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 --parallel $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:

AGENTS.md

@@ -4,5 +4,6 @@ READ ~/Projects/agent-scripts/{AGENTS.MD,TOOLS.MD} BEFORE ANYTHING (skip if file
 
 Tachikoma notes:
 - Keep this repo in sync with Peekaboo; bump the submodule there after changes.
+- Batch git network ops with Peekaboo: commit related changes first, then push/pull repos together so the submodule pointer never races the source repo.
 - Default workflow: `pnpm install`, `pnpm run lint`, `pnpm run test` before publishing.
 - Adapters live under `src/providers`; keep new providers consistent with existing patterns.

CHANGELOG.md

@@ -2,43 +2,58 @@
 
 All notable changes to the Tachikoma project will be documented in this file.
 
-## [0.1.0] - 2026-01-18
+## [Unreleased]
 
 ### Added
-- Core Swift 6 AI SDK with strict concurrency, streaming responses, and typed tool calling.
-- Unified message/content model (text, images, audio) with structured tool results.
-- Provider support for OpenAI (Chat + Responses), Anthropic, xAI (Grok), Google Gemini, Ollama, and OpenAI-compatible endpoints (OpenRouter/Together/Replicate).
-- Config system with credential store + env overrides, model registry, and capability lookup helpers.
-- Test helpers and mock infrastructure for deterministic provider/unit coverage.
+- 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.
 
-All notable changes to the Tachikoma project will be documented in this file.
+### 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`.
 
-The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
-and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+### 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.
 
-## [Unreleased]
+## [0.2.0] - 2026-04-28
 
 ### Added
 - First-class Azure OpenAI provider: deployment-based model case `.azureOpenAI`, Azure-specific URL/header/query wiring (api-version, api-key or bearer token), env overrides (`AZURE_OPENAI_API_KEY`, `AZURE_OPENAI_BEARER_TOKEN`, `AZURE_OPENAI_ENDPOINT`/`RESOURCE`, `AZURE_OPENAI_API_VERSION`), and README guidance.
 - 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.
 - `TachikomaConfiguration` can optionally override the provider factory so test harnesses can inject mock providers without affecting production defaults, improving hermetic test runs.
 - Implemented OpenRouter, Together, Replicate, and Anthropic-compatible providers on top of the shared helpers so aggregator models no longer throw “not yet implemented” errors and honour custom base URLs/headers.
 - `Provider.environmentValue` falls back to classic `getenv` lookups when the modern configuration reader returns no value, ensuring environment overrides succeed on macOS 14 deployments.
+- Provider environment reads now use direct process environment lookups, so test overrides and runtime unsets behave deterministically across SwiftPM/Xcode runs.
 
 ### Fixed
+- MCP bridge conversions now handle embedded resources and resource links from `swift-sdk` 0.11, and test helpers no longer swap mock keys for live environment credentials during provider/audio suites.
 - `retryWithCancellation` now registers token handlers per-attempt and cancels in-flight work, resolving hangs when external cancellation should short-circuit retries.
 - Audio provider tests and helpers consistently force mock mode when exercising stub audio payloads, preventing accidental live API calls that fail to decode fixtures.
 - `TestHelpers` expose discardable configuration helpers and stricter mock-key detection, reducing compiler warnings and flaky skips.
 - OpenAI transcription timestamp tests no longer hit the live API and succeed reliably under both mock and real key configurations.
+- Google provider API-key resolution no longer treats `GOOGLE_APPLICATION_CREDENTIALS` file paths as credential strings.
 - Anthropic OAuth login token exchange now uses the correct request format (JSON body + `state`). Thanks @jonathanglasmeyer.
 
 ### Testing
@@ -48,13 +63,14 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Added provider-level network E2E coverage using local `URLProtocol` stubs plus new OpenAI Responses API tests (request encoding + streaming) so critical serialization paths are exercised without live traffic.
 - `ProviderEndToEndTests` now exercise every provider flavor (OpenRouter/Together/Replicate, OpenAI/Anthropic compatible, etc.), pushing overall line coverage above 40 % while keeping the suite deterministic via URLProtocol stubs.
 
-### Planned Features
-- Enhanced caching with persistence and TTL
-- Bidirectional streaming support
-- Request batching for high-volume usage
-- Advanced error recovery mechanisms
-- Metrics collection and monitoring
-- Distributed caching support
+## [0.1.0] - 2026-01-18
+
+### Added
+- Core Swift 6 AI SDK with strict concurrency, streaming responses, and typed tool calling.
+- Unified message/content model (text, images, audio) with structured tool results.
+- Provider support for OpenAI (Chat + Responses), Anthropic, xAI (Grok), Google Gemini, Ollama, and OpenAI-compatible endpoints (OpenRouter/Together/Replicate).
+- Config system with credential store + env overrides, model registry, and capability lookup helpers.
+- Test helpers and mock infrastructure for deterministic provider/unit coverage.
 
 ## [1.0.0] - 2025-01-XX
 
@@ -71,8 +87,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 #### 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
@@ -80,7 +96,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 - **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
@@ -209,4 +225,4 @@ This is the initial release, so no breaking changes from previous versions.
 
 ## License
 
-This project is licensed under the MIT License. See LICENSE file for details.
+This project is licensed under the MIT License. See LICENSE file for details.

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 |
@@ -48,7 +48,7 @@ Set API keys for your providers:
 ```bash
 export OPENAI_API_KEY='sk-...'         # OpenAI models
 export ANTHROPIC_API_KEY='sk-ant-...'  # Claude models
-export GEMINI_API_KEY='...'            # Gemini models (legacy GOOGLE_API_KEY or GOOGLE_APPLICATION_CREDENTIALS also accepted)
+export GEMINI_API_KEY='...'            # Gemini models (legacy GOOGLE_API_KEY also accepted)
 export MISTRAL_API_KEY='...'           # Mistral models
 export GROQ_API_KEY='gsk-...'          # Groq models
 export X_AI_API_KEY='xai-...'          # Grok models
@@ -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"
 
@@ -213,63 +213,61 @@ 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:
         Set the appropriate environment variable for your provider:
           • OPENAI_API_KEY for OpenAI models
           • ANTHROPIC_API_KEY for Claude models
-          • GEMINI_API_KEY for Gemini models (legacy GOOGLE_API_KEY / GOOGLE_APPLICATION_CREDENTIALS also accepted)
+          • GEMINI_API_KEY for Gemini models (legacy GOOGLE_API_KEY also accepted)
           • MISTRAL_API_KEY for Mistral models
           • GROQ_API_KEY for Groq models
-          • X_AI_API_KEY or XAI_API_KEY for Grok models
+          • X_AI_API_KEY, XAI_API_KEY, or GROK_API_KEY for Grok models
           • Ollama requires local installation (no API key needed)
 
         For detailed documentation, visit: https://github.com/steipete/tachikoma
@@ -314,7 +312,7 @@ struct AICLI {
         self.checkAPIKeyStatus(provider: "Google", envVars: ["GEMINI_API_KEY", "GOOGLE_API_KEY"])
         self.checkAPIKeyStatus(provider: "Mistral", envVars: ["MISTRAL_API_KEY"])
         self.checkAPIKeyStatus(provider: "Groq", envVars: ["GROQ_API_KEY"])
-        self.checkAPIKeyStatus(provider: "Grok", envVars: ["X_AI_API_KEY", "XAI_API_KEY"])
+        self.checkAPIKeyStatus(provider: "Grok", envVars: ["X_AI_API_KEY", "XAI_API_KEY", "GROK_API_KEY"])
 
         // Ollama status
         print("   • Ollama: Local (no API key required)")
@@ -352,7 +350,7 @@ struct AICLI {
             let envVar = provider.environmentVariable.isEmpty ? "API key" : provider.environmentVariable
             if provider == .grok {
                 // Special case for Grok with alternative variables
-                throw CLIError.missingAPIKey("X_AI_API_KEY or XAI_API_KEY")
+                throw CLIError.missingAPIKey("X_AI_API_KEY, XAI_API_KEY, or GROK_API_KEY")
             } else {
                 throw CLIError.missingAPIKey(envVar)
             }
@@ -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
@@ -405,9 +405,15 @@ struct AICLI {
             print("export GEMINI_API_KEY='gk-your-key-here'")
             print("# Legacy names still supported:")
             print("export GOOGLE_API_KEY='gk-your-key-here'")
-            print("# or service-account path:")
-            print("export GOOGLE_APPLICATION_CREDENTIALS='/path/to/service-account.json'")
             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'")
@@ -421,6 +427,7 @@ struct AICLI {
             print("export X_AI_API_KEY='xai-your-key-here'")
             print("# or alternatively:")
             print("export XAI_API_KEY='xai-your-key-here'")
+            print("export GROK_API_KEY='xai-your-key-here'")
             print("Get your key at: https://console.x.ai/")
         case .ollama:
             print("Install Ollama locally:")
@@ -474,7 +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 {
@@ -586,15 +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
@@ -673,7 +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
@@ -802,23 +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:
@@ -826,6 +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"
 ```
 
