,
+            ),
+            reasoningProvider: provider == .minimaxCN ? "minimax-cn" : "minimax",
+            reasoningBaseURL: baseURL,
+        )
+    }
 }
 
 // MARK: - Third-Party Aggregators

Sources/Tachikoma/Providers/ProviderParser.swift

@@ -8,10 +8,10 @@ public enum ProviderParser {
         /// The provider name (e.g., "openai", "anthropic", "ollama")
         public let provider: String
 
-        /// The model name (e.g., "gpt-4", "claude-3", "llava:latest")
+        /// The model name (e.g., "gpt-5.5", "claude-fable-5", "llava:latest")
         public let model: String
 
-        /// The full string representation (e.g., "openai/gpt-4")
+        /// The full string representation (e.g., "openai/gpt-5.5")
         public var fullString: String {
             "\(self.provider)/\(self.model)"
         }
@@ -23,7 +23,7 @@ public enum ProviderParser {
     }
 
     /// Parse a provider string in the format "provider/model"
-    /// - Parameter providerString: String like "openai/gpt-4" or "ollama/llava:latest"
+    /// - Parameter providerString: String like "openai/gpt-5.5" or "ollama/llava:latest"
     /// - Returns: Parsed configuration or nil if invalid format
     public static func parse(_ providerString: String) -> ProviderConfig? {
         // Parse a provider string in the format "provider/model"
@@ -44,7 +44,7 @@ public enum ProviderParser {
     }
 
     /// Parse a comma-separated list of providers
-    /// - Parameter providersString: String like "openai/gpt-4,anthropic/claude-3,ollama/llava:latest"
+    /// - Parameter providersString: String like "openai/gpt-5.5,anthropic/claude-fable-5,ollama/llava:latest"
     /// - Returns: Array of parsed configurations
     public static func parseList(_ providersString: String) -> [ProviderConfig] {
         // Parse a comma-separated list of providers
@@ -54,7 +54,7 @@ public enum ProviderParser {
     }
 
     /// Get the first provider from a comma-separated list
-    /// - Parameter providersString: String like "openai/gpt-4,anthropic/claude-3"
+    /// - Parameter providersString: String like "openai/gpt-5.5,anthropic/claude-fable-5"
     /// - Returns: First parsed configuration or nil if none valid
     public static func parseFirst(_ providersString: String) -> ProviderConfig? {
         // Get the first provider from a comma-separated list
@@ -94,6 +94,8 @@ public enum ProviderParser {
     ///   - hasOpenAI: Whether OpenAI API key is available
     ///   - hasAnthropic: Whether Anthropic API key is available
     ///   - hasGrok: Whether Grok API key is available
+    ///   - hasGoogle: Whether Google/Gemini API key is available
+    ///   - hasMiniMax: Whether MiniMax API key is available
     ///   - hasOllama: Whether Ollama is available (always true as it doesn't require API key)
     ///   - configuredDefault: Optional default from configuration
     ///   - isEnvironmentProvided: Whether the providers string came from environment variable
@@ -103,6 +105,8 @@ public enum ProviderParser {
         hasOpenAI: Bool,
         hasAnthropic: Bool,
         hasGrok: Bool = false,
+        hasGoogle: Bool? = nil,
+        hasMiniMax: Bool = false,
         hasOllama: Bool = true,
         configuredDefault: LanguageModel? = nil,
         isEnvironmentProvided: Bool = false,
@@ -113,14 +117,21 @@ public enum ProviderParser {
         let providers = self.parseList(providersString)
         var environmentModel: LanguageModel?
 
+        let canUseGoogleProvider = hasGoogle ?? true
+        let canFallbackToGoogle = hasGoogle ?? false
+
         for config in providers {
             switch config.provider.lowercased() {
             case "openai" where hasOpenAI:
                 environmentModel = self.parseOpenAIModel(config.model)
             case "anthropic" where hasAnthropic:
                 environmentModel = self.parseAnthropicModel(config.model)
-            case "google", "gemini":
+            case "google" where canUseGoogleProvider, "gemini" where canUseGoogleProvider:
                 environmentModel = self.parseGoogleModel(config.model)
+            case "minimax" where hasMiniMax:
+                environmentModel = self.parseMiniMaxModel(config.model)
+            case "minimax-cn" where hasMiniMax, "minimax_cn" where hasMiniMax, "minimaxi" where hasMiniMax:
+                environmentModel = self.parseMiniMaxCNModel(config.model)
             case "grok" where hasGrok, "xai" where hasGrok:
                 environmentModel = self.parseGrokModel(config.model)
             case "ollama" where hasOllama:
@@ -151,6 +162,8 @@ public enum ProviderParser {
                 hasOpenAI: hasOpenAI,
                 hasAnthropic: hasAnthropic,
                 hasGrok: hasGrok,
+                hasGoogle: canFallbackToGoogle,
+                hasMiniMax: hasMiniMax,
                 hasOllama: hasOllama,
             )
         }
@@ -169,6 +182,8 @@ public enum ProviderParser {
         hasOpenAI: Bool,
         hasAnthropic: Bool,
         hasGrok: Bool = false,
+        hasGoogle: Bool? = nil,
+        hasMiniMax: Bool = false,
         hasOllama: Bool = true,
         configuredDefault: LanguageModel? = nil,
     )
@@ -180,6 +195,8 @@ public enum ProviderParser {
             hasOpenAI: hasOpenAI,
             hasAnthropic: hasAnthropic,
             hasGrok: hasGrok,
+            hasGoogle: hasGoogle,
+            hasMiniMax: hasMiniMax,
             hasOllama: hasOllama,
             configuredDefault: configuredDefault,
             isEnvironmentProvided: false,
@@ -202,32 +219,38 @@ public enum ProviderParser {
     // MARK: - Private Helpers
 
     private static func parseOpenAIModel(_ modelString: String) -> LanguageModel? {
-        switch modelString.lowercased() {
-        case "o4-mini": .openai(.o4Mini)
-        case "gpt-5.2", "gpt5.2", "gpt-5-2", "gpt5-2", "gpt52": .openai(.gpt52)
-        case "gpt-5.2-mini", "gpt5.2-mini", "gpt-5-2-mini", "gpt5-2-mini", "gpt52-mini", "gpt52mini":
+        let normalized = modelString.lowercased()
+        let compact = normalized.replacingOccurrences(of: "-", with: "").replacingOccurrences(of: ".", with: "")
+        guard
+            !normalized.hasPrefix("gpt-4"), !compact.hasPrefix("gpt4"),
+            !normalized.hasPrefix("gpt-3"), !compact.hasPrefix("gpt3"),
+            !normalized.hasPrefix("o3"), !normalized.hasPrefix("o4"),
+            !normalized.hasPrefix("gpt-5.1"), !compact.hasPrefix("gpt51"),
+            !normalized.hasPrefix("gpt-5.2"), !compact.hasPrefix("gpt52"),
+            !normalized.contains("gpt-5-thinking"), !compact.contains("gpt5thinking") else
+        {
+            return nil
+        }
+
+        return switch normalized {
+        case "chat-latest", "chatlatest":
+            .openai(.chatLatest)
+        case "gpt-5-chat-latest", "gpt5-chat-latest", "gpt5chatlatest":
+            .openai(.gpt5ChatLatest)
+        case "gpt-5.5", "gpt5.5", "gpt-5-5", "gpt5-5", "gpt55": .openai(.gpt55)
+        case "gpt-5.5-mini", "gpt5.5-mini", "gpt-5-5-mini", "gpt5-5-mini", "gpt55-mini", "gpt55mini":
             .openai(.gpt5Mini)
-        case "gpt-5.2-nano", "gpt5.2-nano", "gpt-5-2-nano", "gpt5-2-nano", "gpt52-nano", "gpt52nano":
-            .openai(.gpt5Nano)
-        case "gpt-5.1", "gpt5.1", "gpt-5-1", "gpt5-1", "gpt51": .openai(.gpt51)
-        case "gpt-5.1-mini", "gpt5.1-mini", "gpt-5-1-mini", "gpt5-1-mini", "gpt51-mini", "gpt51mini":
-            .openai(.gpt5Mini)
-        case "gpt-5.1-nano", "gpt5.1-nano", "gpt-5-1-nano", "gpt5-1-nano", "gpt51-nano", "gpt51nano":
+        case "gpt-5.5-nano", "gpt5.5-nano", "gpt-5-5-nano", "gpt5-5-nano", "gpt55-nano", "gpt55nano":
             .openai(.gpt5Nano)
+        case "gpt-5.4", "gpt5.4", "gpt-5-4", "gpt5-4", "gpt54": .openai(.gpt54)
+        case "gpt-5.4-mini", "gpt5.4-mini", "gpt-5-4-mini", "gpt5-4-mini", "gpt54-mini", "gpt54mini":
+            .openai(.gpt54Mini)
+        case "gpt-5.4-nano", "gpt5.4-nano", "gpt-5-4-nano", "gpt5-4-nano", "gpt54-nano", "gpt54nano":
+            .openai(.gpt54Nano)
         case "gpt-5", "gpt5": .openai(.gpt5)
         case "gpt-5-pro", "gpt5-pro", "gpt5pro": .openai(.gpt5Pro)
         case "gpt-5-mini", "gpt5-mini", "gpt5mini": .openai(.gpt5Mini)
         case "gpt-5-nano", "gpt5-nano", "gpt5nano": .openai(.gpt5Nano)
-        case "gpt-5-thinking", "gpt5-thinking", "gpt5thinking": .openai(.gpt5Thinking)
-        case "gpt-5-thinking-mini", "gpt5-thinking-mini", "gpt5thinkingmini": .openai(.gpt5ThinkingMini)
-        case "gpt-5-thinking-nano", "gpt5-thinking-nano", "gpt5thinkingnano": .openai(.gpt5ThinkingNano)
-        case "gpt-5-chat-latest", "gpt5-chat-latest": .openai(.gpt5ChatLatest)
-        case "gpt-4.1", "gpt4.1": .openai(.gpt41)
-        case "gpt-4.1-mini", "gpt4.1-mini": .openai(.gpt41Mini)
-        case "gpt-4o", "gpt4o": .openai(.gpt4o)
-        case "gpt-4o-mini", "gpt4o-mini": .openai(.gpt4oMini)
-        case "gpt-4-turbo", "gpt4-turbo": .openai(.gpt4Turbo)
-        case "gpt-3.5-turbo", "gpt35-turbo": .openai(.gpt35Turbo)
         default:
             // Handle custom/fine-tuned models
             .openai(.custom(modelString))
@@ -235,15 +258,34 @@ public enum ProviderParser {
     }
 
     private static func parseAnthropicModel(_ modelString: String) -> LanguageModel? {
-        switch modelString.lowercased() {
+        let normalized = modelString.lowercased()
+        let compact = normalized.replacingOccurrences(of: "-", with: "").replacingOccurrences(of: ".", with: "")
+        guard
+            !normalized.hasPrefix("claude-3"), !compact.hasPrefix("claude3"),
+            normalized != "claude-opus-4-20250514",
+            normalized != "claude-sonnet-4-20250514",
+            !normalized.contains("-thinking") else
+        {
+            return nil
+        }
+
+        return switch normalized {
+        case "claude-fable-5", "claude-fable-5-latest", "fable-5", "fable.5", "fable5", "fable":
+            .anthropic(.fable5)
+        case "claude-opus-4-8", "claude-opus-4.8", "claude-opus-4-8-latest", "opus-4-8", "opus-4.8",
+             "opus48", "claude", "claude-latest", "claude_latest", "claudelatest", "claude-default", "claude_default":
+            .anthropic(.opus48)
+        case "claude-opus-4-7", "claude-opus-4.7", "claude-opus-4-7-latest", "opus-4-7", "opus-4.7", "opus47":
+            .anthropic(.opus47)
         case "claude-opus-4-5", "claude-opus-4.5", "claude-opus-4-5-latest", "opus-4-5", "opus-4.5", "opus45":
             .anthropic(.opus45)
-        case "claude-opus-4-1-20250805", "claude-opus-4-20250514", "claude-opus-4", "opus-4": .anthropic(.opus4)
-        case "claude-opus-4-1-20250805-thinking", "claude-opus-4-20250514-thinking", "claude-opus-4-thinking",
-             "opus-4-thinking": .anthropic(.opus4Thinking)
-        case "claude-sonnet-4-20250514", "claude-sonnet-4", "sonnet-4": .anthropic(.sonnet4)
-        case "claude-sonnet-4-20250514-thinking", "claude-sonnet-4-thinking",
-             "sonnet-4-thinking": .anthropic(.sonnet4Thinking)
+        case "claude-opus-4-1-20250805", "claude-opus-4", "opus-4": .anthropic(.opus4)
+        case "claude-sonnet-4-6", "claude-sonnet-4.6", "sonnet-4-6", "sonnet-4.6", "sonnet46":
+            .anthropic(.sonnet46)
+        case "claude-sonnet-4-5-20250929", "claude-sonnet-4.5", "sonnet-4-5", "sonnet-4.5", "sonnet45":
+            .anthropic(.sonnet45)
+        case "claude-haiku-4-5-20251001", "claude-haiku-4-5", "claude-haiku-4.5", "haiku-4-5", "haiku45":
+            .anthropic(.haiku45)
         default:
             // Handle custom models
             .anthropic(.custom(modelString))
@@ -252,6 +294,12 @@ public enum ProviderParser {
 
     private static func parseGoogleModel(_ modelString: String) -> LanguageModel? {
         switch modelString.lowercased() {
+        case "gemini-3.5-flash", "gemini35flash", "gemini-3-5-flash":
+            .google(.gemini35Flash)
+        case "gemini-3.1-pro-preview", "gemini-3.1-pro", "gemini31pro", "gemini31propreview":
+            .google(.gemini31ProPreview)
+        case "gemini-3.1-flash-lite", "gemini31flashlite", "gemini-3.1-flashlite":
+            .google(.gemini31FlashLite)
         case "gemini-3-flash", "gemini-3-flash-preview", "gemini3flash", "gemini-3flash":
             .google(.gemini3Flash)
         case "gemini-2.5-pro", "gemini25pro", "gemini2.5pro":
@@ -261,7 +309,29 @@ public enum ProviderParser {
         case "gemini-2.5-flash-lite", "gemini25flashlite", "gemini-2.5-flashlite":
             .google(.gemini25FlashLite)
         case "gemini":
-            .google(.gemini3Flash)
+            .google(.gemini35Flash)
+        default:
+            nil
+        }
+    }
+
+    private static func parseMiniMaxModel(_ modelString: String) -> LanguageModel? {
+        switch modelString.lowercased() {
+        case "minimax-m2.7", "minimax-m2-7", "m2.7", "m2-7":
+            .minimax(.m27)
+        case "minimax-m2.7-highspeed", "minimax-m2-7-highspeed", "m2.7-highspeed", "m2-7-highspeed":
+            .minimax(.m27Highspeed)
+        default:
+            nil
+        }
+    }
+
+    private static func parseMiniMaxCNModel(_ modelString: String) -> LanguageModel? {
+        switch modelString.lowercased() {
+        case "minimax-m2.7", "minimax-m2-7", "m2.7", "m2-7":
+            .minimaxCN(.m27)
+        case "minimax-m2.7-highspeed", "minimax-m2-7-highspeed", "m2.7-highspeed", "m2-7-highspeed":
+            .minimaxCN(.m27Highspeed)
         default:
             nil
         }
@@ -269,22 +339,27 @@ public enum ProviderParser {
 
     private static func parseGrokModel(_ modelString: String) -> LanguageModel? {
         switch modelString.lowercased() {
-        case "grok-4-0709": .grok(.grok4)
-        case "grok-4-fast-reasoning": .grok(.grok4FastReasoning)
-        case "grok-4-fast-non-reasoning": .grok(.grok4FastNonReasoning)
-        case "grok-code-fast-1": .grok(.grokCodeFast1)
-        case "grok-3", "grok3": .grok(.grok3)
-        case "grok-3-mini": .grok(.grok3Mini)
-        case "grok-2-1212": .grok(.grok2)
-        case "grok-2-vision-1212": .grok(.grok2Vision)
-        case "grok-2-image-1212": .grok(.grok2Image)
-        case "grok-vision-beta": .grok(.grokVisionBeta)
-        case "grok-beta": .grok(.grokBeta)
+        case "grok-4.3", "grok-4-3", "grok43", "grok-4.3-latest", "grok-4-latest", "grok-4", "grok-latest":
+            return .grok(.grok43)
+        case "grok-4.20-0309-reasoning", "grok-4-20-0309-reasoning":
+            return .grok(.grok420Reasoning)
+        case "grok-4.20-0309-non-reasoning", "grok-4-20-0309-non-reasoning":
+            return .grok(.grok420NonReasoning)
         default:
-            .grok(.custom(modelString))
+            if self.isUnsupportedLegacyGrokModel(modelString) {
+                return nil
+            }
+            return .grok(.custom(modelString))
         }
     }
 
+    private static func isUnsupportedLegacyGrokModel(_ modelString: String) -> Bool {
+        let normalized = modelString.lowercased()
+        return normalized.contains("grok-4.20-multi-agent") ||
+            normalized.contains("grok-4-20-multi-agent") ||
+            normalized.contains("grok420multiagent")
+    }
+
     private static func parseOllamaModel(_ modelString: String) -> LanguageModel? {
         switch modelString.lowercased() {
         // GPT-OSS models
@@ -300,7 +375,6 @@ public enum ProviderParser {
         case "llava:34b": .ollama(.custom("llava:34b"))
         case "mistral-nemo", "mistral-nemo:latest": .ollama(.mistralNemo)
         case "qwen2.5", "qwen2.5:latest": .ollama(.qwen25)
-        case "codellama", "codellama:latest": .ollama(.codellama)
         default:
             .ollama(.custom(modelString))
         }
@@ -310,16 +384,22 @@ public enum ProviderParser {
         hasOpenAI: Bool,
         hasAnthropic: Bool,
         hasGrok: Bool,
+        hasGoogle: Bool,
+        hasMiniMax: Bool,
         hasOllama _: Bool,
     )
         -> LanguageModel
     {
         if hasAnthropic {
-            .anthropic(.opus4)
+            .anthropic(.opus48)
         } else if hasOpenAI {
-            .openai(.gpt5Mini)
+            .openai(.gpt55)
         } else if hasGrok {
-            .grok(.grok4FastReasoning)
+            .grok(.grok43)
+        } else if hasGoogle {
+            .google(.gemini35Flash)
+        } else if hasMiniMax {
+            .minimax(.m27)
         } else {
             .ollama(.llama33)
         }

Sources/Tachikoma/Providers/Together/TogetherProvider.swift

@@ -27,12 +27,13 @@ public final class TogetherProvider: ModelProvider {
             throw TachikomaError.authenticationFailed("TOGETHER_API_KEY not found")
         }
 
+        let isFable = LanguageModel.Anthropic.isFable(modelId: modelId)
         self.capabilities = ModelCapabilities(
             supportsVision: true,
             supportsTools: true,
-            supportsStreaming: true,
-            contextLength: 128_000,
-            maxOutputTokens: 4096,
+            supportsStreaming: !LanguageModel.Anthropic.hasStreamingRefusalRisk(modelId: modelId),
+            contextLength: isFable ? 1_000_000 : 128_000,
+            maxOutputTokens: isFable ? 128_000 : 4096,
         )
     }
 
@@ -55,6 +56,9 @@ public final class TogetherProvider: ModelProvider {
         guard let baseURL, let apiKey else {
             throw TachikomaError.invalidConfiguration("Together provider missing base URL or API key")
         }
+        guard !LanguageModel.Anthropic.hasStreamingRefusalRisk(modelId: self.modelId) else {
+            throw TachikomaError.invalidConfiguration("\(self.modelId) does not support streaming")
+        }
 
         return try await OpenAICompatibleHelper.streamText(
             request: request,

Sources/Tachikoma/Utilities/ResponseCache.swift

@@ -6,18 +6,44 @@ import UIKit
 // MARK: - Cache Key
 
 /// Hashable key for cache entries
+@available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
+public struct CacheProviderIdentity: Hashable, Sendable {
+    public let providerKind: String
+    public let modelId: String
+    public let endpointIdentity: String?
+
+    public init(providerKind: String, modelId: String, baseURL: String?) {
+        self.providerKind = providerKind
+        self.modelId = modelId
+        self.endpointIdentity = ReasoningEndpointIdentity.canonical(baseURL)
+    }
+}
+
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
 struct CacheKey: Hashable {
     let hash: String
     let model: String? // Store model ID for invalidation
+    let isCacheable: Bool
 
-    init(from request: ProviderRequest, model: String? = nil) {
-        self.model = model
+    init(from request: ProviderRequest, providerIdentity: CacheProviderIdentity? = nil, model: String? = nil) {
+        self.model = providerIdentity?.modelId ?? model
         // Create a unique hash from the request
         var hasher = Hasher()
+        if let providerIdentity {
+            hasher.combine(providerIdentity.providerKind)
+            hasher.combine(providerIdentity.modelId)
+            hasher.combine(providerIdentity.endpointIdentity)
+        }
         // Combine message content
         for message in request.messages {
             hasher.combine(message.role.rawValue)
+            hasher.combine(message.channel?.rawValue)
+            hasher.combine(message.metadata?.conversationId)
+            hasher.combine(message.metadata?.turnId)
+            for key in message.metadata?.customData?.keys.sorted() ?? [] {
+                hasher.combine(key)
+                hasher.combine(message.metadata?.customData?[key])
+            }
             for part in message.content {
                 switch part {
                 case let .text(text):
@@ -41,8 +67,32 @@ struct CacheKey: Hashable {
         hasher.combine(request.settings.temperature)
         hasher.combine(request.settings.maxTokens)
         hasher.combine(request.settings.topP)
+        hasher.combine(request.settings.topK)
+        hasher.combine(request.settings.frequencyPenalty)
+        hasher.combine(request.settings.presencePenalty)
+        hasher.combine(request.settings.stopSequences)
+        hasher.combine(request.settings.reasoningEffort?.rawValue)
+        hasher.combine(request.settings.seed)
+        if let stopConditions = request.settings.stopConditions {
+            guard let cacheKey = (stopConditions as? StableCacheKeyStopCondition)?.stableCacheKey else {
+                self.hash = ""
+                self.isCacheable = false
+                return
+            }
+            hasher.combine(cacheKey)
+        }
+        if let providerOptionsData = try? Self.providerOptionsEncoder.encode(request.settings.providerOptions) {
+            hasher.combine(providerOptionsData)
+        }
         self.hash = String(hasher.finalize())
+        self.isCacheable = true
     }
+
+    private static let providerOptionsEncoder: JSONEncoder = {
+        let encoder = JSONEncoder()
+        encoder.outputFormatting = [.sortedKeys]
+        return encoder
+    }()
 }
 
 // MARK: - Response Cache
@@ -75,11 +125,16 @@ public actor ResponseCache {
     public func get(
         for request: ProviderRequest,
         ttlOverride: TimeInterval? = nil,
+        providerIdentity: CacheProviderIdentity? = nil,
     )
         -> ProviderResponse?
     {
         // Get cached response with TTL validation
-        let key = CacheKey(from: request)
+        let key = CacheKey(from: request, providerIdentity: providerIdentity)
+        guard key.isCacheable else {
+            self.statistics.recordMiss()
+            return nil
+        }
 
         guard let entry = cache[key] else {
             self.statistics.recordMiss()
@@ -109,9 +164,13 @@ public actor ResponseCache {
         for request: ProviderRequest,
         ttl: TimeInterval? = nil,
         priority: CachePriority = .normal,
+        providerIdentity: CacheProviderIdentity? = nil,
     ) {
         // Store response with custom TTL and priority
-        let key = CacheKey(from: request)
+        let key = CacheKey(from: request, providerIdentity: providerIdentity)
+        guard key.isCacheable else {
+            return
+        }
 
         // Check memory limit
         if self.shouldEvictForMemory() {
@@ -457,9 +516,33 @@ final class CacheEntry: @unchecked Sendable {
         // Rough estimation based on response content
         let textSize = self.response.text.utf8.count
         let toolCallsSize = (response.toolCalls?.count ?? 0) * 100 // Estimate 100 bytes per tool call
+        let reasoningSize = self.response.reasoning.reduce(0) { total, block in
+            total + block.text.utf8.count +
+                (block.signature?.utf8.count ?? 0) +
+                (block.rawJSON?.utf8.count ?? 0) +
+                block.type.utf8.count
+        }
+        let assistantMessageSize = self.response.assistantMessages.reduce(0) { total, message in
+            let contentSize = message.content.reduce(0) { contentTotal, part in
+                switch part {
+                case let .text(text):
+                    contentTotal + text.utf8.count
+                case let .image(image):
+                    contentTotal + image.mimeType.utf8.count + image.data.utf8.count
+                case let .toolCall(call):
+                    contentTotal + call.id.utf8.count + call.name.utf8.count + 100
+                case let .toolResult(result):
+                    contentTotal + result.toolCallId.utf8.count + 100
+                }
+            }
+            let metadataSize = (message.metadata?.customData ?? [:]).reduce(0) { metadataTotal, pair in
+                metadataTotal + pair.key.utf8.count + pair.value.utf8.count
+            }
+            return total + contentSize + metadataSize + (message.channel?.rawValue.utf8.count ?? 0)
+        }
         let usageSize = 50 // Fixed overhead for usage data
 
-        return textSize + toolCallsSize + usageSize + 100 // 100 bytes overhead
+        return textSize + toolCallsSize + reasoningSize + assistantMessageSize + usageSize + 100
     }
 }
 
@@ -546,6 +629,7 @@ extension ResponseCache {
 public struct CacheAwareProvider<Base: ModelProvider>: ModelProvider {
     let provider: Base
     let cache: ResponseCache
+    private let providerIdentity: CacheProviderIdentity
 
     public var modelId: String {
         self.provider.modelId
@@ -563,11 +647,21 @@ public struct CacheAwareProvider<Base: ModelProvider>: ModelProvider {
         self.provider.capabilities
     }
 
+    init(provider: Base, cache: ResponseCache) {
+        self.provider = provider
+        self.cache = cache
+        self.providerIdentity = CacheProviderIdentity(
+            providerKind: String(reflecting: Base.self),
+            modelId: provider.modelId,
+            baseURL: provider.baseURL,
+        )
+    }
+
     public func generateText(request: ProviderRequest) async throws -> ProviderResponse {
         // Check cache with smart TTL based on request type
         let ttl = self.determineTTL(for: request)
 
-        if let cached = await cache.get(for: request, ttlOverride: ttl) {
+        if let cached = await cache.get(for: request, ttlOverride: ttl, providerIdentity: self.providerIdentity) {
             return cached
         }
 
@@ -575,7 +669,13 @@ public struct CacheAwareProvider<Base: ModelProvider>: ModelProvider {
         let response = try await provider.generateText(request: request)
         let priority = self.determinePriority(for: request)
 
-        await self.cache.store(response, for: request, ttl: ttl, priority: priority)
+        await self.cache.store(
+            response,
+            for: request,
+            ttl: ttl,
+            priority: priority,
+            providerIdentity: self.providerIdentity,
+        )
         return response
     }
 

Sources/Tachikoma/Utilities/UsageTracking.swift

@@ -528,39 +528,38 @@ public struct ModelCostCalculator: Sendable {
         // OpenAI Pricing (as of 2025)
         case let .openai(openaiModel):
             switch openaiModel {
-            case .o4Mini: (1.50, 6.00)
-            case .gpt52: (5.00, 20.00) // GPT-5.2 pricing estimate
-            case .gpt51: (5.00, 20.00) // GPT-5.1 pricing estimate
+            case .chatLatest: (5.00, 30.00) // ChatGPT Instant alias pricing estimate
+            case .gpt5ChatLatest: (1.25, 10.00)
+            case .gpt55: (5.00, 20.00) // GPT-5.5 pricing estimate
+            case .gpt54: (5.00, 20.00) // GPT-5.4 pricing estimate
+            case .gpt54Mini: (1.00, 4.00)
+            case .gpt54Nano: (0.50, 2.00)
             case .gpt5: (5.00, 20.00) // GPT-5 pricing estimate
             case .gpt5Pro: (12.00, 48.00) // Higher reasoning budget
             case .gpt5Mini: (1.00, 4.00) // GPT-5 Mini pricing estimate
             case .gpt5Nano: (0.50, 2.00) // GPT-5 Nano pricing estimate
-            case .gpt5Thinking: (16.00, 64.00) // Extended reasoning premium
-            case .gpt5ThinkingMini: (4.00, 16.00)
-            case .gpt5ThinkingNano: (1.50, 6.00)
-            case .gpt5ChatLatest: (2.50, 10.00)
-            case .gpt41: (2.50, 10.00)
-            case .gpt41Mini: (0.15, 0.60)
-            case .gpt4o: (2.50, 10.00)
-            case .gpt4oMini: (0.15, 0.60)
-            case .gpt4oRealtime: (5.00, 20.00) // Realtime API pricing estimate
-            case .gpt4Turbo: (10.00, 30.00)
-            case .gpt35Turbo: (0.50, 1.50)
             case .custom: (2.50, 10.00) // Default estimate
             }
-        // Anthropic Pricing (as of 2025)
+        // Anthropic Pricing (as of 2026)
         case let .anthropic(anthropicModel):
             switch anthropicModel {
+            case .fable5: (10.00, 50.00)
+            case .opus48: (5.00, 25.00)
+            case .opus47: (5.00, 25.00)
             case .opus45: (5.00, 25.00)
-            case .opus4, .opus4Thinking: (15.00, 75.00)
-            case .sonnet4, .sonnet4Thinking: (3.00, 15.00)
+            case .opus4: (15.00, 75.00)
+            case .sonnet46: (3.00, 15.00)
             case .sonnet45: (4.00, 18.00)
             case .haiku45: (1.20, 6.00)
-            case .custom: (3.00, 15.00) // Default estimate
+            case let .custom(id):
+                id.lowercased().contains("claude-fable-5") ? (10.00, 50.00) : (3.00, 15.00)
             }
-        // Google Pricing (estimates)
+        // Google Pricing (standard tier, as of 2026)
         case let .google(googleModel):
             switch googleModel {
+            case .gemini35Flash: (1.50, 9.00)
+            case .gemini31ProPreview: (1.25, 10.00)
+            case .gemini31FlashLite: (0.10, 0.40)
             case .gemini3Flash: (0.50, 3.00)
             case .gemini25Pro: (1.25, 10.00)
             case .gemini25Flash: (0.30, 2.50)
@@ -569,7 +568,12 @@ public struct ModelCostCalculator: Sendable {
         // Other providers - estimates
         case .mistral: (2.00, 6.00)
         case .groq: (0.27, 0.27) // Groq has very low pricing
-        case .grok: (2.00, 8.00)
+        case let .grok(grokModel):
+            switch grokModel {
+            case .grok43, .grok420MultiAgent, .grok420Reasoning, .grok420NonReasoning: (1.25, 2.50)
+            case .custom: (2.00, 8.00)
+            }
+        case .minimax, .minimaxCN: (0.30, 1.20)
         case .ollama: (0.00, 0.00) // Local inference
         case .lmstudio: (0.00, 0.00) // Local inference
         case .openRouter, .together, .replicate: (1.00, 3.00) // Typical aggregator pricing

Sources/TachikomaAgent/Agent.swift

@@ -26,32 +26,43 @@ public final class Agent<Context>: @unchecked Sendable {
     public private(set) var tools: [AgentTool]
 
     /// Language model used by this agent
-    public var model: LanguageModel
+    public var model: LanguageModel {
+        didSet {
+            self.usesImplicitDefaultModel = false
+        }
+    }
 
     /// Generation settings for the agent
     public var settings: GenerationSettings
 
+    /// Provider configuration for generation and streaming
+    private let configuration: TachikomaConfiguration
+
     /// The context instance passed to tool executions
     private let context: Context
 
     /// Current conversation history
     public private(set) var conversation: Conversation
+    private var usesImplicitDefaultModel: Bool
 
     public init(
         name: String,
         instructions: String,
-        model: LanguageModel = .default,
+        model: LanguageModel? = nil,
         tools: [AgentTool] = [],
         settings: GenerationSettings = .default,
+        configuration: TachikomaConfiguration = .current,
         context: Context,
     ) {
         self.name = name
         self.instructions = instructions
-        self.model = model
+        self.usesImplicitDefaultModel = model == nil
+        self.model = model ?? .default
         self.tools = tools
         self.settings = settings
+        self.configuration = configuration
         self.context = context
-        self.conversation = Conversation()
+        self.conversation = Conversation(configuration: configuration)
 
         // Add system message with instructions
         self.conversation.addSystemMessage(instructions)
@@ -71,66 +82,133 @@ public final class Agent<Context>: @unchecked Sendable {
 
     /// Execute a single message with the agent
     public func execute(_ message: String) async throws -> AgentResponse {
-        // Add user message to conversation
-        self.conversation.addUserMessage(message)
+        let conversation = self.conversation
+        let model = self.model
+        let tools = self.tools
+        let settings = self.settings
 
-        // Generate response using the conversation
-        let result = try await generateText(
-            model: model,
-            messages: conversation.getModelMessages(),
-            tools: self.tools.isEmpty ? nil : self.tools,
-            settings: self.settings,
-            maxSteps: 5, // Allow multi-step tool execution
-        )
+        return try await conversation.withContinuationLock {
+            conversation.addUserMessage(message)
+            let conversationMessages = conversation.messages
+            let modelMessages = conversationMessages.map { $0.toModelMessage() }
+            let snapshotIDs = conversationMessages.map(\.id)
+            let anchorID = conversationMessages.last?.id
+            let result = try await generateText(
+                model: model,
+                messages: modelMessages,
+                tools: tools.isEmpty ? nil : tools,
+                settings: settings,
+                maxSteps: 5, // Allow multi-step tool execution
+                configuration: self.configuration,
+            )
 
-        // Add assistant response to conversation
-        self.conversation.addAssistantMessage(result.text)
-
-        // Add any tool calls and results to conversation
-        for step in result.steps {
-            if !step.toolCalls.isEmpty {
-                for _ in step.toolCalls {
-                    // Tool calls are already added by generateText
+            let didMerge: Bool
+            if result.finishReason == .contentFilter {
+                didMerge = conversation.mergeContentFilterResult(
+                    result.messages,
+                    originalMessages: modelMessages,
+                    afterMessageID: anchorID,
+                    validatingSnapshotIDs: snapshotIDs,
+                )
+            } else if let anchorID {
+                let generatedMessages = Array(result.messages.dropFirst(modelMessages.count))
+                let didMerge = conversation.appendGeneratedMessages(
+                    generatedMessages,
+                    afterMessageID: anchorID,
+                    validatingSnapshotIDs: snapshotIDs,
+                )
+                guard didMerge else {
+                    throw TachikomaError.invalidConfiguration(
+                        "Conversation changed during generation; refusing to merge response",
+                    )
                 }
+                return AgentResponse(
+                    text: result.text,
+                    usage: result.usage,
+                    finishReason: result.finishReason ?? .other,
+                    steps: result.steps,
+                    conversationLength: conversation.messages.count,
+                )
+            } else {
+                didMerge = conversation.messages.isEmpty
             }
-            if !step.toolResults.isEmpty {
-                for _ in step.toolResults {
-                    // Tool results are already added by generateText
-                }
+
+            guard didMerge else {
+                throw TachikomaError.invalidConfiguration(
+                    "Conversation changed during generation; refusing to merge response",
+                )
             }
+
+            return AgentResponse(
+                text: result.text,
+                usage: result.usage,
+                finishReason: result.finishReason ?? .other,
+                steps: result.steps,
+                conversationLength: conversation.messages.count,
+            )
         }
-
-        return AgentResponse(
-            text: result.text,
-            usage: result.usage,
-            finishReason: result.finishReason ?? .other,
-            steps: result.steps,
-            conversationLength: self.conversation.messages.count,
-        )
     }
 