@@ -220,7 +217,7 @@ Set API keys for providers:
 ```bash
 export OPENAI_API_KEY='sk-...'
 export ANTHROPIC_API_KEY='sk-ant-...'
-export GEMINI_API_KEY='...'   # Legacy GOOGLE_API_KEY / GOOGLE_APPLICATION_CREDENTIALS still work
+export GEMINI_API_KEY='...'   # Legacy GOOGLE_API_KEY still works
 export MISTRAL_API_KEY='...'
 export GROQ_API_KEY='gsk-...'
 export X_AI_API_KEY='xai-...'

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")
@@ -301,8 +301,7 @@ struct AgentCLI: AsyncParsableCommand {
     private func loadMessages(from path: String) async throws -> [ModelMessage] {
         let url = URL(fileURLWithPath: path)
         let data = try Data(contentsOf: url)
-        let messages = try JSONDecoder().decode([ModelMessage].self, from: data)
-        return messages
+        return try JSONDecoder().decode([ModelMessage].self, from: data)
     }
 
     private func saveMessages(_ messages: [ModelMessage], to path: String) async throws {
@@ -501,10 +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")
         }
 
@@ -546,10 +548,10 @@ enum CLIError: LocalizedError {
     }
 }
 
-// Extension to add provider helpers
+/// 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 {
@@ -557,6 +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/DemoScript.swift

@@ -4,14 +4,14 @@ import Foundation
 
 // swift-sh Tachikoma ~> 1.0.0
 
-/// This demo script shows how to use Tachikoma in a real application
-/// Run with: swift-sh DemoScript.swift
-/// Or: swift run DemoScript (if added to Package.swift)
+// This demo script shows how to use Tachikoma in a real application
+// Run with: swift-sh DemoScript.swift
+// Or: swift run DemoScript (if added to Package.swift)
 
 print("🕷️  Tachikoma Demo Script")
 print("=" * 40)
 
-// Check environment variables directly
+/// Check environment variables directly
 let env = ProcessInfo.processInfo.environment
 print("Environment check:")
 for key in ["OPENAI_API_KEY", "ANTHROPIC_API_KEY", "X_AI_API_KEY"] {

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)")
@@ -106,7 +106,7 @@ func testRealtimeConfiguration() async throws {
     print("  swift run RealtimeVoiceAssistant --basic")
 }
 
-// Extension for string multiplication
+/// Extension for string multiplication
 extension String {
     static func * (string: String, count: Int) -> String {
         String(repeating: string, count: count)

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

LICENSE

@@ -1,6 +1,6 @@
 MIT License
 
-\g<1>2026 Peter Steinberger
+Copyright (c) 2026 Peter Steinberger
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal

Package.resolved

@@ -1,13 +1,22 @@
 {
-  "originHash" : "71bb7313f33f89ee4c3627d91346c2371cad8918a22c27c75a7b66f6ebde75ef",
+  "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",
       "location" : "https://github.com/mattt/eventsource.git",
       "state" : {
-        "revision" : "ca2a9d90cbe49e09b92f4b6ebd922c03ebea51d0",
-        "version" : "1.3.0"
+        "revision" : "a3a85a85214caf642abaa96ae664e4c772a59f6e",
+        "version" : "1.4.1"
       }
     },
     {
@@ -24,8 +33,8 @@
       "kind" : "remoteSourceControl",
       "location" : "https://github.com/apple/swift-asn1.git",
       "state" : {
-        "revision" : "810496cf121e525d660cd0ea89a758740476b85f",
-        "version" : "1.5.1"
+        "revision" : "a9a5efd40eaf558a2bcd48d64b1d1646be686008",
+        "version" : "1.7.1"
       }
     },
     {
@@ -33,17 +42,26 @@
       "kind" : "remoteSourceControl",
       "location" : "https://github.com/apple/swift-async-algorithms.git",
       "state" : {
-        "revision" : "6c050d5ef8e1aa6342528460db614e9770d7f804",
-        "version" : "1.1.1"
+        "revision" : "d0b4a06d0f173a2f3be27d3ea21b3c3aa18db440",
+        "version" : "1.1.4"
+      }
+    },
+    {
+      "identity" : "swift-atomics",
+      "kind" : "remoteSourceControl",
+      "location" : "https://github.com/apple/swift-atomics.git",
+      "state" : {
+        "revision" : "b601256eab081c0f92f059e12818ac1d4f178ff7",
+        "version" : "1.3.0"
       }
     },
     {
       "identity" : "swift-collections",
       "kind" : "remoteSourceControl",
-      "location" : "https://github.com/apple/swift-collections.git",
+      "location" : "https://github.com/apple/swift-collections",
       "state" : {
-        "revision" : "7b847a3b7008b2dc2f47ca3110d8c782fb2e5c7e",
-        "version" : "1.3.0"
+        "revision" : "a0cb0954ecb21e4e31b0070e6ed5674e8556685a",
+        "version" : "1.6.0"
       }
     },
     {
@@ -51,8 +69,8 @@
       "kind" : "remoteSourceControl",
       "location" : "https://github.com/apple/swift-configuration",
       "state" : {
-        "revision" : "3528deb75256d7dcbb0d71fa75077caae0a8c749",
-        "version" : "1.0.0"
+        "revision" : "be76c4ad929eb6c4bcaf3351799f2adf9e6848a9",
+        "version" : "1.2.0"
       }
     },
     {
@@ -60,8 +78,8 @@
       "kind" : "remoteSourceControl",
       "location" : "https://github.com/apple/swift-crypto.git",
       "state" : {
-        "revision" : "6f70fa9eab24c1fd982af18c281c4525d05e3095",
-        "version" : "4.2.0"
+        "revision" : "1b6b2e274e85105bfa155183145a1dcfd63331f1",
+        "version" : "4.5.0"
       }
     },
     {
@@ -69,8 +87,17 @@
       "kind" : "remoteSourceControl",
       "location" : "https://github.com/apple/swift-log.git",
       "state" : {
-        "revision" : "bc386b95f2a16ccd0150a8235e7c69eab2b866ca",
-        "version" : "1.8.0"
+        "revision" : "92448c359f00ebe36ae97d3bd9086f13c7692b5a",
+        "version" : "1.13.2"
+      }
+    },
+    {
+      "identity" : "swift-nio",
+      "kind" : "remoteSourceControl",
+      "location" : "https://github.com/apple/swift-nio.git",
+      "state" : {
+        "revision" : "a8e036cb8628fcc1ff67dfec6ce8168617172c9b",
+        "version" : "2.101.1"
       }
     },
     {
@@ -87,8 +114,8 @@
       "kind" : "remoteSourceControl",
       "location" : "https://github.com/modelcontextprotocol/swift-sdk.git",
       "state" : {
-        "revision" : "c0407a0b52677cb395d824cac2879b963075ba8c",
-        "version" : "0.10.2"
+        "revision" : "a0ae212ebf6eab5f754c3129608bc5557637e605",
+        "version" : "0.12.1"
       }
     },
     {
@@ -96,8 +123,8 @@
       "kind" : "remoteSourceControl",
       "location" : "https://github.com/swift-server/swift-service-lifecycle",
       "state" : {
-        "revision" : "1de37290c0ab3c5a96028e0f02911b672fd42348",
-        "version" : "2.9.1"
+        "revision" : "9829955b385e5bb88128b73f1b8389e9b9c3191a",
+        "version" : "2.11.0"
       }
     },
     {
@@ -105,8 +132,8 @@
       "kind" : "remoteSourceControl",
       "location" : "https://github.com/apple/swift-system.git",
       "state" : {
-        "revision" : "395a77f0aa927f0ff73941d7ac35f2b46d47c9db",
-        "version" : "1.6.3"
+        "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: [
@@ -49,9 +59,9 @@ 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: "0.10.2"),
+        .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"),
         .package(url: "https://github.com/apple/swift-algorithms", from: "1.2.1"),
         .package(url: "https://github.com/apple/swift-crypto.git", from: "4.2.0"),

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)")
             }
         }
@@ -485,7 +525,6 @@ extension ProviderFactory {
         -> any ModelProvider
     {
         // Create a provider with configuration
-        let provider = try createProvider(for: model, configuration: configuration)
-        return provider
+        try createProvider(for: model, configuration: configuration)
     }
 }

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)
 }
@@ -17,7 +18,7 @@ public final class CustomProviderRegistry: @unchecked Sendable {
 
     private init() {}
 
-    // Load from ~/.<profile>/config.json customProviders
+    /// Load from ~/.<profile>/config.json customProviders
     public func loadFromProfile() {
         guard let json = Self.loadRawConfigJSON() else { return }
         guard let cp = json["customProviders"] as? [String: Any] else { return }
@@ -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
     }
@@ -52,15 +61,12 @@ 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 { "\(self.profileDirectoryPath())/config.json" }
+    private static func profileConfigPath() -> String {
+        "\(self.profileDirectoryPath())/config.json"
+    }
 
     private static func stripJSONComments(from json: String) -> String {
         var result = ""

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)
         }
     }
 }
@@ -220,7 +220,7 @@ protocol EmbeddingProvider: Sendable {
 
 /// Request for embedding generation
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
-struct EmbeddingRequest: Sendable {
+struct EmbeddingRequest {
     let input: EmbeddingInput
     let settings: EmbeddingSettings
 }

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

@@ -1,4 +1,3 @@
-import Configuration
 #if canImport(Darwin)
 import Darwin
 #else
@@ -42,6 +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
 
@@ -63,6 +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"
@@ -79,6 +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"
@@ -95,6 +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"
@@ -105,8 +116,9 @@ public enum Provider: Sendable, Hashable, Codable {
     /// Alternative environment variable names (for compatibility)
     public var alternativeEnvironmentVariables: [String] {
         switch self {
-        case .grok: ["XAI_API_KEY"] // Additional Grok alias
-        case .google: ["GOOGLE_API_KEY", "GOOGLE_APPLICATION_CREDENTIALS"] // Backwards compatibility
+        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: []
         }
@@ -121,6 +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
@@ -140,7 +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
@@ -152,7 +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)
@@ -164,56 +180,21 @@ public enum Provider: Sendable, Hashable, Codable {
 
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
 extension Provider {
-    @available(macOS 15.0, iOS 18.0, watchOS 11.0, tvOS 18.0, *)
-    private static var environmentReader: ConfigReader {
-        enum Holder {
-            static let reader = ConfigReader(
-                provider: EnvironmentVariablesProvider(
-                    secretsSpecifier: .dynamic { key, _ in
-                        let lowercased = key.lowercased()
-                        return lowercased.contains("key") ||
-                            lowercased.contains("token") ||
-                            lowercased.contains("secret")
-                    },
-                ),
-            )
-        }
-        return Holder.reader
-    }
-
     /// Load API key from environment variables
     /// Checks primary environment variable first, then alternatives
     public func loadAPIKeyFromEnvironment() -> String? {
         // Check primary environment variable
-        if !self.environmentVariable.isEmpty {
-            if #available(macOS 15.0, iOS 18.0, watchOS 11.0, tvOS 18.0, *) {
-                if
-                    let key = Self.environmentReader.string(
-                        forKey: ConfigKey([self.environmentVariable]),
-                        isSecret: true,
-                    ),
-                    !key.isEmpty
-                {
-                    return key
-                }
-            } else if
-                let key = ProcessInfo.processInfo.environment[environmentVariable],
-                !key.isEmpty
-            {
-                return key
-            }
+        if
+            !self.environmentVariable.isEmpty,
+            let key = Self.processEnvironmentValue(for: self.environmentVariable),
+            !key.isEmpty
+        {
+            return key
         }
 
         // Check alternative environment variables
         for altVar in self.alternativeEnvironmentVariables {
-            if #available(macOS 15.0, iOS 18.0, watchOS 11.0, tvOS 18.0, *) {
-                if
-                    let key = Self.environmentReader.string(forKey: ConfigKey([altVar]), isSecret: true),
-                    !key.isEmpty
-                {
-                    return key
-                }
-            } else if let key = ProcessInfo.processInfo.environment[altVar], !key.isEmpty {
+            if let key = Self.processEnvironmentValue(for: altVar), !key.isEmpty {
                 return key
             }
         }
@@ -228,19 +209,12 @@ extension Provider {
 
     /// Read an environment value using the shared configuration reader.
     public static func environmentValue(for key: String, isSecret: Bool = false) -> String? {
-        // Read an environment value using the shared configuration reader.
-        if #available(macOS 15.0, iOS 18.0, watchOS 11.0, tvOS 18.0, *) {
-            if
-                let value = environmentReader.string(forKey: ConfigKey([key]), isSecret: isSecret),
-                !value.isEmpty
-            {
-                return value
-            }
-        }
+        _ = isSecret
+        return self.processEnvironmentValue(for: key)
+    }
 
-        guard let pointer = getenv(key) else {
-            return nil
-        }
+    private static func processEnvironmentValue(for key: String) -> String? {
+        guard let pointer = getenv(key) else { return nil }
         return String(cString: pointer)
     }
 }

Sources/Tachikoma/Core/ProviderFeatureParity.swift

@@ -74,7 +74,7 @@ public struct ProviderConfiguration: Sendable {
         self.customHeaders = customHeaders
     }
 
-    // Common configurations for providers
+    /// Common configurations for providers
     public static let openAI = ProviderConfiguration(
         maxTokens: 4096,
         maxContextLength: 128_000,
@@ -111,10 +111,21 @@ public final class ProviderAdapter: EnhancedModelProvider {
     private let baseProvider: ModelProvider
     public let configuration: ProviderConfiguration
 
-    public var modelId: String { self.baseProvider.modelId }
-    public var baseURL: String? { self.baseProvider.baseURL }
-    public var apiKey: String? { self.baseProvider.apiKey }
-    public var capabilities: ModelCapabilities { self.baseProvider.capabilities }
+    public var modelId: String {
+        self.baseProvider.modelId
+    }
+
+    public var baseURL: String? {
+        self.baseProvider.baseURL
+    }
+
+    public var apiKey: String? {
+        self.baseProvider.apiKey
+    }
+
+    public var capabilities: ModelCapabilities {
+        self.baseProvider.capabilities
+    }
 
     public init(provider: ModelProvider, configuration: ProviderConfiguration) {
         self.baseProvider = provider

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/ToolTypes.swift

@@ -27,9 +27,13 @@ public enum AgentValueType: String, Sendable, Codable, CaseIterable {
 // MARK: - Built-in Type Conformances
 
 extension String: AgentToolValue {
-    public static var agentValueType: AgentValueType { .string }
+    public static var agentValueType: AgentValueType {
+        .string
+    }
 
-    public func toJSON() throws -> Any { self }
+    public func toJSON() throws -> Any {
+        self
+    }
 
     public static func fromJSON(_ json: Any) throws -> Self {
         guard let string = json as? String else {
@@ -40,9 +44,13 @@ extension String: AgentToolValue {
 }
 
 extension Int: AgentToolValue {
-    public static var agentValueType: AgentValueType { .integer }
+    public static var agentValueType: AgentValueType {
+        .integer
+    }
 
-    public func toJSON() throws -> Any { self }
+    public func toJSON() throws -> Any {
+        self
+    }
 
     public static func fromJSON(_ json: Any) throws -> Self {
         if let int = json as? Int {
@@ -56,9 +64,13 @@ extension Int: AgentToolValue {
 }
 
 extension Double: AgentToolValue {
-    public static var agentValueType: AgentValueType { .number }
+    public static var agentValueType: AgentValueType {
+        .number
+    }
 
-    public func toJSON() throws -> Any { self }
+    public func toJSON() throws -> Any {
+        self
+    }
 
     public static func fromJSON(_ json: Any) throws -> Self {
         if let double = json as? Double {
@@ -72,9 +84,13 @@ extension Double: AgentToolValue {
 }
 
 extension Bool: AgentToolValue {
-    public static var agentValueType: AgentValueType { .boolean }
+    public static var agentValueType: AgentValueType {
+        .boolean
+    }
 
-    public func toJSON() throws -> Any { self }
+    public func toJSON() throws -> Any {
+        self
+    }
 
     public static func fromJSON(_ json: Any) throws -> Self {
         guard let bool = json as? Bool else {
@@ -87,11 +103,15 @@ extension Bool: AgentToolValue {
 /// Null value type
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
 public struct AgentNullValue: AgentToolValue, Equatable {
-    public static var agentValueType: AgentValueType { .null }
+    public static var agentValueType: AgentValueType {
+        .null
+    }
 
     public init() {}
 
-    public func toJSON() throws -> Any { NSNull() }
+    public func toJSON() throws -> Any {
+        NSNull()
+    }
 
     public static func fromJSON(_ json: Any) throws -> Self {
         if json is NSNull {
@@ -103,7 +123,9 @@ public struct AgentNullValue: AgentToolValue, Equatable {
 
 /// Array conformance
 extension Array: AgentToolValue where Element: AgentToolValue {
-    public static var agentValueType: AgentValueType { .array }
+    public static var agentValueType: AgentValueType {
+        .array
+    }
 
     public func toJSON() throws -> Any {
         try map { try $0.toJSON() }
@@ -119,7 +141,9 @@ extension Array: AgentToolValue where Element: AgentToolValue {
 
 /// Dictionary conformance
 extension Dictionary: AgentToolValue where Key == String, Value: AgentToolValue {
-    public static var agentValueType: AgentValueType { .object }
+    public static var agentValueType: AgentValueType {
+        .object
+    }
 
     public func toJSON() throws -> Any {
         var result: [String: Any] = [:]
@@ -158,7 +182,9 @@ public struct AnyAgentToolValue: AgentToolValue, Equatable, Codable {
         case object([String: AnyAgentToolValue])
     }
 
-    public static var agentValueType: AgentValueType { .object } // Generic type
+    public static var agentValueType: AgentValueType {
+        .object
+    } // Generic type
 
     public init(_ value: some AgentToolValue) throws {
         let json = try value.toJSON()

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
         }
@@ -246,7 +247,7 @@ public struct ModelMessage: Sendable, Codable, Equatable {
         self.metadata = metadata
     }
 
-    // Convenience initializers
+    /// Convenience initializers
     public static func system(_ text: String) -> ModelMessage {
         ModelMessage(role: .system, content: [.text(text)])
     }
@@ -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
@@ -361,12 +369,30 @@ 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
+/// Manual Codable conformance excluding non-codable stopConditions
 extension GenerationSettings: Codable {
     enum CodingKeys: String, CodingKey {
         case maxTokens
@@ -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
     }
 }
@@ -508,7 +538,7 @@ public struct TextStreamDelta: Sendable {
         self.finishReason = finishReason
     }
 
-    // Convenience constructors
+    /// Convenience constructors
     public static func text(_ content: String, channel: ResponseChannel? = nil) -> TextStreamDelta {
         TextStreamDelta(type: .textDelta, content: content, channel: channel)
     }
@@ -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/Core/UnifiedErrors.swift

@@ -151,7 +151,7 @@ public struct RecoverySuggestion: Sendable {
         self.helpURL = helpURL
     }
 
-    // Common recovery suggestions
+    /// Common recovery suggestions
     public static let checkAPIKey = RecoverySuggestion(
         suggestion: "Check that your API key is valid and has the necessary permissions",
         actions: [.validateAPIKey, .regenerateAPIKey],

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
         }
     }
 
@@ -194,7 +198,7 @@ enum AnthropicContent: Codable {
             try Self.encodeAnyDict(self.input, to: inputEncoder)
         }
 
-        // Helper methods for encoding/decoding [String: Any]
+        /// Helper methods for encoding/decoding [String: Any]
         private static func decodeAnyDict(from decoder: Decoder) throws -> [String: Any] {
             let container = try decoder.singleValueContainer()
             return try container.decode([String: JSONValue].self).mapValues { $0.value }
@@ -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/AzureOpenAIProvider.swift

@@ -9,8 +9,13 @@ public final class AzureOpenAIProvider: ModelProvider {
     public let modelId: String
     public let apiVersion: String
     public let capabilities: ModelCapabilities
-    public var baseURL: String? { self.resolvedBaseURL }
-    public var apiKey: String? { self.resolvedAPIKey }
+    public var baseURL: String? {
+        self.resolvedBaseURL
+    }
+
+    public var apiKey: String? {
+        self.resolvedAPIKey
+    }
 
     private let authHeaderName: String
     private let authHeaderValuePrefix: String

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

@@ -12,7 +12,9 @@ public final class GoogleProvider: ModelProvider {
     public let capabilities: ModelCapabilities
 
     private let model: LanguageModel.Google
-    private var apiModelName: String { self.model.apiModelId }
+    private var apiModelName: String {
+        self.model.apiModelId
+    }
 
     public init(model: LanguageModel.Google, configuration: TachikomaConfiguration) throws {
         self.model = model
@@ -212,16 +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(
@@ -251,6 +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 {
@@ -269,12 +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] = [
@@ -293,7 +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

@@ -16,11 +16,22 @@ public actor LMStudioProvider: ModelProvider {
     private let configuredModelId: String
     private let configuredCapabilities: ModelCapabilities
 
-    // Expose as optional for protocol conformance, but it's never actually nil
-    public nonisolated var baseURL: String? { self.actualBaseURL }
-    public nonisolated var apiKey: String? { self.configuredApiKey }
-    public nonisolated var modelId: String { self.configuredModelId }
-    public nonisolated var capabilities: ModelCapabilities { self.configuredCapabilities }
+    /// Expose as optional for protocol conformance, but it's never actually nil
+    public nonisolated var baseURL: String? {
+        self.actualBaseURL
+    }
+
+    public nonisolated var apiKey: String? {
+        self.configuredApiKey
+    }
+
+    public nonisolated var modelId: String {
+        self.configuredModelId
+    }
+
+    public nonisolated var capabilities: ModelCapabilities {
+        self.configuredCapabilities
+    }
 
     private let session: URLSession
     private let encoder = JSONEncoder()
@@ -30,7 +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,
     ) {
@@ -63,8 +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/OpenAIEmbeddingProvider.swift

@@ -12,10 +12,15 @@ struct OpenAIEmbeddingProvider: EmbeddingProvider, ModelProvider {
     let apiKey: String?
     let baseURL: String?
 
-    var modelId: String { self.model.rawValue }
-    var capabilities: ModelCapabilities { ModelCapabilities() }
+    var modelId: String {
+        self.model.rawValue
+    }
 
-    // ModelProvider conformance (not used for embeddings)
+    var capabilities: ModelCapabilities {
+        ModelCapabilities()
+    }
+
+    /// ModelProvider conformance (not used for embeddings)
     func generateText(request _: ProviderRequest) async throws -> ProviderResponse {
         throw TachikomaError.unsupportedOperation("Text generation not supported for embedding models")
     }
@@ -92,17 +97,25 @@ struct OpenAIEmbeddingProvider: EmbeddingProvider, ModelProvider {
     }
 }
 
-// Placeholder providers for other services
+/// Placeholder providers for other services
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
 struct CohereEmbeddingProvider: EmbeddingProvider, ModelProvider {
     let model: EmbeddingModel.CohereEmbedding
     let apiKey: String?
 
-    var modelId: String { self.model.rawValue }
-    var baseURL: String? { nil }
-    var capabilities: ModelCapabilities { ModelCapabilities() }
+    var modelId: String {
+        self.model.rawValue
+    }
 
-    // ModelProvider conformance (not used for embeddings)
+    var baseURL: String? {
+        nil
+    }
+
+    var capabilities: ModelCapabilities {
+        ModelCapabilities()
+    }
+
+    /// ModelProvider conformance (not used for embeddings)
     func generateText(request _: ProviderRequest) async throws -> ProviderResponse {
         throw TachikomaError.unsupportedOperation("Text generation not supported for embedding models")
     }
@@ -121,11 +134,19 @@ struct VoyageEmbeddingProvider: EmbeddingProvider, ModelProvider {
     let model: EmbeddingModel.VoyageEmbedding
     let apiKey: String?
 
-    var modelId: String { self.model.rawValue }
-    var baseURL: String? { nil }
-    var capabilities: ModelCapabilities { ModelCapabilities() }
+    var modelId: String {
+        self.model.rawValue
+    }
 
-    // ModelProvider conformance (not used for embeddings)
+    var baseURL: String? {
+        nil
+    }
+
+    var capabilities: ModelCapabilities {
+        ModelCapabilities()
+    }
+
+    /// ModelProvider conformance (not used for embeddings)
     func generateText(request _: ProviderRequest) async throws -> ProviderResponse {
         throw TachikomaError.unsupportedOperation("Text generation not supported for embedding models")
     }

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

@@ -11,7 +11,7 @@ struct OpenAIResponsesRequest: Encodable {
     let topP: Double?
     let maxOutputTokens: Int?
 
-    // Response format and text configuration
+    /// Response format and text configuration
     let text: TextConfig? // GPT-5 text configuration with verbosity
 
     // Tool configuration
@@ -28,13 +28,13 @@ 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
+    /// Truncation for long inputs
     let truncation: String?
 
-    // Streaming support
+    /// Streaming support
     let stream: Bool?
 
     enum CodingKeys: String, CodingKey {
@@ -62,7 +62,7 @@ struct OpenAIResponsesRequest: Encodable {
 
 /// Text verbosity levels for GPT-5
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
-enum TextVerbosity: String, Codable, Sendable {
+enum TextVerbosity: String, Codable {
     case low
     case medium
     case high
@@ -70,7 +70,7 @@ enum TextVerbosity: String, Codable, Sendable {
 
 /// Internal reasoning effort levels for OpenAI responses provider
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
-enum OpenAIReasoningEffort: String, Codable, Sendable {
+enum OpenAIReasoningEffort: String, Codable {
     case minimal
     case low
     case medium
@@ -79,7 +79,7 @@ enum OpenAIReasoningEffort: String, Codable, Sendable {
 
 /// Reasoning summary modes for reasoning models
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
-enum ReasoningSummary: String, Codable, Sendable {
+enum ReasoningSummary: String, Codable {
     case concise
     case detailed
     case auto
@@ -87,13 +87,13 @@ enum ReasoningSummary: String, Codable, Sendable {
 
 /// Text configuration for GPT-5 models
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
-struct TextConfig: Codable, Sendable {
+struct TextConfig: Codable {
     let verbosity: TextVerbosity?
 }
 
 /// Reasoning configuration for reasoning models
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
-struct ReasoningConfig: Codable, Sendable {
+struct ReasoningConfig: Codable {
     let effort: OpenAIReasoningEffort?
     let summary: ReasoningSummary?
 }
@@ -186,11 +186,11 @@ struct JSONSchemaFormat: Codable {
 
 /// Message format for Responses API
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
-struct ResponsesMessage: Codable, Sendable {
+struct ResponsesMessage: Codable {
     let role: String
     let content: ResponsesContent
 
-    enum ResponsesContent: Codable, Sendable {
+    enum ResponsesContent: Codable {
         case text(String)
         case parts([ResponsesContentPart])
 
@@ -218,7 +218,7 @@ struct ResponsesMessage: Codable, Sendable {
 
 /// Content part for multimodal messages
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
-struct ResponsesContentPart: Codable, Sendable {
+struct ResponsesContentPart: Codable {
     let type: String
     let text: String?
     /// OpenAI Responses API (GPT‑5.x) accepts `image_url` only as a string (URL or data URL).
@@ -226,7 +226,7 @@ struct ResponsesContentPart: Codable, Sendable {
     /// avoid 400s ("expected an image URL, but got an object instead").
     let imageUrl: ImageURL?
 
-    struct ImageURL: Codable, Sendable {
+    struct ImageURL: Codable {
         let url: String
         let detail: String?
     }
@@ -271,12 +271,12 @@ struct ResponsesContentPart: Codable, Sendable {
 
 /// Heterogeneous input entries supported by the Responses API
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
-enum ResponsesInputItem: Encodable, Sendable {
+enum ResponsesInputItem: Encodable {
     case message(ResponsesMessage)
     case functionCall(FunctionCall)
     case functionCallOutput(FunctionCallOutput)
 
-    struct FunctionCall: Encodable, Sendable {
+    struct FunctionCall: Encodable {
         let type: String = "function_call"