     /// Stream a response from the agent
     public func stream(_ message: String) async throws -> AsyncThrowingStream<TextStreamDelta, Error> {
+        let streamingModel = if self.usesImplicitDefaultModel {
+            LanguageModel.defaultStreaming
+        } else if self.model.supportsStreaming {
+            self.model
+        } else {
+            self.model
+        }
+        guard streamingModel.supportsStreaming else {
+            throw TachikomaError.invalidConfiguration("\(self.model.modelId) does not support streaming")
+        }
+
+        let conversation = self.conversation
+        try await conversation.acquireContinuationLock()
+        let gateRelease = AsyncReleaseOnce {
+            await conversation.releaseContinuationLock()
+        }
+
         // Add user message to conversation
-        self.conversation.addUserMessage(message)
+        conversation.addUserMessage(message)
+        let conversationMessages = conversation.messages
+        let modelMessages = conversationMessages.map { $0.toModelMessage() }
+        let snapshotIDs = conversationMessages.map(\.id)
+        let buffersUntilDone = self.settings.streamBuffering == .untilTerminal
 
         // Stream response
-        let streamResult = try await streamText(
-            model: model,
-            messages: conversation.getModelMessages(),
-            tools: self.tools.isEmpty ? nil : self.tools,
-            settings: self.settings,
-            maxSteps: 5,
-        )
+        let streamResult: StreamTextResult
+        do {
+            streamResult = try await streamText(
+                model: streamingModel,
+                messages: modelMessages,
+                tools: self.tools.isEmpty ? nil : self.tools,
+                settings: self.settings,
+                maxSteps: 5,
+                configuration: self.configuration,
+            )
+        } catch {
+            gateRelease.release()
+            throw error
+        }
 
         // Track final message in conversation (this is approximate for streaming)
         return AsyncThrowingStream<TextStreamDelta, Error> { continuation in
-            Task {
+            let producer = Task {
+                defer {
+                    gateRelease.release()
+                }
                 do {
                     var assistantText = ""
+                    var bufferedDeltas: [TextStreamDelta] = []
+                    var didReceiveTerminal = false
 
                     for try await delta in streamResult.stream {
-                        continuation.yield(delta)
+                        try Task.checkCancellation()
+                        if buffersUntilDone {
+                            bufferedDeltas.append(delta)
+                        } else {
+                            continuation.yield(delta)
+                        }
 
                         // Collect assistant text
                         if case .textDelta = delta.type, let content = delta.content {
@@ -138,25 +216,60 @@ public final class Agent<Context>: @unchecked Sendable {
                         }
 
                         if case .done = delta.type {
+                            didReceiveTerminal = true
+                            guard delta.finishReason != .contentFilter else {
+                                let didRollback = conversation.replaceModelMessages(
+                                    modelMessages.droppingLastUserTurn(),
+                                    validatingSnapshotIDs: snapshotIDs,
+                                )
+                                guard didRollback else {
+                                    throw TachikomaError.invalidConfiguration(
+                                        "Conversation changed during streaming; refusing to merge response",
+                                    )
+                                }
+                                assistantText = ""
+                                bufferedDeltas.removeAll()
+                                if buffersUntilDone {
+                                    continuation.yield(delta)
+                                }
+                                continue
+                            }
+                            if buffersUntilDone {
+                                for bufferedDelta in bufferedDeltas {
+                                    continuation.yield(bufferedDelta)
+                                }
+                                bufferedDeltas.removeAll()
+                            }
                             // Add final assistant message to conversation
                             if !assistantText.isEmpty {
-                                self.conversation.addAssistantMessage(assistantText)
+                                conversation.addAssistantMessage(assistantText)
+                                assistantText = ""
                             }
                         }
                     }
+                    if buffersUntilDone, !didReceiveTerminal, !bufferedDeltas.isEmpty {
+                        throw TachikomaError.apiError("Stream ended before provider completion status was received")
+                    }
+                    if !buffersUntilDone, !assistantText.isEmpty {
+                        try Task.checkCancellation()
+                        conversation.addAssistantMessage(assistantText)
+                    }
 
                     continuation.finish()
                 } catch {
                     continuation.finish(throwing: error)
                 }
             }
+            continuation.onTermination = { @Sendable _ in
+                producer.cancel()
+            }
         }
     }
 
     /// Reset the agent's conversation history
     public func resetConversation() {
         // Reset the agent's conversation history
-        self.conversation = Conversation()
+        self.conversation = Conversation(configuration: self.configuration)
         self.conversation.addSystemMessage(self.instructions)
     }
 
@@ -169,7 +282,7 @@ public final class Agent<Context>: @unchecked Sendable {
     public func updateInstructions(_ newInstructions: String) {
         // Create new conversation with updated instructions
         let oldMessages = self.conversation.getModelMessages().filter { $0.role != .system }
-        self.conversation = Conversation()
+        self.conversation = Conversation(configuration: self.configuration)
         self.conversation.addSystemMessage(newInstructions)
 
         // Re-add non-system messages
@@ -179,6 +292,30 @@ public final class Agent<Context>: @unchecked Sendable {
     }
 }
 
+final class AsyncReleaseOnce: @unchecked Sendable {
+    private let lock = NSLock()
+    private var didRelease = false
+    private let operation: @Sendable () async -> Void
+
+    init(operation: @escaping @Sendable () async -> Void) {
+        self.operation = operation
+    }
+
+    func release() {
+        self.lock.lock()
+        guard !self.didRelease else {
+            self.lock.unlock()
+            return
+        }
+        self.didRelease = true
+        self.lock.unlock()
+
+        Task {
+            await self.operation()
+        }
+    }
+}
+
 /// Response from an agent execution
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
 public struct AgentResponse: Sendable {
@@ -537,3 +674,33 @@ public enum SessionStatus: String, Codable, Sendable, CaseIterable {
     case failed
     case cancelled
 }
+
+extension LanguageModel {
+    var requiresTerminalRefusalBuffering: Bool {
+        switch self {
+        case let .anthropic(model):
+            return LanguageModel.Anthropic.hasStreamingRefusalRisk(modelId: model.modelId)
+        case let .anthropicCompatible(modelId, _):
+            return LanguageModel.Anthropic.hasStreamingRefusalRisk(modelId: modelId)
+        case let .openRouter(modelId), let .together(modelId):
+            return LanguageModel.Anthropic.hasStreamingRefusalRisk(modelId: modelId)
+        case let .openaiCompatible(modelId, _):
+            return LanguageModel.Anthropic.hasStreamingRefusalRisk(modelId: modelId)
+        case let .custom(provider):
+            guard
+                let parsed = ProviderParser.parse(provider.modelId),
+                let registeredProvider = CustomProviderRegistry.shared.get(parsed.provider) else
+            {
+                return LanguageModel.Anthropic.hasStreamingRefusalRisk(modelId: provider.modelId)
+            }
+            switch registeredProvider.kind {
+            case .openai:
+                return false
+            case .anthropic:
+                return LanguageModel.Anthropic.hasStreamingRefusalRisk(modelId: parsed.model)
+            }
+        default:
+            return false
+        }
+    }
+}

Sources/TachikomaAgent/Conversation.swift

@@ -3,10 +3,66 @@ import Tachikoma
 
 // MARK: - Conversation Management
 
+private actor ContinuationGate {
+    private struct Waiter {
+        let id: UUID
+        let continuation: CheckedContinuation<Bool, Never>
+    }
+
+    private var isLocked = false
+    private var waiters: [Waiter] = []
+
+    func acquire() async throws {
+        try Task.checkCancellation()
+
+        if !self.isLocked {
+            self.isLocked = true
+            return
+        }
+
+        let id = UUID()
+        let acquired = await withTaskCancellationHandler {
+            await withCheckedContinuation { continuation in
+                self.waiters.append(Waiter(id: id, continuation: continuation))
+            }
+        } onCancel: {
+            Task { await self.cancelWaiter(id) }
+        }
+
+        guard acquired else {
+            throw CancellationError()
+        }
+
+        if Task.isCancelled {
+            self.release()
+            throw CancellationError()
+        }
+    }
+
+    private func cancelWaiter(_ id: UUID) {
+        guard let index = self.waiters.firstIndex(where: { $0.id == id }) else {
+            return
+        }
+
+        let waiter = self.waiters.remove(at: index)
+        waiter.continuation.resume(returning: false)
+    }
+
+    func release() {
+        if self.waiters.isEmpty {
+            self.isLocked = false
+        } else {
+            let waiter = self.waiters.removeFirst()
+            waiter.continuation.resume(returning: true)
+        }
+    }
+}
+
 /// A conversation with an AI model
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
 public final class Conversation: @unchecked Sendable {
     private let lock = NSLock()
+    private let continuationGate = ContinuationGate()
     private var _messages: [ConversationMessage] = []
 
     /// The configuration used by this conversation
@@ -59,7 +115,6 @@ public final class Conversation: @unchecked Sendable {
 
     /// Get messages as ModelMessage array for API compatibility
     public func getModelMessages() -> [ModelMessage] {
-        // Get messages as ModelMessage array for API compatibility
         self.messages.map { $0.toModelMessage() }
     }
 
@@ -72,31 +127,193 @@ public final class Conversation: @unchecked Sendable {
         self.lock.unlock()
     }
 
-    /// Continue the conversation with a model
-    public func continueConversation(using model: Model? = nil, tools _: [AgentTool]? = nil) async throws -> String {
-        // Convert conversation messages to model messages
-        let modelMessages = self.messages.map { conversationMessage in
-            ModelMessage(
-                id: conversationMessage.id,
-                role: ModelMessage.Role(rawValue: conversationMessage.role.rawValue) ?? .user,
-                content: [.text(conversationMessage.content)],
-                timestamp: conversationMessage.timestamp,
+    /// Replace the conversation with lossless ModelMessage history.
+    public func replaceModelMessages(_ modelMessages: [ModelMessage]) {
+        self.lock.lock()
+        self._messages = modelMessages.map { ConversationMessage.from($0) }
+        self.lock.unlock()
+    }
+
+    /// Replace the conversation only if the original snapshot is still current.
+    public func replaceModelMessages(
+        _ modelMessages: [ModelMessage],
+        validatingSnapshotIDs snapshotIDs: [String],
+    )
+        -> Bool
+    {
+        self.lock.lock()
+        defer { self.lock.unlock() }
+
+        guard self._messages.count >= snapshotIDs.count else {
+            return false
+        }
+
+        let currentPrefixIDs = self._messages.prefix(snapshotIDs.count).map(\.id)
+        guard currentPrefixIDs == snapshotIDs else {
+            return false
+        }
+
+        let laterMessages = self._messages.dropFirst(snapshotIDs.count)
+        self._messages = modelMessages.map { ConversationMessage.from($0) } + laterMessages
+        return true
+    }
+
+    /// Replace the original snapshot with generated history while preserving later appends.
+    public func mergeGeneratedMessages(_ modelMessages: [ModelMessage], replacingPrefixCount prefixCount: Int) {
+        self.lock.lock()
+        let laterMessages = self._messages.dropFirst(min(prefixCount, self._messages.count))
+        self._messages = modelMessages.map { ConversationMessage.from($0) } + laterMessages
+        self.lock.unlock()
+    }
+
+    /// Insert generated response messages after the snapshot anchor while preserving concurrent appends.
+    public func appendGeneratedMessages(_ modelMessages: [ModelMessage], afterMessageID messageID: String) {
+        guard !modelMessages.isEmpty else { return }
+
+        self.lock.lock()
+        let conversationMessages = modelMessages.map { ConversationMessage.from($0) }
+        if let index = self._messages.firstIndex(where: { $0.id == messageID }) {
+            self._messages.insert(contentsOf: conversationMessages, at: self._messages.index(after: index))
+        } else {
+            self._messages.append(contentsOf: conversationMessages)
+        }
+        self.lock.unlock()
+    }
+
+    /// Insert generated response messages only if the snapshot prefix is still current.
+    public func appendGeneratedMessages(
+        _ modelMessages: [ModelMessage],
+        afterMessageID messageID: String,
+        validatingSnapshotIDs snapshotIDs: [String],
+    )
+        -> Bool
+    {
+        guard !modelMessages.isEmpty else { return true }
+
+        self.lock.lock()
+        defer { self.lock.unlock() }
+
+        guard self._messages.count >= snapshotIDs.count else {
+            return false
+        }
+
+        let currentPrefixIDs = self._messages.prefix(snapshotIDs.count).map(\.id)
+        guard currentPrefixIDs == snapshotIDs else {
+            return false
+        }
+
+        let conversationMessages = modelMessages.map { ConversationMessage.from($0) }
+        if let index = self._messages.firstIndex(where: { $0.id == messageID }) {
+            self._messages.insert(contentsOf: conversationMessages, at: self._messages.index(after: index))
+        } else {
+            self._messages.append(contentsOf: conversationMessages)
+        }
+        return true
+    }
+
+    /// Merge a refused generation without losing completed tool steps.
+    public func mergeContentFilterResult(
+        _ resultMessages: [ModelMessage],
+        originalMessages: [ModelMessage],
+        afterMessageID _: String?,
+        validatingSnapshotIDs snapshotIDs: [String],
+    )
+        -> Bool
+    {
+        let generatedMessages = Array(resultMessages.dropFirst(originalMessages.count))
+        if !generatedMessages.isEmpty {
+            return self.replaceModelMessages(
+                originalMessages + generatedMessages,
+                validatingSnapshotIDs: snapshotIDs,
             )
         }
 
-        // Generate response using the core API
-        let response = try await generateText(
-            model: model ?? .default,
-            messages: modelMessages,
-            tools: [],
-            settings: .default,
-            configuration: configuration,
+        return self.replaceModelMessages(
+            originalMessages.droppingLastUserTurn(),
+            validatingSnapshotIDs: snapshotIDs,
         )
+    }
 
-        // Add the response to the conversation
-        self.addAssistantMessage(response.text)
+    public func removeMessage(id: String) {
+        self.lock.lock()
+        self._messages.removeAll { $0.id == id }
+        self.lock.unlock()
+    }
 
-        return response.text
+    public func withContinuationLock<T>(_ operation: () async throws -> T) async throws -> T {
+        try await self.acquireContinuationLock()
+        do {
+            let result = try await operation()
+            await self.releaseContinuationLock()
+            return result
+        } catch {
+            await self.releaseContinuationLock()
+            throw error
+        }
+    }
+
+    public func acquireContinuationLock() async throws {
+        try await self.continuationGate.acquire()
+    }
+
+    public func releaseContinuationLock() async {
+        await self.continuationGate.release()
+    }
+
+    /// Continue the conversation with a model
+    public func continueConversation(
+        using model: Model? = nil,
+        tools: [AgentTool]? = nil,
+        maxSteps: Int = 5,
+    ) async throws
+        -> String
+    {
+        try await self.withContinuationLock {
+            let conversationMessages = self.messages
+            let modelMessages = conversationMessages.map { $0.toModelMessage() }
+            let snapshotIDs = conversationMessages.map(\.id)
+            let anchorID = conversationMessages.last?.id
+
+            // Generate response using the core API
+            let response = try await generateText(
+                model: model ?? .default,
+                messages: modelMessages,
+                tools: tools,
+                settings: .default,
+                maxSteps: maxSteps,
+                configuration: configuration,
+            )
+
+            let didMerge: Bool
+            if response.finishReason == .contentFilter {
+                didMerge = self.mergeContentFilterResult(
+                    response.messages,
+                    originalMessages: modelMessages,
+                    afterMessageID: anchorID,
+                    validatingSnapshotIDs: snapshotIDs,
+                )
+            } else if let anchorID {
+                let generatedMessages = Array(response.messages.dropFirst(modelMessages.count))
+                didMerge = self.appendGeneratedMessages(
+                    generatedMessages,
+                    afterMessageID: anchorID,
+                    validatingSnapshotIDs: snapshotIDs,
+                )
+            } else if self.messages.isEmpty {
+                self.replaceModelMessages(response.messages)
+                didMerge = true
+            } else {
+                didMerge = false
+            }
+
+            guard didMerge else {
+                throw TachikomaError.invalidConfiguration(
+                    "Conversation changed during generation; refusing to merge response",
+                )
+            }
+
+            return response.text
+        }
     }
 
     /// Continue the conversation with a model, streaming the response
@@ -106,43 +323,90 @@ public final class Conversation: @unchecked Sendable {
     ) async throws
         -> AsyncThrowingStream<String, Error>
     {
-        // Convert conversation messages to model messages
-        let modelMessages = self.messages.map { conversationMessage in
-            ModelMessage(
-                id: conversationMessage.id,
-                role: ModelMessage.Role(rawValue: conversationMessage.role.rawValue) ?? .user,
-                content: [.text(conversationMessage.content)],
-                timestamp: conversationMessage.timestamp,
-            )
+        try await self.acquireContinuationLock()
+        let gateRelease = AsyncReleaseOnce {
+            await self.releaseContinuationLock()
         }
+        let conversationMessages = self.messages
+        let modelMessages = conversationMessages.map { $0.toModelMessage() }
+        let snapshotIDs = conversationMessages.map(\.id)
+        let resolvedModel = model ?? .defaultStreaming
+        let streamSettings = GenerationSettings.default
+        let buffersUntilDone = streamSettings.streamBuffering == .untilTerminal ||
+            resolvedModel.requiresTerminalRefusalBuffering
 
         // Generate response using the core API
-        let responseStream = try await streamText(
-            model: model ?? .default,
-            messages: modelMessages,
-            tools: tools ?? [], // Use provided tools or empty array
-            settings: .default,
-            configuration: configuration,
-        )
+        let responseStream: StreamTextResult
+        do {
+            responseStream = try await streamText(
+                model: resolvedModel,
+                messages: modelMessages,
+                tools: tools ?? [], // Use provided tools or empty array
+                settings: streamSettings,
+                configuration: self.configuration,
+            )
+        } catch {
+            gateRelease.release()
+            throw error
+        }
 
         // Create a new stream to process the response and update the conversation
         return AsyncThrowingStream<String, Error> { continuation in
-            Task {
+            let producer = Task {
+                defer {
+                    gateRelease.release()
+                }
                 var fullResponse = ""
+                var isContentFiltered = false
+                var bufferedText: [String] = []
+                var didApproveBufferedResponse = !buffersUntilDone
+                var didReceiveTerminal = false
                 do {
                     for try await delta in responseStream.stream {
+                        try Task.checkCancellation()
                         switch delta.type {
                         case .textDelta:
                             if let text = delta.content {
-                                continuation.yield(text)
+                                if buffersUntilDone {
+                                    bufferedText.append(text)
+                                } else {
+                                    continuation.yield(text)
+                                }
                                 fullResponse += text
                             }
+                        case .done where delta.finishReason == .contentFilter:
+                            didReceiveTerminal = true
+                            isContentFiltered = true
+                            let didRollback = self.replaceModelMessages(
+                                modelMessages.droppingLastUserTurn(),
+                                validatingSnapshotIDs: snapshotIDs,
+                            )
+                            guard didRollback else {
+                                throw TachikomaError.invalidConfiguration(
+                                    "Conversation changed during streaming; refusing to merge response",
+                                )
+                            }
+                            fullResponse = ""
+                            bufferedText.removeAll()
+                        case .done:
+                            didReceiveTerminal = true
+                            if buffersUntilDone {
+                                for text in bufferedText {
+                                    continuation.yield(text)
+                                }
+                                didApproveBufferedResponse = true
+                                bufferedText.removeAll()
+                            }
                         default:
                             break
                         }
                     }
+                    if buffersUntilDone, !didReceiveTerminal, !bufferedText.isEmpty {
+                        throw TachikomaError.apiError("Stream ended before provider completion status was received")
+                    }
                     // Add the full response to the conversation
-                    if !fullResponse.isEmpty {
+                    if !isContentFiltered, !fullResponse.isEmpty, didApproveBufferedResponse {
+                        try Task.checkCancellation()
                         self.addAssistantMessage(fullResponse)
                     }
                     continuation.finish()
@@ -150,10 +414,20 @@ public final class Conversation: @unchecked Sendable {
                     continuation.finish(throwing: error)
                 }
             }
+            continuation.onTermination = { @Sendable _ in
+                producer.cancel()
+            }
         }
     }
 }
 
+extension [ModelMessage] {
+    func droppingLastUserTurn() -> [ModelMessage] {
+        guard self.last?.role == .user else { return self }
+        return Array(self.dropLast())
+    }
+}
+
 /// A message in a conversation
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
 public struct ConversationMessage: Sendable, Codable, Equatable {
@@ -161,6 +435,9 @@ public struct ConversationMessage: Sendable, Codable, Equatable {
     public let role: Role
     public let content: String
     public let timestamp: Date
+    public let contentParts: [ModelMessage.ContentPart]?
+    public let channel: ResponseChannel?
+    public let metadata: MessageMetadata?
 
     public enum Role: String, Sendable, Codable, CaseIterable {
         case system
@@ -169,11 +446,22 @@ public struct ConversationMessage: Sendable, Codable, Equatable {
         case tool
     }
 
-    public init(id: String = UUID().uuidString, role: Role, content: String, timestamp: Date = Date()) {
+    public init(
+        id: String = UUID().uuidString,
+        role: Role,
+        content: String,
+        timestamp: Date = Date(),
+        contentParts: [ModelMessage.ContentPart]? = nil,
+        channel: ResponseChannel? = nil,
+        metadata: MessageMetadata? = nil,
+    ) {
         self.id = id
         self.role = role
         self.content = content
         self.timestamp = timestamp
+        self.contentParts = contentParts
+        self.channel = channel
+        self.metadata = metadata
     }
 
     /// Convert to ModelMessage for API compatibility
@@ -189,8 +477,10 @@ public struct ConversationMessage: Sendable, Codable, Equatable {
         return ModelMessage(
             id: self.id,
             role: modelRole,
-            content: [.text(self.content)],
+            content: self.contentParts ?? [.text(self.content)],
             timestamp: self.timestamp,
+            channel: self.channel,
+            metadata: self.metadata,
         )
     }
 
@@ -219,6 +509,9 @@ public struct ConversationMessage: Sendable, Codable, Equatable {
             role: role,
             content: textContent,
             timestamp: modelMessage.timestamp,
+            contentParts: modelMessage.content,
+            channel: modelMessage.channel,
+            metadata: modelMessage.metadata,
         )
     }
 }

Sources/TachikomaAudio/Realtime/Configuration/RealtimeConfiguration.swift

@@ -128,7 +128,7 @@ public struct InputAudioTranscription: Sendable, Codable {
 /// Session configuration with all options
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
 public struct SessionConfiguration: Sendable, Codable {
-    /// Model to use (e.g., "gpt-4o-realtime-preview")
+    /// Model to use (e.g., "gpt-realtime")
     public var model: String
 
     /// Voice for audio responses
@@ -213,7 +213,7 @@ public struct SessionConfiguration: Sendable, Codable {
     }
 
     public init(
-        model: String = "gpt-4o-realtime-preview",
+        model: String = "gpt-realtime",
         voice: RealtimeVoice = .alloy,
         instructions: String? = nil,
         inputAudioFormat: RealtimeAudioFormat = .pcm16,
@@ -242,7 +242,7 @@ public struct SessionConfiguration: Sendable, Codable {
 
     /// Create a default configuration for voice conversations
     public static func voiceConversation(
-        model: String = "gpt-4o-realtime-preview",
+        model: String = "gpt-realtime",
         voice: RealtimeVoice = .alloy,
     )
         -> SessionConfiguration
@@ -258,7 +258,7 @@ public struct SessionConfiguration: Sendable, Codable {
 
     /// Create a configuration for text-only interactions
     public static func textOnly(
-        model: String = "gpt-4o-realtime-preview",
+        model: String = "gpt-realtime",
     )
         -> SessionConfiguration
     {
@@ -273,7 +273,7 @@ public struct SessionConfiguration: Sendable, Codable {
 
     /// Create a configuration with tools
     public static func withTools(
-        model: String = "gpt-4o-realtime-preview",
+        model: String = "gpt-realtime",
         voice: RealtimeVoice = .alloy,
         tools: [RealtimeTool],
     )

Sources/TachikomaAudio/Realtime/RealtimeConversation.swift

@@ -150,7 +150,7 @@ public final class RealtimeConversation: ObservableObject {
 
         // Create session with configuration
         let sessionConfig = SessionConfiguration(
-            model: "gpt-4o-realtime-preview",
+            model: "gpt-realtime",
             voice: .alloy,
             instructions: nil,
             tools: nil,
@@ -167,7 +167,7 @@ public final class RealtimeConversation: ObservableObject {
 
     /// Start the conversation
     public func start(
-        model: LanguageModel.OpenAI = .gpt4oRealtime,
+        model: LanguageModel.OpenAI = .custom("gpt-realtime"),
         voice: RealtimeVoice = .alloy,
         instructions: String? = nil,
         tools: [RealtimeTool]? = nil,
@@ -495,7 +495,7 @@ public final class RealtimeConversation: ObservableObject {
 
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
 public func startRealtimeConversation(
-    model: LanguageModel.OpenAI = .gpt4oRealtime,
+    model: LanguageModel.OpenAI = .custom("gpt-realtime"),
     voice: RealtimeVoice = .alloy,
     instructions: String? = nil,
     tools: [AgentTool]? = nil,
@@ -548,7 +548,7 @@ public final class RealtimeConversation {
     }
 
     public func start(
-        model _: LanguageModel.OpenAI = .gpt4oRealtime,
+        model _: LanguageModel.OpenAI = .custom("gpt-realtime"),
         voice _: RealtimeVoice = .alloy,
         instructions _: String? = nil,
         tools _: [RealtimeTool]? = nil,
@@ -560,7 +560,7 @@ public final class RealtimeConversation {
 // swiftformat:disable wrapMultilineStatementBraces wrapReturnType indent
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
 public func startRealtimeConversation(
-    model _: LanguageModel.OpenAI = .gpt4oRealtime,
+    model _: LanguageModel.OpenAI = .custom("gpt-realtime"),
     voice _: RealtimeVoice = .alloy,
     instructions _: String? = nil,
     tools _: [AgentTool]? = nil,

Sources/TachikomaAudio/Realtime/RealtimeEvents.swift

@@ -188,7 +188,7 @@ public struct RealtimeSessionConfig: Codable, Sendable {
     }
 
     public init(
-        model: String = "gpt-4o-realtime-preview",
+        model: String = "gpt-realtime",
         voice: RealtimeVoice = .alloy,
         instructions: String? = nil,
         tools: [RealtimeTool]? = nil,

Sources/TachikomaAudio/Types/AudioModels.swift

@@ -16,20 +16,16 @@ public enum TranscriptionModel: Sendable, CustomStringConvertible {
 
     public enum OpenAI: String, CaseIterable, Sendable {
         case whisper1 = "whisper-1"
-        case gpt4oTranscribe = "gpt-4o-transcribe"
-        case gpt4oMiniTranscribe = "gpt-4o-mini-transcribe"
 
         public var supportsTimestamps: Bool {
             switch self {
             case .whisper1: true
-            case .gpt4oTranscribe, .gpt4oMiniTranscribe: false
             }
         }
 
         public var supportsLanguageDetection: Bool {
             switch self {
             case .whisper1: true
-            case .gpt4oTranscribe, .gpt4oMiniTranscribe: false
             }
         }
     }
@@ -155,12 +151,10 @@ public enum SpeechModel: Sendable, CustomStringConvertible {
     public enum OpenAI: String, CaseIterable, Sendable {
         case tts1 = "tts-1"
         case tts1HD = "tts-1-hd"
-        case gpt4oMiniTTS = "gpt-4o-mini-tts"
 
         public var supportsVoiceInstructions: Bool {
             switch self {
             case .tts1, .tts1HD: false
-            case .gpt4oMiniTTS: true
             }
         }
 
@@ -226,5 +220,5 @@ public enum SpeechModel: Sendable, CustomStringConvertible {
     public static let `default`: SpeechModel = .openai(.tts1)
     public static let highQuality: SpeechModel = .openai(.tts1HD)
     public static let fast: SpeechModel = .openai(.tts1)
-    public static let expressive: SpeechModel = .openai(.gpt4oMiniTTS)
+    public static let expressive: SpeechModel = .openai(.tts1HD)
 }

Sources/TachikomaConfigCLI/main.swift

@@ -116,7 +116,7 @@ struct TKConfigCLI {
         var mutable = raw
         let timeout = self.parseTimeout(&mutable)
         print("Providers:")
-        for pid in [TKProviderId.openai, .anthropic, .grok, .gemini] {
+        for pid in [TKProviderId.openai, .anthropic, .grok, .gemini, .openrouter] {
             let status = await TKConfigCLI.status(for: pid, timeout: timeout)
             print("  \(pid.displayName): \(status)")
         }
@@ -150,6 +150,8 @@ struct TKConfigCLI {
             }
         case .gemini:
             if let v = env["GEMINI_API_KEY"], !v.isEmpty { return ("GEMINI_API_KEY", v) }
+        case .openrouter:
+            if let v = env["OPENROUTER_API_KEY"], !v.isEmpty { return ("OPENROUTER_API_KEY", v) }
         }
         return nil
     }
@@ -169,6 +171,8 @@ struct TKConfigCLI {
             }
         case .gemini:
             if let v = manager.credentialValue(for: "GEMINI_API_KEY") { return ("GEMINI_API_KEY", v) }
+        case .openrouter:
+            if let v = manager.credentialValue(for: "OPENROUTER_API_KEY") { return ("OPENROUTER_API_KEY", v) }
         }
         return nil
     }

Sources/TachikomaMCP/Client/MCPClientManager.swift

@@ -499,12 +499,7 @@ public final class TachikomaMCPClientManager {
     }
 
     private func profileDirectoryPath() -> String {
-        #if os(Windows)
-        let home = ProcessInfo.processInfo.environment["USERPROFILE"] ?? ""
-        #else
-        let home = ProcessInfo.processInfo.environment["HOME"] ?? ""
-        #endif
-        return "\(home)/\(self.profileDirectoryName)"
+        TachikomaConfiguration.profileDirectoryPath
     }
 
     private func profileConfigPath() -> String {

Sources/TachikomaMCP/README.md

@@ -12,7 +12,7 @@ import Tachikoma
 
 // Use Tachikoma normally without MCP
 let result = try await generateText(
-    model: .openai(.gpt4o),
+    model: .openai(.gpt55),
     messages: messages,
     tools: staticTools
 )
@@ -26,7 +26,7 @@ import TachikomaMCP
 // Now you can use MCP tools alongside static tools
 let mcpTools = try await MCPToolDiscovery.withFilesystem()
 let result = try await generateText(
-    model: .openai(.gpt4o),
+    model: .openai(.gpt55),
     messages: messages,
     tools: staticTools + mcpTools
 )

Tests/TachikomaMCPTests/MCPClientTests.swift

@@ -191,7 +191,7 @@ struct MCPClientTests {
         let imageData = Data("test".utf8)
         let imageResponse = ToolResponse.image(data: imageData, mimeType: "image/png")
         #expect(imageResponse.content.count == 1)
-        if case .image(data: let data, mimeType: let mimeType, annotations: _, _meta: _) = imageResponse.content.first {
+        if case let .image(data: data, mimeType: mimeType, annotations: _, _meta: _) = imageResponse.content.first {
             #expect(data == imageData.base64EncodedString())
             #expect(mimeType == "image/png")
         } else {

Tests/TachikomaMCPTests/MCPToolAdapterTests.swift

@@ -130,7 +130,7 @@ struct MCPToolAdapterTests {
 
         #expect(response.isError == false)
         #expect(response.content.count == 1)
-        if case .image(data: let data, mimeType: let mimeType, annotations: _, _meta: _) = response.content[0] {
+        if case let .image(data: data, mimeType: mimeType, annotations: _, _meta: _) = response.content[0] {
             #expect(data == imageData.base64EncodedString())
             #expect(mimeType == "image/jpeg")
         } else {

Tests/TachikomaTests/Audio/AudioFunctionsTests.swift

@@ -362,7 +362,7 @@ enum AudioFunctionsTests {
         }
 
         @Test
-        func `multiple provider integration`() async throws {
+        func `multiple provider integration`() async {
             await TestHelpers.withStandardTestConfiguration { config in
                 let audioData = TestHelpers.sampleAudioData(configuration: config)
 

Tests/TachikomaTests/Audio/AudioProviderFactoryTests.swift

@@ -81,43 +81,47 @@ struct AudioProviderFactoryTests {
     }
 
     @Test
-    func `AudioConfiguration reads configuration keys before environment`() {
-        let previousOpenAIKey = getenv("OPENAI_API_KEY").flatMap { String(cString: $0) }
-        unsetenv("OPENAI_API_KEY")
-        defer {
-            if let previousOpenAIKey {
-                setenv("OPENAI_API_KEY", previousOpenAIKey, 1)
+    func `AudioConfiguration reads configuration keys before environment`() async {
+        await TestEnvironmentMutex.shared.withLock {
+            let previousOpenAIKey = getenv("OPENAI_API_KEY").flatMap { String(cString: $0) }
+            unsetenv("OPENAI_API_KEY")
+            defer {
+                if let previousOpenAIKey {
+                    setenv("OPENAI_API_KEY", previousOpenAIKey, 1)
+                }
             }
+
+            let configuration = TachikomaConfiguration()
+            configuration.setAPIKey("configured-key", for: "openai")
+
+            let resolvedKey = AudioConfiguration.getAPIKey(for: "openai", configuration: configuration)
+            #expect(resolvedKey == "configured-key")
         }
-
-        let configuration = TachikomaConfiguration()
-        configuration.setAPIKey("configured-key", for: "openai")
-
-        let resolvedKey = AudioConfiguration.getAPIKey(for: "openai", configuration: configuration)
-        #expect(resolvedKey == "configured-key")
     }
 
     @Test
-    func `AudioConfiguration falls back to environment variable`() {
-        let previousOpenAIKey = getenv("OPENAI_API_KEY").flatMap { String(cString: $0) }
-        setenv("OPENAI_API_KEY", "env-key-123", 1)
-        defer {
-            if let previousOpenAIKey {
-                setenv("OPENAI_API_KEY", previousOpenAIKey, 1)
-            } else {
-                unsetenv("OPENAI_API_KEY")
+    func `AudioConfiguration falls back to environment variable`() async {
+        await TestEnvironmentMutex.shared.withLock {
+            let previousOpenAIKey = getenv("OPENAI_API_KEY").flatMap { String(cString: $0) }
+            setenv("OPENAI_API_KEY", "env-key-123", 1)
+            defer {
+                if let previousOpenAIKey {
+                    setenv("OPENAI_API_KEY", previousOpenAIKey, 1)
+                } else {
+                    unsetenv("OPENAI_API_KEY")
+                }
             }
-        }
 