         let callId: String
         let name: String
@@ -290,7 +290,7 @@ enum ResponsesInputItem: Encodable, Sendable {
         }
     }
 
-    struct FunctionCallOutput: Encodable, Sendable {
+    struct FunctionCallOutput: Encodable {
         let type: String = "function_call_output"
         let callId: String
         let output: String
@@ -473,26 +473,32 @@ struct ResponsesTool: Codable {
 
 /// Response from OpenAI Responses API (GPT-5 format)
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
-struct OpenAIResponsesResponse: Codable, Sendable {
+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"
     }
 
-    // GPT-5 output format
-    struct ResponsesOutput: Codable, Sendable {
+    struct IncompleteDetails: Codable {
+        let reason: String?
+    }
+
+    /// GPT-5 output format
+    struct ResponsesOutput: Codable {
         let id: String
         let type: String
         let status: String?
@@ -510,21 +516,35 @@ struct OpenAIResponsesResponse: Codable, Sendable {
             case toolCall = "tool_call"
         }
 
-        struct OutputContent: Codable, Sendable {
+        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)
-    struct ResponsesChoice: Codable, Sendable {
+    /// Alternate choices format.
+    struct ResponsesChoice: Codable {
         let index: Int
         let message: ResponsesOutputMessage
         let finishReason: String?
@@ -538,7 +558,7 @@ struct OpenAIResponsesResponse: Codable, Sendable {
         }
     }
 
-    struct ResponsesOutputMessage: Codable, Sendable {
+    struct ResponsesOutputMessage: Codable {
         let role: String
         let content: String?
         let toolCalls: [ResponsesToolCall]?
@@ -552,18 +572,18 @@ struct OpenAIResponsesResponse: Codable, Sendable {
         }
     }
 
-    struct ResponsesToolCall: Codable, Sendable {
+    struct ResponsesToolCall: Codable {
         let id: String
         let type: String
         let function: ResponsesToolFunction
 
-        struct ResponsesToolFunction: Codable, Sendable {
+        struct ResponsesToolFunction: Codable {
             let name: String
             let arguments: String
         }
     }
 
-    struct ResponsesUsage: Codable, Sendable {
+    struct ResponsesUsage: Codable {
         let inputTokens: Int?
         let outputTokens: Int?
         let totalTokens: Int?
@@ -584,7 +604,7 @@ struct OpenAIResponsesResponse: Codable, Sendable {
             case outputTokensDetails = "output_tokens_details"
         }
 
-        struct TokenDetails: Codable, Sendable {
+        struct TokenDetails: Codable {
             let cachedTokens: Int?
 
             enum CodingKeys: String, CodingKey {
@@ -592,7 +612,7 @@ struct OpenAIResponsesResponse: Codable, Sendable {
             }
         }
 
-        struct OutputTokenDetails: Codable, Sendable {
+        struct OutputTokenDetails: Codable {
             let reasoningTokens: Int?
 
             enum CodingKeys: String, CodingKey {
@@ -601,7 +621,7 @@ struct OpenAIResponsesResponse: Codable, Sendable {
         }
     }
 
-    struct ResponsesMetadata: Codable, Sendable {
+    struct ResponsesMetadata: Codable {
         let responseId: String?
         let reasoningItemIds: [String]?
 
@@ -614,16 +634,16 @@ struct OpenAIResponsesResponse: Codable, Sendable {
 
 // 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, Sendable {
+struct OpenAIResponsesStreamChunk: Codable {
     let id: String
     let object: String
     let created: Int
     let model: String
     let choices: [StreamChoice]
 
-    struct StreamChoice: Codable, Sendable {
+    struct StreamChoice: Codable {
         let index: Int
         let delta: StreamDelta
         let finishReason: String?
@@ -635,7 +655,7 @@ struct OpenAIResponsesStreamChunk: Codable, Sendable {
         }
     }
 
-    struct StreamDelta: Codable, Sendable {
+    struct StreamDelta: Codable {
         let role: String?
         let content: String?
         let toolCalls: [StreamToolCall]?
@@ -647,13 +667,13 @@ struct OpenAIResponsesStreamChunk: Codable, Sendable {
         }
     }
 
-    struct StreamToolCall: Codable, Sendable {
+    struct StreamToolCall: Codable {
         let index: Int
         let id: String?
         let type: String?
         let function: StreamToolFunction?
 
-        struct StreamToolFunction: Codable, Sendable {
+        struct StreamToolFunction: Codable {
             let name: String?
             let arguments: 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
     }
 }
 
@@ -140,7 +155,7 @@ enum OpenAIChatMessageContent: Codable {
         let url: String
     }
 
-    // Provide custom Codable to match OpenAI schema (flattened objects with type field)
+    /// Provide custom Codable to match OpenAI schema (flattened objects with type field)
     enum CodingKeys: String, CodingKey {
         case type
         case text
@@ -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"
         }
     }
 
@@ -326,8 +344,8 @@ struct OpenAIErrorResponse: Codable {
     }
 }
 
-// Helper type for Either content
-enum Either<Left, Right>: Codable where Left: Codable, Right: Codable {
+/// Helper type for Either content
+enum Either<Left: Codable, Right: Codable>: Codable {
     case left(Left)
     case right(Right)
 

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/Tools/ToolExtensions.swift

@@ -97,9 +97,17 @@ public struct AnyAgentTool: Sendable {
     private let _schema: AgentToolSchema
     private let _execute: @Sendable (AnyAgentToolValue, ToolExecutionContext) async throws -> AnyAgentToolValue
 
-    public var name: String { self._name }
-    public var description: String { self._description }
-    public var schema: AgentToolSchema { self._schema }
+    public var name: String {
+        self._name
+    }
+
+    public var description: String {
+        self._description
+    }
+
+    public var schema: AgentToolSchema {
+        self._schema
+    }
 
     public init<T: AgentToolProtocol>(_ tool: T) {
         self._name = tool.name

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
     }
 }
 
@@ -546,17 +629,39 @@ 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 }
-    public var baseURL: String? { self.provider.baseURL }
-    public var apiKey: String? { self.provider.apiKey }
-    public var capabilities: ModelCapabilities { self.provider.capabilities }
+    public var modelId: String {
+        self.provider.modelId
+    }
+
+    public var baseURL: String? {
+        self.provider.baseURL
+    }
+
+    public var apiKey: String? {
+        self.provider.apiKey
+    }
+
+    public var capabilities: ModelCapabilities {
+        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
         }
 
@@ -564,7 +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

@@ -267,7 +267,7 @@ public struct UsageOperation: Sendable, Codable {
         self.type = type
     }
 
-    // For backward compatibility, provide a computed property to reconstruct model info
+    /// For backward compatibility, provide a computed property to reconstruct model info
     public var modelDescription: String {
         "\(self.providerName)/\(self.modelId)"
     }
@@ -393,10 +393,21 @@ public struct UsageReport: Sendable {
     public let endDate: Date
     public let sessions: [UsageSession]
 
-    public var totalSessions: Int { self.sessions.count }
-    public var totalOperations: Int { self.sessions.reduce(0) { $0 + $1.operations.count } }
-    public var totalTokens: Int { self.sessions.reduce(0) { $0 + $1.totalTokens } }
-    public var totalCost: Double { self.sessions.reduce(0) { $0 + $1.totalCost } }
+    public var totalSessions: Int {
+        self.sessions.count
+    }
+
+    public var totalOperations: Int {
+        self.sessions.reduce(0) { $0 + $1.operations.count }
+    }
+
+    public var totalTokens: Int {
+        self.sessions.reduce(0) { $0 + $1.totalTokens }
+    }
+
+    public var totalCost: Double {
+        self.sessions.reduce(0) { $0 + $1.totalCost }
+    }
 
     public let providerBreakdown: [String: ProviderUsage]
     public let modelBreakdown: [String: ModelUsage]
@@ -517,39 +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)
@@ -558,7 +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)
-        let trackedStream = AsyncThrowingStream<TextStreamDelta, Error> { continuation in
-            Task {
+        return AsyncThrowingStream<TextStreamDelta, Error> { continuation in
+            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 {
@@ -138,27 +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()
+            }
         }
-
-        return trackedStream
     }
 
     /// 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)
     }
 
@@ -171,7 +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
@@ -181,6 +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 {
@@ -539,3 +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
-        let processedStream = AsyncThrowingStream<String, Error> { continuation in
-            Task {
+        return AsyncThrowingStream<String, Error> { continuation in
+            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()
@@ -150,9 +414,17 @@ public final class Conversation: @unchecked Sendable {
                     continuation.finish(throwing: error)
                 }
             }
+            continuation.onTermination = { @Sendable _ in
+                producer.cancel()
+            }
         }
+    }
+}
 
-        return processedStream
+extension [ModelMessage] {
+    func droppingLastUserTurn() -> [ModelMessage] {
+        guard self.last?.role == .user else { return self }
+        return Array(self.dropLast())
     }
 }
 
@@ -163,6 +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
@@ -171,11 +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
@@ -191,8 +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,
         )
     }
 
@@ -221,6 +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/TachikomaAgent/ToolBuilder.swift

@@ -150,7 +150,7 @@ public func createTool(
     )
 }
 
-// Helper struct for type-safe tool creation (moved outside to avoid nesting in generic function)
+/// Helper struct for type-safe tool creation (moved outside to avoid nesting in generic function)
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
 private struct ConcreteAgentTool<I: AgentToolValue, O: AgentToolValue>: AgentToolProtocol {
     typealias Input = I

Sources/TachikomaAudio/AudioFunctions.swift

@@ -238,22 +238,30 @@ public func transcribeBatch(
         of: (Int, TranscriptionResult).self,
         returning: [TranscriptionResult].self,
     ) { group in
-        let semaphore = AsyncSemaphore(value: concurrency)
+        let limit = min(max(concurrency, 1), audioURLs.count)
+        var nextIndex = 0
 
-        for (index, url) in audioURLs.indexed() {
+        func submit(_ index: Int) {
+            let url = audioURLs[index]
             group.addTask {
-                await semaphore.wait()
-                defer { Task { await semaphore.signal() } }
-
                 let audio = try AudioData(contentsOf: url)
                 let result = try await transcribe(audio, using: model, language: language, configuration: configuration)
                 return (index, result)
             }
         }
 
+        while nextIndex < limit {
+            submit(nextIndex)
+            nextIndex += 1
+        }
+
         var results: [(Int, TranscriptionResult)] = []
-        for try await result in group {
+        while let result = try await group.next() {
             results.append(result)
+            if nextIndex < audioURLs.count {
+                submit(nextIndex)
+                nextIndex += 1
+            }
         }
 
         // Sort by original index to maintain order
@@ -282,21 +290,29 @@ public func generateSpeechBatch(
     -> [SpeechResult]
 {
     try await withThrowingTaskGroup(of: (Int, SpeechResult).self, returning: [SpeechResult].self) { group in
-        let semaphore = AsyncSemaphore(value: concurrency)
+        let limit = min(max(concurrency, 1), texts.count)
+        var nextIndex = 0
 
-        for (index, text) in texts.indexed() {
+        func submit(_ index: Int) {
+            let text = texts[index]
             group.addTask {
-                await semaphore.wait()
-                defer { Task { await semaphore.signal() } }
-
                 let result = try await generateSpeech(text, using: model, voice: voice, configuration: configuration)
                 return (index, result)
             }
         }
 
+        while nextIndex < limit {
+            submit(nextIndex)
+            nextIndex += 1
+        }
+
         var results: [(Int, SpeechResult)] = []
-        for try await result in group {
+        while let result = try await group.next() {
             results.append(result)
+            if nextIndex < texts.count {
+                submit(nextIndex)
+                nextIndex += 1
+            }
         }
 
         // Sort by original index to maintain order
@@ -378,34 +394,3 @@ public func capabilities(
 }
 
 // MARK: - Helper Types
-
-/// Simple semaphore for controlling concurrency
-@available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
-actor AsyncSemaphore {
-    private var value: Int
-    private var waiters: [CheckedContinuation<Void, Never>] = []
-
-    init(value: Int) {
-        self.value = value
-    }
-
-    func wait() async {
-        if self.value > 0 {
-            self.value -= 1
-            return
-        }
-
-        await withCheckedContinuation { continuation in
-            self.waiters.append(continuation)
-        }
-    }
-
-    func signal() {
-        if self.waiters.isEmpty {
-            self.value += 1
-        } else {
-            let waiter = self.waiters.removeFirst()
-            waiter.resume()
-        }
-    }
-}

Sources/TachikomaAudio/Realtime/Audio/AudioManager.swift

@@ -50,7 +50,7 @@ public final class RealtimeAudioManager: NSObject {
     private let audioSession = AVAudioSession.sharedInstance()
     #endif
 
-    // Audio format for API (24kHz PCM16)
+    /// Audio format for API (24kHz PCM16)
     private let apiFormat = AVAudioFormat(
         commonFormat: .pcmFormatInt16,
         sampleRate: 24000,
@@ -297,7 +297,7 @@ public final class RealtimeAudioManager: NSObject {
         #endif
     }
 
-    /// Set audio input device (macOS only)
+    // Set audio input device (macOS only)
     #if os(macOS)
     public func setInputDevice(_: AudioDeviceID) throws {
         // Note: Setting audio device on macOS requires more complex Core Audio API

Sources/TachikomaAudio/Realtime/Audio/AudioProcessor.swift

@@ -304,9 +304,7 @@ public final class RealtimeAudioProcessor: @unchecked Sendable {
         }
 
         let mantissa = UInt8((s >> (exponent + 3)) & 0x0F)
-        let encoded = ~(sign | (exponent << 4) | mantissa)
-
-        return encoded
+        return ~(sign | (exponent << 4) | mantissa)
     }
 
     private func decodeUlaw(_ ulaw: UInt8) -> Int16 {

Sources/TachikomaAudio/Realtime/Audio/AudioStreamPipeline.swift

@@ -275,7 +275,7 @@ public protocol AudioStreamPipelineDelegate: AnyObject {
     func audioStreamPipeline(didEncounterError error: Error) async
 }
 
-// Default implementation for optional methods
+/// Default implementation for optional methods
 @available(macOS 13.0, iOS 16.0, watchOS 9.0, tvOS 16.0, *)
 extension AudioStreamPipelineDelegate {
     // Pipeline started

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

@@ -135,7 +135,7 @@ public final class RealtimeConversation: ObservableObject {
     private var audioBuffer = Data()
     private let audioChunkSize = 1024 * 4 // 4KB chunks
 
-    // Background tasks
+    /// Background tasks
     private var eventProcessingTask: Task<Void, Never>?
 
     // MARK: - Initialization
@@ -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

@@ -39,7 +39,9 @@ public enum RealtimeClientEvent: RealtimeEventProtocol {
         }
     }
 
-    public var eventId: String? { UUID().uuidString }
+    public var eventId: String? {
+        UUID().uuidString
+    }
 }
 
 // MARK: - Server Events (Events sent from server to client)
@@ -111,7 +113,9 @@ public enum RealtimeServerEvent: RealtimeEventProtocol {
         }
     }
 
-    public var eventId: String? { nil }
+    public var eventId: String? {
+        nil
+    }
 }
 
 // MARK: - Session Events
@@ -184,7 +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/Realtime/Tools/AgentToolWrapper.swift

@@ -111,7 +111,7 @@ public struct AgentToolWrapper: RealtimeExecutableTool {
         }
     }
 
-    // Helper function to convert Any to AnyAgentToolValue
+    /// Helper function to convert Any to AnyAgentToolValue
     private func convertAnyToToolArgument(_ value: Any) -> AnyAgentToolValue {
         do {
             return try AnyAgentToolValue.fromJSON(value)
@@ -121,7 +121,7 @@ public struct AgentToolWrapper: RealtimeExecutableTool {
         }
     }
 
-    // Helper function to convert AnyAgentToolValue to Any
+    /// Helper function to convert AnyAgentToolValue to Any
     private func convertToolArgumentToAny(_ arg: AnyAgentToolValue) -> Any {
         do {
             return try arg.toJSON()

Sources/TachikomaAudio/Recording/AudioRecorder.swift

@@ -88,7 +88,7 @@ public final class AudioRecorder: ObservableObject, AudioRecorderProtocol {
     private let channels: AVAudioChannelCount = 1
     private let bitDepth: Int = 16
 
-    // Maximum recording duration (5 minutes by default)
+    /// Maximum recording duration (5 minutes by default)
     public var maxRecordingDuration: TimeInterval = 300
 
     // MARK: - Initialization
@@ -211,8 +211,7 @@ public final class AudioRecorder: ObservableObject, AudioRecorderProtocol {
             recordingURL = nil
         }
 
-        let audioData = try AudioData(contentsOf: url)
-        return audioData
+        return try AudioData(contentsOf: url)
     }
 
     /// Cancel recording without returning data
@@ -418,9 +417,17 @@ private let logger = Logger(label: "tachikoma.audio.recorder")
 public final class AudioRecorder: AudioRecorderProtocol {
     public init() {}
 
-    public var isRecording: Bool { false }
-    public var isAvailable: Bool { false }
-    public var recordingDuration: TimeInterval { 0 }
+    public var isRecording: Bool {
+        false
+    }
+
+    public var isAvailable: Bool {
+        false
+    }
+
+    public var recordingDuration: TimeInterval {
+        0
+    }
 
     public func startRecording() async throws {
         throw AudioRecordingError.audioEngineError("Audio recording is unavailable on this platform")

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
             }
         }
     }
@@ -39,8 +35,13 @@ public enum TranscriptionModel: Sendable, CustomStringConvertible {
         case whisperLargeV3Turbo = "whisper-large-v3-turbo"
         case distilWhisperLargeV3En = "distil-whisper-large-v3-en"
 
-        public var supportsTimestamps: Bool { true }
-        public var supportsLanguageDetection: Bool { true }
+        public var supportsTimestamps: Bool {
+            true
+        }
+
+        public var supportsLanguageDetection: Bool {
+            true
+        }
     }
 
     public enum Deepgram: String, CaseIterable, Sendable {
@@ -49,17 +50,30 @@ public enum TranscriptionModel: Sendable, CustomStringConvertible {
         case enhanced
         case base
 
-        public var supportsTimestamps: Bool { true }
-        public var supportsLanguageDetection: Bool { true }
-        public var supportsSummarization: Bool { true }
+        public var supportsTimestamps: Bool {
+            true
+        }
+
+        public var supportsLanguageDetection: Bool {
+            true
+        }
+
+        public var supportsSummarization: Bool {
+            true
+        }
     }
 
     public enum ElevenLabs: String, CaseIterable, Sendable {
         case scribeV1 = "scribe_v1"
         case scribeV1Experimental = "scribe_v1_experimental"
 
-        public var supportsTimestamps: Bool { false }
-        public var supportsLanguageDetection: Bool { true }
+        public var supportsTimestamps: Bool {
+            false
+        }
+
+        public var supportsLanguageDetection: Bool {
+            true
+        }
     }
 
     // MARK: - Model Properties
@@ -137,12 +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
             }
         }
 
@@ -160,8 +172,13 @@ public enum SpeechModel: Sendable, CustomStringConvertible {
         case multilingualV2 = "eleven_multilingual_v2"
         case englishV1 = "eleven_english_v1"
 
-        public var supportsVoiceCloning: Bool { true }
-        public var supportedFormats: [AudioFormat] { [.mp3, .wav, .pcm] }
+        public var supportsVoiceCloning: Bool {
+            true
+        }
+
+        public var supportedFormats: [AudioFormat] {
+            [.mp3, .wav, .pcm]
+        }
     }
 
     // MARK: - Model Properties
@@ -203,5 +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/TachikomaAudio/Types/AudioTypes.swift

@@ -98,7 +98,7 @@ public enum VoiceOption: Sendable, Hashable {
     case nova
     case shimmer
 
-    // Custom voice (provider-specific)
+    /// Custom voice (provider-specific)
     case custom(String)
 
     public var stringValue: String {

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/Bridge/MCPContentBridge.swift

@@ -0,0 +1,167 @@
+import Foundation
+import MCP
+import Tachikoma
+
+enum MCPContentBridge {
+    private struct ResourceLinkDetails {
+        let title: String?
+        let description: String?
+        let mimeType: String?
+    }
+
+    static func summary(for content: MCP.Tool.Content) -> String {
+        switch content {
+        case let .text(text, _, _):
+            return text
+        case let .image(data, mimeType, _, _):
+            return "[Image: \(mimeType), size: \(data.count) bytes]"
+        case let .resource(resource, _, _):
+            if let text = resource.text {
+                return text
+            }
+            if let blob = resource.blob {
+                return "[Resource: \(resource.uri), blob size: \(blob.count) bytes]"
+            }
+            return "[Resource: \(resource.uri)]"
+        case let .resourceLink(uri, _, _, _, mimeType, _):
+            if let mimeType {
+                return "[Resource Link: \(uri), type: \(mimeType)]"
+            }
+            return "[Resource Link: \(uri)]"
+        case let .audio(data, mimeType, _, _):
+            return "[Audio: \(mimeType), size: \(data.count) bytes]"
+        }
+    }
+
+    static func convert(_ content: MCP.Tool.Content) -> AnyAgentToolValue {
+        switch content {
+        case let .text(text, _, _):
+            AnyAgentToolValue(string: text)
+        case let .image(data, mimeType, _, _):
+            AnyAgentToolValue(object: [
+                "type": AnyAgentToolValue(string: "image"),
+                "mimeType": AnyAgentToolValue(string: mimeType),
+                "data": AnyAgentToolValue(string: data),
+            ])
+        case let .resource(resource, annotations, meta):
+            AnyAgentToolValue(object: self.resourceObject(
+                resource: resource,
+                annotations: annotations,
+                meta: meta,
+            ))
+        case let .resourceLink(uri, name, title, description, mimeType, annotations):
+            AnyAgentToolValue(object: self.resourceLinkObject(
+                uri: uri,
+                name: name,
+                details: ResourceLinkDetails(
+                    title: title,
+                    description: description,
+                    mimeType: mimeType,
+                ),
+                annotations: annotations,
+            ))
+        case let .audio(data, mimeType, _, _):
+            AnyAgentToolValue(object: [
+                "type": AnyAgentToolValue(string: "audio"),
+                "mimeType": AnyAgentToolValue(string: mimeType),
+                "data": AnyAgentToolValue(string: data),
+            ])
+        }
+    }
+
+    private static func resourceObject(
+        resource: Resource.Content,
+        annotations: Resource.Annotations?,
+        meta: Metadata?,
+    )
+        -> [String: AnyAgentToolValue]
+    {
+        var resourceDict: [String: AnyAgentToolValue] = [
+            "type": AnyAgentToolValue(string: "resource"),
+            "uri": AnyAgentToolValue(string: resource.uri),
+        ]
+
+        if let mimeType = resource.mimeType {
+            resourceDict["mimeType"] = AnyAgentToolValue(string: mimeType)
+        } else {
+            resourceDict["mimeType"] = AnyAgentToolValue(null: ())
+        }
+
+        if let text = resource.text {
+            resourceDict["text"] = AnyAgentToolValue(string: text)
+        } else {
+            resourceDict["text"] = AnyAgentToolValue(null: ())
+        }
+
+        if let blob = resource.blob {
+            resourceDict["blob"] = AnyAgentToolValue(string: blob)
+        }
+
+        if let annotations {
+            resourceDict["annotations"] = self.convert(annotations)
+        }
+
+        if let meta {
+            resourceDict["meta"] = self.convert(meta)
+        }
+
+        return resourceDict
+    }
+
+    private static func resourceLinkObject(
+        uri: String,
+        name: String,
+        details: ResourceLinkDetails,
+        annotations: Resource.Annotations?,
+    )
+        -> [String: AnyAgentToolValue]
+    {
+        var resourceDict: [String: AnyAgentToolValue] = [
+            "type": AnyAgentToolValue(string: "resourceLink"),
+            "uri": AnyAgentToolValue(string: uri),
+            "name": AnyAgentToolValue(string: name),
+        ]
+
+        if let title = details.title {
+            resourceDict["title"] = AnyAgentToolValue(string: title)
+        }
+
+        if let description = details.description {
+            resourceDict["description"] = AnyAgentToolValue(string: description)
+        }
+
+        if let mimeType = details.mimeType {
+            resourceDict["mimeType"] = AnyAgentToolValue(string: mimeType)
+        }
+
+        if let annotations {
+            resourceDict["annotations"] = self.convert(annotations)
+        }
+
+        return resourceDict
+    }
+
+    private static func convert(_ annotations: Resource.Annotations) -> AnyAgentToolValue {
+        var dict: [String: AnyAgentToolValue] = [:]
+
+        if let audience = annotations.audience {
+            dict["audience"] = AnyAgentToolValue(array: audience.map {
+                AnyAgentToolValue(string: $0.rawValue)
+            })
+        }
+
+        if let priority = annotations.priority {
+            dict["priority"] = AnyAgentToolValue(double: priority)
+        }
+
+        if let lastModified = annotations.lastModified {
+            dict["lastModified"] = AnyAgentToolValue(string: lastModified)
+        }
+
+        return AnyAgentToolValue(object: dict)
+    }
+
+    private static func convert(_ metadata: Metadata) -> AnyAgentToolValue {
+        AnyAgentToolValue(object: metadata.fields.mapValues { $0.toAnyAgentToolValue() })
+    }
+}

Sources/TachikomaMCP/Bridge/MCPToolAdapter.swift

@@ -185,7 +185,7 @@ public enum MCPToolAdapter {
     private static func convertResponse(_ response: ToolResponse) -> AnyAgentToolValue {
         if response.isError {
             let errorMessage = response.content.compactMap { content -> String? in
-                if case let .text(text) = content {
+                if case let .text(text, _, _) = content {
                     return text
                 }
                 return nil
@@ -249,34 +249,6 @@ public enum MCPToolAdapter {
 
     /// Convert MCP Tool.Content to AnyAgentToolValue
     private static func convertContent(_ content: MCP.Tool.Content) -> AnyAgentToolValue {
-        // Convert MCP Tool.Content to AnyAgentToolValue
-        switch content {
-        case let .text(text):
-            return AnyAgentToolValue(string: text)
-        case let .image(data, mimeType, _):
-            return AnyAgentToolValue(object: [
-                "type": AnyAgentToolValue(string: "image"),
-                "mimeType": AnyAgentToolValue(string: mimeType),
-                "data": AnyAgentToolValue(string: data),
-            ])
-        case let .resource(uri, mimeType, text):
-            var resourceDict: [String: AnyAgentToolValue] = [
-                "type": AnyAgentToolValue(string: "resource"),
-                "uri": AnyAgentToolValue(string: uri),
-                "mimeType": AnyAgentToolValue(string: mimeType),
-            ]
-            if let text {
-                resourceDict["text"] = AnyAgentToolValue(string: text)
-            } else {
-                resourceDict["text"] = AnyAgentToolValue(null: ())
-            }
-            return AnyAgentToolValue(object: resourceDict)
-        case let .audio(data, mimeType):
-            return AnyAgentToolValue(object: [
-                "type": AnyAgentToolValue(string: "audio"),
-                "mimeType": AnyAgentToolValue(string: mimeType),
-                "data": AnyAgentToolValue(string: data),
-            ])
-        }
+        MCPContentBridge.convert(content)
     }
 }

Sources/TachikomaMCP/Bridge/MCPToolProvider.swift

@@ -161,7 +161,7 @@ public final class MCPToolProvider: DynamicToolProvider {
         // If there's an error, return it as a string
         if response.isError {
             let errorMessage = response.content.compactMap { content -> String? in
-                if case let .text(text) = content {
+                if case let .text(text, _, _) = content {
                     return text
                 }
                 return nil