-        let expectedKey = getenv("OPENAI_API_KEY").map { String(cString: $0) }
-        let resolvedKey = AudioConfiguration.getAPIKey(
-            for: "openai",
-            configuration: TachikomaConfiguration(loadFromEnvironment: false),
-        )
-        if let expectedKey {
-            #expect(resolvedKey == expectedKey)
-        } else {
-            Issue.record("Expected environment override for OPENAI_API_KEY")
+            let expectedKey = getenv("OPENAI_API_KEY").map { String(cString: $0) }
+            let resolvedKey = AudioConfiguration.getAPIKey(
+                for: "openai",
+                configuration: TachikomaConfiguration(loadFromEnvironment: false),
+            )
+            if let expectedKey {
+                #expect(resolvedKey == expectedKey)
+            } else {
+                Issue.record("Expected environment override for OPENAI_API_KEY")
+            }
         }
     }
 }

Tests/TachikomaTests/Auth/AuthManagerTests.swift

@@ -44,6 +44,7 @@ struct AuthManagerTests {
         unsetenv("ANTHROPIC_ACCESS_TOKEN")
         unsetenv("GEMINI_API_KEY")
         unsetenv("GOOGLE_API_KEY")
+        unsetenv("OPENROUTER_API_KEY")
     }
 
     @Test
@@ -79,6 +80,29 @@ struct AuthManagerTests {
         }
     }
 
+    @Test
+    func `openrouter resolves env and credential keys as bearer auth`() async throws {
+        try await self.withIsolatedAuthState {
+            self.resetAuthEnv()
+            try TKAuthManager.shared.setCredential(key: "OPENROUTER_API_KEY", value: "credential-openrouter-key")
+
+            guard case let .bearer(credentialToken, _)? = TKAuthManager.shared.resolveAuth(for: .openrouter) else {
+                Issue.record("Expected OpenRouter bearer auth from credentials")
+                return
+            }
+            #expect(credentialToken == "credential-openrouter-key")
+
+            setenv("OPENROUTER_API_KEY", "env-openrouter-key", 1)
+            defer { unsetenv("OPENROUTER_API_KEY") }
+
+            guard case let .bearer(envToken, _)? = TKAuthManager.shared.resolveAuth(for: .openrouter) else {
+                Issue.record("Expected OpenRouter bearer auth from environment")
+                return
+            }
+            #expect(envToken == "env-openrouter-key")
+        }
+    }
+
     @Test
     @MainActor
     func `validate success mock`() async throws {

Tests/TachikomaTests/Core/AsyncSequenceTests.swift

@@ -17,7 +17,7 @@ struct AsyncSequenceTests {
 
         let result = StreamTextResult(
             stream: testStream,
-            model: .openai(.gpt4o),
+            model: .openai(.gpt55),
             settings: .default,
         )
 
@@ -86,7 +86,7 @@ struct AsyncSequenceTests {
 
         let result = StreamObjectResult(
             objectStream: testStream,
-            model: .openai(.gpt4o),
+            model: .openai(.gpt55),
             settings: .default,
             schema: TestData.self,
         )
@@ -117,7 +117,7 @@ struct AsyncSequenceTests {
 
         let result = StreamTextResult(
             stream: testStream,
-            model: .openai(.gpt4o),
+            model: .openai(.gpt55),
             settings: .default,
         )
 
@@ -149,7 +149,7 @@ struct AsyncSequenceTests {
 
         let result = StreamTextResult(
             stream: testStream,
-            model: .openai(.gpt4o),
+            model: .openai(.gpt55),
             settings: .default,
         )
 
@@ -183,7 +183,7 @@ struct AsyncSequenceTests {
 
         let result = StreamTextResult(
             stream: testStream,
-            model: .openai(.gpt4o),
+            model: .openai(.gpt55),
             settings: .default,
         )
 
@@ -218,7 +218,7 @@ struct AsyncSequenceTests {
 
         let result = StreamTextResult(
             stream: testStream,
-            model: .openai(.gpt4o),
+            model: .openai(.gpt55),
             settings: .default,
         )
 
@@ -262,7 +262,7 @@ struct AsyncSequenceTests {
 
         let result = StreamObjectResult(
             objectStream: testStream,
-            model: .openai(.gpt4o),
+            model: .openai(.gpt55),
             settings: .default,
             schema: TestItem.self,
         )

Tests/TachikomaTests/Core/ConfigurationArchitectureTests.swift

@@ -55,7 +55,7 @@ struct ConfigurationArchitectureTests {
         // Example 1: Zero configuration
         _ = {
             Task {
-                _ = try await generate("What is 2+2?", using: .openai(.gpt4o))
+                _ = try await generate("What is 2+2?", using: .openai(.gpt55))
             }
         }
 
@@ -66,7 +66,7 @@ struct ConfigurationArchitectureTests {
             )
 
             Task {
-                _ = try await generate("Hello", using: .openai(.gpt4o))
+                _ = try await generate("Hello", using: .openai(.gpt55))
             }
         }
 
@@ -79,7 +79,7 @@ struct ConfigurationArchitectureTests {
 
                 _ = try await generate(
                     "Test prompt",
-                    using: .openai(.gpt4o),
+                    using: .openai(.gpt55),
                     configuration: testConfig,
                 )
             }

Tests/TachikomaTests/Core/ConfigurationEnvironmentTests.swift

@@ -33,4 +33,17 @@ struct ConfigurationEnvironmentTests {
         let configuration = TachikomaConfiguration(loadFromEnvironment: true)
         #expect(configuration.getBaseURL(for: .openai) == "https://env.example.com")
     }
+
+    @Test
+    func `TachikomaConfiguration picks up ANTHROPIC_BASE_URL from environment`() {
+        let key = "ANTHROPIC_BASE_URL"
+        setenv(key, "https://env.example.com", 1)
+        defer { unsetenv(key) }
+
+        let rawValue = Provider.environmentValue(for: key)
+        #expect(rawValue == "https://env.example.com")
+
+        let configuration = TachikomaConfiguration(loadFromEnvironment: true)
+        #expect(configuration.getBaseURL(for: .anthropic) == "https://env.example.com")
+    }
 }

Tests/TachikomaTests/Core/CredentialLoadingTests.swift

@@ -0,0 +1,164 @@
+#if canImport(Darwin)
+import Darwin
+#else
+import Glibc
+#endif
+import Foundation
+import Testing
+@testable import Tachikoma
+
+@Suite(.serialized)
+struct CredentialLoadingTests {
+    @Test
+    func `OAuth tokens are not loaded as OpenAI API keys`() async throws {
+        try await self.withIsolatedCredentials(
+            """
+            OPENAI_ACCESS_TOKEN=access-token
+            OPENAI_REFRESH_TOKEN=refresh-token
+            OPENAI_ACCESS_EXPIRES=4102444800
+            """,
+        ) {
+            let config = TachikomaConfiguration(loadFromEnvironment: true)
+            #expect(config.getAPIKey(for: .openai) == nil)
+        }
+    }
+
+    @Test
+    func `OpenAI API key credential is preferred over OAuth token noise`() async throws {
+        try await self.withIsolatedCredentials(
+            """
+            OPENAI_ACCESS_TOKEN=access-token
+            OPENAI_API_KEY=api-key
+            OPENAI_REFRESH_TOKEN=refresh-token
+            """,
+        ) {
+            let config = TachikomaConfiguration(loadFromEnvironment: true)
+            #expect(config.getAPIKey(for: .openai) == "api-key")
+        }
+    }
+
+    @Test
+    func `Absolute profile path credentials load without HOME`() async throws {
+        #if !os(Windows)
+        try await TestEnvironmentMutex.shared.withLock {
+            let originalProfileDirectory = TachikomaConfiguration.profileDirectoryName
+            let profilePath = FileManager.default.temporaryDirectory
+                .appendingPathComponent("tachikoma-absolute-credentials-\(UUID().uuidString)")
+                .path
+            let credentialPath = "\(profilePath)/credentials"
+            let savedHome = getenv("HOME").map { String(cString: $0) }
+            let savedEnvironment = self.unsetOpenAIEnvironment() + [("HOME", savedHome)]
+
+            TachikomaConfiguration.profileDirectoryName = profilePath
+            try FileManager.default.createDirectory(atPath: profilePath, withIntermediateDirectories: true)
+            try "OPENAI_API_KEY=absolute-api-key\n".write(toFile: credentialPath, atomically: true, encoding: .utf8)
+            unsetenv("HOME")
+
+            defer {
+                self.restoreEnvironment(savedEnvironment)
+                TachikomaConfiguration.profileDirectoryName = originalProfileDirectory
+                try? FileManager.default.removeItem(atPath: profilePath)
+            }
+
+            let config = TachikomaConfiguration(loadFromEnvironment: true)
+            #expect(config.getAPIKey(for: .openai) == "absolute-api-key")
+        }
+        #endif
+    }
+
+    @Test
+    func `MiniMax China credentials save and reload with canonical env name`() async throws {
+        #if !os(Windows)
+        try await TestEnvironmentMutex.shared.withLock {
+            let originalProfileDirectory = TachikomaConfiguration.profileDirectoryName
+            let profilePath = FileManager.default.temporaryDirectory
+                .appendingPathComponent("tachikoma-minimax-cn-credentials-\(UUID().uuidString)")
+                .path
+            let credentialPath = "\(profilePath)/credentials"
+            let savedEnvironment = self.savedEnvironment(for: ["MINIMAX_CN_API_KEY", "MINIMAX_API_KEY"])
+
+            TachikomaConfiguration.profileDirectoryName = profilePath
+            for (key, _) in savedEnvironment {
+                unsetenv(key)
+            }
+
+            defer {
+                self.restoreEnvironment(savedEnvironment)
+                TachikomaConfiguration.profileDirectoryName = originalProfileDirectory
+                try? FileManager.default.removeItem(atPath: profilePath)
+            }
+
+            let config = TachikomaConfiguration(loadFromEnvironment: false)
+            config.setAPIKey("cn-api-key", for: .minimaxCN)
+            try config.saveCredentials()
+
+            let savedCredentials = try String(contentsOfFile: credentialPath, encoding: .utf8)
+            #expect(savedCredentials.contains("MINIMAX_CN_API_KEY=cn-api-key"))
+            #expect(!savedCredentials.contains("MINIMAX-CN_API_KEY"))
+
+            let reloaded = TachikomaConfiguration(loadFromEnvironment: true)
+            #expect(reloaded.getAPIKey(for: .minimaxCN) == "cn-api-key")
+        }
+        #endif
+    }
+
+    @Test
+    func `MiniMax China availability accepts configured shared MiniMax key`() {
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setAPIKey("shared-minimax-key", for: .minimax)
+
+        #expect(config.getAPIKey(for: .minimaxCN) == "shared-minimax-key")
+        #expect(config.hasAPIKey(for: .minimaxCN))
+    }
+
+    private func withIsolatedCredentials<T: Sendable>(
+        _ credentials: String,
+        _ body: @Sendable () throws -> T,
+    ) async throws
+        -> T
+    {
+        try await TestEnvironmentMutex.shared.withLock {
+            let originalProfileDirectory = TachikomaConfiguration.profileDirectoryName
+            let profileDirectory = ".tachikoma-credential-tests-\(UUID().uuidString)"
+            let homeDirectory = try #require(ProcessInfo.processInfo.environment["HOME"])
+            let profilePath = "\(homeDirectory)/\(profileDirectory)"
+            let credentialPath = "\(profilePath)/credentials"
+            let savedEnvironment = self.unsetOpenAIEnvironment()
+
+            TachikomaConfiguration.profileDirectoryName = profileDirectory
+            try FileManager.default.createDirectory(atPath: profilePath, withIntermediateDirectories: true)
+            try credentials.write(toFile: credentialPath, atomically: true, encoding: .utf8)
+
+            defer {
+                self.restoreEnvironment(savedEnvironment)
+                TachikomaConfiguration.profileDirectoryName = originalProfileDirectory
+                try? FileManager.default.removeItem(atPath: profilePath)
+            }
+
+            return try body()
+        }
+    }
+
+    private func unsetOpenAIEnvironment() -> [(String, String?)] {
+        let keys = ["OPENAI_API_KEY", "OPENAI_ACCESS_TOKEN", "OPENAI_REFRESH_TOKEN", "OPENAI_ACCESS_EXPIRES"]
+        let saved = self.savedEnvironment(for: keys)
+        keys.forEach { unsetenv($0) }
+        return saved
+    }
+
+    private func savedEnvironment(for keys: [String]) -> [(String, String?)] {
+        keys.map { key in
+            (key, getenv(key).map { String(cString: $0) })
+        }
+    }
+
+    private func restoreEnvironment(_ saved: [(String, String?)]) {
+        for (key, value) in saved {
+            if let value {
+                setenv(key, value, 1)
+            } else {
+                unsetenv(key)
+            }
+        }
+    }
+}

Tests/TachikomaTests/Core/CustomProviderRegistryTests.swift

@@ -38,7 +38,8 @@ struct CustomProviderRegistryTests {
                 /* secret token */
                 "headers": {
                   "Authorization": "Bearer ${WEATHER_TOKEN}"
-                }
+                },
+                "apiKey": "${WEATHER_API_KEY}"
               },
               "models": {
                 "fast": { "name": "weather-fast" }
@@ -46,7 +47,10 @@ struct CustomProviderRegistryTests {
             },
             "claude-proxy": {
               "type": "anthropic",
-              "options": { "baseURL": "https://anthropic.local" }
+              "options": {
+                "baseURL": "https://anthropic.local",
+                "apiKey": "claude-provider-key"
+              }
             }
           }
         }
@@ -68,6 +72,7 @@ struct CustomProviderRegistryTests {
             #endif
             TachikomaConfiguration.profileDirectoryName = originalProfileDir
             unsetenv("WEATHER_TOKEN")
+            unsetenv("WEATHER_API_KEY")
             self.resetRegistry(
                 forProfile: originalProfileDir,
                 originalHome: originalHome,
@@ -78,6 +83,7 @@ struct CustomProviderRegistryTests {
         TachikomaConfiguration.profileDirectoryName = profileDirName
         self.setHomeEnvironment(to: tempHome.path)
         setenv("WEATHER_TOKEN", "sk-test-weather", 1)
+        setenv("WEATHER_API_KEY", "sk-provider-key", 1)
 
         CustomProviderRegistry.shared.loadFromProfile()
         let providers = CustomProviderRegistry.shared.list()
@@ -87,14 +93,57 @@ struct CustomProviderRegistryTests {
         let weather = try #require(providers["weather-ai"])
         #expect(weather.kind == .openai)
         #expect(weather.baseURL == "https://api.example.com")
+        #expect(weather.apiKey == "sk-provider-key")
         #expect(weather.headers["Authorization"] == "Bearer sk-test-weather")
         #expect(weather.models["fast"] == "weather-fast")
 
         let claude = try #require(CustomProviderRegistry.shared.get("claude-proxy"))
         #expect(claude.kind == .anthropic)
+        #expect(claude.apiKey == "claude-provider-key")
         #expect(claude.headers.isEmpty)
         #expect(claude.models.isEmpty)
 
+        let dynamicProvider = DynamicCustomProvider(modelId: "weather-ai/fast")
+        let resolved = try ProviderFactory.createProvider(
+            for: .custom(provider: dynamicProvider),
+            configuration: TachikomaConfiguration(loadFromEnvironment: false),
+        )
+        let compatibleProvider = try #require(resolved as? OpenAICompatibleProvider)
+        #expect(compatibleProvider.apiKey == "sk-provider-key")
+        #expect(compatibleProvider.additionalHeaders["Authorization"] == "Bearer sk-test-weather")
+
+        let claudeProvider = DynamicCustomProvider(modelId: "claude-proxy/sonnet")
+        let resolvedClaude = try ProviderFactory.createProvider(
+            for: .custom(provider: claudeProvider),
+            configuration: TachikomaConfiguration(loadFromEnvironment: false),
+        )
+        let compatibleClaude = try #require(resolvedClaude as? AnthropicCompatibleProvider)
+        #expect(compatibleClaude.apiKey == "claude-provider-key")
+
+        let fableModel = LanguageModel.custom(
+            provider: DynamicCustomProvider(modelId: "claude-proxy/claude-fable-5"),
+        )
+        #expect(fableModel.supportsStreaming == false)
+        let resolvedFable = try ProviderFactory.createProvider(
+            for: fableModel,
+            configuration: TachikomaConfiguration(loadFromEnvironment: false),
+        )
+        let compatibleFable = try #require(resolvedFable as? AnthropicCompatibleProvider)
+        #expect(compatibleFable.capabilities.supportsStreaming == false)
+
+        let directFableProvider = DynamicCustomProvider(
+            modelId: "claude-fable-5",
+            capabilities: ModelCapabilities(supportsStreaming: false),
+        )
+        let directFableModel = LanguageModel.custom(provider: directFableProvider)
+        #expect(directFableModel.supportsStreaming == false)
+
+        let unrelatedFableNamedProvider = DynamicCustomProvider(
+            modelId: "local-claude-fable-5-benchmark",
+            capabilities: ModelCapabilities(supportsStreaming: true),
+        )
+        #expect(LanguageModel.custom(provider: unrelatedFableNamedProvider).supportsStreaming == true)
+
         #expect(CustomProviderRegistry.shared.get("missing") == nil)
     }
 
@@ -145,3 +194,25 @@ struct CustomProviderRegistryTests {
         #endif
     }
 }
+
+private final class DynamicCustomProvider: ModelProvider {
+    let modelId: String
+    let baseURL: String? = nil
+    let apiKey: String? = nil
+    let capabilities: ModelCapabilities
+
+    init(modelId: String, capabilities: ModelCapabilities = ModelCapabilities()) {
+        self.modelId = modelId
+        self.capabilities = capabilities
+    }
+
+    func generateText(request _: ProviderRequest) async throws -> ProviderResponse {
+        ProviderResponse(text: "")
+    }
+
+    func streamText(request _: ProviderRequest) async throws -> AsyncThrowingStream<TextStreamDelta, any Error> {
+        AsyncThrowingStream { continuation in
+            continuation.finish()
+        }
+    }
+}

Tests/TachikomaTests/Core/GrokModelCatalogTests.swift

@@ -4,17 +4,9 @@ import Testing
 
 struct GrokModelCatalogTests {
     private static let catalog: [Model.Grok] = [
-        .grok4,
-        .grok4FastReasoning,
-        .grok4FastNonReasoning,
-        .grokCodeFast1,
-        .grok3,
-        .grok3Mini,
-        .grok2,
-        .grok2Vision,
-        .grok2Image,
-        .grokVisionBeta,
-        .grokBeta,
+        .grok43,
+        .grok420Reasoning,
+        .grok420NonReasoning,
     ]
 
     private func requireModernPlatforms(_ body: () throws -> Void) rethrows {
@@ -39,6 +31,7 @@ struct GrokModelCatalogTests {
                 let parsed = try ModelSelector.parseModel(model.modelId)
                 #expect(parsed == .grok(model))
             }
+            #expect(try ModelSelector.parseModel("grok-4.3-latest") == .grok(.grok43))
         }
     }
 
@@ -52,13 +45,69 @@ struct GrokModelCatalogTests {
     }
 
     @Test
-    func `Vision capability only flips on for vision/image Grok models`() {
+    func `Grok model vision support matches current xAI catalog`() {
         self.requireModernPlatforms {
-            let visionModels: Set<Model.Grok> = [.grok2Vision, .grok2Image, .grokVisionBeta]
+            #expect(Model.grok(.grok43).supportsVision)
+            #expect(Model.grok(.grok420Reasoning).supportsVision)
+            #expect(Model.grok(.grok420NonReasoning).supportsVision)
+            #expect(Model.grok(.grok420MultiAgent).supportsVision == false)
+            #expect(Model.grok(.grok420MultiAgent).supportsTools == false)
+        }
+    }
 
-            for model in Self.catalog {
-                let languageModel = Model.grok(model)
-                #expect(languageModel.supportsVision == visionModels.contains(model))
+    @Test
+    func `ModelSelector preserves server-redirected Grok identifiers`() throws {
+        try self.requireModernPlatforms {
+            for id in [
+                "grok-4-0709",
+                "grok-3",
+                "grok-2-1212",
+                "grok-4-fast",
+                "grok-code-fast-1",
+            ] {
+                let parsed = try ModelSelector.parseModel(id)
+                #expect(parsed == .grok(.custom(id)))
+            }
+        }
+    }
+
+    @Test
+    func `ModelSelector keeps provider-qualified Grok slugs on xAI`() throws {
+        try self.requireModernPlatforms {
+            let parsed = try ModelSelector.parseModel("xai/grok-code-fast-1")
+
+            #expect(parsed == .grok(.custom("grok-code-fast-1")))
+        }
+    }
+
+    @Test
+    func `ModelSelector rejects unsupported Grok multi-agent identifiers`() {
+        self.requireModernPlatforms {
+            for id in [
+                "grok-4.20-multi-agent-0309",
+                "grok420multiagent",
+                "xai/grok-4.20-multi-agent",
+            ] {
+                #expect(throws: ModelValidationError.self) {
+                    _ = try ModelSelector.parseModel(id)
+                }
+            }
+        }
+    }
+
+    @Test
+    func `Grok provider rejects multi-agent until Responses routing exists`() {
+        self.requireModernPlatforms {
+            let config = TachikomaConfiguration(apiKeys: ["grok": "test-key"])
+
+            #expect(throws: TachikomaError.self) {
+                _ = try ProviderFactory.createProvider(for: .grok(.grok420MultiAgent), configuration: config)
+            }
+            #expect(throws: TachikomaError.self) {
+                _ = try GrokProvider(model: .grok420MultiAgent, configuration: config)
+            }
+            #expect(throws: TachikomaError.self) {
+                _ = try GrokProvider(model: .custom("grok420multiagent"), configuration: config)
             }
         }
     }

Tests/TachikomaTests/Core/LanguageModelCoverageTests.swift

@@ -6,6 +6,7 @@ struct LanguageModelCoverageTests {
     func `OpenAI enum exposes properties`() {
         let models = LanguageModel.OpenAI.allCases
         #expect(!models.isEmpty)
+        #expect(models.contains(.gpt5ChatLatest))
         for model in models {
             #expect(!model.modelId.isEmpty)
             _ = model.supportsVision
@@ -52,6 +53,13 @@ struct LanguageModelCoverageTests {
             #expect(model.contextLength > 0)
         }
 
+        for model in LanguageModel.MiniMax.allCases {
+            #expect(!model.modelId.isEmpty)
+            _ = model.supportsVision
+            _ = model.supportsTools
+            #expect(model.contextLength > 0)
+        }
+
         for model in LanguageModel.Ollama.allCases {
             #expect(!model.modelId.isEmpty)
             _ = model.supportsVision
@@ -68,12 +76,14 @@ struct LanguageModelCoverageTests {
     @Test
     func `LanguageModel top level switches`() {
         let baseModels: [LanguageModel] = [
-            .openai(.gpt51),
-            .anthropic(.opus45),
-            .google(.gemini25Flash),
-            .mistral(.large2),
-            .groq(.mixtral8x7b),
-            .grok(.grok4),
+            .openai(.gpt55),
+            .anthropic(.fable5),
+            .google(.gemini35Flash),
+            .mistral(.medium35),
+            .groq(.llama3370b),
+            .grok(.grok43),
+            .minimax(.m27),
+            .minimaxCN(.m27),
             .ollama(.llama33),
             .lmstudio(.gptOSS20B),
             .openRouter(modelId: "openrouter/alpha"),

Tests/TachikomaTests/Core/MinimalModernAPITests.swift

@@ -9,9 +9,9 @@ struct MinimalModernAPITests {
     @Test
     func `Model enum construction`() {
         // Test that model enums can be constructed
-        let openaiModel = Model.openai(.gpt4o)
-        let anthropicModel = Model.anthropic(.opus45)
-        _ = Model.grok(.grok4)
+        let openaiModel = Model.openai(.gpt55)
+        let anthropicModel = Model.anthropic(.opus48)
+        _ = Model.grok(.grok43)
         _ = Model.ollama(.llama33)
 
         // Test that they can be used in a switch statement
@@ -35,13 +35,131 @@ struct MinimalModernAPITests {
         let defaultModel = Model.default
         // Should compile without errors
         switch defaultModel {
-        case .anthropic(.opus45):
+        case .anthropic(.opus48):
             break // Expected default
         default:
-            Issue.record("Expected default to be Anthropic Opus 4.5")
+            Issue.record("Expected default to be Anthropic Opus 4.8")
         }
     }
 
+    @Test
+    func `Streaming default value`() {
+        #expect(Model.default.supportsStreaming == false)
+        #expect(Model.defaultStreaming == .openai(.gpt55))
+        #expect(Model.defaultStreaming.supportsStreaming == true)
+    }
+
+    @Test
+    func `Agent default model preserves execution default`() {
+        let agent = Agent(name: "test", instructions: "test", context: ())
+
+        #expect(agent.model == .default)
+    }
+
+    @Test
+    func `Agent stream uses streaming fallback for execution default`() async throws {
+        let seenModel = MinimalModelBox()
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setProviderFactoryOverride { model, _ in
+            seenModel.model = model
+            return MinimalStreamingProvider(deltas: [
+                .text("ok"),
+                .done(finishReason: .stop),
+            ])
+        }
+        let agent = Agent(name: "test", instructions: "test", configuration: config, context: ())
+
+        let stream = try await agent.stream("hi")
+        var received = ""
+        for try await delta in stream where delta.type == .textDelta {
+            received += delta.content ?? ""
+        }
+
+        #expect(agent.model == .default)
+        #expect(seenModel.model == .openai(.gpt55))
+        #expect(!received.isEmpty)
+    }
+
+    @Test
+    func `Agent stream rejects explicit execution default`() async throws {
+        let agent = Agent(name: "test", instructions: "test", model: .default, context: ())
+
+        await #expect(throws: TachikomaError.self) {
+            _ = try await agent.stream("hi")
+        }
+    }
+
+    @Test
+    func `Agent stream rejects nonstreaming model after mutation`() async throws {
+        let agent = Agent(name: "test", instructions: "test", context: ())
+        agent.model = .anthropic(.fable5)
+
+        await #expect(throws: TachikomaError.self) {
+            _ = try await agent.stream("hi")
+        }
+    }
+
+    @Test
+    func `Agent stream flushes buffered text on natural completion`() async throws {
+        let provider = MinimalStreamingProvider(deltas: [
+            .text("ok"),
+        ])
+        let agent = Agent(
+            name: "test",
+            instructions: "test",
+            model: .custom(provider: provider),
+            context: (),
+        )
+
+        let stream = try await agent.stream("hi")
+        var received = ""
+        for try await delta in stream where delta.type == .textDelta {
+            received += delta.content ?? ""
+        }
+
+        #expect(received == "ok")
+    }
+
+    @Test
+    func `Agent stream flushes buffered compatible text when done has no finish reason`() async throws {
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setProviderFactoryOverride { _, _ in
+            MinimalStreamingProvider(deltas: [
+                .text("ok"),
+                .done(),
+            ])
+        }
+        let agent = Agent(
+            name: "test",
+            instructions: "test",
+            model: .openaiCompatible(modelId: "gpt-compatible", baseURL: "https://example.test"),
+            configuration: config,
+            context: (),
+        )
+
+        let stream = try await agent.stream("hi")
+        var received = ""
+        for try await delta in stream where delta.type == .textDelta {
+            received += delta.content ?? ""
+        }
+
+        #expect(received == "ok")
+        #expect(agent.conversation.messages.map(\.content) == ["test", "hi", "ok"])
+    }
+
+    @Test
+    func `Agent conversation uses agent configuration`() async throws {
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setProviderFactoryOverride { _, _ in
+            MinimalStaticProvider(response: ProviderResponse(text: "configured", finishReason: .stop))
+        }
+        let agent = Agent(name: "test", instructions: "test", configuration: config, context: ())
+
+        let text = try await agent.conversation.continueConversation(using: .openai(.gpt55))
+
+        #expect(text == "configured")
+    }
+
     // MARK: - Tool System Tests
 