@@ -184,33 +184,6 @@ public final class MCPToolProvider: DynamicToolProvider {
     }
 
     private func convertContentToAnyAgentToolValue(_ content: MCP.Tool.Content) -> AnyAgentToolValue {
-        switch content {
-        case let .text(text):
-            return AnyAgentToolValue(string: text)
-        case let .image(data, mimeType, _):
-            return AnyAgentToolValue(object: [
-                "type": AnyAgentToolValue(string: "image"),
-                "mimeType": AnyAgentToolValue(string: mimeType),
-                "data": AnyAgentToolValue(string: data),
-            ])
-        case let .resource(uri, mimeType, text):
-            var resourceDict: [String: AnyAgentToolValue] = [
-                "type": AnyAgentToolValue(string: "resource"),
-                "uri": AnyAgentToolValue(string: uri),
-                "mimeType": AnyAgentToolValue(string: mimeType),
-            ]
-            if let text {
-                resourceDict["text"] = AnyAgentToolValue(string: text)
-            } else {
-                resourceDict["text"] = AnyAgentToolValue(null: ())
-            }
-            return AnyAgentToolValue(object: resourceDict)
-        case let .audio(data, mimeType):
-            return AnyAgentToolValue(object: [
-                "type": AnyAgentToolValue(string: "audio"),
-                "mimeType": AnyAgentToolValue(string: mimeType),
-                "data": AnyAgentToolValue(string: data),
-            ])
-        }
+        MCPContentBridge.convert(content)
     }
 }

Sources/TachikomaMCP/Bridge/TypeConversions.swift

@@ -114,7 +114,7 @@ extension ToolResponse {
         // If there's an error, return error message
         if isError {
             let errorMessage = content.compactMap { content -> String? in
-                if case let .text(text) = content {
+                if case let .text(text, _, _) = content {
                     return text
                 }
                 return nil
@@ -125,18 +125,7 @@ extension ToolResponse {
 
         // Convert the first content item to a result
         if let firstContent = content.first {
-            switch firstContent {
-            case let .text(text):
-                return AnyAgentToolValue(string: text)
-            case let .image(data, mimeType, _):
-                // For images, return a descriptive string
-                return AnyAgentToolValue(string: "[Image: \(mimeType), size: \(data.count) bytes]")
-            case let .resource(uri, _, text):
-                // For resources, return the text content if available
-                return AnyAgentToolValue(string: text ?? "[Resource: \(uri)]")
-            case let .audio(data, mimeType):
-                return AnyAgentToolValue(string: "[Audio: \(mimeType), size: \(data.count) bytes]")
-            }
+            return AnyAgentToolValue(string: MCPContentBridge.summary(for: firstContent))
         }
 
         // No content
@@ -148,7 +137,7 @@ extension ToolResponse {
         // If there's an error, return it as a string
         if isError {
             let errorMessage = content.compactMap { content -> String? in
-                if case let .text(text) = content {
+                if case let .text(text, _, _) = content {
                     return text
                 }
                 return nil
@@ -171,31 +160,6 @@ extension ToolResponse {
     }
 
     private func convertContentToAnyAgentToolValue(_ content: MCP.Tool.Content) -> AnyAgentToolValue {
-        switch content {
-        case let .text(text):
-            return AnyAgentToolValue(string: text)
-        case let .image(data, mimeType, _):
-            return AnyAgentToolValue(object: [
-                "type": AnyAgentToolValue(string: "image"),
-                "mimeType": AnyAgentToolValue(string: mimeType),
-                "data": AnyAgentToolValue(string: data),
-            ])
-        case let .resource(uri, mimeType, text):
-            var resourceDict: [String: AnyAgentToolValue] = [
-                "type": AnyAgentToolValue(string: "resource"),
-                "uri": AnyAgentToolValue(string: uri),
-                "mimeType": AnyAgentToolValue(string: mimeType),
-            ]
-            if let text {
-                resourceDict["text"] = AnyAgentToolValue(string: text)
-            }
-            return AnyAgentToolValue(object: resourceDict)
-        case let .audio(data, mimeType):
-            return AnyAgentToolValue(object: [
-                "type": AnyAgentToolValue(string: "audio"),
-                "mimeType": AnyAgentToolValue(string: mimeType),
-                "data": AnyAgentToolValue(string: data),
-            ])
-        }
+        MCPContentBridge.convert(content)
     }
 }

Sources/TachikomaMCP/Client/HTTPTransport.swift

@@ -4,7 +4,7 @@ import FoundationNetworking
 #endif
 import Logging
 
-// Actor to manage mutable state for Sendable conformance
+/// Actor to manage mutable state for Sendable conformance
 private actor HTTPTransportState {
     var urlSession: URLSession?
     var baseURL: URL?
@@ -18,10 +18,21 @@ private actor HTTPTransportState {
         self.headers = headers
     }
 
-    func getSession() -> URLSession? { self.urlSession }
-    func getBaseURL() -> URL? { self.baseURL }
-    func getTimeout() -> TimeInterval { self.requestTimeout }
-    func getHeaders() -> [String: String] { self.headers }
+    func getSession() -> URLSession? {
+        self.urlSession
+    }
+
+    func getBaseURL() -> URL? {
+        self.baseURL
+    }
+
+    func getTimeout() -> TimeInterval {
+        self.requestTimeout
+    }
+
+    func getHeaders() -> [String: String] {
+        self.headers
+    }
 }
 
 /// HTTP transport for MCP communication

Sources/TachikomaMCP/Client/MCPClient.swift

@@ -2,7 +2,7 @@ import Foundation
 import Logging
 import MCP
 
-// Shared JSON-RPC types for HTTP transport
+/// Shared JSON-RPC types for HTTP transport
 struct HTTPJSONRPCRequest<P: Encodable>: Encodable {
     let jsonrpc = "2.0"
     let method: String
@@ -56,7 +56,7 @@ public struct MCPServerConfig: Sendable, Codable {
     }
 }
 
-// Actor to manage mutable state for Sendable conformance
+/// Actor to manage mutable state for Sendable conformance
 private actor MCPClientState {
     var transport: (any MCPTransport)?
     var tools: [Tool] = []
@@ -258,7 +258,7 @@ struct InitializeParams: Codable {
     }
 }
 
-// Some servers use snake_case for protocol_version in initialize
+/// Some servers use snake_case for protocol_version in initialize
 struct InitializeParamsSnake: Codable {
     let protocolVersion: String
     let clientInfo: ClientInfo

Sources/TachikomaMCP/Client/MCPClientManager.swift

@@ -189,7 +189,7 @@ public final class TachikomaMCPClientManager {
         return all
     }
 
-    // Execute a tool on a specific server
+    /// Execute a tool on a specific server
     public func executeTool(
         serverName: String,
         toolName: String,
@@ -222,7 +222,9 @@ public final class TachikomaMCPClientManager {
 
     // MARK: Health/Info (lightweight)
 
-    public func getServerNames() -> [String] { Array(self.effectiveConfigs.keys).sorted() }
+    public func getServerNames() -> [String] {
+        Array(self.effectiveConfigs.keys).sorted()
+    }
 
     public func getServerInfo(name: String) async -> (config: MCPServerConfig, connected: Bool)? {
         guard let cfg = effectiveConfigs[name] else { return nil }
@@ -497,12 +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/Client/SSETransport.swift

@@ -5,7 +5,7 @@ import FoundationNetworking
 import Logging
 import MCP
 
-// Internal state for SSE transport
+/// Internal state for SSE transport
 private actor SSEState {
     var transport: HTTPClientTransport?
     var baseURL: URL?
@@ -16,29 +16,60 @@ private actor SSEState {
     var timeoutTasks: [Int: Task<Void, Never>] = [:]
     var requestTimeoutNs: UInt64 = 30_000_000_000 // default 30s
 
-    func setTransport(_ t: HTTPClientTransport?) { self.transport = t }
-    func getTransport() -> HTTPClientTransport? { self.transport }
-    func setBaseURL(_ url: URL?) { self.baseURL = url }
-    func setHeaders(_ h: [String: String]) { self.headers = h }
-    func setEndpoint(_ url: URL?) { self.endpointURL = url }
-    func getEndpoint() -> URL? { self.endpointURL }
-    func getBaseURL() -> URL? { self.baseURL }
+    func setTransport(_ t: HTTPClientTransport?) {
+        self.transport = t
+    }
 
-    func getNextId() -> Int { defer { nextId += 1 }
+    func getTransport() -> HTTPClientTransport? {
+        self.transport
+    }
+
+    func setBaseURL(_ url: URL?) {
+        self.baseURL = url
+    }
+
+    func setHeaders(_ h: [String: String]) {
+        self.headers = h
+    }
+
+    func setEndpoint(_ url: URL?) {
+        self.endpointURL = url
+    }
+
+    func getEndpoint() -> URL? {
+        self.endpointURL
+    }
+
+    func getBaseURL() -> URL? {
+        self.baseURL
+    }
+
+    func getNextId() -> Int {
+        defer { nextId += 1 }
         return self.nextId
     }
 
-    func addPending(_ id: Int, _ cont: CheckedContinuation<Data, Swift.Error>) { self.pendingRequests[id] = cont }
-    func removePending(_ id: Int) -> CheckedContinuation<Data, Swift.Error>? { self.pendingRequests
-        .removeValue(forKey: id)
+    func addPending(_ id: Int, _ cont: CheckedContinuation<Data, Swift.Error>) {
+        self.pendingRequests[id] = cont
+    }
+
+    func removePending(_ id: Int) -> CheckedContinuation<Data, Swift.Error>? {
+        self.pendingRequests
+            .removeValue(forKey: id)
     }
 
     func setTimeout(_ seconds: TimeInterval) {
         self.requestTimeoutNs = UInt64((seconds > 0 ? seconds : 30) * 1_000_000_000)
     }
 
-    func addTimeoutTask(_ id: Int, _ task: Task<Void, Never>) { self.timeoutTasks[id] = task }
-    func cancelTimeout(_ id: Int) { self.timeoutTasks.removeValue(forKey: id)?.cancel() }
+    func addTimeoutTask(_ id: Int, _ task: Task<Void, Never>) {
+        self.timeoutTasks[id] = task
+    }
+
+    func cancelTimeout(_ id: Int) {
+        self.timeoutTasks.removeValue(forKey: id)?.cancel()
+    }
+
     func cancelAll(_ error: Swift.Error) {
         for (_, c) in self.pendingRequests {
             c.resume(throwing: error)
@@ -99,7 +130,7 @@ public final class SSETransport: MCPTransport {
         await self.state.setTransport(nil)
     }
 
-    // Expose underlying swift-sdk HTTP transport for advanced usage
+    /// Expose underlying swift-sdk HTTP transport for advanced usage
     public func underlyingSDKTransport() async -> HTTPClientTransport? {
         await self.state.getTransport()
     }

Sources/TachikomaMCP/Client/StdioTransport.swift

@@ -8,7 +8,7 @@ import Logging
 import MCP
 import Tachikoma
 
-// Actor to manage mutable state for Sendable conformance
+/// Actor to manage mutable state for Sendable conformance
 private actor StdioTransportState {
     var process: Process?
     var inputPipe: Pipe?

Sources/TachikomaMCP/Protocol/MCPTool.swift

@@ -19,7 +19,7 @@ public protocol MCPTool: Sendable {
 
 /// Wrapper for tool arguments received from MCP
 public struct ToolArguments: Sendable {
-    // Execute the tool with the given arguments
+    /// Execute the tool with the given arguments
     private let raw: Value
 
     public init(raw: [String: Any]) {
@@ -174,7 +174,7 @@ private func ValueToAny(_ value: Value) -> Any {
     }
 }
 
-// Helper function to convert Any to Value
+/// Helper function to convert Any to Value
 private func convertToValue(_ value: Any) -> Value {
     switch value {
     case let string as String:
@@ -213,7 +213,7 @@ public struct ToolResponse: Sendable {
     public static func text(_ text: String, meta: Value? = nil) -> ToolResponse {
         // Create a text response
         ToolResponse(
-            content: [.text(text)],
+            content: [.text(text: text, annotations: nil, _meta: nil)],
             isError: false,
             meta: meta,
         )
@@ -223,7 +223,7 @@ public struct ToolResponse: Sendable {
     public static func error(_ message: String, meta: Value? = nil) -> ToolResponse {
         // Create an error response
         ToolResponse(
-            content: [.text(message)],
+            content: [.text(text: message, annotations: nil, _meta: nil)],
             isError: true,
             meta: meta,
         )
@@ -233,7 +233,7 @@ public struct ToolResponse: Sendable {
     public static func image(data: Data, mimeType: String = "image/png", meta: Value? = nil) -> ToolResponse {
         // Create an image response
         ToolResponse(
-            content: [.image(data: data.base64EncodedString(), mimeType: mimeType, metadata: nil)],
+            content: [.image(data: data.base64EncodedString(), mimeType: mimeType, annotations: nil, _meta: nil)],
             isError: false,
             meta: meta,
         )
@@ -250,5 +250,5 @@ public struct ToolResponse: Sendable {
     }
 }
 
-// Type alias for convenience
+/// Type alias for convenience
 public typealias MCPContent = MCP.Tool.Content

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

@@ -4,10 +4,9 @@ import Testing
 @testable import Tachikoma
 @testable import TachikomaMCP
 
-@Suite("MCP Client Tests")
 struct MCPClientTests {
-    @Test("MCPServerConfig initialization with all parameters")
-    func serverConfigFullInit() {
+    @Test
+    func `MCPServerConfig initialization with all parameters`() {
         let config = MCPServerConfig(
             transport: "stdio",
             command: "npx",
@@ -29,8 +28,8 @@ struct MCPClientTests {
         #expect(config.description == "Test server")
     }
 
-    @Test("MCPServerConfig initialization with minimal parameters")
-    func serverConfigMinimalInit() {
+    @Test
+    func `MCPServerConfig initialization with minimal parameters`() {
         let config = MCPServerConfig(command: "test")
 
         #expect(config.transport == "stdio")
@@ -43,8 +42,8 @@ struct MCPClientTests {
         #expect(config.description == nil)
     }
 
-    @Test("MCPClient initialization")
-    func clientInit() {
+    @Test
+    func `MCPClient initialization`() {
         let config = MCPServerConfig(
             command: "test-command",
             args: ["arg1"],
@@ -54,8 +53,8 @@ struct MCPClientTests {
         _ = MCPClient(name: "test-client", config: config)
     }
 
-    @Test("MCPError descriptions")
-    func errorDescriptions() {
+    @Test
+    func `MCPError descriptions`() {
         #expect(MCPError.serverDisabled.errorDescription == "MCP server is disabled")
         #expect(MCPError.notConnected.errorDescription == "MCP client is not connected")
         #expect(MCPError.invalidResponse.errorDescription == "Invalid response from MCP server")
@@ -64,8 +63,8 @@ struct MCPClientTests {
         #expect(MCPError.executionFailed("error").errorDescription == "Execution failed: error")
     }
 
-    @Test("ToolArguments convenience methods")
-    func toolArgumentsConvenienceMethods() {
+    @Test
+    func `ToolArguments convenience methods`() {
         let args = ToolArguments(raw: [
             "text": "hello",
             "number": 42,
@@ -106,8 +105,8 @@ struct MCPClientTests {
         #expect(emptyArgs.isEmpty == true)
     }
 
-    @Test("ToolArguments from Value")
-    func toolArgumentsFromValue() {
+    @Test
+    func `ToolArguments from Value`() {
         let value = Value.object([
             "name": .string("test"),
             "count": .int(42),
@@ -121,8 +120,8 @@ struct MCPClientTests {
         #expect(args.getBool("active") == true)
     }
 
-    @Test("ToolArguments raw dictionary preserves nested structures")
-    func toolArgumentsRawDictionary() {
+    @Test
+    func `ToolArguments raw dictionary preserves nested structures`() {
         let value = Value.object([
             "text": .string("hello"),
             "number": .int(5),
@@ -167,13 +166,13 @@ struct MCPClientTests {
         #expect(dictionary["none"] is NSNull)
     }
 
-    @Test("ToolResponse creation methods")
-    func toolResponseCreation() {
+    @Test
+    func `ToolResponse creation methods`() {
         // Text response
         let textResponse = ToolResponse.text("Success")
         #expect(textResponse.content.count == 1)
         #expect(textResponse.isError == false)
-        if case let .text(text) = textResponse.content.first {
+        if case .text(text: let text, annotations: _, _meta: _) = textResponse.content.first {
             #expect(text == "Success")
         } else {
             Issue.record("Expected text content")
@@ -182,7 +181,7 @@ struct MCPClientTests {
         // Error response
         let errorResponse = ToolResponse.error("Failed")
         #expect(errorResponse.isError == true)
-        if case let .text(text) = errorResponse.content.first {
+        if case .text(text: let text, annotations: _, _meta: _) = errorResponse.content.first {
             #expect(text == "Failed")
         } else {
             Issue.record("Expected text content")
@@ -192,7 +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 let .image(data, mimeType, _) = 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 {
@@ -201,14 +200,14 @@ struct MCPClientTests {
 
         // Multi-content response
         let multiResponse = ToolResponse.multiContent([
-            .text("Part 1"),
-            .text("Part 2"),
+            .text(text: "Part 1", annotations: nil, _meta: nil),
+            .text(text: "Part 2", annotations: nil, _meta: nil),
         ])
         #expect(multiResponse.content.count == 2)
     }
 
-    @Test("ToolResponse with metadata")
-    func toolResponseWithMetadata() {
+    @Test
+    func `ToolResponse with metadata`() {
         let meta = Value.object([
             "executionTime": .double(1.5),
             "status": .string("ok"),
@@ -225,15 +224,15 @@ struct MCPClientTests {
         }
     }
 
-    @Test("MCPToolProvider initialization")
-    func toolProviderInit() {
+    @Test
+    func `MCPToolProvider initialization`() {
         let config = MCPServerConfig(command: "test")
         let client = MCPClient(name: "test", config: config)
         _ = MCPToolProvider(client: client)
     }
 
-    @Test("MCPToolProvider as DynamicToolProvider")
-    func toolProviderAsDynamicToolProvider() {
+    @Test
+    func `MCPToolProvider as DynamicToolProvider`() {
         let config = MCPServerConfig(command: "test")
         let client = MCPClient(name: "test-server", config: config)
         let provider = MCPToolProvider(client: client)
@@ -242,8 +241,8 @@ struct MCPClientTests {
         _ = provider as DynamicToolProvider
     }
 
-    @Test("Tool metadata structure")
-    func toolMetadata() {
+    @Test
+    func `Tool metadata structure`() {
         let tool = MCP.Tool(
             name: "test-tool",
             description: "A test tool",
@@ -269,8 +268,8 @@ struct MCPClientTests {
         }
     }
 
-    @Test("Value encoding and decoding")
-    func valueEncodingDecoding() throws {
+    @Test
+    func `Value encoding and decoding`() throws {
         let originalValue = Value.object([
             "string": .string("test"),
             "number": .int(42),
@@ -292,8 +291,8 @@ struct MCPClientTests {
         #expect(originalValue == decodedValue)
     }
 
-    @Test("ToolArguments decoding")
-    func toolArgumentsDecoding() throws {
+    @Test
+    func `ToolArguments decoding`() throws {
         struct TestArgs: Decodable {
             let name: String
             let count: Int
@@ -341,40 +340,39 @@ private struct MockMCPTool: MCPTool {
     }
 }
 
-@Suite("Mock Tool Tests")
 struct MockToolTests {
-    @Test("Mock tool execution with valid arguments")
-    func mockToolValidExecution() async throws {
+    @Test
+    func `Mock tool execution with valid arguments`() async throws {
         let tool = MockMCPTool()
 
         let args = ToolArguments(raw: ["message": "Hello World"])
         let response = try await tool.execute(arguments: args)
 
         #expect(response.isError == false)
-        if case let .text(text) = response.content.first {
+        if case .text(text: let text, annotations: _, _meta: _) = response.content.first {
             #expect(text == "Received: Hello World")
         } else {
             Issue.record("Expected text response")
         }
     }
 
-    @Test("Mock tool execution with missing arguments")
-    func mockToolMissingArguments() async throws {
+    @Test
+    func `Mock tool execution with missing arguments`() async throws {
         let tool = MockMCPTool()
 
         let args = ToolArguments(raw: [:])
         let response = try await tool.execute(arguments: args)
 
         #expect(response.isError == true)
-        if case let .text(text) = response.content.first {
+        if case .text(text: let text, annotations: _, _meta: _) = response.content.first {
             #expect(text == "Missing required 'message' parameter")
         } else {
             Issue.record("Expected error text")
         }
     }
 
-    @Test("Mock tool schema validation")
-    func mockToolSchema() {
+    @Test
+    func `Mock tool schema validation`() {
         let tool = MockMCPTool()
 
         #expect(tool.name == "mock_tool")

Tests/TachikomaMCPTests/MCPToolAdapterTests.swift

@@ -3,10 +3,9 @@ import Tachikoma
 import Testing
 @testable import TachikomaMCP
 
-@Suite("MCP Tool Adapter Tests")
 struct MCPToolAdapterTests {
-    @Test("ToolArguments getString")
-    func toolArgumentsGetString() {
+    @Test
+    func `ToolArguments getString`() {
         let args = ToolArguments(raw: [
             "name": "Alice",
             "count": 42,
@@ -21,8 +20,8 @@ struct MCPToolAdapterTests {
         #expect(args.getString("missing") == nil)
     }
 
-    @Test("ToolArguments getNumber")
-    func toolArgumentsGetNumber() {
+    @Test
+    func `ToolArguments getNumber`() {
         let args = ToolArguments(raw: [
             "int": 42,
             "double": 3.14,
@@ -37,8 +36,8 @@ struct MCPToolAdapterTests {
         #expect(args.getNumber("missing") == nil)
     }
 
-    @Test("ToolArguments getInt")
-    func toolArgumentsGetInt() {
+    @Test
+    func `ToolArguments getInt`() {
         let args = ToolArguments(raw: [
             "int": 42,
             "double": 3.14,
@@ -53,8 +52,8 @@ struct MCPToolAdapterTests {
         #expect(args.getInt("missing") == nil)
     }
 
-    @Test("ToolArguments getBool")
-    func toolArgumentsGetBool() {
+    @Test
+    func `ToolArguments getBool`() {
         let args = ToolArguments(raw: [
             "bool": true,
             "stringTrue": "true",
@@ -75,8 +74,8 @@ struct MCPToolAdapterTests {
         #expect(args.getBool("missing") == nil)
     }
 
-    @Test("ToolArguments getStringArray")
-    func toolArgumentsGetStringArray() {
+    @Test
+    func `ToolArguments getStringArray`() {
         let args = ToolArguments(raw: [
             "array": ["a", "b", "c"],
             "mixed": ["string", 123, true],
@@ -89,8 +88,8 @@ struct MCPToolAdapterTests {
         #expect(args.getStringArray("missing") == nil)
     }
 
-    @Test("ToolArguments isEmpty")
-    func toolArgumentsIsEmpty() {
+    @Test
+    func `ToolArguments isEmpty`() {
         let emptyArgs = ToolArguments(raw: [:])
         #expect(emptyArgs.isEmpty == true)
 
@@ -98,40 +97,40 @@ struct MCPToolAdapterTests {
         #expect(nonEmptyArgs.isEmpty == false)
     }
 
-    @Test("ToolResponse text creation")
-    func toolResponseText() {
+    @Test
+    func `ToolResponse text creation`() {
         let response = ToolResponse.text("Hello, world!")
 
         #expect(response.isError == false)
         #expect(response.content.count == 1)
-        if case let .text(text) = response.content[0] {
+        if case .text(text: let text, annotations: _, _meta: _) = response.content[0] {
             #expect(text == "Hello, world!")
         } else {
             Issue.record("Expected text content")
         }
     }
 
-    @Test("ToolResponse error creation")
-    func toolResponseError() {
+    @Test
+    func `ToolResponse error creation`() {
         let response = ToolResponse.error("Something went wrong")
 
         #expect(response.isError == true)
         #expect(response.content.count == 1)
-        if case let .text(text) = response.content[0] {
+        if case .text(text: let text, annotations: _, _meta: _) = response.content[0] {
             #expect(text == "Something went wrong")
         } else {
             Issue.record("Expected text content")
         }
     }
 
-    @Test("ToolResponse image creation")
-    func toolResponseImage() {
+    @Test
+    func `ToolResponse image creation`() {
         let imageData = Data([0xFF, 0xD8, 0xFF]) // JPEG header
         let response = ToolResponse.image(data: imageData, mimeType: "image/jpeg")
 
         #expect(response.isError == false)
         #expect(response.content.count == 1)
-        if case let .image(data, mimeType, _) = 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/TachikomaMCPTests/SchemaBuilderTests.swift

@@ -2,10 +2,9 @@ import MCP
 import Testing
 @testable import TachikomaMCP
 
-@Suite("Schema Builder Tests")
 struct SchemaBuilderTests {
-    @Test("Build string schema")
-    func stringSchema() {
+    @Test
+    func `Build string schema`() {
         let schema = SchemaBuilder.string(
             description: "User name",
             enum: ["Alice", "Bob"],
@@ -32,8 +31,8 @@ struct SchemaBuilderTests {
         }
     }
 
-    @Test("Build boolean schema")
-    func booleanSchema() {
+    @Test
+    func `Build boolean schema`() {
         let schema = SchemaBuilder.boolean(
             description: "Is active",
             default: true,
@@ -49,8 +48,8 @@ struct SchemaBuilderTests {
         #expect(dict["default"] == .bool(true))
     }
 
-    @Test("Build number schema")
-    func numberSchema() {
+    @Test
+    func `Build number schema`() {
         let schema = SchemaBuilder.number(
             description: "Temperature",
             minimum: 0.0,
@@ -70,8 +69,8 @@ struct SchemaBuilderTests {
         #expect(dict["default"] == .double(25.0))
     }
 
-    @Test("Build integer schema")
-    func integerSchema() {
+    @Test
+    func `Build integer schema`() {
         let schema = SchemaBuilder.integer(
             description: "Count",
             minimum: 1,
@@ -91,8 +90,8 @@ struct SchemaBuilderTests {
         #expect(dict["default"] == .int(10))
     }
 
-    @Test("Build array schema")
-    func arraySchema() {
+    @Test
+    func `Build array schema`() {
         let itemSchema = SchemaBuilder.string()
         let schema = SchemaBuilder.array(
             items: itemSchema,
@@ -115,8 +114,8 @@ struct SchemaBuilderTests {
         #expect(dict["uniqueItems"] == .bool(true))
     }
 
-    @Test("Build object schema")
-    func objectSchema() {
+    @Test
+    func `Build object schema`() {
         let schema = SchemaBuilder.object(
             properties: [
                 "name": SchemaBuilder.string(description: "User name"),
@@ -151,8 +150,8 @@ struct SchemaBuilderTests {
         }
     }
 
-    @Test("Build nullable schema")
-    func testNullableSchema() {
+    @Test
+    func `Build nullable schema`() {
         let baseSchema = SchemaBuilder.string(description: "Optional value")
         let nullableSchema = SchemaBuilder.nullable(baseSchema)
 
@@ -168,8 +167,8 @@ struct SchemaBuilderTests {
         }
     }
 
-    @Test("Build oneOf schema")
-    func oneOfSchema() {
+    @Test
+    func `Build oneOf schema`() {
         let schemas = [
             SchemaBuilder.string(),
             SchemaBuilder.integer(),
@@ -191,8 +190,8 @@ struct SchemaBuilderTests {
         }
     }
 
-    @Test("Build anyOf schema")
-    func anyOfSchema() {
+    @Test
+    func `Build anyOf schema`() {
         let schemas = [
             SchemaBuilder.string(),
             SchemaBuilder.integer(),

Tests/TachikomaMCPTests/TypeConversionTests.swift

@@ -4,10 +4,9 @@ import Testing
 @testable import Tachikoma
 @testable import TachikomaMCP
 
-@Suite("Type Conversion Tests")
 struct TypeConversionTests {
-    @Test("AnyAgentToolValue to JSON conversion")
-    func anyAgentToolValueToJSON() throws {
+    @Test
+    func `AnyAgentToolValue to JSON conversion`() throws {
         // String
         let stringVal = AnyAgentToolValue(string: "hello")
         let stringJSON = try stringVal.toJSON()
@@ -50,8 +49,8 @@ struct TypeConversionTests {
         #expect(objectJSON?["count"] as? Int == 5)
     }
 
-    @Test("Any to AnyAgentToolValue conversion")
-    func anyToAnyAgentToolValue() throws {
+    @Test
+    func `Any to AnyAgentToolValue conversion`() {
         // String
         let stringVal = AnyAgentToolValue.from("hello")
         #expect(stringVal.stringValue == "hello")
@@ -99,8 +98,8 @@ struct TypeConversionTests {
         #expect(dateVal.stringValue != nil)
     }
 
-    @Test("AnyAgentToolValue to Value conversion")
-    func anyAgentToolValueToValue() {
+    @Test
+    func `AnyAgentToolValue to Value conversion`() {
         // String
         let stringVal = AnyAgentToolValue(string: "hello")
         let stringValue = stringVal.toValue()
@@ -151,8 +150,8 @@ struct TypeConversionTests {
         }
     }
 
-    @Test("Value to AnyAgentToolValue conversion")
-    func valueToAnyAgentToolValue() {
+    @Test
+    func `Value to AnyAgentToolValue conversion`() {
         // String
         let stringValue = Value.string("hello")
         let stringVal = stringValue.toAnyAgentToolValue()
@@ -203,8 +202,8 @@ struct TypeConversionTests {
         }
     }
 
-    @Test("ToolArguments initialization from AgentToolArguments")
-    func toolArgumentsFromAgentToolArguments() throws {