     @Test
@@ -126,4 +244,757 @@ extension MinimalModernAPITests {
         #expect(message.role == .user)
         #expect(message.content == "Test")
     }
+
+    @Test
+    func `Conversation preserves signed thinking messages`() {
+        let conversation = Conversation()
+        let signedThinking = ModelMessage(
+            role: .assistant,
+            content: [.text("private reasoning")],
+            channel: .thinking,
+            metadata: .init(customData: [
+                "anthropic.thinking.signature": "sig",
+                "anthropic.thinking.type": "thinking",
+            ]),
+        )
+
+        conversation.replaceModelMessages([.user("hi"), signedThinking, .assistant("hello")])
+
+        let messages = conversation.getModelMessages()
+        #expect(messages.count == 3)
+        #expect(messages[1] == signedThinking)
+        #expect(conversation.messages[1].content == "private reasoning")
+    }
+
+    @Test
+    func `Conversation merge preserves messages appended after snapshot`() {
+        let conversation = Conversation()
+        conversation.addUserMessage("original")
+        let snapshotCount = conversation.messages.count
+        conversation.addUserMessage("concurrent")
+
+        conversation.mergeGeneratedMessages(
+            [.user("original"), .assistant("generated")],
+            replacingPrefixCount: snapshotCount,
+        )
+
+        let messages = conversation.getModelMessages()
+        #expect(messages.map(\.role) == [.user, .assistant, .user])
+        if case let .text(text) = messages[2].content.first {
+            #expect(text == "concurrent")
+        } else {
+            Issue.record("Expected preserved concurrent user message")
+        }
+    }
+
+    @Test
+    func `Conversation refusal rollback preserves messages appended after snapshot`() {
+        let conversation = Conversation()
+        conversation.addUserMessage("blocked")
+        let snapshotIDs = conversation.messages.map(\.id)
+        conversation.addUserMessage("concurrent")
+
+        let didReplace = conversation.replaceModelMessages([], validatingSnapshotIDs: snapshotIDs)
+
+        #expect(didReplace == true)
+        #expect(conversation.messages.map(\.content) == ["concurrent"])
+    }
+
+    @Test
+    func `Conversation lock removes cancelled waiters`() async throws {
+        let conversation = Conversation()
+        let probe = ConversationLockProbe()
+
+        let first = Task {
+            try await conversation.withContinuationLock {
+                await probe.markFirstStarted()
+                await probe.waitForRelease()
+            }
+        }
+
+        await probe.waitUntilFirstStarted()
+
+        let second = Task {
+            try await conversation.withContinuationLock {
+                await probe.markSecondRan()
+            }
+        }
+
+        try await Task.sleep(nanoseconds: 10_000_000)
+        second.cancel()
+
+        do {
+            try await second.value
+            Issue.record("Expected queued waiter to be cancelled")
+        } catch is CancellationError {
+            // Expected
+        }
+
+        await probe.releaseFirst()
+        try await first.value
+
+        try await conversation.withContinuationLock {
+            await probe.markThirdRan()
+        }
+
+        #expect(await probe.secondRan == false)
+        #expect(await probe.thirdRan == true)
+    }
+
+    @Test
+    func `Conversation append generated messages preserves concurrent appends`() {
+        let conversation = Conversation()
+        conversation.addUserMessage("original")
+        let anchorID = conversation.messages[0].id
+        conversation.addUserMessage("concurrent")
+
+        conversation.appendGeneratedMessages([.assistant("generated")], afterMessageID: anchorID)
+
+        #expect(conversation.messages.map(\.content) == ["original", "generated", "concurrent"])
+    }
+
+    @Test
+    func `Conversation continue persists generated message from empty history`() async throws {
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setProviderFactoryOverride { _, _ in
+            MinimalStaticProvider(response: ProviderResponse(text: "hello", finishReason: .stop))
+        }
+        let conversation = Conversation(configuration: config)
+
+        let text = try await conversation.continueConversation(using: .anthropic(.opus48))
+
+        #expect(text == "hello")
+        #expect(conversation.messages.map(\.content) == ["hello"])
+    }
+
+    @Test
+    func `Conversation continue rolls back refused trailing user turn`() async throws {
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setProviderFactoryOverride { _, _ in
+            MinimalStaticProvider(response: ProviderResponse(text: "Refused by policy", finishReason: .contentFilter))
+        }
+        let conversation = Conversation(configuration: config)
+        conversation.addUserMessage("blocked")
+
+        let text = try await conversation.continueConversation(using: .anthropic(.fable5))
+
+        #expect(text.isEmpty)
+        #expect(conversation.messages.isEmpty)
+    }
+
+    @Test
+    func `Conversation continue preserves completed tool history after late refusal`() async throws {
+        let provider = MinimalSequenceProvider(responses: [
+            ProviderResponse(
+                text: "",
+                finishReason: .toolCalls,
+                toolCalls: [AgentToolCall(id: "call-1", name: "side_effect", arguments: [:])],
+            ),
+            ProviderResponse(text: "Refused by policy", finishReason: .contentFilter),
+        ])
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setProviderFactoryOverride { _, _ in provider }
+        let conversation = Conversation(configuration: config)
+        conversation.addUserMessage("do it")
+
+        let text = try await conversation.continueConversation(
+            using: .anthropic(.fable5),
+            tools: [sideEffectTool],
+            maxSteps: 2,
+        )
+
+        #expect(text.isEmpty)
+        let messages = conversation.getModelMessages()
+        #expect(messages.map(\.role) == [.user, .assistant, .tool])
+        #expect(messages[0].content == [.text("do it")])
+        #expect(messages[1].content.contains { part in
+            if case let .toolCall(toolCall) = part {
+                return toolCall.id == "call-1"
+            }
+            return false
+        })
+        #expect(messages[2].content.contains { part in
+            if case let .toolResult(toolResult) = part {
+                return toolResult.toolCallId == "call-1"
+            }
+            return false
+        })
+    }
+
+    @Test
+    func `Agent stream rejects non-streaming model before mutating conversation`() async throws {
+        let agent = Agent(
+            name: "test",
+            instructions: "test",
+            model: .anthropic(.fable5),
+            context: (),
+        )
+
+        await #expect(throws: TachikomaError.self) {
+            _ = try await agent.stream("hi")
+        }
+
+        #expect(agent.conversation.messages.map(\.content) == ["test"])
+    }
+
+    @Test
+    func `Conversation streaming rolls back refused trailing user turn`() async throws {
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setProviderFactoryOverride { _, _ in
+            MinimalStreamingProvider(deltas: [
+                .text("partial"),
+                .done(finishReason: .contentFilter),
+            ])
+        }
+        let conversation = Conversation(configuration: config)
+        conversation.addUserMessage("blocked")
+
+        let stream = try await conversation.continueConversationStreaming(using: .openai(.gpt55))
+        var received = ""
+        for try await chunk in stream {
+            received += chunk
+        }
+
+        #expect(received == "partial")
+        #expect(conversation.messages.isEmpty)
+    }
+
+    @Test
+    func `Conversation streaming flushes buffered text on natural completion`() async throws {
+        let provider = MinimalStreamingProvider(deltas: [
+            .text("ok"),
+        ])
+        let conversation = Conversation(configuration: TachikomaConfiguration(loadFromEnvironment: false))
+
+        let stream = try await conversation.continueConversationStreaming(using: .custom(provider: provider))
+        var received = ""
+        for try await chunk in stream {
+            received += chunk
+        }
+
+        #expect(received == "ok")
+        #expect(conversation.messages.map(\.content) == ["ok"])
+    }
+
+    @Test
+    func `Conversation streaming flushes buffered compatible text when done has no finish reason`() async throws {
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setProviderFactoryOverride { _, _ in
+            MinimalStreamingProvider(deltas: [
+                .text("ok"),
+                .done(),
+            ])
+        }
+        let conversation = Conversation(configuration: config)
+
+        let stream = try await conversation.continueConversationStreaming(
+            using: .openaiCompatible(modelId: "gpt-compatible", baseURL: "https://example.test"),
+        )
+        var received = ""
+        for try await chunk in stream {
+            received += chunk
+        }
+
+        #expect(received == "ok")
+        #expect(conversation.messages.map(\.content) == ["ok"])
+    }
+
+    @Test
+    func `Conversation streaming flushes compatible text when stream ends without done`() async throws {
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setProviderFactoryOverride { _, _ in
+            MinimalStreamingProvider(deltas: [
+                .text("partial"),
+            ])
+        }
+        let conversation = Conversation(configuration: config)
+
+        let stream = try await conversation.continueConversationStreaming(
+            using: .openaiCompatible(modelId: "gpt-compatible", baseURL: "https://example.test"),
+        )
+
+        var received = ""
+        for try await chunk in stream {
+            received += chunk
+        }
+
+        #expect(received == "partial")
+        #expect(conversation.messages.map(\.content) == ["partial"])
+    }
+}
+
+@Suite(.serialized)
+private struct AgentRefusalTests {
+    @Test
+    func `Agent execute rolls back refused user turn`() async throws {
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setProviderFactoryOverride { _, _ in
+            MinimalStaticProvider(response: ProviderResponse(text: "Refused by policy", finishReason: .contentFilter))
+        }
+
+        let agent = Agent(
+            name: "test",
+            instructions: "test",
+            model: .anthropic(.fable5),
+            configuration: config,
+            context: (),
+        )
+
+        let response = try await agent.execute("blocked")
+
+        #expect(response.text.isEmpty)
+        #expect(response.finishReason == .contentFilter)
+        #expect(agent.conversation.messages.map(\.content) == ["test"])
+    }
+
+    @Test
+    func `Agent stream stays incremental by default when terminal content filter arrives`() async throws {
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setProviderFactoryOverride { _, _ in
+            MinimalStreamingProvider(deltas: [
+                .text("partial"),
+                .done(finishReason: .contentFilter),
+            ])
+        }
+
+        let agent = Agent(
+            name: "test",
+            instructions: "test",
+            model: .openai(.gpt55),
+            configuration: config,
+            context: (),
+        )
+
+        let stream = try await agent.stream("blocked")
+        var received: [TextStreamDelta] = []
+        for try await delta in stream {
+            received.append(delta)
+        }
+
+        #expect(received.contains { $0.type == .textDelta && $0.content == "partial" })
+        #expect(received.contains { $0.type == .done && $0.finishReason == .contentFilter })
+        #expect(agent.conversation.messages.map(\.content) == ["test"])
+    }
+
+    @Test
+    func `Agent stream explicit terminal buffering errors when stream ends without done`() async throws {
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setProviderFactoryOverride { _, _ in
+            MinimalStreamingProvider(deltas: [
+                .text("partial"),
+            ])
+        }
+
+        let agent = Agent(
+            name: "test",
+            instructions: "test",
+            model: .openaiCompatible(modelId: "gpt-compatible", baseURL: "https://example.test"),
+            settings: GenerationSettings(streamBuffering: .untilTerminal),
+            configuration: config,
+            context: (),
+        )
+
+        let stream = try await agent.stream("hi")
+        do {
+            for try await _ in stream {}
+            Issue.record("Expected missing terminal status error")
+        } catch let error as TachikomaError {
+            guard case let .apiError(message) = error else {
+                Issue.record("Expected apiError, got \(error)")
+                return
+            }
+            #expect(message.contains("completion status"))
+        }
+
+        #expect(!agent.conversation.messages.map(\.content).contains("partial"))
+    }
+
+    @Test
+    func `Agent stream explicit terminal buffering suppresses Azure OpenAI refusals`() async throws {
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setProviderFactoryOverride { _, _ in
+            MinimalStreamingProvider(deltas: [
+                .text("partial"),
+                .done(finishReason: .contentFilter),
+            ])
+        }
+
+        let agent = Agent(
+            name: "test",
+            instructions: "test",
+            model: .azureOpenAI(deployment: "gpt-compatible", endpoint: "https://example.openai.azure.com"),
+            settings: GenerationSettings(streamBuffering: .untilTerminal),
+            configuration: config,
+            context: (),
+        )
+
+        let stream = try await agent.stream("blocked")
+        var received: [TextStreamDelta] = []
+        for try await delta in stream {
+            received.append(delta)
+        }
+
+        #expect(!received.contains { $0.type == .textDelta && $0.content == "partial" })
+        #expect(received.contains { $0.type == .done && $0.finishReason == .contentFilter })
+        #expect(agent.conversation.messages.map(\.content) == ["test"])
+    }
+
+    @Test
+    func `Agent stream explicit terminal buffering suppresses Google refusals`() async throws {
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setProviderFactoryOverride { _, _ in
+            MinimalStreamingProvider(deltas: [
+                .text("partial"),
+                .done(finishReason: .contentFilter),
+            ])
+        }
+
+        let agent = Agent(
+            name: "test",
+            instructions: "test",
+            model: .google(.gemini25Flash),
+            settings: GenerationSettings(streamBuffering: .untilTerminal),
+            configuration: config,
+            context: (),
+        )
+
+        let stream = try await agent.stream("blocked")
+        var received: [TextStreamDelta] = []
+        for try await delta in stream {
+            received.append(delta)
+        }
+
+        #expect(!received.contains { $0.type == .textDelta && $0.content == "partial" })
+        #expect(received.contains { $0.type == .done && $0.finishReason == .contentFilter })
+        #expect(agent.conversation.messages.map(\.content) == ["test"])
+    }
+
+    @Test
+    func `Agent stream explicit terminal buffering suppresses registered custom OpenAI refusals`() async throws {
+        try await self.withRegisteredCustomProvider(
+            """
+            {
+              "customProviders": {
+                "proxy": {
+                  "type": "openai",
+                  "options": { "baseURL": "https://example.test/v1" }
+                }
+              }
+            }
+            """,
+        ) {
+            let config = TachikomaConfiguration(loadFromEnvironment: false)
+            config.setProviderFactoryOverride { _, _ in
+                MinimalStreamingProvider(
+                    modelId: "proxy/gpt-compatible",
+                    deltas: [
+                        .text("partial"),
+                        .done(finishReason: .contentFilter),
+                    ],
+                )
+            }
+
+            let agent = Agent(
+                name: "test",
+                instructions: "test",
+                model: .custom(provider: MinimalStreamingProvider(modelId: "proxy/gpt-compatible", deltas: [])),
+                settings: GenerationSettings(streamBuffering: .untilTerminal),
+                configuration: config,
+                context: (),
+            )
+
+            let stream = try await agent.stream("blocked")
+            var received: [TextStreamDelta] = []
+            for try await delta in stream {
+                received.append(delta)
+            }
+
+            #expect(!received.contains { $0.type == .textDelta && $0.content == "partial" })
+            #expect(received.contains { $0.type == .done && $0.finishReason == .contentFilter })
+            #expect(agent.conversation.messages.map(\.content) == ["test"])
+        }
+    }
+
+    @Test
+    func `Agent stream releases continuation gate when consumer stops early`() async throws {
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setProviderFactoryOverride { _, _ in
+            StallingStreamingProvider()
+        }
+        let agent = Agent(
+            name: "test",
+            instructions: "test",
+            model: .custom(provider: StallingStreamingProvider()),
+            configuration: config,
+            context: (),
+        )
+
+        do {
+            let stream = try await agent.stream("first")
+            var iterator = stream.makeAsyncIterator()
+            let firstDelta = try await iterator.next()
+            #expect(firstDelta?.type == .textDelta)
+            #expect(firstDelta?.content == "partial")
+        }
+        try await Task.sleep(nanoseconds: 10_000_000)
+
+        let response = try await withTimeout(0.2) {
+            try await agent.execute("second")
+        }
+        #expect(response.text == "after")
+    }
+
+    @Test
+    func `Agent execute preserves completed tool history after late refusal`() async throws {
+        let provider = MinimalSequenceProvider(responses: [
+            ProviderResponse(
+                text: "",
+                finishReason: .toolCalls,
+                toolCalls: [AgentToolCall(id: "call-1", name: "side_effect", arguments: [:])],
+            ),
+            ProviderResponse(text: "Refused by policy", finishReason: .contentFilter),
+        ])
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setProviderFactoryOverride { _, _ in provider }
+
+        let agent = Agent(
+            name: "test",
+            instructions: "test",
+            model: .anthropic(.fable5),
+            tools: [sideEffectTool],
+            configuration: config,
+            context: (),
+        )
+
+        let response = try await agent.execute("do it")
+
+        #expect(response.text.isEmpty)
+        #expect(response.finishReason == .contentFilter)
+        let messages = agent.conversation.getModelMessages()
+        #expect(messages.map(\.role) == [.system, .user, .assistant, .tool])
+        #expect(messages[1].content == [.text("do it")])
+        #expect(messages[2].content.contains { part in
+            if case let .toolCall(toolCall) = part {
+                return toolCall.id == "call-1"
+            }
+            return false
+        })
+        #expect(messages[3].content.contains { part in
+            if case let .toolResult(toolResult) = part {
+                return toolResult.toolCallId == "call-1"
+            }
+            return false
+        })
+    }
+
+    private func withRegisteredCustomProvider(
+        _ configJSON: String,
+        operation: () async throws -> Void,
+    ) async throws {
+        let originalProfile = TachikomaConfiguration.profileDirectoryName
+        let tempProfile = FileManager.default.temporaryDirectory.appendingPathComponent(UUID().uuidString)
+        let emptyProfile = FileManager.default.temporaryDirectory.appendingPathComponent(UUID().uuidString)
+
+        try FileManager.default.createDirectory(at: tempProfile, withIntermediateDirectories: true)
+        try FileManager.default.createDirectory(at: emptyProfile, withIntermediateDirectories: true)
+        try configJSON.write(to: tempProfile.appendingPathComponent("config.json"), atomically: true, encoding: .utf8)
+        try #"{"customProviders":{}}"#.write(
+            to: emptyProfile.appendingPathComponent("config.json"),
+            atomically: true,
+            encoding: .utf8,
+        )
+
+        TachikomaConfiguration.profileDirectoryName = tempProfile.path
+        CustomProviderRegistry.shared.loadFromProfile()
+
+        do {
+            try await operation()
+            TachikomaConfiguration.profileDirectoryName = emptyProfile.path
+            CustomProviderRegistry.shared.loadFromProfile()
+            TachikomaConfiguration.profileDirectoryName = originalProfile
+        } catch {
+            TachikomaConfiguration.profileDirectoryName = emptyProfile.path
+            CustomProviderRegistry.shared.loadFromProfile()
+            TachikomaConfiguration.profileDirectoryName = originalProfile
+            throw error
+        }
+    }
+}
+
+private struct StallingStreamingProvider: ModelProvider {
+    let modelId = "stalling-streaming"
+    let baseURL: String? = nil
+    let apiKey: String? = nil
+    let capabilities = ModelCapabilities(supportsStreaming: true)
+
+    func generateText(request _: ProviderRequest) async throws -> ProviderResponse {
+        ProviderResponse(text: "after")
+    }
+
+    func streamText(request _: ProviderRequest) async throws -> AsyncThrowingStream<TextStreamDelta, Error> {
+        AsyncThrowingStream { continuation in
+            continuation.yield(.text("partial"))
+        }
+    }
+}
+
+private actor ConversationLockProbe {
+    var secondRan = false
+    var thirdRan = false
+    private var firstStarted = false
+    private var firstStartedWaiters: [CheckedContinuation<Void, Never>] = []
+    private var releaseWaiters: [CheckedContinuation<Void, Never>] = []
+
+    func markFirstStarted() {
+        self.firstStarted = true
+        let waiters = self.firstStartedWaiters
+        self.firstStartedWaiters.removeAll()
+        for waiter in waiters {
+            waiter.resume()
+        }
+    }
+
+    func waitUntilFirstStarted() async {
+        if self.firstStarted {
+            return
+        }
+
+        await withCheckedContinuation { continuation in
+            self.firstStartedWaiters.append(continuation)
+        }
+    }
+
+    func waitForRelease() async {
+        await withCheckedContinuation { continuation in
+            self.releaseWaiters.append(continuation)
+        }
+    }
+
+    func releaseFirst() {
+        let waiters = self.releaseWaiters
+        self.releaseWaiters.removeAll()
+        for waiter in waiters {
+            waiter.resume()
+        }
+    }
+
+    func markSecondRan() {
+        self.secondRan = true
+    }
+
+    func markThirdRan() {
+        self.thirdRan = true
+    }
+}
+
+private final class MinimalModelBox: @unchecked Sendable {
+    private let lock = NSLock()
+    private var _model: LanguageModel?
+
+    var model: LanguageModel? {
+        get {
+            self.lock.lock()
+            defer { self.lock.unlock() }
+            return self._model
+        }
+        set {
+            self.lock.lock()
+            self._model = newValue
+            self.lock.unlock()
+        }
+    }
+}
+
+private struct MinimalStaticProvider: ModelProvider {
+    let modelId = "minimal-static"
+    let baseURL: String? = nil
+    let apiKey: String? = nil
+    let capabilities = ModelCapabilities()
+    let response: ProviderResponse
+
+    func generateText(request _: ProviderRequest) async throws -> ProviderResponse {
+        self.response
+    }
+
+    func streamText(request _: ProviderRequest) async throws -> AsyncThrowingStream<TextStreamDelta, Error> {
+        AsyncThrowingStream { continuation in
+            continuation.finish()
+        }
+    }
+}
+
+private struct MinimalStreamingProvider: ModelProvider {
+    let modelId: String
+    let baseURL: String? = nil
+    let apiKey: String? = nil
+    let capabilities = ModelCapabilities(supportsStreaming: true)
+    let deltas: [TextStreamDelta]
+
+    init(modelId: String = "minimal-streaming", deltas: [TextStreamDelta]) {
+        self.modelId = modelId
+        self.deltas = deltas
+    }
+
+    func generateText(request _: ProviderRequest) async throws -> ProviderResponse {
+        ProviderResponse(text: "")
+    }
+
+    func streamText(request _: ProviderRequest) async throws -> AsyncThrowingStream<TextStreamDelta, any Error> {
+        AsyncThrowingStream { continuation in
+            for delta in self.deltas {
+                continuation.yield(delta)
+            }
+            continuation.finish()
+        }
+    }
+}
+
+private struct MinimalSequenceProvider: ModelProvider {
+    let modelId = "minimal-sequence"
+    let baseURL: String? = nil
+    let apiKey: String? = nil
+    let capabilities = ModelCapabilities()
+    private let queue: MinimalResponseQueue
+
+    init(responses: [ProviderResponse]) {
+        self.queue = MinimalResponseQueue(responses: responses)
+    }
+
+    func generateText(request _: ProviderRequest) async throws -> ProviderResponse {
+        self.queue.next()
+    }
+
+    func streamText(request _: ProviderRequest) async throws -> AsyncThrowingStream<TextStreamDelta, Error> {
+        AsyncThrowingStream { continuation in
+            continuation.finish()
+        }
+    }
+}
+
+private final class MinimalResponseQueue: @unchecked Sendable {
+    private let lock = NSLock()
+    private var responses: [ProviderResponse]
+
+    init(responses: [ProviderResponse]) {
+        self.responses = responses
+    }
+
+    func next() -> ProviderResponse {
+        self.lock.lock()
+        defer { self.lock.unlock() }
+
+        if self.responses.count > 1 {
+            return self.responses.removeFirst()
+        }
+        return self.responses[0]
+    }
+}
+
+private let sideEffectTool = Tachikoma.createTool(
+    name: "side_effect",
+    description: "Records an external action",
+    parameters: [],
+    required: [],
+) { _ in
+    AnyAgentToolValue(string: "done")
 }

Tests/TachikomaTests/Core/ModelCapabilitiesTests.swift

@@ -7,8 +7,11 @@ enum ModelCapabilitiesTests {
         @Test
         func `GPT-5 models exclude temperature and topP`() {
             let models: [LanguageModel] = [
-                .openai(.gpt52),
-                .openai(.gpt51),
+                .openai(.gpt5ChatLatest),
+                .openai(.gpt55),
+                .openai(.gpt54),
+                .openai(.gpt54Mini),
+                .openai(.gpt54Nano),
                 .openai(.gpt5),
                 .openai(.gpt5Mini),
                 .openai(.gpt5Nano),
@@ -27,34 +30,24 @@ enum ModelCapabilitiesTests {
         }
 
         @Test
-        func `Gemini 3 Flash supports thinking config options`() {
-            let capabilities = ModelCapabilityRegistry.shared.capabilities(for: .google(.gemini3Flash))
+        func `chat-latest does not advertise audio input`() {
+            #expect(LanguageModel.openai(.chatLatest).supportsVision)
+            #expect(LanguageModel.openai(.chatLatest).supportsTools)
+            #expect(LanguageModel.openai(.chatLatest).supportsAudioInput == false)
+        }
+
+        @Test
+        func `Gemini 3_5 Flash supports thinking config options`() {
+            let capabilities = ModelCapabilityRegistry.shared.capabilities(for: .google(.gemini35Flash))
             #expect(capabilities.supportsTopK)
             #expect(capabilities.supportedProviderOptions.supportsThinkingConfig)
             #expect(capabilities.supportedProviderOptions.supportsSafetySettings)
         }
 
         @Test
-        func `Reasoning models have forced temperature`() {
-            let model = LanguageModel.openai(.o4Mini)
-            let capabilities = ModelCapabilityRegistry.shared.capabilities(for: model)
-
-            #expect(!capabilities.supportsTemperature)
-            #expect(!capabilities.supportsTopP)
-            #expect(capabilities.forcedTemperature == 1.0)
-            #expect(capabilities.excludedParameters.contains("temperature"))
-            #expect(capabilities.excludedParameters.contains("topP"))
-            #expect(capabilities.supportedProviderOptions.supportsReasoningEffort)
-            #expect(capabilities.supportedProviderOptions.supportsPreviousResponseId)
-        }
-
-        @Test
-        func `GPT-4 models support standard parameters`() {
+        func `Custom OpenAI models support standard parameters`() {
             let models: [LanguageModel] = [
-                .openai(.gpt4o),
-                .openai(.gpt4oMini),
-                .openai(.gpt41),
-                .openai(.gpt4Turbo),
+                .openai(.custom("custom-openai")),
             ]
 
             for model in models {
@@ -65,17 +58,16 @@ enum ModelCapabilitiesTests {
                 #expect(capabilities.supportsMaxTokens)
                 #expect(capabilities.supportsFrequencyPenalty)
                 #expect(capabilities.supportsPresencePenalty)
-                #expect(capabilities.supportedProviderOptions.supportsParallelToolCalls)
-                #expect(capabilities.supportedProviderOptions.supportsResponseFormat)
-                #expect(capabilities.supportedProviderOptions.supportsLogprobs)
             }
         }
 
         @Test
         func `Claude models support thinking`() {
             let models: [LanguageModel] = [
+                .anthropic(.fable5),
+                .anthropic(.opus47),
                 .anthropic(.opus4),
-                .anthropic(.sonnet4),
+                .anthropic(.sonnet46),
                 .anthropic(.sonnet45),
                 .anthropic(.haiku45),
             ]
@@ -88,9 +80,37 @@ enum ModelCapabilitiesTests {
             }
         }
 
+        @Test
+        func `Claude Fable 5 and Opus 4_7 plus 4_8 advertise adaptive thinking without sampling options`() {
+            for model in [LanguageModel.anthropic(.fable5), .anthropic(.opus47), .anthropic(.opus48)] {
+                let capabilities = ModelCapabilityRegistry.shared.capabilities(for: model)
+
+                #expect(!capabilities.supportsTemperature)
+                #expect(!capabilities.supportsTopP)
+                #expect(!capabilities.supportsTopK)
+                #expect(capabilities.excludedParameters.contains("temperature"))
+                #expect(capabilities.excludedParameters.contains("topP"))
+                #expect(capabilities.excludedParameters.contains("topK"))
+                #expect(capabilities.supportedProviderOptions.supportsThinking)
+                #expect(capabilities.supportedProviderOptions.supportsCacheControl)
+            }
+        }
+
+        @Test
+        func `Custom Anthropic models keep thinking options by default`() {
+            let capabilities = ModelCapabilityRegistry.shared
+                .capabilities(for: .anthropic(.custom("claude-opus-latest")))
+
+            #expect(capabilities.supportedProviderOptions.supportsThinking)
+            #expect(capabilities.supportedProviderOptions.supportsCacheControl)
+        }
+
         @Test
         func `Google models support topK and thinking`() {
             let models: [LanguageModel] = [
+                .google(.gemini35Flash),
+                .google(.gemini31ProPreview),
+                .google(.gemini31FlashLite),
                 .google(.gemini25Pro),
                 .google(.gemini25Flash),
                 .google(.gemini25FlashLite),
@@ -108,8 +128,9 @@ enum ModelCapabilitiesTests {
         @Test
         func `Mistral models support safe mode`() {
             let models: [LanguageModel] = [
-                .mistral(.large2),
-                .mistral(.codestral),
+                .mistral(.largeLatest),
+                .mistral(.medium35),
+                .mistral(.codestralLatest),
             ]
 
             for model in models {
@@ -122,8 +143,8 @@ enum ModelCapabilitiesTests {
         @Test
         func `Groq models support speed level`() {
             let models: [LanguageModel] = [
-                .groq(.llama3170b),
-                .groq(.llama370b),
+                .groq(.llama3370b),
+                .groq(.llama4Maverick),
             ]
 
             for model in models {
@@ -136,8 +157,8 @@ enum ModelCapabilitiesTests {
         @Test
         func `Grok models support fun mode`() {
             let models: [LanguageModel] = [
-                .grok(.grok4),
-                .grok(.grok3),
+                .grok(.grok43),
+                .grok(.grok420Reasoning),
             ]
 
             for model in models {
@@ -151,7 +172,7 @@ enum ModelCapabilitiesTests {
 
     struct SettingsValidationTests {
         @Test
-        func `Validate settings for GPT-5.1`() {
+        func `Validate settings for GPT-5.5`() {
             let settings = GenerationSettings(
                 maxTokens: 1000,
                 temperature: 0.7,
@@ -166,7 +187,7 @@ enum ModelCapabilitiesTests {
                 ),
             )
 
-            let validated = settings.validated(for: .openai(.gpt51))
+            let validated = settings.validated(for: .openai(.gpt55))
 
             #expect(validated.maxTokens == 1000)
             #expect(validated.temperature == nil) // Excluded
@@ -178,7 +199,20 @@ enum ModelCapabilitiesTests {
         }
 
         @Test
-        func `Validate settings for O3 with forced temperature`() {
+        func `Validate settings preserves stream buffering mode`() {
+            let settings = GenerationSettings(
+                temperature: 0.7,
+                streamBuffering: .untilTerminal,
+            )
+
+            let validated = settings.validated(for: .openai(.gpt55))
+
+            #expect(validated.temperature == nil)
+            #expect(validated.streamBuffering == .untilTerminal)
+        }
+
+        @Test
+        func `Validate settings for GPT-5 strips unsupported options`() {
             let settings = GenerationSettings(
                 temperature: 0.5,
                 topP: 0.8,
@@ -190,16 +224,16 @@ enum ModelCapabilitiesTests {
                 ),
             )
 
-            let validated = settings.validated(for: LanguageModel.openai(.o4Mini))
+            let validated = settings.validated(for: LanguageModel.openai(.gpt55))
 
-            #expect(validated.temperature == 1.0) // Forced to 1.0
+            #expect(validated.temperature == nil) // Excluded
             #expect(validated.topP == nil) // Excluded
-            #expect(validated.providerOptions.openai?.reasoningEffort == .high) // Kept
-            #expect(validated.providerOptions.openai?.verbosity == nil) // Removed
+            #expect(validated.providerOptions.openai?.reasoningEffort == nil) // Removed
+            #expect(validated.providerOptions.openai?.verbosity == .medium) // Kept
         }
 
         @Test
-        func `Validate settings for GPT-4`() {
+        func `Validate settings for custom OpenAI model`() {
             let settings = GenerationSettings(
                 maxTokens: 2000,
                 temperature: 0.8,
@@ -216,17 +250,13 @@ enum ModelCapabilitiesTests {
                 ),
             )
 
-            let validated = settings.validated(for: .openai(.gpt4o))
+            let validated = settings.validated(for: .openai(.custom("custom-openai")))
 
             #expect(validated.maxTokens == 2000)
             #expect(validated.temperature == 0.8)
             #expect(validated.topP == 0.95)
             #expect(validated.frequencyPenalty == 0.2)
             #expect(validated.presencePenalty == 0.1)
-            #expect(validated.providerOptions.openai?.parallelToolCalls == true)
-            #expect(validated.providerOptions.openai?.responseFormat == .json)
-            #expect(validated.providerOptions.openai?.logprobs == true)
-            #expect(validated.providerOptions.openai?.topLogprobs == 3)
         }
 
         @Test
@@ -252,6 +282,72 @@ enum ModelCapabilitiesTests {
             // OpenAI options remain unfiltered (they won't be used by Anthropic provider)
             #expect(validated.providerOptions.openai?.verbosity == .high)
         }
+
+        @Test
+        func `Validate Anthropic options keeps adaptive thinking for Opus 4_8`() {
+            let settings = GenerationSettings(
+                temperature: 0.7,
+                topP: 0.9,
+                topK: 40,
+                providerOptions: .init(
+                    anthropic: .init(
+                        thinking: .enabled(budgetTokens: 3000),
+                        cacheControl: .persistent,
+                    ),
+                ),
+            )
+
+            let validated = settings.validated(for: LanguageModel.anthropic(.opus48))
+
+            #expect(validated.temperature == nil)
+            #expect(validated.topP == nil)
+            #expect(validated.topK == nil)
+            #expect(validated.providerOptions.anthropic?.thinking != nil)
+            #expect(validated.providerOptions.anthropic?.cacheControl == .persistent)
+        }
+
+        @Test
+        func `Validate Anthropic-compatible Fable strips unsupported sampling`() {
+            let settings = GenerationSettings(
+                temperature: 0.7,
+                topP: 0.9,
+                topK: 40,
+                providerOptions: .init(
+                    anthropic: .init(thinking: .adaptive),
+                ),
+            )
+
+            let validated = settings.validated(for: LanguageModel.anthropicCompatible(
+                modelId: "claude-fable-5",
+                baseURL: "https://example.test",
+            ))
+
+            #expect(validated.temperature == nil)
+            #expect(validated.topP == nil)
+            #expect(validated.topK == nil)
+            #expect(validated.providerOptions.anthropic?.thinking != nil)
+        }
+
+        @Test
+        func `Validate direct custom Fable strips unsupported sampling`() {
+            let settings = GenerationSettings(
+                temperature: 0.7,
+                topP: 0.9,
+                topK: 40,
+                providerOptions: .init(
+                    anthropic: .init(thinking: .adaptive),
+                ),
+            )
+
+            let validated = settings.validated(
+                for: LanguageModel.anthropic(.custom("anthropic.claude-fable-5")),
+            )
+
+            #expect(validated.temperature == nil)
+            #expect(validated.topP == nil)
+            #expect(validated.topK == nil)
+            #expect(validated.providerOptions.anthropic?.thinking != nil)
+        }
     }
 
     struct CustomModelTests {
@@ -307,8 +403,8 @@ enum ModelCapabilitiesTests {
         @Test
         func `Concurrent capability access`() async {
             let models: [LanguageModel] = [
-                .openai(.gpt51),
-                .openai(.gpt4o),
+                .openai(.gpt54),
+                .openai(.gpt55),
                 .anthropic(.opus4),
                 .google(.gemini25Flash),
             ]

Tests/TachikomaTests/Core/ModelParsingTests.swift

@@ -9,21 +9,53 @@ struct ModelParsingTests {
     }
 
     @Test
-    func `parse GPT-5.1 base model`() {
-        let parsed = LanguageModel.parse(from: "gpt-5.1")
-        #expect(parsed == .openai(.gpt51))
+    func `parse GPT-5.5 base model`() {
+        let parsed = LanguageModel.parse(from: "gpt-5.5")
+        #expect(parsed == .openai(.gpt55))
     }
 
     @Test
-    func `parse GPT-5.2 base model`() {
-        let parsed = LanguageModel.parse(from: "gpt-5.2")
-        #expect(parsed == .openai(.gpt52))
+    func `parse chat latest OpenAI alias`() throws {
+        #expect(LanguageModel.parse(from: "chat-latest") == .openai(.chatLatest))
+        #expect(LanguageModel.parse(from: "gpt-5-chat-latest") == .openai(.gpt5ChatLatest))
+        #expect(LanguageModel.parse(from: "openai/chat-latest") == .openai(.chatLatest))
+        #expect(LanguageModel.parse(from: "openai/gpt-5-chat-latest") == .openai(.gpt5ChatLatest))
+        #expect(try ModelSelector.parseModel("openai/chat-latest") == .openai(.chatLatest))
+        #expect(try ModelSelector.parseModel("openai/gpt-5-chat-latest") == .openai(.gpt5ChatLatest))
     }
 
     @Test
-    func `parse GPT-5.1 nano alias`() {
-        let parsed = LanguageModel.parse(from: "gpt51-nano")
-        #expect(parsed == .openai(.gpt5Nano))
+    func `parse GPT-5.4 base model`() {
+        let parsed = LanguageModel.parse(from: "gpt-5.4")
+        #expect(parsed == .openai(.gpt54))
+    }
+
+    @Test
+    func `parse GPT-5.4 nano alias`() {
+        let parsed = LanguageModel.parse(from: "gpt54-nano")
+        #expect(parsed == .openai(.gpt54Nano))
+    }
+
+    @Test
+    func `LanguageModel rejects retired OpenAI ids`() {
+        for model in ["gpt-4o", "gpt-4.1", "gpt-5.1", "gpt-5.2", "gpt-5-thinking"] {
+            #expect(LanguageModel.parse(from: model) == nil)
+        }
+    }
+
+    @Test
+    func `parse Claude Fable 5 model id`() throws {
+        #expect(LanguageModel.parse(from: "claude-fable-5") == .anthropic(.fable5))
+        #expect(LanguageModel.parse(from: "fable") == .anthropic(.fable5))
+        #expect(try ModelSelector.parseModel("fable5") == .anthropic(.fable5))
+        #expect(LanguageModel.parse(from: "my-fable5-7b") == nil)
+    }
+
+    @Test
+    func `parse Claude Opus 4.8 model id`() {
+        let parsed = LanguageModel.parse(from: "claude-opus-4-8")
+        #expect(parsed == .anthropic(.opus48))
+        #expect(LanguageModel.parse(from: "my-opus48-distill") == nil)
     }
 
     @Test
@@ -33,20 +65,172 @@ struct ModelParsingTests {
     }
 
     @Test
-    func `parse shorthand Claude alias`() {
+    func `parse shorthand Claude alias`() throws {
         let parsed = LanguageModel.parse(from: "claude")
-        #expect(parsed == .anthropic(.sonnet45))
+        #expect(parsed == .anthropic(.opus48))
+        #expect(try ModelSelector.parseModel("anthropic") == .anthropic(.opus48))
     }
 
     @Test
-    func `parse Gemini 3 Flash model id`() {
-        let parsed = LanguageModel.parse(from: "gemini-3-flash")
-        #expect(parsed == .google(.gemini3Flash))
+    func `parse Gemini 3.5 Flash model id`() {
+        let parsed = LanguageModel.parse(from: "gemini-3.5-flash")
+        #expect(parsed == .google(.gemini35Flash))
     }
 
     @Test
     func `parse shorthand Gemini alias`() {
         let parsed = LanguageModel.parse(from: "gemini")
-        #expect(parsed == .google(.gemini3Flash))
+        #expect(parsed == .google(.gemini35Flash))
+    }
+
+    @Test
+    func `parse provider qualified latest hosted models`() throws {
+        #expect(LanguageModel.parse(from: "anthropic/claude-fable-5") == .anthropic(.fable5))
+        #expect(LanguageModel.parse(from: "anthropic/claude-opus-4-8") == .anthropic(.opus48))
+        #expect(LanguageModel.parse(from: "google/gemini-3.5-flash") == .google(.gemini35Flash))
+        #expect(LanguageModel.parse(from: "xai/grok-4.3-latest") == .grok(.grok43))
+        #expect(LanguageModel.parse(from: "grok-4-latest") == .grok(.grok43))
+        #expect(LanguageModel.parse(from: "grok-4") == .grok(.grok43))
+        #expect(LanguageModel.parse(from: "xai/grok-code-fast-1") == .grok(.custom("grok-code-fast-1")))
+        #expect(try ModelSelector.parseModel("grok-4") == .grok(.grok43))
+    }
+
+    @Test
+    func `parse rejects provider-qualified hosted model mismatches`() {
+        #expect(LanguageModel.parse(from: "openai/claude") == nil)
+        #expect(LanguageModel.parse(from: "google/claude") == nil)
+        #expect(LanguageModel.parse(from: "xai/gemini-3.5-flash") == nil)
+        #expect(LanguageModel.parse(from: "anthropic/gpt-5.5") == nil)
+    }
+
+    @Test
+    func `ModelSelector keeps generic slash IDs as OpenRouter models`() throws {
+        #expect(try ModelSelector
+            .parseModel("anthropic/claude-opus-4-8") == .openRouter(modelId: "anthropic/claude-opus-4-8"))
+        #expect(try ModelSelector
+            .parseModel("google/gemini-3.5-flash") == .openRouter(modelId: "google/gemini-3.5-flash"))
+        #expect(try ModelSelector.parseModel("xai/grok-4.3-latest") == .grok(.grok43))
+        #expect(try ModelSelector.parseModel("openai/claude") == .openRouter(modelId: "openai/claude"))
+    }
+
+    @Test
+    func `parse MiniMax model ids`() throws {
+        #expect(LanguageModel.parse(from: "MiniMax-M2.7") == .minimax(.m27))
+        #expect(LanguageModel.parse(from: "minimax/m2.7") == .minimax(.m27))
+        #expect(try ModelSelector.parseModel("minimax/m2-7") == .minimax(.m27))
+        #expect(LanguageModel.parse(from: "minimax/MiniMax-M2.7-highspeed") == .minimax(.m27Highspeed))
+        #expect(LanguageModel.parse(from: "minimax/m2.7-highspeed") == .minimax(.m27Highspeed))
+        #expect(try ModelSelector.parseModel("minimax/m2-7-highspeed") == .minimax(.m27Highspeed))
+        #expect(LanguageModel.parse(from: "minimax") == .minimax(.m27))
+        #expect(LanguageModel.parse(from: "minimax-cn/MiniMax-M2.7") == .minimaxCN(.m27))
+        #expect(LanguageModel.parse(from: "minimax-cn/m2.7-highspeed") == .minimaxCN(.m27Highspeed))
+        #expect(try ModelSelector.parseModel("minimax-cn/m2-7") == .minimaxCN(.m27))
+        #expect(try ModelSelector.parseModel("minimax_cn/m2.7") == .minimaxCN(.m27))
+        #expect(LanguageModel.parse(from: "minimaxi/m2.7") == .minimaxCN(.m27))
+        #expect(LanguageModel.parse(from: "minimax-cn") == .minimaxCN(.m27))
+    }
+
+    @Test
+    func `parse OpenRouter model ids`() throws {
+        #expect(LanguageModel
+            .parse(from: "openrouter/xiaomi/mimo-v2.5-pro") == .openRouter(modelId: "xiaomi/mimo-v2.5-pro"))
+        #expect(LanguageModel.parse(from: "xiaomi/mimo-v2.5-pro") == .openRouter(modelId: "xiaomi/mimo-v2.5-pro"))
+        #expect(try ModelSelector.parseModel("xiaomi/mimo-v2.5-pro") == .openRouter(modelId: "xiaomi/mimo-v2.5-pro"))
+    }
+
+    @Test
+    func `parse custom Ollama Qwen vision model without falling back to Llama`() {
+        let parsed = LanguageModel.parse(from: "qwen2.5vl:3b")
+        #expect(parsed == .ollama(.custom("qwen2.5vl:3b")))
+        #expect(parsed?.modelId == "qwen2.5vl:3b")
+        #expect(parsed?.supportsVision == true)
+        #expect(parsed?.supportsTools == false)
+    }
+
+    @Test
+    func `parse provider-qualified custom Ollama model`() {
+        let parsed = LanguageModel.parse(from: "ollama/qwen2.5vl:3b")
+        #expect(parsed == .ollama(.custom("qwen2.5vl:3b")))
+        #expect(parsed?.modelId == "qwen2.5vl:3b")
+    }
+
+    @Test
+    func `parse local provider shortcuts`() {
+        #expect(LanguageModel.parse(from: "ollama") == .ollama(.llama33))
+        #expect(LanguageModel.parse(from: "lmstudio") == .lmstudio(.gptOSS120B))
+        #expect(LanguageModel.parse(from: "lmstudio/openai/gpt-oss-120b") == .lmstudio(.gptOSS120B))
+        #expect(LanguageModel.parse(from: "lmstudio/custom-local-model") == .lmstudio(.custom("custom-local-model")))
+    }
+
+    @Test
+    func `ModelSelector parses local provider selections`() throws {
+        if #available(macOS 13.0, iOS 16.0, tvOS 16.0, watchOS 9.0, *) {
+            #expect(try ModelSelector.parseModel("lmstudio") == .lmstudio(.gptOSS120B))
+            #expect(try ModelSelector.parseModel("lmstudio/openai/gpt-oss-120b") == .lmstudio(.gptOSS120B))
+            #expect(try ModelSelector.parseModel("lm-studio/custom-local") == .lmstudio(.custom("custom-local")))
+            #expect(ModelSelector.availableModels(for: "lmstudio").contains("openai/gpt-oss-120b"))
+        }
+    }
+
+    @Test
+    func `ProviderParser keeps configured Google model behavior`() {
+        if #available(macOS 13.0, iOS 16.0, tvOS 16.0, watchOS 9.0, *) {
+            let model = ProviderParser.determineDefaultModel(
+                from: "google/gemini-3.1-pro-preview",
+                hasOpenAI: false,
+                hasAnthropic: false,
+            )
+
+            #expect(model == .google(.gemini31ProPreview))
+        }
+    }
+
+    @Test
+    func `ProviderParser keeps keyless fallback local by default`() {
+        if #available(macOS 13.0, iOS 16.0, tvOS 16.0, watchOS 9.0, *) {
+            let model = ProviderParser.determineDefaultModel(
+                from: "",
+                hasOpenAI: false,
+                hasAnthropic: false,
+            )
+
+            #expect(model == .ollama(.llama33))
+        }
+    }
+
+    @Test
+    func `ProviderParser accepts MiniMax China provider aliases`() {
+        if #available(macOS 13.0, iOS 16.0, tvOS 16.0, watchOS 9.0, *) {
+            for provider in ["minimax-cn", "minimax_cn", "minimaxi"] {
+                let model = ProviderParser.determineDefaultModel(
+                    from: "\(provider)/m2.7",
+                    hasOpenAI: false,
+                    hasAnthropic: false,
+                    hasMiniMax: true,
+                )
+
+                #expect(model == .minimaxCN(.m27))
+            }
+        }
+    }
+
+    @Test
+    func `ModelSelector rejects legacy OpenAI before Ollama fallback`() throws {
+        if #available(macOS 13.0, iOS 16.0, tvOS 16.0, watchOS 9.0, *) {
+            for model in ["gpt-4o", "gpt-4.1", "gpt-3.5-turbo", "o4-mini", "o3-mini", "gpt-5.2"] {
+                #expect(throws: ModelValidationError.self) {
+                    _ = try ModelSelector.parseModel(model)
+                }
+            }
+        }
+    }
+
+    @Test
+    func `ModelSelector rejects Claude 3 before Ollama fallback`() throws {
+        if #available(macOS 13.0, iOS 16.0, tvOS 16.0, watchOS 9.0, *) {
+            #expect(throws: ModelValidationError.self) {
+                _ = try ModelSelector.parseModel("claude-3-sonnet")
+            }
+        }
     }
 }

Tests/TachikomaTests/Core/OpenAICompatibleHelperTests.swift

@@ -121,6 +121,337 @@ struct OpenAICompatibleHelperTests {
         #expect(deltas == "Hello world")
     }
 
+    @Test
+    func `streamText maps content filter finish reasons`() async throws {
+        let request = ProviderRequest(
+            messages: [ModelMessage(role: .user, content: [.text("blocked")])],
+        )
+
+        let deltas = try await withMockedSession { urlRequest in
+            let sse = """
+            data: {\"id\":\"chunk_1\",\"choices\":[{\"delta\":{\"content\":\"partial\"},\"index\":0,\"finish_reason\":null}]}
+
+            data: {\"id\":\"chunk_2\",\"choices\":[{\"delta\":{},\"index\":0,\"finish_reason\":\"content_filter\"}]}
+
+            data: [DONE]
+
+            """.utf8Data()
+            let response = HTTPURLResponse(
+                url: urlRequest.url!,
+                statusCode: 200,
+                httpVersion: nil,
+                headerFields: ["Content-Type": "text/event-stream"],
+            )!
+            return (response, sse)
+        } operation: { session in
+            let stream = try await OpenAICompatibleHelper.streamText(
+                request: request,
+                modelId: "compatible-model",
+                baseURL: "https://mock.compatible",
+                apiKey: "sk-test",
+                providerName: "TestProvider",
+                session: session,
+            )
+
+            var deltas: [TextStreamDelta] = []
+            for try await delta in stream {
+                deltas.append(delta)
+            }
+            return deltas
+        }
+
+        #expect(deltas.contains { $0.type == .textDelta && $0.content == "partial" })
+        #expect(deltas.contains { $0.type == .done && $0.finishReason == .contentFilter })
+    }
+
+    @Test
+    func `OpenAI-compatible provider forwards configured headers`() async throws {
+        let request = ProviderRequest(
+            messages: [ModelMessage(role: .user, content: [.text("ping")])],
+        )
+
+        try await self.withMockedSession { urlRequest in
+            #expect(urlRequest.value(forHTTPHeaderField: "client_id") == "proxy-client")
+            #expect(urlRequest.value(forHTTPHeaderField: "client_secret") == "proxy-secret")
+            return self.jsonResponse(for: urlRequest, data: Self.chatCompletionPayload(text: "pong"))
+        } operation: { session in
+            let configuration = TachikomaConfiguration(apiKeys: ["openai_compatible": "sk-test"])
+            let provider = try OpenAICompatibleProvider(
+                modelId: "compatible-model",
+                baseURL: "https://mock.compatible",
+                configuration: configuration,
+                additionalHeaders: [
+                    "client_id": "proxy-client",
+                    "client_secret": "proxy-secret",
+                ],
+                session: session,
+            )
+
+            let response = try await provider.generateText(request: request)
+            #expect(response.text == "pong")
+        }
+    }
+
+    @Test
+    func `generateText decodes OpenRouter reasoning details`() async throws {
+        let response = try await withMockedSession { urlRequest in
+            let reasoningDetails: [[String: String]] = [["type": "reasoning.encrypted", "data": "sealed"]]
+            let toolCall: [String: Any] = [
+                "id": "call-1",
+                "type": "function",
+                "function": ["name": "lookup", "arguments": "{}"],
+            ]
+            let toolCalls = [toolCall]
+            let choice: [String: Any] = [
+                "index": 0,
+                "message": [
+                    "role": "assistant",
+                    "content": NSNull(),
+                    "reasoning_details": reasoningDetails,
+                    "tool_calls": toolCalls,
+                ],
+                "finish_reason": "tool_calls",
+            ]
+            let payload: [String: Any] = [
+                "id": "chatcmpl-test",
+                "object": "chat.completion",
+                "created": 1_700_000_000,
+                "model": "anthropic/claude-fable-5",
+                "choices": [choice],
+            ]
+            return try self.jsonResponse(for: urlRequest, data: JSONSerialization.data(withJSONObject: payload))
+        } operation: { session in
+            try await OpenAICompatibleHelper.generateText(
+                request: ProviderRequest(messages: [.user("hi")]),
+                modelId: "anthropic/claude-fable-5",
+                baseURL: "https://mock.compatible",
+                apiKey: "sk-test",
+                providerName: "OpenRouter",
+                session: session,
+            )
+        }
+
+        let reasoning = try #require(response.reasoning.first)
+        #expect(reasoning.type == "openrouter_reasoning_details")
+        #expect(reasoning.rawJSON?.contains("reasoning.encrypted") == true)
+        #expect(response.toolCalls?.first?.id == "call-1")
+    }
+
+    @Test
+    func `generateText strips unsupported Fable sampling for OpenRouter route`() async throws {
+        let capture = CapturedRequest()
+        let request = ProviderRequest(
+            messages: [ModelMessage(role: .user, content: [.text("ping")])],
+            settings: GenerationSettings(maxTokens: 128, temperature: 0.7),
+        )
+
+        _ = try await self.withMockedSession { urlRequest in
+            capture.body = self.bodyData(from: urlRequest)
+            return self.jsonResponse(for: urlRequest, data: Self.chatCompletionPayload(text: "pong"))
+        } operation: { session in
+            try await OpenAICompatibleHelper.generateText(
+                request: request,
+                modelId: "anthropic/claude-fable-5",
+                baseURL: "https://mock.compatible",
+                apiKey: "sk-test",
+                providerName: "OpenRouter",
+                session: session,
+            )
+        }
+
+        let bodyJSON = try #require(capture.body).jsonObject()
+        #expect(bodyJSON["temperature"] == nil)
+        #expect(bodyJSON["max_tokens"] as? Int == 128)
+    }
+
+    @Test
+    func `generateText replays OpenRouter reasoning details on assistant tool messages`() async throws {
+        let capture = CapturedRequest()
+        let rawReasoning = #"[{"type":"reasoning.encrypted","data":"sealed"}]"#
+        let call = AgentToolCall(id: "call-1", name: "lookup", arguments: [:])
+        let request = try ProviderRequest(messages: [
+            .user("hi"),
+            ModelMessage(
+                role: .assistant,
+                content: [.text("")],
+                channel: .thinking,
+                metadata: .init(customData: [
+                    "openrouter.reasoning_details": rawReasoning,
+                    "tachikoma.reasoning.provider": "openrouter",
+                    "tachikoma.reasoning.model": "anthropic/claude-fable-5",
+                    "tachikoma.reasoning.base_url": #require(ReasoningEndpointIdentity
+                        .canonical("https://mock.compatible")),
+                ]),
+            ),
+            ModelMessage(role: .assistant, content: [.toolCall(call)]),
+            ModelMessage(
+                role: .tool,
+                content: [.toolResult(.success(toolCallId: "call-1", result: AnyAgentToolValue(string: "ok")))],
+            ),
+        ])
+
+        _ = try await self.withMockedSession { urlRequest in
+            capture.body = self.bodyData(from: urlRequest)
+            return self.jsonResponse(for: urlRequest, data: Self.chatCompletionPayload(text: "done"))
+        } operation: { session in
+            try await OpenAICompatibleHelper.generateText(
+                request: request,
+                modelId: "anthropic/claude-fable-5",
+                baseURL: "https://mock.compatible",
+                apiKey: "sk-test",
+                providerName: "OpenRouter",
+                session: session,
+            )
+        }
+
+        let bodyJSON = try #require(capture.body).jsonObject()
+        let messages = try #require(bodyJSON["messages"] as? [[String: Any]])
+        let assistant = try #require(messages.first { $0["role"] as? String == "assistant" })
+        let details = try #require(assistant["reasoning_details"] as? [[String: Any]])
+        #expect(details.first?["type"] as? String == "reasoning.encrypted")
+        #expect(details.first?["data"] as? String == "sealed")
+        #expect(assistant["tool_calls"] != nil)
+    }
+
+    @Test
+    func `generateText replays OpenRouter reasoning details on reasoning-only assistant boundary`() async throws {
+        let capture = CapturedRequest()
+        let rawReasoning = #"[{"type":"reasoning.encrypted","data":"sealed"}]"#
+        let request = try ProviderRequest(messages: [
+            .user("first"),
+            ModelMessage(
+                role: .assistant,
+                content: [.text("")],
+                channel: .thinking,
+                metadata: .init(customData: [
+                    "openrouter.reasoning_details": rawReasoning,
+                    "tachikoma.reasoning.provider": "openrouter",
+                    "tachikoma.reasoning.model": "anthropic/claude-fable-5",
+                    "tachikoma.reasoning.base_url": #require(ReasoningEndpointIdentity
+                        .canonical("https://mock.compatible")),
+                ]),
+            ),
+            ModelMessage(
+                role: .assistant,
+                content: [.text("")],
+                metadata: .init(customData: ["tachikoma.internal.boundary": "reasoning_only"]),
+            ),
+            .user("next"),
+        ])
+
+        _ = try await self.withMockedSession { urlRequest in
+            capture.body = self.bodyData(from: urlRequest)
+            return self.jsonResponse(for: urlRequest, data: Self.chatCompletionPayload(text: "done"))
+        } operation: { session in
+            try await OpenAICompatibleHelper.generateText(
+                request: request,
+                modelId: "anthropic/claude-fable-5",
+                baseURL: "https://mock.compatible",
+                apiKey: "sk-test",
+                providerName: "OpenRouter",
+                session: session,
+            )
+        }
+
+        let bodyJSON = try #require(capture.body).jsonObject()
+        let messages = try #require(bodyJSON["messages"] as? [[String: Any]])
+        let assistantIndex = try #require(messages.firstIndex { $0["role"] as? String == "assistant" })
+        let assistant = messages[assistantIndex]
+        let details = try #require(assistant["reasoning_details"] as? [[String: Any]])
+        #expect(details.first?["data"] as? String == "sealed")
+        let nextMessage = try #require(messages.indices
+            .contains(assistantIndex + 1) ? messages[assistantIndex + 1] : nil)
+        #expect(nextMessage["role"] as? String == "user")
+    }
+
+    @Test
+    func `generateText does not replay OpenRouter reasoning from another endpoint`() async throws {
+        let capture = CapturedRequest()
+        let rawReasoning = #"[{"type":"reasoning.encrypted","data":"sealed"}]"#
+        let call = AgentToolCall(id: "call-1", name: "lookup", arguments: [:])
+        let request = try ProviderRequest(messages: [
+            .user("hi"),
+            ModelMessage(
+                role: .assistant,
+                content: [.text("")],
+                channel: .thinking,
+                metadata: .init(customData: [
+                    "openrouter.reasoning_details": rawReasoning,
+                    "tachikoma.reasoning.provider": "openrouter",
+                    "tachikoma.reasoning.model": "anthropic/claude-fable-5",
+                    "tachikoma.reasoning.base_url": #require(ReasoningEndpointIdentity
+                        .canonical("https://other.example.test")),
+                ]),
+            ),
+            ModelMessage(role: .assistant, content: [.toolCall(call)]),
+            ModelMessage(
+                role: .tool,
+                content: [.toolResult(.success(toolCallId: "call-1", result: AnyAgentToolValue(string: "ok")))],
+            ),
+        ])
+
+        _ = try await self.withMockedSession { urlRequest in
+            capture.body = self.bodyData(from: urlRequest)
+            return self.jsonResponse(for: urlRequest, data: Self.chatCompletionPayload(text: "done"))
+        } operation: { session in
+            try await OpenAICompatibleHelper.generateText(
+                request: request,
+                modelId: "anthropic/claude-fable-5",
+                baseURL: "https://mock.compatible",
+                apiKey: "sk-test",
+                providerName: "OpenRouter",
+                session: session,
+            )
+        }
+
+        let bodyJSON = try #require(capture.body).jsonObject()
+        let messages = try #require(bodyJSON["messages"] as? [[String: Any]])
+        let assistantMessages = messages.filter { $0["role"] as? String == "assistant" }
+        #expect(assistantMessages.allSatisfy { $0["reasoning_details"] == nil })
+    }
+
+    @Test
+    func `generateText drops unmatched OpenRouter reasoning instead of serializing it as text`() async throws {
+        let capture = CapturedRequest()
+        let request = try ProviderRequest(messages: [
+            .user("hi"),
+            ModelMessage(
+                role: .assistant,
+                content: [.text("private reasoning")],
+                channel: .thinking,
+                metadata: .init(customData: [
+                    "openrouter.reasoning": "private reasoning",
+                    "tachikoma.reasoning.provider": "openrouter",
+                    "tachikoma.reasoning.model": "other-model",
+                    "tachikoma.reasoning.base_url": #require(ReasoningEndpointIdentity
+                        .canonical("https://mock.compatible")),
+                ]),
+            ),
+            .assistant("visible"),
+        ])
+
+        _ = try await self.withMockedSession { urlRequest in
+            capture.body = self.bodyData(from: urlRequest)
+            return self.jsonResponse(for: urlRequest, data: Self.chatCompletionPayload(text: "done"))
+        } operation: { session in
+            try await OpenAICompatibleHelper.generateText(
+                request: request,
+                modelId: "anthropic/claude-fable-5",
+                baseURL: "https://mock.compatible",
+                apiKey: "sk-test",
+                providerName: "OpenRouter",
+                session: session,
+            )
+        }
+
+        let bodyJSON = try #require(capture.body).jsonObject()
+        let messages = try #require(bodyJSON["messages"] as? [[String: Any]])
+        let assistantMessages = messages.filter { $0["role"] as? String == "assistant" }
+        #expect(assistantMessages.count == 1)
+        #expect(assistantMessages.first?["content"] as? String == "visible")
+        #expect(try String(data: #require(capture.body), encoding: .utf8)?.contains("private reasoning") == false)
+    }
+
     @Test
     func `non-200 responses surface TachikomaError.apiError`() async {
         await self.withMockedSession { urlRequest in

Tests/TachikomaTests/Core/ProviderOptionsTests.swift

@@ -207,6 +207,24 @@ enum ProviderOptionsTests {
             }
         }
 
+        @Test
+        func `Encode and decode Anthropic adaptive thinking mode`() throws {
+            let original = AnthropicOptions(
+                thinking: .adaptive,
+            )
+
+            let encoder = JSONEncoder()
+            let data = try encoder.encode(original)
+
+            let decoder = JSONDecoder()
+            let decoded = try decoder.decode(AnthropicOptions.self, from: data)
+
+            guard case .adaptive = decoded.thinking else {
+                Issue.record("Expected thinking to be adaptive")
+                return
+            }
+        }
+
         @Test
         func `Encode and decode provider options container`() throws {
             let original = ProviderOptions(

Tests/TachikomaTests/Core/ProviderTests.swift

@@ -50,6 +50,8 @@ enum ProviderTests {
             #expect(Provider.groq.identifier == "groq")
             #expect(Provider.mistral.identifier == "mistral")
             #expect(Provider.google.identifier == "google")
+            #expect(Provider.minimax.identifier == "minimax")
+            #expect(Provider.minimaxCN.identifier == "minimax-cn")
             #expect(Provider.ollama.identifier == "ollama")
             #expect(Provider.azureOpenAI.identifier == "azure-openai")
         }
@@ -68,6 +70,8 @@ enum ProviderTests {
             #expect(Provider.groq.displayName == "Groq")
             #expect(Provider.mistral.displayName == "Mistral")
             #expect(Provider.google.displayName == "Google")
+            #expect(Provider.minimax.displayName == "MiniMax")
+            #expect(Provider.minimaxCN.displayName == "MiniMax China")
             #expect(Provider.ollama.displayName == "Ollama")
             #expect(Provider.azureOpenAI.displayName == "Azure OpenAI")
             #expect(Provider.custom("test").displayName == "Test")
@@ -81,6 +85,8 @@ enum ProviderTests {
             #expect(Provider.groq.environmentVariable == "GROQ_API_KEY")
             #expect(Provider.mistral.environmentVariable == "MISTRAL_API_KEY")
             #expect(Provider.google.environmentVariable == "GEMINI_API_KEY")
+            #expect(Provider.minimax.environmentVariable == "MINIMAX_API_KEY")
+            #expect(Provider.minimaxCN.environmentVariable == "MINIMAX_CN_API_KEY")
             #expect(Provider.ollama.environmentVariable == "OLLAMA_API_KEY")
             #expect(Provider.azureOpenAI.environmentVariable == "AZURE_OPENAI_API_KEY")
             #expect(Provider.custom("test").environmentVariable.isEmpty)
@@ -90,6 +96,7 @@ enum ProviderTests {
         func `Alternative environment variables`() {
             #expect(Provider.grok.alternativeEnvironmentVariables == ["XAI_API_KEY", "GROK_API_KEY"])
             #expect(Provider.google.alternativeEnvironmentVariables == ["GOOGLE_API_KEY"])
+            #expect(Provider.minimaxCN.alternativeEnvironmentVariables == ["MINIMAX_API_KEY"])
             #expect(Provider.openai.alternativeEnvironmentVariables.isEmpty)
             #expect(Provider.anthropic.alternativeEnvironmentVariables.isEmpty)
             #expect(Provider.azureOpenAI.alternativeEnvironmentVariables == [
@@ -106,6 +113,8 @@ enum ProviderTests {
             #expect(Provider.groq.defaultBaseURL == "https://api.groq.com/openai/v1")
             #expect(Provider.mistral.defaultBaseURL == "https://api.mistral.ai/v1")
             #expect(Provider.google.defaultBaseURL == "https://generativelanguage.googleapis.com/v1beta")
+            #expect(Provider.minimax.defaultBaseURL == "https://api.minimax.io/anthropic")
+            #expect(Provider.minimaxCN.defaultBaseURL == "https://api.minimaxi.com/anthropic")
             #expect(Provider.ollama.defaultBaseURL == "http://localhost:11434")
             #expect(Provider.azureOpenAI.defaultBaseURL == nil)
             #expect(Provider.custom("test").defaultBaseURL == nil)
@@ -119,6 +128,8 @@ enum ProviderTests {
             #expect(Provider.groq.requiresAPIKey == true)
             #expect(Provider.mistral.requiresAPIKey == true)
             #expect(Provider.google.requiresAPIKey == true)
+            #expect(Provider.minimax.requiresAPIKey == true)
+            #expect(Provider.minimaxCN.requiresAPIKey == true)
             #expect(Provider.ollama.requiresAPIKey == false) // Ollama typically doesn't require API key
             #expect(Provider.azureOpenAI.requiresAPIKey == true)
             #expect(Provider.custom("test").requiresAPIKey == true) // Assume custom providers need keys
@@ -134,6 +145,9 @@ enum ProviderTests {
             #expect(Provider.from(identifier: "groq") == .groq)
             #expect(Provider.from(identifier: "mistral") == .mistral)
             #expect(Provider.from(identifier: "google") == .google)
+            #expect(Provider.from(identifier: "minimax") == .minimax)
+            #expect(Provider.from(identifier: "minimax-cn") == .minimaxCN)
+            #expect(Provider.from(identifier: "minimaxi") == .minimaxCN)
             #expect(Provider.from(identifier: "ollama") == .ollama)
             #expect(Provider.from(identifier: "azure-openai") == .azureOpenAI)
         }
@@ -172,6 +186,8 @@ enum ProviderTests {
                 .groq,
                 .mistral,
                 .google,
+                .minimax,
+                .minimaxCN,
                 .ollama,
                 .azureOpenAI,
             ]
@@ -194,6 +210,18 @@ enum ProviderTests {
             #expect(provider.alternativeEnvironmentVariables == ["XAI_API_KEY", "GROK_API_KEY"])
         }
 
+        @Test
+        func `MiniMax China falls back to MiniMax environment variable`() async {
+            let resolved = await withTemporaryEnvironment([
+                "MINIMAX_CN_API_KEY": nil,
+                "MINIMAX_API_KEY": "shared-minimax-key",
+            ]) {
+                Provider.minimaxCN.loadAPIKeyFromEnvironment()
+            }
+
+            #expect(resolved == "shared-minimax-key")
+        }
+
         @Test
         func `Custom providers don't have environment variables`() {
             let customProvider = Provider.custom("test")

Tests/TachikomaTests/Core/StopConditionsIntegrationTests.swift

@@ -327,7 +327,7 @@ private struct MockTextProvider: ModelProvider {
 
 private struct MockOpenAIProvider: ModelProvider {
     var modelId: String {
-        "gpt-4"
+        "gpt-5.5"
     }
 
     var baseURL: String? {

Tests/TachikomaTests/Core/StopConditionsTests.swift

@@ -180,6 +180,30 @@ struct StopConditionsTests {
         #expect(!collectedText.contains("ignored"))
     }
 
+    @Test
+    func `Stop conditions finish immediately after local match`() async throws {
+        let stream = AsyncThrowingStream<TextStreamDelta, Error> { continuation in
+            Task {
+                continuation.yield(TextStreamDelta(type: .textDelta, content: "STOP"))
+                try? await Task.sleep(nanoseconds: 2_000_000_000)
+                continuation.yield(TextStreamDelta(type: .textDelta, content: "late"))
+                continuation.yield(TextStreamDelta(type: .done, finishReason: .length))
+                continuation.finish()
+            }
+        }
+
+        let start = Date()
+        var received: [TextStreamDelta] = []
+        for try await delta in stream.stopWhen(StringStopCondition("STOP")) {
+            received.append(delta)
+        }
+
+        #expect(Date().timeIntervalSince(start) < 0.5)
+        #expect(received.map(\.content).compactMap(\.self) == ["STOP"])
+        #expect(received.last?.type == .done)
+        #expect(received.last?.finishReason == .stop)
+    }
+
     // MARK: - Builder Pattern Tests
 
     @Test

Tests/TachikomaTests/Core/StreamObjectTests.swift

@@ -89,12 +89,12 @@ struct StreamObjectTests {
 
         let result = StreamObjectResult(
             objectStream: testStream,
-            model: .openai(.gpt4o),
+            model: .openai(.gpt55),
             settings: .default,
             schema: TestPerson.self,
         )
 
-        #expect(result.model == .openai(.gpt4o))
+        #expect(result.model == .openai(.gpt55))
         #expect(result.schema == TestPerson.self)
     }
 
@@ -157,7 +157,7 @@ struct StreamObjectTests {
 
         let result = StreamObjectResult(
             objectStream: testStream,
-            model: .openai(.gpt4o),
+            model: .openai(.gpt55),
             settings: .default,
             schema: TestPerson.self,
         )
@@ -193,7 +193,7 @@ struct StreamObjectTests {
 
         let result = StreamObjectResult(
             objectStream: testStream,
-            model: .openai(.gpt4o),
+            model: .openai(.gpt55),
             settings: .default,
             schema: TestPerson.self,
         )
@@ -221,7 +221,7 @@ struct StreamObjectTests {
 
         let result = StreamObjectResult(
             objectStream: testStream,
-            model: .openai(.gpt4o),
+            model: .openai(.gpt55),
             settings: .default,
             schema: TestPerson.self,
         )

Tests/TachikomaTests/Core/StreamTransformTests.swift

@@ -231,7 +231,7 @@ struct StreamTransformTests {
 
         let result = StreamTextResult(
             stream: stream,
-            model: .openai(.gpt4o),
+            model: .openai(.gpt55),
             settings: .default,
         )
 
@@ -264,7 +264,7 @@ struct StreamTransformTests {
 
         let result = StreamTextResult(
             stream: stream,
-            model: .openai(.gpt4o),
+            model: .openai(.gpt55),
             settings: .default,
         )
 
@@ -291,7 +291,7 @@ struct StreamTransformTests {
 
         let result = StreamTextResult(
             stream: stream,
-            model: .openai(.gpt4o),
+            model: .openai(.gpt55),
             settings: .default,
         )
 

Tests/TachikomaTests/Core/TachikomaConfigurationTests.swift

@@ -367,35 +367,35 @@ struct TachikomaConfigurationTests {
                 Task {
                     // Uses default (.current)
                     _ = try await generateText(
-                        model: .openai(.gpt4o),
+                        model: .openai(.gpt55),
                         messages: [.user("Test")],
                     )
 
                     // Uses explicit
                     _ = try await generateText(
-                        model: .openai(.gpt4o),
+                        model: .openai(.gpt55),
                         messages: [.user("Test")],
                         configuration: explicitConfig,
                     )
 
                     // Stream functions
                     _ = try await streamText(
-                        model: .openai(.gpt4o),
+                        model: .openai(.gpt55),
                         messages: [.user("Test")],
                     )
 
                     _ = try await streamText(
-                        model: .openai(.gpt4o),
+                        model: .openai(.gpt55),
                         messages: [.user("Test")],
                         configuration: explicitConfig,
                     )
 
                     // Convenience functions
-                    _ = try await generate("Test", using: .openai(.gpt4o))
-                    _ = try await generate("Test", using: .openai(.gpt4o), configuration: explicitConfig)
+                    _ = try await generate("Test", using: .openai(.gpt55))
+                    _ = try await generate("Test", using: .openai(.gpt55), configuration: explicitConfig)
 
-                    _ = try await stream("Test", using: .openai(.gpt4o))
-                    _ = try await stream("Test", using: .openai(.gpt4o), configuration: explicitConfig)
+                    _ = try await stream("Test", using: .openai(.gpt55))
+                    _ = try await stream("Test", using: .openai(.gpt55), configuration: explicitConfig)
                 }
             }
 
@@ -559,9 +559,9 @@ struct TachikomaConfigurationTests {
                 return DummyProvider()
             }
 
-            let provider = try config.makeProvider(for: .openai(.gpt4o))
+            let provider = try config.makeProvider(for: .openai(.gpt55))
             #expect(provider is DummyProvider)
-            #expect(capturedModel == .openai(.gpt4o))
+            #expect(capturedModel == .openai(.gpt55))
         }
 
         @Test
@@ -569,8 +569,8 @@ struct TachikomaConfigurationTests {
             let config = TachikomaConfiguration(loadFromEnvironment: false)
             config.setAPIKey("mock-key", for: .openai)
 
-            let provider = try config.makeProvider(for: .openai(.gpt4o))
-            #expect(provider is OpenAIProvider)
+            let provider = try config.makeProvider(for: .openai(.gpt55))
+            #expect(provider is OpenAIResponsesProvider)
         }
     }
 }

Tests/TachikomaTests/GrokDebugTest.swift

@@ -31,7 +31,7 @@ func `Debug Grok streaming issue`() async throws {
     // Test with minimal setup
     let stream = try await stream(
         "Say hello",
-        using: .grok(.grok3),
+        using: .grok(.grok43),
     )
 
     var receivedContent = false

Tests/TachikomaTests/HarmonyFeatures/ResponseCacheTests.swift

@@ -59,6 +59,67 @@ struct ResponseCacheTests {
         #expect(cached?.finishReason == .stop)
     }
 
+    @Test
+    func `ResponseCache keys include reasoning metadata`() async {
+        let cache = ResponseCache()
+        let response = ProviderResponse(text: "cached", usage: nil, finishReason: .stop)
+
+        func request(signature: String) -> ProviderRequest {
+            ProviderRequest(
+                messages: [
+                    .user("Hello"),
+                    ModelMessage(
+                        role: .assistant,
+                        content: [.text("thinking")],
+                        channel: .thinking,
+                        metadata: .init(customData: [
+                            "anthropic.thinking.signature": signature,
+                            "anthropic.thinking.type": "thinking",
+                        ]),
+                    ),
+                    .assistant("Hi"),
+                ],
+                tools: nil,
+                settings: .default,
+            )
+        }
+
+        await cache.store(response, for: request(signature: "sig-a"))
+
+        #expect(await cache.get(for: request(signature: "sig-a"))?.text == "cached")
+        #expect(await cache.get(for: request(signature: "sig-b")) == nil)
+    }
+
+    @Test
+    func `CacheEntry size includes reasoning and assistant messages`() {
+        let small = CacheEntry(response: ProviderResponse(text: "ok"))
+        let largePayload = String(repeating: "x", count: 4096)
+        let large = CacheEntry(response: ProviderResponse(
+            text: "ok",
+            reasoning: [
+                ProviderReasoningBlock(text: largePayload, signature: largePayload, type: "thinking"),
+                ProviderReasoningBlock(
+                    text: "",
+                    type: "openrouter_reasoning_details",
+                    rawJSON: largePayload,
+                ),
+            ],
+            assistantMessages: [
+                ModelMessage(
+                    role: .assistant,
+                    content: [.text(largePayload)],
+                    channel: .thinking,
+                    metadata: .init(customData: [
+                        "anthropic.thinking.model": "claude-fable-5",
+                        "anthropic.thinking.signature": largePayload,
+                    ]),
+                ),
+            ],
+        ))
+
+        #expect(large.estimatedMemorySize() > small.estimatedMemorySize() + 12000)
+    }
+
     @Test
     func `ResponseCache cache miss`() async {
         let cache = ResponseCache()
@@ -267,6 +328,99 @@ struct ResponseCacheTests {
         #expect(key1.hash != key3.hash)
     }
 
+    @Test
+    func `CacheKey includes reasoning effort and Anthropic thinking options`() {
+        let messages = [ModelMessage.user("Hello")]
+        let lowEffort = ProviderRequest(
+            messages: messages,
+            settings: GenerationSettings(
+                reasoningEffort: .low,
+                providerOptions: .init(anthropic: .init(thinking: .adaptive)),
+            ),
+        )
+        let highEffort = ProviderRequest(
+            messages: messages,
+            settings: GenerationSettings(
+                reasoningEffort: .high,
+                providerOptions: .init(anthropic: .init(thinking: .adaptive)),
+            ),
+        )
+        let disabledThinking = ProviderRequest(
+            messages: messages,
+            settings: GenerationSettings(
+                reasoningEffort: .low,
+                providerOptions: .init(anthropic: .init(thinking: .disabled)),
+            ),
+        )
+
+        #expect(CacheKey(from: lowEffort).hash != CacheKey(from: highEffort).hash)
+        #expect(CacheKey(from: lowEffort).hash != CacheKey(from: disabledThinking).hash)
+    }
+
+    @Test
+    func `CacheKey includes string stop condition values`() {
+        let endRequest = ProviderRequest(
+            messages: [ModelMessage.user("Hello")],
+            settings: GenerationSettings(stopConditions: StringStopCondition("END")),
+        )
+        let stopRequest = ProviderRequest(
+            messages: [ModelMessage.user("Hello")],
+            settings: GenerationSettings(stopConditions: StringStopCondition("STOP")),
+        )
+
+        #expect(CacheKey(from: endRequest).hash != CacheKey(from: stopRequest).hash)
+    }
+
+    @Test
+    func `CacheKey encodes composite stop conditions without delimiter collisions`() async {
+        let cache = ResponseCache()
+        let splitRequest = ProviderRequest(
+            messages: [ModelMessage.user("Hello")],
+            settings: GenerationSettings(stopConditions: AnyStopCondition(
+                StringStopCondition("a"),
+                StringStopCondition("b"),
+            )),
+        )
+        let joinedRequest = ProviderRequest(
+            messages: [ModelMessage.user("Hello")],
+            settings: GenerationSettings(stopConditions: AnyStopCondition(
+                StringStopCondition("a,string:true:b"),
+            )),
+        )
+
+        #expect(CacheKey(from: splitRequest).hash != CacheKey(from: joinedRequest).hash)
+
+        await cache.store(ProviderResponse(text: "split", finishReason: .stop), for: splitRequest)
+        let joinedCached = await cache.get(for: joinedRequest)
+
+        #expect(joinedCached == nil)
+    }
+
+    @Test
+    func `CacheKey marks custom stop conditions uncacheable`() {
+        let request = ProviderRequest(
+            messages: [ModelMessage.user("Hello")],
+            settings: GenerationSettings(stopConditions: PredicateStopCondition { _, _ in false }),
+        )
+
+        let key = CacheKey(from: request)
+        #expect(key.isCacheable == false)
+    }
+
+    @Test
+    func `ResponseCache skips custom stop condition entries`() async {
+        let cache = ResponseCache()
+        let request = ProviderRequest(
+            messages: [ModelMessage.user("Hello")],
+            settings: GenerationSettings(stopConditions: PredicateStopCondition { _, _ in false }),
+        )
+
+        await cache.store(ProviderResponse(text: "cached", finishReason: .stop), for: request)
+        let cached = await cache.get(for: request)
+
+        #expect(cached == nil)
+    }
+
     @Test
     func `CacheKey includes tools in hash`() {
         let tool1 = AgentTool(
@@ -317,7 +471,7 @@ struct ResponseCacheTests {
 
         // Create a mock provider
         let mockProvider = ResponseCacheMockProvider(
-            model: .openai(.gpt4o),
+            model: .openai(.gpt55),
             response: ProviderResponse(text: "Cached response", usage: nil, finishReason: .stop),
         )
 
@@ -336,7 +490,7 @@ struct ResponseCacheTests {
         // Use a simple counter that can be modified in the closure
         let callCount = Box(value: 0)
         var mockProvider = ResponseCacheMockProvider(
-            model: .openai(.gpt4o),
+            model: .openai(.gpt55),
             response: ProviderResponse(text: "Response", usage: nil, finishReason: .stop),
         )
         mockProvider.onGenerateText = { _ in
@@ -362,13 +516,45 @@ struct ResponseCacheTests {
         #expect(callCount.value == 1) // Provider not called again
     }
 
+    @Test
+    func `CachedProvider keys include provider endpoint identity`() async throws {
+        let cache = ResponseCache()
+        let callCountA = Box(value: 0)
+        let callCountB = Box(value: 0)
+        var providerA = ResponseCacheMockProvider(
+            model: .openaiCompatible(modelId: "shared-model", baseURL: "https://gateway.test/v1?tenant=a"),
+            response: ProviderResponse(text: "tenant-a", usage: nil, finishReason: .stop),
+            mockModelId: "shared-model",
+            mockBaseURL: "https://gateway.test/v1?tenant=a",
+        )
+        var providerB = ResponseCacheMockProvider(
+            model: .openaiCompatible(modelId: "shared-model", baseURL: "https://gateway.test/v1?tenant=b"),
+            response: ProviderResponse(text: "tenant-b", usage: nil, finishReason: .stop),
+            mockModelId: "shared-model",
+            mockBaseURL: "https://gateway.test/v1?tenant=b",
+        )
+        providerA.onGenerateText = { _ in callCountA.value += 1 }
+        providerB.onGenerateText = { _ in callCountB.value += 1 }
+
+        let cachedA = await cache.wrapProvider(providerA)
+        let cachedB = await cache.wrapProvider(providerB)
+        let request = ProviderRequest(messages: [ModelMessage.user("Test")], tools: nil, settings: .default)
+
+        #expect(try await cachedA.generateText(request: request).text == "tenant-a")
+        #expect(try await cachedB.generateText(request: request).text == "tenant-b")
+        #expect(try await cachedA.generateText(request: request).text == "tenant-a")
+        #expect(try await cachedB.generateText(request: request).text == "tenant-b")
+        #expect(callCountA.value == 1)
+        #expect(callCountB.value == 1)
+    }
+
     @Test
     func `CachedProvider doesn't cache streaming`() async throws {
         let cache = ResponseCache()
 
         let callCount = Box(value: 0)
         var mockProvider = ResponseCacheMockProvider(
-            model: .openai(.gpt4o),
+            model: .openai(.gpt55),
             response: ProviderResponse(text: "Test", usage: nil, finishReason: .stop),
         )
         mockProvider.onStreamText = { _ in
@@ -397,15 +583,17 @@ struct ResponseCacheTests {
 private struct ResponseCacheMockProvider: ModelProvider {
     let model: LanguageModel
     let response: ProviderResponse
+    let mockModelId: String
+    let mockBaseURL: String?
     var onGenerateText: (@Sendable (ProviderRequest) -> Void)?
     var onStreamText: (@Sendable (ProviderRequest) -> Void)?
 
     var modelId: String {
-        "mock-model"
+        self.mockModelId
     }
 
     var baseURL: String? {
-        nil
+        self.mockBaseURL
     }
 
     var apiKey: String? {
@@ -419,11 +607,15 @@ private struct ResponseCacheMockProvider: ModelProvider {
     init(
         model: LanguageModel,
         response: ProviderResponse,
+        mockModelId: String = "mock-model",
+        mockBaseURL: String? = nil,
         onGenerateText: (@Sendable (ProviderRequest) -> Void)? = nil,
         onStreamText: (@Sendable (ProviderRequest) -> Void)? = nil,
     ) {
         self.model = model
         self.response = response
+        self.mockModelId = mockModelId
+        self.mockBaseURL = mockBaseURL
         self.onGenerateText = onGenerateText
         self.onStreamText = onStreamText
     }

Tests/TachikomaTests/IntegrationTests.swift

@@ -29,7 +29,7 @@ struct IntegrationTests {
 
         let streamResult = StreamTextResult(
             stream: textStream,
-            model: .openai(.gpt4o),
+            model: .openai(.gpt55),
             settings: .default,
         )
 

Tests/TachikomaTests/LMStudioProviderTests.swift

@@ -7,7 +7,7 @@ struct LMStudioProviderTests {
     func `Provider initialization`() {
         let provider = LMStudioProvider(
             baseURL: "http://localhost:1234/v1",
-            modelId: "gpt-oss-120b",
+            modelId: "openai/gpt-oss-120b",
             apiKey: nil,
         )
 
@@ -18,7 +18,7 @@ struct LMStudioProviderTests {
         let capabilities = provider.capabilities
 
         #expect(baseURL == "http://localhost:1234/v1")
-        #expect(modelId == "gpt-oss-120b")
+        #expect(modelId == "openai/gpt-oss-120b")
         #expect(apiKey == nil)
         #expect(capabilities.supportsTools == true)
         #expect(capabilities.supportsStreaming == true)
@@ -28,11 +28,11 @@ struct LMStudioProviderTests {
     func `Model enum integration`() {
         let model1 = LanguageModel.lmstudio(.gptOSS120B)
         let model2 = LanguageModel.lmstudio(.gptOSS20B)
-        let model3 = LanguageModel.lmstudio(.current)
+        let model3 = LanguageModel.lmstudio(.llama3370B)
 
-        #expect(model1.modelId == "gpt-oss-120b")
-        #expect(model2.modelId == "gpt-oss-20b")
-        #expect(model3.modelId == "current")
+        #expect(model1.modelId == "openai/gpt-oss-120b")
+        #expect(model2.modelId == "openai/gpt-oss-20b")
+        #expect(model3.modelId == "meta/llama-3.3-70b")
 
         #expect(model1.supportsTools == true)
         #expect(model1.contextLength == 128_000)
@@ -47,7 +47,7 @@ struct LMStudioProviderTests {
         #expect(model2.providerName == "LMStudio")
 
         #expect(model1.modelId == "gpt-oss:120b")
-        #expect(model2.modelId == "gpt-oss-120b")
+        #expect(model2.modelId == "openai/gpt-oss-120b")
     }
 
     @Test
@@ -59,7 +59,7 @@ struct LMStudioProviderTests {
         let provider = try ProviderFactory.createProvider(for: model, configuration: config)
 
         let modelId = provider.modelId
-        #expect(modelId == "gpt-oss-120b")
+        #expect(modelId == "openai/gpt-oss-120b")
 
         // Should work without API key (local model)
         #expect(provider is LMStudioProvider)

Tests/TachikomaTests/Providers/AnthropicInterleavedDefaultsTests.swift

@@ -1,4 +1,7 @@
 import Foundation
+#if canImport(FoundationNetworking)
+import FoundationNetworking
+#endif
 import Testing
 @testable import Tachikoma
 
@@ -22,6 +25,19 @@ struct AnthropicInterleavedDefaultsTests {
         #expect(parts.contains("fine-grained-tool-streaming-2025-05-14"))
     }
 
+    @Test
+    func `Endpoint identity includes routing query without exposing raw values`() {
+        let tenantA = ReasoningEndpointIdentity.canonical("https://gateway.test/v1?tenant=a")
+        let tenantB = ReasoningEndpointIdentity.canonical("https://gateway.test/v1?tenant=b")
+
+        #expect(tenantA != tenantB)
+        #expect(tenantA?.hasPrefix("sha256:") == true)
+        #expect(tenantA?.contains("tenant") == false)
+        #expect(tenantA?.contains("gateway") == false)
+        #expect(ReasoningEndpointIdentity.canonical("https://gateway.test/v1/?tenant=a") == tenantA)
+        #expect(ReasoningEndpointIdentity.canonical("https://user:secret@gateway.test/v1?tenant=a#frag") == tenantA)
+    }
+
     @Test
     func `Provider request includes beta header and thinking payload`() throws {
         let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
@@ -57,6 +73,358 @@ struct AnthropicInterleavedDefaultsTests {
         #expect(thinking["budget_tokens"] as? Int == 12000)
     }
 
+    @Test
+    func `Opus 4_7 request strips unsupported sampling and uses adaptive thinking`() throws {
+        let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
+        let provider = try AnthropicProvider(model: .opus47, configuration: config)
+
+        let settings = GenerationSettings(
+            maxTokens: 64,
+            temperature: 0.7,
+            topP: 0.9,
+            topK: 40,
+            providerOptions: .init(anthropic: .init(thinking: .enabled(budgetTokens: 12000))),
+        )
+
+        let request = ProviderRequest(
+            messages: [.user("hi")],
+            settings: settings,
+        )
+
+        let urlRequest = try provider.makeURLRequest(for: request, stream: false)
+        let body = try #require(urlRequest.httpBody)
+        let json = try #require(try JSONSerialization.jsonObject(with: body) as? [String: Any])
+
+        #expect(json["model"] as? String == "claude-opus-4-7")
+        #expect(json["temperature"] == nil)
+        #expect(json["top_p"] == nil)
+        #expect(json["top_k"] == nil)
+        let thinking = try #require(json["thinking"] as? [String: Any])
+        #expect(thinking["type"] as? String == "adaptive")
+        #expect(thinking["budget_tokens"] == nil)
+        let outputConfig = try #require(json["output_config"] as? [String: Any])
+        #expect(outputConfig["effort"] as? String == "medium")
+        #expect(json["max_tokens"] as? Int == 64)
+    }
+
+    @Test
+    func `Opus 4_8 request strips unsupported sampling and uses adaptive thinking`() throws {
+        let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
+        let provider = try AnthropicProvider(model: .opus48, configuration: config)
+
+        let settings = GenerationSettings(
+            maxTokens: 64,
+            temperature: 0.7,
+            topP: 0.9,
+            topK: 40,
+            reasoningEffort: .low,
+            providerOptions: .init(anthropic: .init(thinking: .enabled(budgetTokens: 12000))),
+        )
+
+        let request = ProviderRequest(
+            messages: [.user("hi")],
+            settings: settings,
+        )
+
+        let urlRequest = try provider.makeURLRequest(for: request, stream: false)
+        let body = try #require(urlRequest.httpBody)
+        let json = try #require(try JSONSerialization.jsonObject(with: body) as? [String: Any])
+
+        #expect(json["model"] as? String == "claude-opus-4-8")
+        #expect(json["temperature"] == nil)
+        #expect(json["top_p"] == nil)
+        #expect(json["top_k"] == nil)
+        let thinking = try #require(json["thinking"] as? [String: Any])
+        #expect(thinking["type"] as? String == "adaptive")
+        #expect(thinking["budget_tokens"] == nil)
+        let outputConfig = try #require(json["output_config"] as? [String: Any])
+        #expect(outputConfig["effort"] as? String == "low")
+        #expect(json["max_tokens"] as? Int == 64)
+    }
+
+    @Test
+    func `Fable 5 request omits thinking config and uses effort output config`() throws {
+        let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
+        let provider = try AnthropicProvider(model: .fable5, configuration: config)
+
+        let settings = GenerationSettings(
+            maxTokens: 128_000,
+            temperature: 0.7,
+            topP: 0.9,
+            topK: 40,
+            reasoningEffort: .high,
+            providerOptions: .init(anthropic: .init(thinking: .adaptive)),
+        )
+
+        let request = ProviderRequest(
+            messages: [.user("hi")],
+            settings: settings,
+        )
+
+        let urlRequest = try provider.makeURLRequest(for: request, stream: false)
+        let body = try #require(urlRequest.httpBody)
+        let json = try #require(try JSONSerialization.jsonObject(with: body) as? [String: Any])
+
+        #expect(json["model"] as? String == "claude-fable-5")
+        #expect(json["temperature"] == nil)
+        #expect(json["top_p"] == nil)
+        #expect(json["top_k"] == nil)
+        #expect(json["thinking"] == nil)
+        let outputConfig = try #require(json["output_config"] as? [String: Any])
+        #expect(outputConfig["effort"] as? String == "high")
+        #expect(json["max_tokens"] as? Int == 128_000)
+        #expect(urlRequest.value(forHTTPHeaderField: "anthropic-beta") == nil)
+    }
+
+    @Test
+    func `Fable 5 request uses model-aware default output budget`() throws {
+        let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
+        let provider = try AnthropicProvider(model: .fable5, configuration: config)
+
+        let request = ProviderRequest(messages: [.user("hi")])
+        let urlRequest = try provider.makeURLRequest(for: request, stream: false)
+        let body = try #require(urlRequest.httpBody)
+        let json = try #require(try JSONSerialization.jsonObject(with: body) as? [String: Any])
+
+        #expect(json["max_tokens"] as? Int == 16384)
+        #expect(urlRequest.timeoutInterval == 1800)
+    }
+
+    @Test
+    func `Fable 5 long output requests extend non-streaming timeout`() throws {
+        let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
+        let provider = try AnthropicProvider(model: .fable5, configuration: config)
+
+        let urlRequest = try provider.makeURLRequest(
+            for: ProviderRequest(
+                messages: [.user("long")],
+                settings: GenerationSettings(maxTokens: 128_000),
+            ),
+            stream: false,
+        )
+
+        #expect(urlRequest.timeoutInterval == 1800)
+    }
+
+    @Test
+    func `Opus long output requests extend non-streaming timeout`() throws {
+        let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
+
+        for model in [LanguageModel.Anthropic.opus47, .opus48] {
+            let provider = try AnthropicProvider(model: model, configuration: config)
+            let urlRequest = try provider.makeURLRequest(
+                for: ProviderRequest(
+                    messages: [.user("long")],
+                    settings: GenerationSettings(maxTokens: 128_000),
+                ),
+                stream: false,
+            )
+
+            #expect(urlRequest.timeoutInterval == 1800)
+        }
+    }
+
+    @Test
+    func `Custom Fable model id uses Fable request defaults`() throws {
+        let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
+        let provider = try AnthropicProvider(model: .custom("claude-fable-5"), configuration: config)
+
+        let request = ProviderRequest(messages: [.user("hi")])
+        let urlRequest = try provider.makeURLRequest(for: request, stream: false)
+        let body = try #require(urlRequest.httpBody)
+        let json = try #require(try JSONSerialization.jsonObject(with: body) as? [String: Any])
+
+        #expect(provider.capabilities.supportsStreaming == false)
+        #expect(provider.capabilities.contextLength == 1_000_000)
+        #expect(provider.capabilities.maxOutputTokens == 128_000)
+        #expect(LanguageModel.anthropic(.custom("claude-fable-5")).supportsStreaming == false)
+        #expect(LanguageModel.anthropic(.custom("claude-fable-5")).contextLength == 1_000_000)
+        #expect(LanguageModel.Anthropic.custom("claude-fable-5").maxOutputTokens == 128_000)
+        #expect(json["model"] as? String == "claude-fable-5")
+        #expect(json["thinking"] == nil)
+        #expect(json["max_tokens"] as? Int == 16384)
+    }
+
+    @Test
+    func `Qualified custom Fable model id uses Fable request defaults`() throws {
+        let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
+        let provider = try AnthropicProvider(model: .custom("anthropic.claude-fable-5"), configuration: config)
+
+        let request = ProviderRequest(messages: [.user("hi")])
+        let urlRequest = try provider.makeURLRequest(for: request, stream: false)
+        let body = try #require(urlRequest.httpBody)
+        let json = try #require(try JSONSerialization.jsonObject(with: body) as? [String: Any])
+
+        #expect(provider.capabilities.supportsStreaming == false)
+        #expect(provider.capabilities.contextLength == 1_000_000)
+        #expect(provider.capabilities.maxOutputTokens == 128_000)
+        #expect(LanguageModel.anthropic(.custom("anthropic.claude-fable-5")).contextLength == 1_000_000)
+        #expect(LanguageModel.Anthropic.custom("anthropic.claude-fable-5").maxOutputTokens == 128_000)
+        #expect(json["model"] as? String == "anthropic.claude-fable-5")
+        #expect(json["thinking"] == nil)
+        #expect(json["max_tokens"] as? Int == 16384)
+    }
+
+    @Test
+    func `Fable 5 rejects disabled thinking mode`() throws {
+        let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
+        let provider = try AnthropicProvider(model: .fable5, configuration: config)
+        let settings = GenerationSettings(
+            maxTokens: 64,
+            providerOptions: .init(anthropic: .init(thinking: .disabled)),
+        )
+
+        #expect(throws: TachikomaError.self) {
+            _ = try provider.makeURLRequest(
+                for: ProviderRequest(messages: [.user("hi")], settings: settings),
+                stream: false,
+            )
+        }
+    }
+
+    @Test
+    func `Custom Fable model id rejects disabled thinking mode`() throws {
+        let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
+        let provider = try AnthropicProvider(model: .custom("claude-fable-5"), configuration: config)
+        let settings = GenerationSettings(
+            maxTokens: 64,
+            providerOptions: .init(anthropic: .init(thinking: .disabled)),
+        )
+
+        #expect(throws: TachikomaError.self) {
+            _ = try provider.makeURLRequest(
+                for: ProviderRequest(messages: [.user("hi")], settings: settings),
+                stream: false,
+            )
+        }
+    }
+
+    @Test
+    func `Opus reasoning effort is kept when thinking is disabled`() throws {
+        let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
+        let provider = try AnthropicProvider(model: .opus48, configuration: config)
+
+        let settings = GenerationSettings(
+            maxTokens: 64,
+            reasoningEffort: .low,
+            providerOptions: .init(anthropic: .init(thinking: .disabled)),
+        )
+
+        let request = ProviderRequest(
+            messages: [.user("hi")],
+            settings: settings,
+        )
+
+        let urlRequest = try provider.makeURLRequest(for: request, stream: false)
+        let body = try #require(urlRequest.httpBody)
+        let json = try #require(try JSONSerialization.jsonObject(with: body) as? [String: Any])
+        let outputConfig = try #require(json["output_config"] as? [String: Any])
+
+        #expect(json["thinking"] == nil)
+        #expect(outputConfig["effort"] as? String == "low")
+    }
+
+    @Test
+    func `Opus effort is sent without thinking when reasoning effort is configured`() throws {
+        let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
+        let provider = try AnthropicProvider(model: .opus48, configuration: config)
+
+        let settings = GenerationSettings(
+            maxTokens: 64,
+            reasoningEffort: .low,
+        )
+
+        let request = ProviderRequest(
+            messages: [.user("hi")],
+            settings: settings,
+        )
+
+        let urlRequest = try provider.makeURLRequest(for: request, stream: false)
+        let body = try #require(urlRequest.httpBody)
+        let json = try #require(try JSONSerialization.jsonObject(with: body) as? [String: Any])
+        let outputConfig = try #require(json["output_config"] as? [String: Any])
+
+        #expect(json["thinking"] == nil)
+        #expect(outputConfig["effort"] as? String == "low")
+    }
+
+    @Test
+    func `Unsupported adaptive thinking is omitted for older Claude models`() throws {
+        let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
+        let provider = try AnthropicProvider(model: .opus45, configuration: config)
+
+        let settings = GenerationSettings(
+            maxTokens: 64,
+            providerOptions: .init(anthropic: .init(thinking: .adaptive)),
+        )
+
+        let request = ProviderRequest(
+            messages: [.user("hi")],
+            settings: settings,
+        )
+
+        let urlRequest = try provider.makeURLRequest(for: request, stream: false)
+        let body = try #require(urlRequest.httpBody)
+        let json = try #require(try JSONSerialization.jsonObject(with: body) as? [String: Any])
+
+        #expect(json["thinking"] == nil)
+        #expect(json["output_config"] == nil)
+    }
+
+    @Test
+    func `Sonnet 4_6 request keeps adaptive thinking payload`() throws {
+        let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
+        let provider = try AnthropicProvider(model: .sonnet46, configuration: config)
+
+        let settings = GenerationSettings(
+            maxTokens: 64,
+            temperature: 0.7,
+            reasoningEffort: .medium,
+            providerOptions: .init(anthropic: .init(thinking: .adaptive)),
+        )
+
+        let request = ProviderRequest(
+            messages: [.user("hi")],
+            settings: settings,
+        )
+
+        let urlRequest = try provider.makeURLRequest(for: request, stream: false)
+        let body = try #require(urlRequest.httpBody)
+        let json = try #require(try JSONSerialization.jsonObject(with: body) as? [String: Any])
+        let thinking = try #require(json["thinking"] as? [String: Any])
+        let outputConfig = try #require(json["output_config"] as? [String: Any])
+
+        #expect(json["model"] as? String == "claude-sonnet-4-6")
+        #expect(json["temperature"] == nil)
+        #expect(thinking["type"] as? String == "adaptive")
+        #expect(outputConfig["effort"] as? String == "medium")
+    }
+
+    @Test
+    func `Custom Anthropic request keeps thinking payload`() throws {
+        let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
+        let provider = try AnthropicProvider(model: .custom("claude-opus-4-5-latest"), configuration: config)
+
+        let settings = GenerationSettings(
+            maxTokens: 64,
+            providerOptions: .init(anthropic: .init(thinking: .enabled(budgetTokens: 12000))),
+        )
+
+        let request = ProviderRequest(
+            messages: [.user("hi")],
+            settings: settings,
+        )
+
+        let urlRequest = try provider.makeURLRequest(for: request, stream: false)
+        let body = try #require(urlRequest.httpBody)
+        let json = try #require(try JSONSerialization.jsonObject(with: body) as? [String: Any])
+        let thinking = try #require(json["thinking"] as? [String: Any])
+
+        #expect(json["model"] as? String == "claude-opus-4-5-latest")
+        #expect(thinking["type"] as? String == "enabled")
+        #expect(thinking["budget_tokens"] as? Int == 12000)
+    }
+
     @Test
     func `Provider respects custom baseURL`() throws {
         let config = TachikomaConfiguration(
@@ -70,6 +438,24 @@ struct AnthropicInterleavedDefaultsTests {
         #expect(urlRequest.url?.absoluteString == "https://entropic.example/v1/messages")
     }
 
+    @Test
+    func `Provider includes additional proxy headers`() throws {
+        let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
+        let provider = try AnthropicProvider(
+            model: .opus45,
+            configuration: config,
+            additionalHeaders: [
+                "client_id": "proxy-client",
+                "client_secret": "proxy-secret",
+            ],
+        )
+
+        let request = ProviderRequest(messages: [.user("hi")])
+        let urlRequest = try provider.makeURLRequest(for: request, stream: false)
+        #expect(urlRequest.value(forHTTPHeaderField: "client_id") == "proxy-client")
+        #expect(urlRequest.value(forHTTPHeaderField: "client_secret") == "proxy-secret")
+    }
+
     @Test
     func `Stream delta decodes thinking_delta payload`() throws {
         let data = try #require("{\"type\":\"thinking_delta\",\"thinking\":\"ok\"}".data(using: .utf8))
@@ -87,6 +473,29 @@ struct AnthropicInterleavedDefaultsTests {
         #expect(delta.signature == "sig")
     }
 
+    @Test
+    func `Stream delta decodes message_delta stop reason without delta type`() throws {
+        let data = try #require(
+            "{\"stop_reason\":\"refusal\",\"stop_sequence\":null}".data(using: .utf8),
+        )
+        let delta = try JSONDecoder().decode(AnthropicStreamDelta.self, from: data)
+        #expect(delta.type.isEmpty)
+        #expect(delta.stopReason == "refusal")
+    }
+
+    @Test
+    func `Stream event decodes partial usage with stop reason`() throws {
+        let data = try #require(
+            #"{"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"output_tokens":42}}"#
+                .data(using: .utf8),
+        )
+        let event = try JSONDecoder().decode(AnthropicStreamEvent.self, from: data)
+
+        #expect(event.delta?.stopReason == "end_turn")
+        #expect(event.usage?.inputTokens == 0)
+        #expect(event.usage?.outputTokens == 42)
+    }
+
     @Test
     func `Signed thinking blocks are preserved for assistant messages`() throws {
         let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
@@ -125,7 +534,125 @@ struct AnthropicInterleavedDefaultsTests {
     }
 
     @Test
-    func `Redacted thinking blocks preserve signature without text`() throws {
+    func `Fable 5 preserves signed thinking history while omitting request thinking field`() throws {
+        let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
+        let provider = try AnthropicProvider(model: .fable5, configuration: config)
+        let signedThinking = try ModelMessage(
+            role: .assistant,
+            content: [.text("fable thinking")],
+            channel: .thinking,
+            metadata: .init(customData: [
+                "anthropic.thinking.model": "claude-fable-5",
+                "anthropic.thinking.signature": "sig-fable",
+                "anthropic.thinking.type": "thinking",
+                "tachikoma.reasoning.provider": "anthropic",
+                "tachikoma.reasoning.model": "claude-fable-5",
+                "tachikoma.reasoning.base_url": #require(ReasoningEndpointIdentity
+                    .canonical("https://api.anthropic.com")),
+            ]),
+        )
+
+        let request = ProviderRequest(
+            messages: [.user("hi"), signedThinking, .assistant("hello"), .user("continue")],
+            settings: GenerationSettings(maxTokens: 64),
+        )
+
+        let urlRequest = try provider.makeURLRequest(for: request, stream: false)
+        let body = try #require(urlRequest.httpBody)
+        let json = try #require(try JSONSerialization.jsonObject(with: body) as? [String: Any])
+        let messages = try #require(json["messages"] as? [[String: Any]])
+        let assistant = try #require(messages[1]["content"] as? [[String: Any]])
+
+        #expect(json["thinking"] == nil)
+        #expect(assistant.first?["type"] as? String == "thinking")
+        #expect(assistant.first?["thinking"] as? String == "fable thinking")
+        #expect(assistant.first?["signature"] as? String == "sig-fable")
+    }
+
+    @Test
+    func `Fable 5 drops mismatched signed thinking history in direct provider requests`() throws {
+        let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
+        let provider = try AnthropicProvider(model: .fable5, configuration: config)
+        let signedThinking = try ModelMessage(
+            role: .assistant,
+            content: [.text("foreign thinking")],
+            channel: .thinking,
+            metadata: .init(customData: [
+                "anthropic.thinking.model": "claude-fable-5",
+                "anthropic.thinking.signature": "sig-foreign",
+                "anthropic.thinking.type": "thinking",
+                "tachikoma.reasoning.provider": "anthropic",
+                "tachikoma.reasoning.model": "claude-fable-5",
+                "tachikoma.reasoning.base_url": #require(ReasoningEndpointIdentity
+                    .canonical("https://other.example.test")),
+            ]),
+        )
+
+        let request = ProviderRequest(
+            messages: [.user("hi"), signedThinking, .assistant("hello"), .user("continue")],
+            settings: GenerationSettings(maxTokens: 64),
+        )
+
+        let urlRequest = try provider.makeURLRequest(for: request, stream: false)
+        let body = try #require(urlRequest.httpBody)
+        let json = try #require(try JSONSerialization.jsonObject(with: body) as? [String: Any])
+        let messages = try #require(json["messages"] as? [[String: Any]])
+        let assistant = try #require(messages[1]["content"] as? [[String: Any]])
+
+        #expect(assistant.count == 1)
+        #expect(assistant.first?["type"] as? String == "text")
+        #expect(assistant.first?["text"] as? String == "hello")
+        #expect(String(data: body, encoding: .utf8)?.contains("foreign thinking") == false)
+    }
+
+    @Test
+    func `Fable 5 rejects assistant prefill requests`() throws {
+        let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
+        let provider = try AnthropicProvider(model: .fable5, configuration: config)
+
+        #expect(throws: TachikomaError.self) {
+            _ = try provider.makeURLRequest(
+                for: ProviderRequest(messages: [.user("hi"), .assistant("prefill")]),
+                stream: false,
+            )
+        }
+    }
+
+    @Test
+    func `Anthropic refusal stop reason maps to content filter`() {
+        #expect(AnthropicProvider.mapFinishReason("refusal") == .contentFilter)
+        #expect(AnthropicProvider.mapFinishReason("model_context_window_exceeded") == .length)
+    }
+
+    @Test
+    func `Anthropic refusal response decodes stop details explanation`() throws {
+        let data = """
+        {
+          "id": "msg_test",
+          "type": "message",
+          "role": "assistant",
+          "content": [],
+          "model": "claude-fable-5",
+          "stop_reason": "refusal",
+          "stop_details": {
+            "category": "cyber",
+            "explanation": "I cannot help with that request."
+          },
+          "usage": {
+            "input_tokens": 10,
+            "output_tokens": 0
+          }
+        }
+        """.data(using: .utf8)!
+
+        let response = try JSONDecoder().decode(AnthropicMessageResponse.self, from: data)
+
+        #expect(response.stopDetails?.category == "cyber")
+        #expect(response.stopDetails?.explanation == "I cannot help with that request.")
+    }
+
+    @Test
+    func `Redacted thinking blocks preserve opaque data`() throws {
         let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
         let provider = try AnthropicProvider(model: .opus45, configuration: config)
 
@@ -136,10 +663,9 @@ struct AnthropicInterleavedDefaultsTests {
 
         let redacted = ModelMessage(
             role: .assistant,
-            content: [.text("")],
+            content: [.text("opaque-redacted-data")],
             channel: .thinking,
             metadata: .init(customData: [
-                "anthropic.thinking.signature": "sig-redacted",
                 "anthropic.thinking.type": "redacted_thinking",
             ]),
         )
@@ -156,8 +682,232 @@ struct AnthropicInterleavedDefaultsTests {
 
         let assistant = try #require(messages[1]["content"] as? [[String: Any]])
         #expect(assistant.first?["type"] as? String == "redacted_thinking")
-        #expect((assistant.first?["redacted_thinking"] as? String)?.isEmpty == true)
-        #expect(assistant.first?["signature"] as? String == "sig-redacted")
+        #expect(assistant.first?["data"] as? String == "opaque-redacted-data")
+        #expect(assistant.first?["signature"] == nil)
+    }
+
+    @Test
+    func `Redacted thinking response decodes opaque data`() throws {
+        let data = try #require(
+            """
+            {"type":"redacted_thinking","data":"opaque-redacted-data"}
+            """.data(using: .utf8),
+        )
+
+        let content = try JSONDecoder().decode(AnthropicResponseContent.self, from: data)
+        guard case let .redactedThinking(redacted) = content else {
+            Issue.record("Expected redacted thinking content")
+            return
+        }
+        #expect(redacted.data == "opaque-redacted-data")
+    }
+
+    @Test
+    func `Consecutive thinking blocks are preserved in order`() throws {
+        let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
+        let provider = try AnthropicProvider(model: .fable5, configuration: config)
+
+        let signedThinking = try ModelMessage(
+            role: .assistant,
+            content: [.text("signed")],
+            channel: .thinking,
+            metadata: .init(customData: [
+                "anthropic.thinking.model": "claude-fable-5",
+                "anthropic.thinking.signature": "sig",
+                "anthropic.thinking.type": "thinking",
+                "tachikoma.reasoning.provider": "anthropic",
+                "tachikoma.reasoning.model": "claude-fable-5",
+                "tachikoma.reasoning.base_url": #require(ReasoningEndpointIdentity
+                    .canonical("https://api.anthropic.com")),
+            ]),
+        )
+        let redactedThinking = try ModelMessage(
+            role: .assistant,
+            content: [.text("opaque")],
+            channel: .thinking,
+            metadata: .init(customData: [
+                "anthropic.thinking.model": "claude-fable-5",
+                "anthropic.thinking.type": "redacted_thinking",
+                "tachikoma.reasoning.provider": "anthropic",
+                "tachikoma.reasoning.model": "claude-fable-5",
+                "tachikoma.reasoning.base_url": #require(ReasoningEndpointIdentity
+                    .canonical("https://api.anthropic.com")),
+            ]),
+        )
+
+        let request = ProviderRequest(
+            messages: [.user("hi"), signedThinking, redactedThinking, .assistant("hello"), .user("continue")],
+            settings: GenerationSettings(maxTokens: 64),
+        )
+
+        let urlRequest = try provider.makeURLRequest(for: request, stream: false)
+        let body = try #require(urlRequest.httpBody)
+        let json = try #require(try JSONSerialization.jsonObject(with: body) as? [String: Any])
+        let messages = try #require(json["messages"] as? [[String: Any]])
+        let assistant = try #require(messages[1]["content"] as? [[String: Any]])
+
+        #expect(assistant.count == 3)
+        #expect(assistant[0]["type"] as? String == "thinking")
+        #expect(assistant[0]["thinking"] as? String == "signed")
+        #expect(assistant[0]["signature"] as? String == "sig")
+        #expect(assistant[1]["type"] as? String == "redacted_thinking")
+        #expect(assistant[1]["data"] as? String == "opaque")
+        #expect(assistant[2]["type"] as? String == "text")
+        #expect(assistant[2]["text"] as? String == "hello")
+    }
+
+    @Test
+    func `Current Anthropic models expose documented output caps`() {
+        #expect(LanguageModel.Anthropic.fable5.maxOutputTokens == 128_000)
+        #expect(LanguageModel.Anthropic.opus47.maxOutputTokens == 128_000)
+        #expect(LanguageModel.Anthropic.opus48.maxOutputTokens == 128_000)
+        #expect(LanguageModel.Anthropic.sonnet46.maxOutputTokens == 64000)
+        #expect(LanguageModel.Anthropic.haiku45.maxOutputTokens == 64000)
+    }
+
+    @Test
+    func `Fable and Opus 4_8 streaming are disabled until rollback is supported`() async throws {
+        let config = TachikomaConfiguration(apiKeys: ["anthropic": "test-key"])
+        let provider = try AnthropicProvider(model: .fable5, configuration: config)
+        let opusProvider = try AnthropicProvider(model: .opus48, configuration: config)
+
+        #expect(provider.capabilities.supportsStreaming == false)
+        #expect(LanguageModel.anthropic(.fable5).supportsStreaming == false)
+        #expect(opusProvider.capabilities.supportsStreaming == false)
+        #expect(LanguageModel.anthropic(.opus47).supportsStreaming == true)
+        #expect(LanguageModel.anthropic(.opus48).supportsStreaming == false)
+        #expect(LanguageModel.anthropic(.sonnet46).supportsStreaming == true)
+        #expect(LanguageModel.anthropic(.sonnet45).supportsStreaming == true)
+        #expect(LanguageModel.anthropic(.haiku45).supportsStreaming == true)
+        await #expect(throws: TachikomaError.self) {
+            _ = try await provider.streamText(request: ProviderRequest(messages: [.user("hi")]))
+        }
+        await #expect(throws: TachikomaError.self) {
+            _ = try await opusProvider.streamText(request: ProviderRequest(messages: [.user("hi")]))
+        }
+    }
+
+    @Test
+    func `Opus 4_8 detection avoids substring false positives`() {
+        #expect(LanguageModel.Anthropic.isOpus48(modelId: "claude-opus-4-8") == true)
+        #expect(LanguageModel.Anthropic.isOpus48(modelId: "anthropic/claude-opus-4.8") == true)
+        #expect(LanguageModel.Anthropic.isOpus48(modelId: "my-opus48-distill") == false)
+        #expect(LanguageModel.Anthropic.isOpus48(modelId: "opus480") == false)
+    }
+
+    @Test
+    func `Fable detection avoids substring false positives`() {
+        #expect(LanguageModel.Anthropic.isFable(modelId: "claude-fable-5") == true)
+        #expect(LanguageModel.Anthropic.isFable(modelId: "anthropic/claude-fable-5") == true)
+        #expect(LanguageModel.Anthropic.isFable(modelId: "vendor/claude-fable-50") == false)
+        #expect(LanguageModel.Anthropic.isFable(modelId: "my-claude-fable-5-distill") == false)
+    }
+
+    @Test
+    func `Anthropic-compatible provider tags native thinking with wrapper identity`() async throws {
+        let sessionConfig = URLSessionConfiguration.ephemeral
+        sessionConfig.protocolClasses = [AnthropicIdentityURLProtocol.self]
+        let provider = try AnthropicProvider(
+            model: .custom("claude-fable-5"),
+            configuration: TachikomaConfiguration(apiKeys: ["anthropic": "test-key"]),
+            reasoningProvider: "anthropic-compatible",
+            reasoningModelId: "claude-fable-5",
+            reasoningBaseURL: "https://user:secret@example.test/path?token=secret#frag",
+            urlSession: URLSession(configuration: sessionConfig),
+        )
+
+        let response = try await provider.generateText(request: ProviderRequest(messages: [.user("hi")]))
+        let thinking = try #require(response.assistantMessages.first { $0.channel == .thinking })
+        #expect(thinking.metadata?.customData?["tachikoma.reasoning.provider"] == "anthropic-compatible")
+        #expect(thinking.metadata?.customData?["tachikoma.reasoning.model"] == "claude-fable-5")
+        let endpointIdentity = thinking.metadata?.customData?["tachikoma.reasoning.base_url"]
+        #expect(endpointIdentity == ReasoningEndpointIdentity.canonical("https://example.test/path?token=secret"))
+        #expect(endpointIdentity?.hasPrefix("sha256:") == true)
+        #expect(endpointIdentity?.contains("path") == false)
+        #expect(endpointIdentity?.contains("secret") == false)
+        #expect(endpointIdentity?.contains("token") == false)
+        #expect(thinking.metadata?.customData?["anthropic.thinking.signature"] == "sig")
+    }
+
+    @Test
+    func `Compatible refusal-prone Anthropic streaming and capabilities are disabled`() async throws {
+        let config = TachikomaConfiguration(apiKeys: ["anthropic_compatible": "test-key"])
+        let provider = try AnthropicCompatibleProvider(
+            modelId: "claude-fable-5",
+            baseURL: "https://example.test",
+            configuration: config,
+        )
+        let opusProvider = try AnthropicCompatibleProvider(
+            modelId: "claude-opus-4-8",
+            baseURL: "https://example.test",
+            configuration: config,
+        )
+        let overriddenProvider = try AnthropicCompatibleProvider(
+            modelId: "claude-fable-5",
+            baseURL: "https://example.test",
+            configuration: config,
+            capabilities: ModelCapabilities(supportsStreaming: true),
+        )
+
+        #expect(provider.capabilities.supportsStreaming == false)
+        #expect(opusProvider.capabilities.supportsStreaming == false)
+        #expect(overriddenProvider.capabilities.supportsStreaming == false)
+        #expect(provider.capabilities.contextLength == 1_000_000)
+        #expect(provider.capabilities.maxOutputTokens == 128_000)
+        #expect(LanguageModel.anthropicCompatible(
+            modelId: "claude-fable-5",
+            baseURL: "https://example.test",
+        ).supportsStreaming == false)
+        #expect(LanguageModel.anthropicCompatible(
+            modelId: "claude-opus-4-8",
+            baseURL: "https://example.test",
+        ).supportsStreaming == false)
+        #expect(LanguageModel.openaiCompatible(
+            modelId: "claude-fable-5",
+            baseURL: "https://example.test",
+        ).supportsStreaming == false)
+        #expect(LanguageModel.anthropicCompatible(
+            modelId: "claude-fable-5",
+            baseURL: "https://example.test",
+        ).contextLength == 1_000_000)
+        #expect(LanguageModel.anthropicCompatible(
+            modelId: "anthropic.claude-fable-5",
+            baseURL: "https://example.test",
+        ).contextLength == 1_000_000)
+        let openAICompatibleProvider = try OpenAICompatibleProvider(
+            modelId: "claude-fable-5",
+            baseURL: "https://example.test",
+            configuration: TachikomaConfiguration(apiKeys: ["openai_compatible": "test-key"]),
+        )
+        let openRouterProvider = try OpenRouterProvider(
+            modelId: "anthropic/claude-fable-5",
+            configuration: TachikomaConfiguration(apiKeys: ["openrouter": "test-key"]),
+        )
+        let togetherProvider = try TogetherProvider(
+            modelId: "anthropic/claude-fable-5",
+            configuration: TachikomaConfiguration(apiKeys: ["together": "test-key"]),
+        )
+        #expect(openAICompatibleProvider.capabilities.supportsStreaming == false)
+        #expect(openRouterProvider.capabilities.supportsStreaming == false)
+        #expect(togetherProvider.capabilities.supportsStreaming == false)
+        #expect(openAICompatibleProvider.capabilities.contextLength == 1_000_000)
+        #expect(openAICompatibleProvider.capabilities.maxOutputTokens == 128_000)
+        #expect(openRouterProvider.capabilities.contextLength == 1_000_000)
+        #expect(openRouterProvider.capabilities.maxOutputTokens == 128_000)
+        #expect(togetherProvider.capabilities.contextLength == 1_000_000)
+        #expect(togetherProvider.capabilities.maxOutputTokens == 128_000)
+        await #expect(throws: TachikomaError.self) {
+            _ = try await provider.streamText(request: ProviderRequest(messages: [.user("hi")]))
+        }
+        await #expect(throws: TachikomaError.self) {
+            _ = try await openAICompatibleProvider.streamText(request: ProviderRequest(messages: [.user("hi")]))
+        }
+        await #expect(throws: TachikomaError.self) {
+            _ = try await openRouterProvider.streamText(request: ProviderRequest(messages: [.user("hi")]))
+        }
+        await #expect(throws: TachikomaError.self) {
+            _ = try await togetherProvider.streamText(request: ProviderRequest(messages: [.user("hi")]))
+        }
     }
 
     @Test
@@ -187,3 +937,49 @@ struct AnthropicInterleavedDefaultsTests {
         #expect(assistant.first?["type"] as? String == "text")
     }
 }
+
+private final class AnthropicIdentityURLProtocol: URLProtocol {
+    override class func canInit(with _: URLRequest) -> Bool {
+        true
+    }
+
+    override class func canonicalRequest(for request: URLRequest) -> URLRequest {
+        request
+    }
+
+    override func startLoading() {
+        guard
+            let url = self.request.url,
+            let response = HTTPURLResponse(
+                url: url,
+                statusCode: 200,
+                httpVersion: nil,
+                headerFields: ["Content-Type": "application/json"],
+            ) else
+        {
+            self.client?.urlProtocol(self, didFailWithError: TachikomaError.invalidInput("Missing mock response"))
+            return
+        }
+
+        let body = """
+        {
+          "id": "msg_test",
+          "type": "message",
+          "role": "assistant",
+          "model": "claude-fable-5",
+          "content": [
+            {"type": "thinking", "thinking": "private", "signature": "sig"},
+            {"type": "text", "text": "ok"}
+          ],
+          "stop_reason": "end_turn",
+          "usage": {"input_tokens": 1, "output_tokens": 2}
+        }
+        """.data(using: .utf8) ?? Data()
+
+        self.client?.urlProtocol(self, didReceive: response, cacheStoragePolicy: .notAllowed)
+        self.client?.urlProtocol(self, didLoad: body)
+        self.client?.urlProtocolDidFinishLoading(self)
+    }
+
+    override func stopLoading() {}
+}

Tests/TachikomaTests/Providers/Compatible/AzureOpenAIProviderTests.swift

@@ -20,7 +20,7 @@ private final class AzureTestURLProtocol: URLProtocol {
     static let responseBody: Data = """
     {
       "id": "chatcmpl-azure",
-      "model": "gpt-4o",
+      "model": "gpt-5.5",
       "choices": [
         {
           "index": 0,
@@ -85,7 +85,7 @@ struct AzureOpenAIProviderTests {
         await AzureTestURLProtocol.reset()
 
         let provider = try AzureOpenAIProvider(
-            deploymentId: "gpt-4o",
+            deploymentId: "gpt-5.5",
             resource: "my-aoai",
             apiVersion: "2025-04-01-preview",
             endpoint: nil,
@@ -99,7 +99,7 @@ struct AzureOpenAIProviderTests {
         #expect(response.text == "hello azure")
 
         let sentRequest = await AzureTestURLProtocol.fetchLastRequest()
-        #expect(sentRequest?.url?.path == "/openai/deployments/gpt-4o/chat/completions")
+        #expect(sentRequest?.url?.path == "/openai/deployments/gpt-5.5/chat/completions")
 
         if let components = sentRequest?.url.flatMap({ URLComponents(url: $0, resolvingAgainstBaseURL: false) }) {
             let apiVersion = components.queryItems?.first { $0.name == "api-version" }?.value
@@ -122,7 +122,7 @@ struct AzureOpenAIProviderTests {
         await AzureTestURLProtocol.reset()
 
         let provider = try AzureOpenAIProvider(
-            deploymentId: "gpt-4o-mini",
+            deploymentId: "gpt-5-mini",
             resource: nil,
             apiVersion: "2025-04-01-preview",
             endpoint: nil,

Tests/TachikomaTests/Providers/Integration/ProviderIntegrationTests.swift

@@ -3,7 +3,7 @@ import Foundation
 import Testing
 @testable import Tachikoma
 
-@Suite(.enabled(if: ProcessInfo.processInfo.environment["INTEGRATION_TESTS"] != nil))
+@Suite(.serialized, .enabled(if: ProcessInfo.processInfo.environment["INTEGRATION_TESTS"] != nil))
 struct ProviderIntegrationTests {
     // MARK: - Test Configuration
 
@@ -14,50 +14,108 @@ struct ProviderIntegrationTests {
         static let streamMessage = "Count from 1 to 3"
     }
 
-    private static func hasEnv(_ name: String) -> Bool {
-        guard let value = ProcessInfo.processInfo.environment[name] else {
-            return false
+    private struct LiveCredentials {
+        var openAI: String?
+        var anthropic: String?
+        var google: String?
+        var mistral: String?
+        var groq: String?
+        var grok: String?
+
+        static func capture() -> Self {
+            let environment = ProcessInfo.processInfo.environment
+            return Self(
+                openAI: Self.validKey(environment["OPENAI_API_KEY"]),
+                anthropic: Self.validKey(environment["ANTHROPIC_API_KEY"]),
+                google: Self.validKey(environment["GEMINI_API_KEY"]) ?? Self.validKey(environment["GOOGLE_API_KEY"]),
+                mistral: Self.validKey(environment["MISTRAL_API_KEY"]),
+                groq: Self.validKey(environment["GROQ_API_KEY"]),
+                grok: Self.validKey(environment["X_AI_API_KEY"])
+                    ?? Self.validKey(environment["XAI_API_KEY"])
+                    ?? Self.validKey(environment["GROK_API_KEY"]),
+            )
+        }
+
+        private static func validKey(_ value: String?) -> String? {
+            guard let key = value?.trimmingCharacters(in: .whitespacesAndNewlines), !key.isEmpty else {
+                return nil
+            }
+            let lowercased = key.lowercased()
+            guard
+                key != "env-key",
+                key != "cred-key",
+                key != "test-key",
+                !lowercased.hasPrefix("test-") else
+            {
+                return nil
+            }
+            return key
         }
-        return !value.isEmpty
     }
 
+    private static let liveCredentials = LiveCredentials.capture()
+
     private static var hasOpenAIKey: Bool {
-        hasEnv("OPENAI_API_KEY")
+        liveCredentials.openAI != nil
     }
 
     private static var hasAnthropicKey: Bool {
-        hasEnv("ANTHROPIC_API_KEY")
+        liveCredentials.anthropic != nil
     }
 
     private static var hasGoogleKey: Bool {
-        hasEnv("GEMINI_API_KEY") || hasEnv("GOOGLE_API_KEY")
+        liveCredentials.google != nil
     }
 
     private static var hasMistralKey: Bool {
-        hasEnv("MISTRAL_API_KEY")
+        liveCredentials.mistral != nil
     }
 
     private static var hasGroqKey: Bool {
-        hasEnv("GROQ_API_KEY")
+        liveCredentials.groq != nil
     }
 
     private static var hasGrokKey: Bool {
-        hasEnv("X_AI_API_KEY") || hasEnv("XAI_API_KEY") || hasEnv("GROK_API_KEY")
+        liveCredentials.grok != nil
+    }
+
+    private static func liveConfiguration() -> TachikomaConfiguration {
+        let credentials = Self.liveCredentials
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        if let openAI = credentials.openAI {
+            config.setAPIKey(openAI, for: .openai)
+        }
+        if let anthropic = credentials.anthropic {
+            config.setAPIKey(anthropic, for: .anthropic)
+        }
+        if let google = credentials.google {
+            config.setAPIKey(google, for: .google)
+        }
+        if let mistral = credentials.mistral {
+            config.setAPIKey(mistral, for: .mistral)
+        }
+        if let groq = credentials.groq {
+            config.setAPIKey(groq, for: .groq)
+        }
+        if let grok = credentials.grok {
+            config.setAPIKey(grok, for: .grok)
+        }
+        return config
     }
 
     // MARK: - OpenAI Integration Tests
 
     @Test(.enabled(if: Self.hasOpenAIKey))
     func `OpenAI Provider - Real API Call`() async throws {
-        let model = Model.openai(.gpt4oMini)
-        let config = TachikomaConfiguration()
+        let model = Model.openai(.gpt5Mini)
+        let config = Self.liveConfiguration()
         do {
             _ = try ProviderFactory.createProvider(for: model, configuration: config)
 
             let response = try await generate(
                 TestConfig.shortMessage,
                 using: model,
-                maxTokens: 50,
+                maxTokens: 300,
                 temperature: 0.0,
                 configuration: config,
             )
@@ -72,8 +130,8 @@ struct ProviderIntegrationTests {
 
     @Test(.enabled(if: Self.hasOpenAIKey))
     func `OpenAI Provider - Tool Calling`() async throws {
-        let model = Model.openai(.gpt4oMini)
-        let config = TachikomaConfiguration()
+        let model = Model.openai(.gpt5Mini)
+        let config = Self.liveConfiguration()
 
         do {
             let provider = try ProviderFactory.createProvider(for: model, configuration: config)
@@ -117,8 +175,8 @@ struct ProviderIntegrationTests {
 
     @Test(.enabled(if: Self.hasOpenAIKey))
     func `OpenAI Provider - Streaming`() async throws {
-        let model = Model.openai(.gpt4oMini)
-        let config = TachikomaConfiguration()
+        let model = Model.openai(.gpt5Mini)
+        let config = Self.liveConfiguration()
 
         do {
             let provider = try ProviderFactory.createProvider(for: model, configuration: config)
@@ -128,7 +186,7 @@ struct ProviderIntegrationTests {
                     ModelMessage(role: .user, content: [.text(TestConfig.streamMessage)]),
                 ],
                 tools: nil,
-                settings: .init(maxTokens: 100, temperature: 0.0),
+                settings: .init(maxTokens: 300, temperature: 0.0),
             )
 
             let stream = try await provider.streamText(request: request)
@@ -164,13 +222,15 @@ struct ProviderIntegrationTests {
 
     @Test(.enabled(if: Self.hasAnthropicKey))
     func `Anthropic Provider - Real API Call`() async throws {
-        let model = Model.anthropic(.sonnet4)
+        let model = Model.anthropic(.sonnet46)
+        let config = Self.liveConfiguration()
         do {
             let response = try await generate(
                 TestConfig.shortMessage,
                 using: model,
                 maxTokens: 50,
                 temperature: 0.0,
+                configuration: config,
             )
 
             if !(response.lowercased().contains("hello") && response.contains("Tachikoma")) {
@@ -183,8 +243,8 @@ struct ProviderIntegrationTests {
 
     @Test(.enabled(if: Self.hasAnthropicKey))
     func `Anthropic Provider - Tool Calling`() async throws {
-        let model = Model.anthropic(.sonnet4)
-        let config = TachikomaConfiguration()
+        let model = Model.anthropic(.sonnet46)
+        let config = Self.liveConfiguration()
 
         do {
             let provider = try ProviderFactory.createProvider(for: model, configuration: config)
@@ -261,13 +321,15 @@ struct ProviderIntegrationTests {
 
     @Test(.enabled(if: Self.hasGrokKey))
     func `Grok Provider - Real API Call`() async throws {
-        let model = Model.grok(.grok3)
+        let model = Model.grok(.grok43)
+        let config = Self.liveConfiguration()
         do {
             let response = try await generate(
                 TestConfig.shortMessage,
                 using: model,
                 maxTokens: 50,
                 temperature: 0.0,
+                configuration: config,
             )
             if !(response.lowercased().contains("hello") && response.contains("Tachikoma")) {
                 Self.warn("Grok integration returned: \(response.prefix(120))…")
@@ -282,12 +344,14 @@ struct ProviderIntegrationTests {
     @Test(.enabled(if: Self.hasGoogleKey))
     func `Google Provider - Real API Call`() async throws {
         let model = Model.google(.gemini25Flash)
+        let config = Self.liveConfiguration()
         do {
             let response = try await generate(
                 TestConfig.shortMessage,
                 using: model,
                 maxTokens: 50,
                 temperature: 0.0,
+                configuration: config,
             )
             if !(response.lowercased().contains("hello") && response.contains("Tachikoma")) {
                 Self.warn("Google integration returned: \(response.prefix(120))…")
@@ -301,8 +365,15 @@ struct ProviderIntegrationTests {
 
     @Test(.enabled(if: Self.hasMistralKey))
     func `Mistral Provider - Real API Call`() async throws {
-        let model = Model.mistral(.small)
-        let response = try await generate(TestConfig.shortMessage, using: model, maxTokens: 50, temperature: 0.0)
+        let model = Model.mistral(.smallLatest)
+        let config = Self.liveConfiguration()
+        let response = try await generate(
+            TestConfig.shortMessage,
+            using: model,
+            maxTokens: 50,
+            temperature: 0.0,
+            configuration: config,
+        )
 
         #expect(response.lowercased().contains("hello"))
         #expect(response.contains("Tachikoma"))
@@ -312,8 +383,15 @@ struct ProviderIntegrationTests {
 
     @Test(.enabled(if: Self.hasGroqKey))
     func `Groq Provider - Real API Call`() async throws {
-        let model = Model.groq(.llama38b)
-        let response = try await generate(TestConfig.shortMessage, using: model, maxTokens: 50, temperature: 0.0)
+        let model = Model.groq(.llama318b)
+        let config = Self.liveConfiguration()
+        let response = try await generate(
+            TestConfig.shortMessage,
+            using: model,
+            maxTokens: 50,
+            temperature: 0.0,
+            configuration: config,
+        )
 
         #expect(response.lowercased().contains("hello"))
         #expect(response.contains("Tachikoma"))
@@ -323,12 +401,12 @@ struct ProviderIntegrationTests {
 
     @Test(.enabled(if: Self.hasOpenAIKey))
     func `Multi-Modal Provider - Vision Support`() async throws {
-        let model = Model.openai(.gpt4o)
-        let config = TachikomaConfiguration()
+        let model = Model.openai(.gpt55)
+        let config = Self.liveConfiguration()
         let provider = try ProviderFactory.createProvider(for: model, configuration: config)
 
-        // Create a simple base64 encoded 1x1 red pixel PNG
-        let redPixelPNG = "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mP8/5+hHgAHggJ/PchI7wAAAABJRU5ErkJggg=="
+        // Create a simple base64 encoded 16x16 red square PNG
+        let redPixelPNG = "iVBORw0KGgoAAAANSUhEUgAAABAAAAAQCAIAAACQkWg2AAAAF0lEQVR4nGP4z8BAEiJN9aiGUQ1DSgMAkPn/Afnh+ngAAAAASUVORK5CYII="
 
         let imageContent = ModelMessage.ContentPart.ImageContent(
             data: redPixelPNG,
@@ -343,13 +421,13 @@ struct ProviderIntegrationTests {
                 ]),
             ],
             tools: nil,
-            settings: .init(maxTokens: 50, temperature: 0.0),
+            settings: .init(maxTokens: 300, temperature: 0.0),
         )
 
         let response = try await provider.generateText(request: request)
 
         let normalized = response.text.lowercased()
-        #expect(normalized.contains("red") || normalized.contains("yellow"))
+        #expect(normalized.contains("red"))
     }
 
     // MARK: - Helper Methods

Tests/TachikomaTests/Providers/Integration/ProviderSystemTests.swift

@@ -9,10 +9,10 @@ struct ProviderSystemTests {
     @Test
     func `Provider Factory - OpenAI Provider Creation`() async throws {
         try await TestHelpers.withTestConfiguration(apiKeys: ["openai": "test-key"]) { config in
-            let model = Model.openai(.gpt4o)
+            let model = Model.openai(.gpt55)
             let provider = try ProviderFactory.createProvider(for: model, configuration: config)
 
-            #expect(provider.modelId == "gpt-4o")
+            #expect(provider.modelId == "gpt-5.5")
             #expect(provider.capabilities.supportsVision == true)
             #expect(provider.capabilities.supportsTools == true)
             #expect(provider.capabilities.supportsStreaming == true)
@@ -22,12 +22,27 @@ struct ProviderSystemTests {
     @Test
     func `Provider Factory - Anthropic Provider Creation`() async throws {
         try await TestHelpers.withTestConfiguration(apiKeys: ["anthropic": "test-key"]) { config in
-            let model = Model.anthropic(.opus4)
+            let model = Model.anthropic(.fable5)
             let provider = try ProviderFactory.createProvider(for: model, configuration: config)
 
-            #expect(provider.modelId == "claude-opus-4-1-20250805")
+            #expect(provider.modelId == "claude-fable-5")
             #expect(provider.capabilities.supportsVision == true)
             #expect(provider.capabilities.supportsTools == true)
+            #expect(provider.capabilities.supportsStreaming == false)
+            #expect(provider.capabilities.contextLength == 1_000_000)
+            #expect(provider.capabilities.maxOutputTokens == 128_000)
+        }
+    }
+
+    @Test
+    func `Provider Factory - MiniMax Provider Creation`() async throws {
+        try await TestHelpers.withTestConfiguration(apiKeys: ["minimax": "test-key"]) { config in
+            let model = Model.minimax(.m27)
+            let provider = try ProviderFactory.createProvider(for: model, configuration: config)
+
+            #expect(provider.modelId == "MiniMax-M2.7")
+            #expect(provider.capabilities.supportsVision == false)
+            #expect(provider.capabilities.supportsTools == true)
             #expect(provider.capabilities.supportsStreaming == true)
         }
     }
@@ -35,10 +50,10 @@ struct ProviderSystemTests {
     @Test
     func `Provider Factory - Grok Provider Creation`() async throws {
         try await TestHelpers.withTestConfiguration(apiKeys: ["grok": "test-key"]) { config in
-            let model = Model.grok(.grok4FastReasoning)
+            let model = Model.grok(.grok43)
             let provider = try ProviderFactory.createProvider(for: model, configuration: config)
 
-            #expect(provider.modelId == "grok-4-fast-reasoning")
+            #expect(provider.modelId == "grok-4.3")
             #expect(provider.capabilities.supportsTools == true)
             #expect(provider.capabilities.supportsStreaming == true)
         }
@@ -81,20 +96,39 @@ struct ProviderSystemTests {
 
             let previousOpenAI = getenv("OPENAI_API_KEY").flatMap { String(cString: $0) }
             let previousAnthropic = getenv("ANTHROPIC_API_KEY").flatMap { String(cString: $0) }
+            let previousMiniMax = getenv("MINIMAX_API_KEY").flatMap { String(cString: $0) }
+            let previousAnthropicCompatible = getenv("ANTHROPIC_COMPATIBLE_API_KEY").flatMap { String(cString: $0) }
+            let previousGeneric = getenv("API_KEY").flatMap { String(cString: $0) }
             unsetenv("OPENAI_API_KEY")
             unsetenv("ANTHROPIC_API_KEY")
+            unsetenv("MINIMAX_API_KEY")
+            setenv("ANTHROPIC_COMPATIBLE_API_KEY", "generic-compatible-key", 1)
+            setenv("API_KEY", "generic-key", 1)
             defer {
                 if let previousOpenAI { setenv("OPENAI_API_KEY", previousOpenAI, 1) }
                 if let previousAnthropic { setenv("ANTHROPIC_API_KEY", previousAnthropic, 1) }
+                // swiftlint:disable:next statement_position
+                if let previousMiniMax { setenv("MINIMAX_API_KEY", previousMiniMax, 1) }
+                else { unsetenv("MINIMAX_API_KEY") }
+                if let previousAnthropicCompatible {
+                    setenv("ANTHROPIC_COMPATIBLE_API_KEY", previousAnthropicCompatible, 1)
+                } else {
+                    unsetenv("ANTHROPIC_COMPATIBLE_API_KEY")
+                }
+                if let previousGeneric { setenv("API_KEY", previousGeneric, 1) } else { unsetenv("API_KEY") }
             }
 
             #expect(throws: TachikomaError.self) {
-                try OpenAIProvider(model: .gpt4o, configuration: config)
+                try OpenAIProvider(model: .gpt55, configuration: config)
             }
 
             #expect(throws: TachikomaError.self) {
                 try AnthropicProvider(model: .opus4, configuration: config)
             }
+
+            #expect(throws: TachikomaError.self) {
+                try ProviderFactory.createProvider(for: .minimax(.m27), configuration: config)
+            }
         }
     }
 
@@ -102,17 +136,15 @@ struct ProviderSystemTests {
 
     @Test
     func `Model Capabilities - Vision Support`() {
-        #expect(Model.openai(.gpt4o).supportsVision == true)
-        #expect(Model.openai(.gpt4oMini).supportsVision == true)
-        #expect(Model.openai(.gpt41).supportsVision == false)
+        #expect(Model.openai(.gpt55).supportsVision == true)
+        #expect(Model.openai(.gpt5Mini).supportsVision == true)
+        #expect(Model.openai(.custom("text-only-openai")).supportsVision == false)
 
+        #expect(Model.anthropic(.fable5).supportsVision == true)
         #expect(Model.anthropic(.opus4).supportsVision == true)
-        #expect(Model.anthropic(.sonnet4).supportsVision == true)
+        #expect(Model.anthropic(.sonnet46).supportsVision == true)
 
-        #expect(Model.grok(.grok2Vision).supportsVision == true)
-        #expect(Model.grok(.grok2Image).supportsVision == true)
-        #expect(Model.grok(.grok2).supportsVision == false)
-        #expect(Model.grok(.grok4).supportsVision == false)
+        #expect(Model.grok(.grok43).supportsVision == true)
 
         #expect(Model.ollama(.llava).supportsVision == true)
         #expect(Model.ollama(.llama33).supportsVision == false)
@@ -121,13 +153,14 @@ struct ProviderSystemTests {
 
     @Test
     func `Model Capabilities - Tool Support`() {
-        #expect(Model.openai(.gpt4o).supportsTools == true)
-        #expect(Model.openai(.gpt41).supportsTools == true)
+        #expect(Model.openai(.gpt55).supportsTools == true)
+        #expect(Model.openai(.gpt55).supportsTools == true)
 
+        #expect(Model.anthropic(.fable5).supportsTools == true)
         #expect(Model.anthropic(.opus4).supportsTools == true)
-        #expect(Model.anthropic(.sonnet4).supportsTools == true)
+        #expect(Model.anthropic(.sonnet46).supportsTools == true)
 
-        #expect(Model.grok(.grok4).supportsTools == true)
+        #expect(Model.grok(.grok43).supportsTools == true)
 
         #expect(Model.ollama(.llama33).supportsTools == true)
         #expect(Model.ollama(.llava).supportsTools == false) // Vision models don't support tools
@@ -136,9 +169,28 @@ struct ProviderSystemTests {
 
     @Test
     func `Model Capabilities - Streaming Support`() {
-        #expect(Model.openai(.gpt4o).supportsStreaming == true)
+        #expect(Model.openai(.gpt55).supportsStreaming == true)
         #expect(Model.anthropic(.opus4).supportsStreaming == true)
-        #expect(Model.grok(.grok4).supportsStreaming == true)
+        #expect(Model.anthropic(.opus47).supportsStreaming == true)
+        #expect(Model.anthropic(.opus48).supportsStreaming == false)
+        #expect(Model.anthropic(.fable5).supportsStreaming == false)
+        #expect(Model.openRouter(modelId: "anthropic/claude-fable-5").supportsStreaming == false)
+        #expect(Model.openRouter(modelId: "fable5").supportsStreaming == false)
+        #expect(Model.openRouter(modelId: "anthropic/claude-opus-4-8").supportsStreaming == false)
+        #expect(Model.together(modelId: "claude-fable-5").supportsStreaming == false)
+        #expect(Model.openaiCompatible(
+            modelId: "anthropic/claude-fable-5",
+            baseURL: "https://example.test",
+        ).supportsStreaming == false)
+        #expect(Model.openaiCompatible(
+            modelId: "fable5",
+            baseURL: "https://example.test",
+        ).supportsStreaming == false)
+        #expect(Model.openaiCompatible(
+            modelId: "sonnet4-local",
+            baseURL: "https://example.test",
+        ).supportsStreaming == true)
+        #expect(Model.grok(.grok43).supportsStreaming == true)
         #expect(Model.ollama(.llama33).supportsStreaming == true)
     }
 

Tests/TachikomaTests/Providers/OpenAIResponsesProviderTests.swift

@@ -2,6 +2,11 @@ import Foundation
 #if canImport(FoundationNetworking)
 import FoundationNetworking
 #endif
+#if canImport(Darwin)
+import Darwin
+#else
+import Glibc
+#endif
 import Testing
 @testable import Tachikoma
 
@@ -18,8 +23,11 @@ struct OpenAIResponsesProviderTests {
         let config = self.openAIConfig()
 
         let gpt5Models: [LanguageModel.OpenAI] = [
-            .gpt52,
-            .gpt51,
+            .chatLatest,
+            .gpt55,
+            .gpt54,
+            .gpt54Mini,
+            .gpt54Nano,
             .gpt5,
             .gpt5Mini,
             .gpt5Nano,
@@ -39,8 +47,8 @@ struct OpenAIResponsesProviderTests {
     }
 
     @Test
-    func `GPT-5.1 text.verbosity parameter is set correctly`() throws {
-        // Test that the text.verbosity parameter is properly configured for GPT-5.1
+    func `GPT-5.5 text.verbosity parameter is set correctly`() throws {
+        // Test that the text.verbosity parameter is properly configured for GPT-5.5
         let config = self.openAIConfig()
 
         // Skip if no API key
@@ -49,7 +57,7 @@ struct OpenAIResponsesProviderTests {
         }
 
         let provider = try OpenAIResponsesProvider(
-            model: .gpt51,
+            model: .gpt55,
             configuration: config,
         )
 
@@ -64,26 +72,27 @@ struct OpenAIResponsesProviderTests {
 
         // We can't directly test the internal request building without making it public
         // But we can verify the provider is configured correctly
-        #expect(provider.modelId == "gpt-5.1")
+        #expect(provider.modelId == "gpt-5.5")
         #expect(provider.capabilities.supportsTools == true)
         #expect(provider.capabilities.supportsVision == true)
     }
 
     @Test
-    func `Reasoning models use Responses API`() throws {
-        // Test that reasoning-oriented models also use the OpenAIResponsesProvider
+    func `GPT-5 models use Responses API`() throws {
+        // Test that GPT-5 models use the OpenAIResponsesProvider
         let config = self.openAIConfig()
 
-        let reasoningModels: [LanguageModel.OpenAI] = [
-            .o4Mini,
-            .gpt52,
-            .gpt51,
+        let responsesModels: [LanguageModel.OpenAI] = [
+            .chatLatest,
+            .gpt55,
+            .gpt54,
+            .gpt54Mini,
+            .gpt54Nano,
             .gpt5,
             .gpt5Mini,
-            .gpt5Thinking,
         ]
 
-        for model in reasoningModels {
+        for model in responsesModels {
             let provider = try ProviderFactory.createProvider(
                 for: .openai(model),
                 configuration: config,
@@ -97,11 +106,10 @@ struct OpenAIResponsesProviderTests {
     }
 
     @Test
-    func `Legacy models use standard OpenAI provider`() throws {
-        // Test that non-GPT-5/reasoning models use the standard OpenAIProvider
+    func `Custom OpenAI models use standard OpenAI provider`() throws {
         let config = self.openAIConfig()
 
-        let legacyModels: [LanguageModel.OpenAI] = [.gpt4o, .gpt4oMini, .gpt41]
+        let legacyModels: [LanguageModel.OpenAI] = [.custom("custom-openai")]
 
         for model in legacyModels {
             let provider = try ProviderFactory.createProvider(
@@ -111,7 +119,7 @@ struct OpenAIResponsesProviderTests {
 
             #expect(
                 provider is OpenAIProvider,
-                "Legacy model \(model) should use OpenAIProvider",
+                "Custom model \(model) should use OpenAIProvider",
             )
         }
     }
@@ -196,6 +204,7 @@ struct OpenAIResponsesProviderTests {
             choices: nil,
             usage: nil,
             metadata: nil,
+            incompleteDetails: nil,
         )
 
         let providerResponse = try OpenAIResponsesProvider.convertToProviderResponse(response)
@@ -208,6 +217,159 @@ struct OpenAIResponsesProviderTests {
         #expect(providerResponse.finishReason == .toolCalls)
     }
 
+    @Test
+    func `GPT-5 incomplete content filter response maps finish reason`() throws {
+        let output = OpenAIResponsesResponse.ResponsesOutput(
+            id: "out_1",
+            type: "message",
+            status: "incomplete",
+            content: [
+                .init(type: "output_text", text: "blocked partial", toolCall: nil),
+            ],
+            role: "assistant",
+            toolCall: nil,
+        )
+
+        let response = try JSONDecoder().decode(OpenAIResponsesResponse.self, from: #require("""
+        {
+          "id": "resp_1",
+          "object": "response",
+          "created_at": 0,
+          "status": "incomplete",
+          "model": "gpt-5",
+          "output": [
+            {
+              "id": "out_1",
+              "type": "message",
+              "status": "incomplete",
+              "role": "assistant",
+              "content": [
+                { "type": "output_text", "text": "blocked partial" }
+              ]
+            }
+          ],
+          "incomplete_details": { "reason": "content_filter" }
+        }
+        """.data(using: .utf8)))
+
+        let providerResponse = try OpenAIResponsesProvider.convertToProviderResponse(response)
+
+        #expect(output.status == "incomplete")
+        #expect(providerResponse.text.isEmpty)
+        #expect(providerResponse.finishReason == .contentFilter)
+    }
+
+    @Test
+    func `GPT-5 incomplete content filter discards parsed tool calls`() throws {
+        let toolCall = OpenAIResponsesResponse.ResponsesToolCall(
+            id: "call_1",
+            type: "function",
+            function: .init(name: "see", arguments: "{\"mode\":\"screen\"}"),
+        )
+        let output = OpenAIResponsesResponse.ResponsesOutput(
+            id: "out_1",
+            type: "message",
+            status: "incomplete",
+            content: [
+                .init(type: "output_text", text: "blocked partial", toolCall: nil),
+                .init(type: "tool_call", text: nil, toolCall: toolCall),
+            ],
+            role: "assistant",
+            toolCall: nil,
+        )
+        let response = OpenAIResponsesResponse(
+            id: "resp_1",
+            object: "response",
+            createdAt: 0,
+            created: nil,
+            status: "incomplete",
+            model: "gpt-5",
+            output: [output],
+            choices: nil,
+            usage: nil,
+            metadata: nil,
+            incompleteDetails: .init(reason: "content_filter"),
+        )
+
+        let providerResponse = try OpenAIResponsesProvider.convertToProviderResponse(response)
+
+        #expect(providerResponse.text.isEmpty)
+        #expect(providerResponse.toolCalls == nil)
+        #expect(providerResponse.finishReason == .contentFilter)
+    }
+
+    @Test
+    func `GPT-5 completed refusal output maps to content filter`() throws {
+        let output = OpenAIResponsesResponse.ResponsesOutput(
+            id: "out_1",
+            type: "message",
+            status: "completed",
+            content: [
+                .init(type: "refusal", refusal: "I cannot help with that."),
+            ],
+            role: "assistant",
+            toolCall: nil,
+        )
+        let response = OpenAIResponsesResponse(
+            id: "resp_1",
+            object: "response",
+            createdAt: 0,
+            created: nil,
+            status: "completed",
+            model: "gpt-5",
+            output: [output],
+            choices: nil,
+            usage: nil,
+            metadata: nil,
+            incompleteDetails: nil,
+        )
+
+        let providerResponse = try OpenAIResponsesProvider.convertToProviderResponse(response)
+
+        #expect(providerResponse.text.isEmpty)
+        #expect(providerResponse.toolCalls == nil)
+        #expect(providerResponse.finishReason == .contentFilter)
+    }
+
+    @Test
+    func `Alternate choices content filter suppresses text and tool calls`() throws {
+        let response = try JSONDecoder().decode(OpenAIResponsesResponse.self, from: #require("""
+        {
+          "id": "chatcmpl_1",
+          "object": "chat.completion",
+          "created": 0,
+          "model": "gpt-5",
+          "choices": [
+            {
+              "index": 0,
+              "message": {
+                "role": "assistant",
+                "content": "blocked partial",
+                "tool_calls": [
+                  {
+                    "id": "call_1",
+                    "type": "function",
+                    "function": {
+                      "name": "see",
+                      "arguments": "{\\"mode\\":\\"screen\\"}"
+                    }
+                  }
+                ]
+              },
+              "finish_reason": "content_filter",
+              "logprobs": null
+            }
+          ]
+        }
+        """.data(using: .utf8)))
+
+        let providerResponse = try OpenAIResponsesProvider.convertToProviderResponse(response)
+
+        #expect(providerResponse.text.isEmpty)
+        #expect(providerResponse.toolCalls == nil)
+        #expect(providerResponse.finishReason == .contentFilter)
+    }
+
     @Test
     func `Responses provider hits /v1/responses and encodes body`() async throws {
         let config = TachikomaConfiguration(loadFromEnvironment: false)
@@ -239,6 +401,67 @@ struct OpenAIResponsesProviderTests {
         }
     }
 
+    @Test
+    func `chat-latest Responses payload omits GPT-5 reasoning controls`() async throws {
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setAPIKey("live-openai", for: .openai)
+
+        try await self.withMockedSession { request in
+            let body = try #require(Self.bodyData(from: request))
+            let json = try JSONSerialization.jsonObject(with: body) as? [String: Any]
+            #expect(json?["model"] as? String == "chat-latest")
+            #expect(json?["reasoning"] == nil)
+            #expect(json?["text"] == nil)
+
+            return NetworkMocking.jsonResponse(for: request, data: Self.responsesPayload(text: "pong"))
+        } operation: { session in
+            let provider = try OpenAIResponsesProvider(model: .chatLatest, configuration: config, session: session)
+            _ = try await provider.generateText(request: self.sampleRequest)
+        }
+    }
+
+    @Test
+    func `GPT-5 Chat Responses payload preserves model ID`() async throws {
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setAPIKey("live-openai", for: .openai)
+
+        try await self.withMockedSession { request in
+            let body = try #require(Self.bodyData(from: request))
+            let json = try JSONSerialization.jsonObject(with: body) as? [String: Any]
+            #expect(json?["model"] as? String == "gpt-5-chat-latest")
+            #expect(json?["reasoning"] == nil)
+            #expect(json?["text"] == nil)
+
+            return NetworkMocking.jsonResponse(for: request, data: Self.responsesPayload(text: "pong"))
+        } operation: { session in
+            let provider = try OpenAIResponsesProvider(
+                model: .gpt5ChatLatest,
+                configuration: config,
+                session: session,
+            )
+            #expect(provider.capabilities.contextLength == 128_000)
+            #expect(provider.capabilities.maxOutputTokens == 16384)
+            _ = try await provider.generateText(request: self.sampleRequest)
+        }
+    }
+
+    @Test
+    func `Responses provider resolves OAuth access token`() async throws {
+        try await self.withIsolatedAuthState {
+            try TKAuthManager.shared.setCredential(key: "OPENAI_ACCESS_TOKEN", value: "oauth-access-token")
+            let config = TachikomaConfiguration(loadFromEnvironment: false)
+
+            try await self.withMockedSession { request in
+                #expect(request.value(forHTTPHeaderField: "Authorization") == "Bearer oauth-access-token")
+                return NetworkMocking.jsonResponse(for: request, data: Self.responsesPayload(text: "oauth ok"))
+            } operation: { session in
+                let provider = try OpenAIResponsesProvider(model: .gpt5Mini, configuration: config, session: session)
+                let response = try await provider.generateText(request: self.sampleRequest)
+                #expect(response.text.contains("oauth ok"))
+            }
+        }
+    }
+
     @Test
     func `Responses payload uses data URL string for images`() async throws {
         let config = self.openAIConfig()
@@ -450,27 +673,235 @@ struct OpenAIResponsesProviderTests {
             #expect(request.url?.path == "/v1/responses")
             let payload = Self.responsesStreamPayload(chunks: [
                 Self.streamChunkJSON(content: "Hello", finishReason: nil),
-                Self.streamChunkJSON(content: "Hello world", finishReason: nil),
+                Self.streamChunkJSON(content: " world", finishReason: nil),
                 Self.streamChunkJSON(content: nil, finishReason: "stop"),
             ])
             return NetworkMocking.streamResponse(for: request, data: payload)
         } operation: { session in
-            let provider = try OpenAIResponsesProvider(model: .o4Mini, configuration: config, session: session)
+            let provider = try OpenAIResponsesProvider(model: .gpt55, configuration: config, session: session)
             let stream = try await provider.streamText(request: self.sampleRequest)
 
             var collected = ""
+            var receivedDone = false
             for try await delta in stream {
                 switch delta.type {
                 case .textDelta:
                     collected.append(delta.content ?? "")
                 case .done:
-                    break
+                    receivedDone = true
                 case .toolCall, .toolResult, .reasoning:
                     break
                 }
             }
 
             #expect(collected == "Hello world")
+            #expect(receivedDone)
+        }
+    }
+
+    @Test
+    func `Responses provider marks completed tool streams as tool calls`() async throws {
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setAPIKey("live-openai", for: .openai)
+
+        try await self.withMockedSession { request in
+            #expect(request.url?.path == "/v1/responses")
+            let payload = Self.responsesStreamPayload(chunks: [
+                Self.streamEventJSON([
+                    "type": "response.output_item.added",
+                    "item": [
+                        "id": "item_1",
+                        "type": "function_call",
+                        "name": "lookup",
+                    ],
+                ]),
+                Self.streamEventJSON([
+                    "type": "response.function_call_arguments.done",
+                    "item_id": "item_1",
+                    "arguments": #"{"query":"weather"}"#,
+                ]),
+                Self.streamEventJSON(["type": "response.completed"]),
+            ])
+            return NetworkMocking.streamResponse(for: request, data: payload)
+        } operation: { session in
+            let provider = try OpenAIResponsesProvider(model: .gpt55, configuration: config, session: session)
+            let stream = try await provider.streamText(request: self.sampleRequest)
+
+            var sawToolCall = false
+            var finishReason: FinishReason?
+            for try await delta in stream {
+                if delta.type == .toolCall {
+                    sawToolCall = true
+                }
+                if delta.type == .done {
+                    finishReason = delta.finishReason
+                }
+            }
+
+            #expect(sawToolCall)
+            #expect(finishReason == .toolCalls)
+        }
+    }
+
+    @Test
+    func `Responses provider maps incomplete content filter stream finish reason`() async throws {
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setAPIKey("live-openai", for: .openai)
+
+        try await self.withMockedSession { request in
+            #expect(request.url?.path == "/v1/responses")
+            let payload = Self.responsesStreamPayload(chunks: [
+                Self.streamChunkJSON(content: "partial", finishReason: nil),
+                Self.streamEventJSON([
+                    "type": "response.incomplete",
+                    "response": [
+                        "incomplete_details": ["reason": "content_filter"],
+                    ],
+                ]),
+            ])
+            return NetworkMocking.streamResponse(for: request, data: payload)
+        } operation: { session in
+            let provider = try OpenAIResponsesProvider(model: .gpt55, configuration: config, session: session)
+            let stream = try await provider.streamText(request: self.sampleRequest)
+
+            var collected = ""
+            var finishReason: FinishReason?
+            for try await delta in stream {
+                if case .textDelta = delta.type {
+                    collected.append(delta.content ?? "")
+                }
+                if delta.type == .done {
+                    finishReason = delta.finishReason
+                }
+            }
+
+            #expect(collected == "partial")
+            #expect(finishReason == .contentFilter)
+        }
+    }
+
+    @Test
+    func `Responses provider maps refusal stream events to content filter`() async throws {
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setAPIKey("live-openai", for: .openai)
+
+        try await self.withMockedSession { request in
+            #expect(request.url?.path == "/v1/responses")
+            let payload = Self.responsesStreamPayload(chunks: [
+                Self.streamEventJSON([
+                    "type": "response.refusal.delta",
+                    "delta": "no",
+                ]),
+                Self.streamEventJSON(["type": "response.refusal.done"]),
+                Self.streamEventJSON(["type": "response.completed"]),
+            ])
+            return NetworkMocking.streamResponse(for: request, data: payload)
+        } operation: { session in
+            let provider = try OpenAIResponsesProvider(model: .gpt55, configuration: config, session: session)
+            let stream = try await provider.streamText(request: self.sampleRequest)
+
+            var finishReason: FinishReason?
+            for try await delta in stream where delta.type == .done {
+                finishReason = delta.finishReason
+            }
+
+            #expect(finishReason == .contentFilter)
+        }
+    }
+
+    @Test
+    func `Responses provider throws on failed stream event`() async throws {
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setAPIKey("live-openai", for: .openai)
+
+        try await self.withMockedSession { request in
+            #expect(request.url?.path == "/v1/responses")
+            let payload = Self.responsesStreamPayload(chunks: [
+                Self.streamEventJSON([
+                    "type": "response.failed",
+                    "response": [
+                        "error": [
+                            "message": "stream failed after partial output",
+                        ],
+                    ],
+                ]),
+            ])
+            return NetworkMocking.streamResponse(for: request, data: payload)
+        } operation: { session in
+            let provider = try OpenAIResponsesProvider(model: .gpt55, configuration: config, session: session)
+            let stream = try await provider.streamText(request: self.sampleRequest)
+
+            do {
+                for try await _ in stream {}
+                Issue.record("Expected stream failure")
+            } catch let error as TachikomaError {
+                guard case let .apiError(message) = error else {
+                    Issue.record("Expected apiError, got \(error)")
+                    return
+                }
+                #expect(message.contains("response.failed"))
+                #expect(message.contains("stream failed after partial output"))
+            }
+        }
+    }
+
+    @Test
+    func `Responses provider throws on error stream event`() async throws {
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setAPIKey("live-openai", for: .openai)
+
+        try await self.withMockedSession { request in
+            #expect(request.url?.path == "/v1/responses")
+            let payload = Self.responsesStreamPayload(chunks: [
+                Self.streamEventJSON([
+                    "type": "error",
+                    "message": "top-level stream error",
+                ]),
+            ])
+            return NetworkMocking.streamResponse(for: request, data: payload)
+        } operation: { session in
+            let provider = try OpenAIResponsesProvider(model: .gpt55, configuration: config, session: session)
+            let stream = try await provider.streamText(request: self.sampleRequest)
+
+            do {
+                for try await _ in stream {}
+                Issue.record("Expected stream failure")
+            } catch let error as TachikomaError {
+                guard case let .apiError(message) = error else {
+                    Issue.record("Expected apiError, got \(error)")
+                    return
+                }
+                #expect(message.contains("error"))
+                #expect(message.contains("top-level stream error"))
+            }
+        }
+    }
+
+    @Test
+    func `chat-latest streams Responses event deltas`() async throws {
+        let config = TachikomaConfiguration(loadFromEnvironment: false)
+        config.setAPIKey("live-openai", for: .openai)
+
+        try await self.withMockedSession { request in
+            #expect(request.url?.path == "/v1/responses")
+            let payload = Self.responsesStreamPayload(chunks: [
+                Self.streamChunkJSON(content: "Hello", finishReason: nil),
+                Self.streamChunkJSON(content: " latest", finishReason: nil),
+                Self.streamChunkJSON(content: nil, finishReason: "stop"),
+            ])
+            return NetworkMocking.streamResponse(for: request, data: payload)
+        } operation: { session in
+            let provider = try OpenAIResponsesProvider(model: .chatLatest, configuration: config, session: session)
+            let stream = try await provider.streamText(request: self.sampleRequest)
+
+            var collected = ""
+            for try await delta in stream {
+                if case .textDelta = delta.type {
+                    collected.append(delta.content ?? "")
+                }
+            }
+
+            #expect(collected == "Hello latest")
         }
     }
 
@@ -544,33 +975,27 @@ struct OpenAIResponsesProviderTests {
     }
 
     private static func streamChunkJSON(content: String?, finishReason: String?) -> String {
-        var delta: [String: Any] = [
-            "role": "assistant",
-        ]
         if let content {
-            delta["content"] = content
-        }
-
-        var choice: [String: Any] = [
-            "index": 0,
-            "delta": delta,
-        ]
-        if let finishReason {
-            choice["finish_reason"] = finishReason
+            let chunk: [String: Any] = [
+                "type": "response.output_text.delta",
+                "delta": content,
+            ]
+            let data = try! JSONSerialization.data(withJSONObject: chunk)
+            return String(data: data, encoding: .utf8)!
         }
 
         let chunk: [String: Any] = [
-            "id": "resp_stream",
-            "object": "response",
-            "created": 1_700_000_000,
-            "model": "o4-mini",
-            "choices": [choice],
+            "type": finishReason == nil ? "response.output_text.done" : "response.completed",
         ]
-
         let data = try! JSONSerialization.data(withJSONObject: chunk)
         return String(data: data, encoding: .utf8)!
     }
 
+    private static func streamEventJSON(_ event: [String: Any]) -> String {
+        let data = try! JSONSerialization.data(withJSONObject: event)
+        return String(data: data, encoding: .utf8)!
+    }
+
     private func withMockedSession<T>(
         handler: @Sendable @escaping (URLRequest) throws -> (HTTPURLResponse, Data),
         operation: (URLSession) async throws -> T,
@@ -597,6 +1022,39 @@ struct OpenAIResponsesProviderTests {
             enableMockOverride: false,
         )
     }
+
+    private func withIsolatedAuthState<T: Sendable>(
+        _ body: @Sendable () async throws -> T,
+    ) async rethrows
+        -> T
+    {
+        try await TestEnvironmentMutex.shared.withLock {
+            let originalProfileDirectory = TachikomaConfiguration.profileDirectoryName
+            let profileDirectory = ".tachikoma-responses-auth-tests-\(UUID().uuidString)"
+            let profilePath = NSString(string: "~/" + profileDirectory).expandingTildeInPath
+            let previousIgnoreEnvironment = TKAuthManager.shared.setIgnoreEnvironment(false)
+            let previousIgnoreCredentialStore = TKAuthManager.shared.setIgnoreCredentialStore(false)
+            let savedOpenAIKey = getenv("OPENAI_API_KEY").map { String(cString: $0) }
+
+            TachikomaConfiguration.profileDirectoryName = profileDirectory
+            unsetenv("OPENAI_API_KEY")
+            try? FileManager.default.removeItem(atPath: profilePath)
+
+            defer {
+                if let savedOpenAIKey {
+                    setenv("OPENAI_API_KEY", savedOpenAIKey, 1)
+                } else {
+                    unsetenv("OPENAI_API_KEY")
+                }
+                TKAuthManager.shared.setIgnoreEnvironment(previousIgnoreEnvironment)
+                TKAuthManager.shared.setIgnoreCredentialStore(previousIgnoreCredentialStore)
+                TachikomaConfiguration.profileDirectoryName = originalProfileDirectory
+                try? FileManager.default.removeItem(atPath: profilePath)
+            }
+
+            return try await body()
+        }
+    }
 }
 
 private final class ResponsesTestURLProtocol: URLProtocol {

Tests/TachikomaTests/Providers/ProviderEndToEndTests.swift

@@ -57,7 +57,7 @@ struct ProviderEndToEndTests {
             let config = Self.makeConfiguration { config in
                 config.setAPIKey("sk-live-openai", for: .openai)
             }
-            let provider = try OpenAIProvider(model: .gpt4o, configuration: config)
+            let provider = try OpenAIProvider(model: .gpt55, configuration: config)
             let response = try await provider.generateText(request: Self.basicRequest)
             #expect(response.text == "OpenAI chat success")
         }
@@ -74,7 +74,7 @@ struct ProviderEndToEndTests {
             let config = Self.makeConfiguration { config in
                 config.setAPIKey("live-anthropic", for: .anthropic)
             }
-            let provider = try AnthropicProvider(model: .sonnet4, configuration: config)
+            let provider = try AnthropicProvider(model: .sonnet46, configuration: config)
             let response = try await provider.generateText(request: Self.basicRequest)
             #expect(response.text == "Claude says hello")
         }
@@ -97,21 +97,141 @@ struct ProviderEndToEndTests {
         }
     }
 
+    @Test
+    func `Google provider encodes tool results as user function responses`() async throws {
+        let toolCall = AgentToolCall(
+            id: "call_weather",
+            name: "get_weather",
+            arguments: ["location": AnyAgentToolValue(string: "Vienna")],
+        )
+        let toolResult = AgentToolResult.success(
+            toolCallId: "call_weather",
+            result: AnyAgentToolValue(object: ["temperature": AnyAgentToolValue(int: 21)]),
+        )
+        let providerRequest = ProviderRequest(
+            messages: [
+                .user("Weather?"),
+                ModelMessage(role: .assistant, content: [.text("Checking."), .toolCall(toolCall)]),
+                ModelMessage(role: .tool, content: [.toolResult(toolResult)]),
+            ],
+            settings: .init(maxTokens: 32),
+        )
+
+        try await NetworkMocking.withMockedNetwork { request in
+            let body = try #require(self.bodyData(from: request))
+            let json = try #require(JSONSerialization.jsonObject(with: body) as? [String: Any])
+            let contents = try #require(json["contents"] as? [[String: Any]])
+
+            #expect(contents.count == 3)
+            #expect(contents.compactMap { $0["role"] as? String } == ["user", "model", "user"])
+
+            let modelParts = try #require(contents[1]["parts"] as? [[String: Any]])
+            #expect(modelParts.count == 2)
+            #expect(modelParts[1]["functionCall"] != nil)
+
+            let toolParts = try #require(contents[2]["parts"] as? [[String: Any]])
+            let functionResponse = try #require(toolParts.first?["functionResponse"] as? [String: Any])
+            #expect(functionResponse["id"] as? String == "call_weather")
+            #expect(functionResponse["name"] as? String == "get_weather")
+
+            return NetworkMocking.streamResponse(for: request, data: Self.googleStreamPayload(text: "Done"))
+        } operation: {
+            let config = Self.makeConfiguration { config in
+                config.setAPIKey("google-live", for: .google)
+            }
+            let provider = try GoogleProvider(model: .gemini25Flash, configuration: config)
+            let response = try await provider.generateText(request: providerRequest)
+            #expect(response.text.contains("Done"))
+        }
+    }
+
+    @Test
+    func `Google provider drops orphan required tool parameters`() async throws {
+        let tool = AgentTool(
+            name: "search",
+            description: "Search files",
+            parameters: AgentToolParameters(
+                properties: [
+                    "query": AgentToolParameterProperty(
+                        name: "query",
+                        type: .string,
+                        description: "Search query",
+                    ),
+                ],
+                required: ["query", "mode"],
+            ),
+        ) { _ in
+            AnyAgentToolValue(string: "ok")
+        }
+        let orphanOnlyTool = AgentTool(
+            name: "noop",
+            description: "No-op",
+            parameters: AgentToolParameters(
+                properties: [
+                    "reason": AgentToolParameterProperty(
+                        name: "reason",
+                        type: .string,
+                        description: "Reason",
+                    ),
+                ],
+                required: ["missing"],
+            ),
+        ) { _ in
+            AnyAgentToolValue(string: "ok")
+        }
+
+        let providerRequest = ProviderRequest(
+            messages: [ModelMessage(role: .user, content: [.text("Find it")])],
+            tools: [tool, orphanOnlyTool],
+            settings: .init(maxTokens: 32),
+        )
+
+        try await NetworkMocking.withMockedNetwork { request in
+            let body = try #require(self.bodyData(from: request))
+            let json = try #require(JSONSerialization.jsonObject(with: body) as? [String: Any])
+            let tools = try #require(json["tools"] as? [[String: Any]])
+            let declarations = try #require(tools.first?["functionDeclarations"] as? [[String: Any]])
+            var parametersByName: [String: [String: Any]] = [:]
+            for declaration in declarations {
+                let name = try #require(declaration["name"] as? String)
+                let parameters = try #require(declaration["parameters"] as? [String: Any])
+                #expect(!parametersByName.keys.contains(name))
+                parametersByName[name] = parameters
+            }
+            let searchParameters = try #require(parametersByName["search"])
+            let noopParameters = try #require(parametersByName["noop"])
+
+            #expect(searchParameters["properties"] is [String: Any])
+            #expect(searchParameters["required"] as? [String] == ["query"])
+            #expect(noopParameters["properties"] is [String: Any])
+            #expect(noopParameters["required"] == nil)
+
+            return NetworkMocking.streamResponse(for: request, data: Self.googleStreamPayload(text: "Done"))
+        } operation: {
+            let config = Self.makeConfiguration { config in
+                config.setAPIKey("google-live", for: .google)
+            }
+            let provider = try GoogleProvider(model: .gemini25Flash, configuration: config)
+            let response = try await provider.generateText(request: providerRequest)
+            #expect(response.text.contains("Done"))
+        }
+    }
+
     // MARK: - OpenAI-compatible providers
 
     @Test
     func `Mistral provider uses OpenAI-compatible flow`() async throws {
-        try await self.assertOpenAICompatibleProvider(.mistral(.small), provider: .mistral)
+        try await self.assertOpenAICompatibleProvider(.mistral(.smallLatest), provider: .mistral)
     }
 
     @Test
     func `Groq provider uses OpenAI-compatible flow`() async throws {
-        try await self.assertOpenAICompatibleProvider(.groq(.llama38b), provider: .groq)
+        try await self.assertOpenAICompatibleProvider(.groq(.llama318b), provider: .groq)
     }
 
     @Test
     func `Grok provider uses OpenAI-compatible flow`() async throws {
-        try await self.assertOpenAICompatibleProvider(.grok(.grok4FastReasoning), provider: .grok)
+        try await self.assertOpenAICompatibleProvider(.grok(.grok43), provider: .grok)
     }
 
     @Test
@@ -204,6 +324,7 @@ struct ProviderEndToEndTests {
         try await NetworkMocking.withMockedNetwork { request in
             self.expectPath(request, endsWith: "/chat/completions")
             #expect(request.value(forHTTPHeaderField: "HTTP-Referer") == "https://peekaboo.app")
+            #expect(request.value(forHTTPHeaderField: "X-OpenRouter-Title") == "Peekaboo")
             return NetworkMocking.jsonResponse(for: request, data: Self.chatCompletionPayload(text: "OpenRouter reply"))
         } operation: {
             let config = Self.makeConfiguration { config in
@@ -292,6 +413,110 @@ struct ProviderEndToEndTests {
         }
     }
 
+    @Test
+    func `Anthropic-compatible provider accepts auth override`() async throws {
+        try await NetworkMocking.withMockedNetwork { request in
+            self.expectPath(request, endsWith: "/messages")
+            #expect(request.value(forHTTPHeaderField: "Authorization") == "Bearer compat-token")
+            #expect(request.value(forHTTPHeaderField: "x-api-key") == nil)
+            return NetworkMocking.jsonResponse(for: request, data: Self.anthropicPayload(text: "Compat bearer"))
+        } operation: {
+            let provider = try AnthropicCompatibleProvider(
+                modelId: "claude-compat-4",
+                baseURL: "https://compat.anthropic.test",
+                configuration: Self.makeConfiguration { _ in },
+                auth: .bearer("compat-token", betaHeader: nil),
+            )
+            let response = try await provider.generateText(request: Self.basicRequest)
+            #expect(response.text == "Compat bearer")
+        }
+    }
+
+    @Test
+    func `MiniMax provider uses bearer auth`() async throws {
+        try await NetworkMocking.withMockedNetwork { request in
+            self.expectPath(request, endsWith: "/messages")
+            #expect(request.value(forHTTPHeaderField: "Authorization") == "Bearer live-minimax")
+            #expect(request.value(forHTTPHeaderField: "x-api-key") == nil)
+            return NetworkMocking.jsonResponse(for: request, data: Self.anthropicPayload(text: "MiniMax ok"))
+        } operation: {
+            let config = Self.makeConfiguration { config in
+                config.setAPIKey("live-minimax", for: .minimax)
+            }
+            let provider = try ProviderFactory.createProvider(for: .minimax(.m27), configuration: config)
+            let response = try await provider.generateText(request: Self.basicRequest)
+            #expect(response.text == "MiniMax ok")
+        }
+    }
+
+    @Test
+    func `MiniMax reasoning metadata is bound to configured endpoint`() async throws {
+        let baseURL = "https://minimax-proxy.test/anthropic?tenant=a"
+        try await NetworkMocking.withMockedNetwork { request in
+            #expect(request.url?.host == "minimax-proxy.test")
+            #expect(request.value(forHTTPHeaderField: "Authorization") == "Bearer live-minimax")
+            return NetworkMocking.jsonResponse(
+                for: request,
+                data: Self.anthropicPayloadWithThinking(
+                    text: "MiniMax ok",
+                    thinking: "native-thought",
+                    signature: "sig-mm",
+                ),
+            )
+        } operation: {
+            let config = Self.makeConfiguration { config in
+                config.setAPIKey("live-minimax", for: .minimax)
+                config.setBaseURL(baseURL, for: .minimax)
+            }
+            let provider = try ProviderFactory.createProvider(for: .minimax(.m27), configuration: config)
+            let response = try await provider.generateText(request: Self.basicRequest)
+            let thinkingMessage = try #require(response.assistantMessages.first { $0.channel == .thinking })
+            let metadata = try #require(thinkingMessage.metadata?.customData)
+
+            #expect(metadata["tachikoma.reasoning.provider"] == "minimax")
+            #expect(metadata["tachikoma.reasoning.model"] == "MiniMax-M2.7")
+            #expect(metadata["anthropic.thinking.signature"] == "sig-mm")
+            #expect(metadata["tachikoma.reasoning.base_url"] == ReasoningEndpointIdentity.canonical(baseURL))
+        }
+    }
+
+    @Test
+    func `MiniMax China provider uses China endpoint and bearer auth`() async throws {
+        try await NetworkMocking.withMockedNetwork { request in
+            #expect(request.url?.host == "api.minimaxi.com")
+            self.expectPath(request, endsWith: "/messages")
+            #expect(request.value(forHTTPHeaderField: "Authorization") == "Bearer live-minimax-cn")
+            #expect(request.value(forHTTPHeaderField: "x-api-key") == nil)
+            return NetworkMocking.jsonResponse(for: request, data: Self.anthropicPayload(text: "MiniMax China ok"))
+        } operation: {
+            let config = Self.makeConfiguration { config in
+                config.setAPIKey("live-minimax-cn", for: .minimaxCN)
+            }
+            let provider = try ProviderFactory.createProvider(for: .minimaxCN(.m27), configuration: config)
+            let response = try await provider.generateText(request: Self.basicRequest)
+            #expect(response.text == "MiniMax China ok")
+        }
+    }
+
+    @Test
+    func `MiniMax China provider falls back to MiniMax API key`() async throws {
+        try await NetworkMocking.withMockedNetwork { request in
+            #expect(request.url?.host == "api.minimaxi.com")
+            #expect(request.value(forHTTPHeaderField: "Authorization") == "Bearer shared-minimax")
+            return NetworkMocking.jsonResponse(
+                for: request,
+                data: Self.anthropicPayload(text: "MiniMax China fallback ok"),
+            )
+        } operation: {
+            let config = Self.makeConfiguration { config in
+                config.setAPIKey("shared-minimax", for: .minimax)
+            }
+            let provider = try ProviderFactory.createProvider(for: .minimaxCN(.m27), configuration: config)
+            let response = try await provider.generateText(request: Self.basicRequest)
+            #expect(response.text == "MiniMax China fallback ok")
+        }
+    }
+
     // MARK: - Helpers
 
     private func assertOpenAICompatibleProvider(_ model: LanguageModel, provider: Provider) async throws {
@@ -362,7 +587,7 @@ struct ProviderEndToEndTests {
             "id": "chatcmpl-123",
             "object": "chat.completion",
             "created": 1_723_000_000,
-            "model": "gpt-4o",
+            "model": "gpt-5.5",
             "choices": [
                 [
                     "index": 0,
@@ -387,7 +612,26 @@ struct ProviderEndToEndTests {
             "content": [
                 ["type": "text", "text": text],
             ],
-            "model": "claude-sonnet-4-20250514",
+            "model": "claude-sonnet-4-6",
+            "stop_reason": "end_turn",
+            "usage": [
+                "input_tokens": 12,
+                "output_tokens": 6,
+            ],
+        ]
+        return try! JSONSerialization.data(withJSONObject: dict)
+    }
+
+    private static func anthropicPayloadWithThinking(text: String, thinking: String, signature: String) -> Data {
+        let dict: [String: Any] = [
+            "id": "msg_1",
+            "type": "message",
+            "role": "assistant",
+            "content": [
+                ["type": "thinking", "thinking": thinking, "signature": signature],
+                ["type": "text", "text": text],
+            ],
+            "model": "MiniMax-M2.7",
             "stop_reason": "end_turn",
             "usage": [
                 "input_tokens": 12,