+    @Test
+    func `ToolArguments initialization from AgentToolArguments`() {
         let agentArgs = AgentToolArguments([
             "text": AnyAgentToolValue(string: "hello"),
             "number": AnyAgentToolValue(int: 42),
@@ -229,8 +228,8 @@ struct TypeConversionTests {
         }
     }
 
-    @Test("ToolResponse to AnyAgentToolValue conversion via toAgentToolResult")
-    func toolResponseToAgentToolResult() {
+    @Test
+    func `ToolResponse to AnyAgentToolValue conversion via toAgentToolResult`() {
         // Text response
         let textResponse = ToolResponse.text("Success message")
         let textResult = textResponse.toAgentToolResult()
@@ -257,8 +256,8 @@ struct TypeConversionTests {
         #expect(emptyResult.stringValue == "Success")
     }
 
-    @Test("ToolResponse to AnyAgentToolValue conversion")
-    func toolResponseToAnyAgentToolValue() {
+    @Test
+    func `ToolResponse to AnyAgentToolValue conversion`() {
         // Single text content
         let textResponse = ToolResponse.text("Hello")
         let textVal = textResponse.toAnyAgentToolValue()
@@ -266,8 +265,8 @@ struct TypeConversionTests {
 
         // Multiple content items
         let multiResponse = ToolResponse.multiContent([
-            .text("Part 1"),
-            .text("Part 2"),
+            .text(text: "Part 1", annotations: nil, _meta: nil),
+            .text(text: "Part 2", annotations: nil, _meta: nil),
         ])
         let multiVal = multiResponse.toAnyAgentToolValue()
         if let elements = multiVal.arrayValue {
@@ -280,7 +279,7 @@ struct TypeConversionTests {
 
         // Image content
         let imageResponse = ToolResponse(content: [
-            .image(data: "base64data", mimeType: "image/png", metadata: nil),
+            .image(data: "base64data", mimeType: "image/png", annotations: nil, _meta: nil),
         ])
         let imageVal = imageResponse.toAnyAgentToolValue()
         if let props = imageVal.objectValue {
@@ -293,7 +292,13 @@ struct TypeConversionTests {
 
         // Resource content
         let resourceResponse = ToolResponse(content: [
-            .resource(uri: "https://example.com", mimeType: "text/html", text: "content"),
+            .resource(
+                resource: .text(
+                    "content",
+                    uri: "https://example.com",
+                    mimeType: "text/html",
+                ),
+            ),
         ])
         let resourceVal = resourceResponse.toAnyAgentToolValue()
         if let props = resourceVal.objectValue {
@@ -304,10 +309,32 @@ struct TypeConversionTests {
         } else {
             Issue.record("Expected object for resource content")
         }
+
+        // Resource link content
+        let resourceLinkResponse = ToolResponse(content: [
+            .resourceLink(
+                uri: "https://example.com/help",
+                name: "help-doc",
+                title: "Help",
+                description: "Operator help",
+                mimeType: "text/markdown",
+            ),
+        ])
+        let resourceLinkVal = resourceLinkResponse.toAnyAgentToolValue()
+        if let props = resourceLinkVal.objectValue {
+            #expect(props["type"]?.stringValue == "resourceLink")
+            #expect(props["uri"]?.stringValue == "https://example.com/help")
+            #expect(props["name"]?.stringValue == "help-doc")
+            #expect(props["title"]?.stringValue == "Help")
+            #expect(props["description"]?.stringValue == "Operator help")
+            #expect(props["mimeType"]?.stringValue == "text/markdown")
+        } else {
+            Issue.record("Expected object for resource link content")
+        }
     }
 
-    @Test("Round-trip conversions")
-    func roundTripConversions() throws {
+    @Test
+    func `Round-trip conversions`() throws {
         // AnyAgentToolValue -> Value -> AnyAgentToolValue
         let originalVal = AnyAgentToolValue(object: [
             "string": AnyAgentToolValue(string: "test"),

Tests/TachikomaTests/AgentToolValueTests.swift

@@ -2,12 +2,11 @@ import Foundation
 import Testing
 @testable import Tachikoma
 
-@Suite("AgentToolValue Protocol System Tests")
 struct AgentToolValueTests {
     // MARK: - Basic Type Conformance Tests
 
-    @Test("String conforms to AgentToolValue")
-    func stringConformance() throws {
+    @Test
+    func `String conforms to AgentToolValue`() throws {
         let value = "Hello, World!"
         let json = try value.toJSON()
         #expect(json as? String == "Hello, World!")
@@ -17,8 +16,8 @@ struct AgentToolValueTests {
         #expect(String.agentValueType == .string)
     }
 
-    @Test("Int conforms to AgentToolValue")
-    func intConformance() throws {
+    @Test
+    func `Int conforms to AgentToolValue`() throws {
         let value = 42
         let json = try value.toJSON()
         #expect(json as? Int == 42)
@@ -32,8 +31,8 @@ struct AgentToolValueTests {
         #expect(fromDouble == 42)
     }
 
-    @Test("Double conforms to AgentToolValue")
-    func doubleConformance() throws {
+    @Test
+    func `Double conforms to AgentToolValue`() throws {
         let value = 3.141_59
         let json = try value.toJSON()
         #expect(json as? Double == 3.141_59)
@@ -47,8 +46,8 @@ struct AgentToolValueTests {
         #expect(fromInt == 42.0)
     }
 
-    @Test("Bool conforms to AgentToolValue")
-    func boolConformance() throws {
+    @Test
+    func `Bool conforms to AgentToolValue`() throws {
         let value = true
         let json = try value.toJSON()
         #expect(json as? Bool == true)
@@ -58,8 +57,8 @@ struct AgentToolValueTests {
         #expect(Bool.agentValueType == .boolean)
     }
 
-    @Test("AgentNullValue works correctly")
-    func testNullValue() throws {
+    @Test
+    func `AgentNullValue works correctly`() throws {
         let value = AgentNullValue()
         let json = try value.toJSON()
         #expect(json is NSNull)
@@ -69,8 +68,8 @@ struct AgentToolValueTests {
         #expect(AgentNullValue.agentValueType == .null)
     }
 
-    @Test("Array conforms to AgentToolValue")
-    func arrayConformance() throws {
+    @Test
+    func `Array conforms to AgentToolValue`() throws {
         let value = ["apple", "banana", "cherry"]
         let json = try value.toJSON()
         let jsonArray = json as? [Any]
@@ -82,8 +81,8 @@ struct AgentToolValueTests {
         #expect([String].agentValueType == .array)
     }
 
-    @Test("Dictionary conforms to AgentToolValue")
-    func dictionaryConformance() throws {
+    @Test
+    func `Dictionary conforms to AgentToolValue`() throws {
         let value = ["name": "John", "city": "NYC"]
         let json = try value.toJSON()
         let jsonDict = json as? [String: Any]
@@ -97,8 +96,8 @@ struct AgentToolValueTests {
 
     // MARK: - AnyAgentToolValue Tests
 
-    @Test("AnyAgentToolValue wraps basic types")
-    func anyAgentToolValueBasicTypes() throws {
+    @Test
+    func `AnyAgentToolValue wraps basic types`() {
         // String
         let stringValue = AnyAgentToolValue(string: "test")
         #expect(stringValue.stringValue == "test")
@@ -126,8 +125,8 @@ struct AgentToolValueTests {
         #expect(nullValue.stringValue == nil)
     }
 
-    @Test("AnyAgentToolValue wraps complex types")
-    func anyAgentToolValueComplexTypes() throws {
+    @Test
+    func `AnyAgentToolValue wraps complex types`() {
         // Array
         let array = [
             AnyAgentToolValue(string: "a"),
@@ -152,8 +151,8 @@ struct AgentToolValueTests {
         #expect(objectValue.objectValue?["active"]?.boolValue == true)
     }
 
-    @Test("AnyAgentToolValue JSON conversion")
-    func anyAgentToolValueJSON() throws {
+    @Test
+    func `AnyAgentToolValue JSON conversion`() throws {
         // Test fromJSON with various types
         let stringValue = try AnyAgentToolValue.fromJSON("hello")
         #expect(stringValue.stringValue == "hello")
@@ -180,8 +179,8 @@ struct AgentToolValueTests {
         #expect(dictValue.objectValue?["num"]?.intValue == 123)
     }
 
-    @Test("AnyAgentToolValue Codable conformance")
-    func anyAgentToolValueCodable() throws {
+    @Test
+    func `AnyAgentToolValue Codable conformance`() throws {
         struct TestContainer: Codable {
             let value: AnyAgentToolValue
         }
@@ -223,8 +222,8 @@ struct AgentToolValueTests {
 
     // MARK: - AgentToolCall and AgentToolResult Tests
 
-    @Test("AgentToolCall uses AnyAgentToolValue")
-    func agentToolCall() throws {
+    @Test
+    func `AgentToolCall uses AnyAgentToolValue`() {
         let arguments = [
             "prompt": AnyAgentToolValue(string: "Hello"),
             "temperature": AnyAgentToolValue(double: 0.7),
@@ -244,8 +243,8 @@ struct AgentToolValueTests {
         #expect(toolCall.arguments["maxTokens"]?.intValue == 100)
     }
 
-    @Test("AgentToolCall legacy init with Any")
-    func agentToolCallLegacyInit() throws {
+    @Test
+    func `AgentToolCall legacy init with Any`() throws {
         let arguments: [String: Any] = [
             "text": "Hello",
             "count": 42,
@@ -263,8 +262,8 @@ struct AgentToolValueTests {
         #expect(toolCall.arguments["enabled"]?.boolValue == true)
     }
 
-    @Test("AgentToolResult uses AnyAgentToolValue")
-    func agentToolResult() {
+    @Test
+    func `AgentToolResult uses AnyAgentToolValue`() {
         let successResult = AgentToolResult.success(
             toolCallId: "call_123",
             result: AnyAgentToolValue(string: "Success!"),
@@ -286,8 +285,8 @@ struct AgentToolValueTests {
 
     // MARK: - AgentToolArguments Tests
 
-    @Test("AgentToolArguments accessor methods")
-    func agentToolArgumentsAccessors() throws {
+    @Test
+    func `AgentToolArguments accessor methods`() throws {
         let args = AgentToolArguments([
             "string": AnyAgentToolValue(string: "text"),
             "number": AnyAgentToolValue(double: 42.5),
@@ -323,8 +322,8 @@ struct AgentToolValueTests {
         #expect(args.optionalBooleanValue("missing") == nil)
     }
 
-    @Test("AgentToolArguments error handling")
-    func agentToolArgumentsErrors() throws {
+    @Test
+    func `AgentToolArguments error handling`() throws {
         let args = AgentToolArguments([
             "text": AnyAgentToolValue(string: "hello"),
         ])
@@ -342,15 +341,17 @@ struct AgentToolValueTests {
 
     // MARK: - Type-Safe Tool Protocol Tests
 
-    @Test("AgentToolProtocol implementation")
-    func agentToolProtocol() async throws {
+    @Test
+    func `AgentToolProtocol implementation`() async throws {
         // Define a concrete tool
         struct WeatherTool: AgentToolProtocol {
             struct Input: AgentToolValue, Equatable {
                 let location: String
                 let units: String
 
-                static var agentValueType: AgentValueType { .object }
+                static var agentValueType: AgentValueType {
+                    .object
+                }
 
                 func toJSON() throws -> Any {
                     ["location": self.location, "units": self.units]
@@ -372,7 +373,9 @@ struct AgentToolValueTests {
                 let temperature: Double
                 let conditions: String
 
-                static var agentValueType: AgentValueType { .object }
+                static var agentValueType: AgentValueType {
+                    .object
+                }
 
                 func toJSON() throws -> Any {
                     ["temperature": self.temperature, "conditions": self.conditions]
@@ -390,8 +393,14 @@ struct AgentToolValueTests {
                 }
             }
 
-            var name: String { "get_weather" }
-            var description: String { "Get current weather" }
+            var name: String {
+                "get_weather"
+            }
+
+            var description: String {
+                "Get current weather"
+            }
+
             var schema: AgentToolSchema {
                 AgentToolSchema(
                     properties: [
@@ -434,8 +443,8 @@ struct AgentToolValueTests {
 
     // MARK: - Edge Cases and Error Handling
 
-    @Test("Handle integer/double ambiguity")
-    func integerDoubleAmbiguity() throws {
+    @Test
+    func `Handle integer/double ambiguity`() throws {
         // Test that whole numbers can be treated as integers
         let wholeDouble = try AnyAgentToolValue.fromJSON(42.0)
         #expect(wholeDouble.intValue == 42)
@@ -452,8 +461,8 @@ struct AgentToolValueTests {
         #expect(largeInt.intValue == safeInteger)
     }
 
-    @Test("Handle nested structures")
-    func nestedStructures() throws {
+    @Test
+    func `Handle nested structures`() throws {
         let nested = [
             "level1": [
                 "level2": [
@@ -469,8 +478,8 @@ struct AgentToolValueTests {
         #expect(level3?["value"]?.stringValue == "deep")
     }
 
-    @Test("Handle mixed-type arrays")
-    func mixedTypeArrays() throws {
+    @Test
+    func `Handle mixed-type arrays`() throws {
         let mixed: [Any] = ["string", 42, true, NSNull(), ["nested": "object"]]
         let value = try AnyAgentToolValue.fromJSON(mixed)
 

Tests/TachikomaTests/AsyncErgonomicsTests.swift

@@ -2,16 +2,15 @@ import Foundation
 import Testing
 @testable import Tachikoma
 
-@Suite("Async Ergonomics Tests")
 struct AsyncErgonomicsTests {
-    @Test("Timeout error description")
-    func timeoutErrorDescription() throws {
+    @Test
+    func `Timeout error description`() {
         let error = TimeoutError(timeout: 5.5)
         #expect(error.errorDescription == "Operation timed out after 5.5 seconds")
     }
 
-    @Test("Cancellation token basic operations")
-    func cancellationTokenBasic() async throws {
+    @Test
+    func `Cancellation token basic operations`() async {
         let token = CancellationToken()
 
         #expect(await token.cancelled == false)
@@ -25,8 +24,8 @@ struct AsyncErgonomicsTests {
         #expect(await token.cancelled == true)
     }
 
-    @Test("Cancellation token with handlers")
-    func cancellationTokenHandlers() async throws {
+    @Test
+    func `Cancellation token with handlers`() async throws {
         let token = CancellationToken()
 
         class Flag: @unchecked Sendable {
@@ -49,8 +48,8 @@ struct AsyncErgonomicsTests {
         #expect(flag.value == true)
     }
 
-    @Test("Retry configuration defaults")
-    func retryConfigurationDefaults() throws {
+    @Test
+    func `Retry configuration defaults`() {
         let config = RetryConfiguration.default
 
         #expect(config.maxAttempts == 3)
@@ -60,8 +59,8 @@ struct AsyncErgonomicsTests {
         #expect(config.timeout == nil)
     }
 
-    @Test("Retry configuration presets")
-    func retryConfigurationPresets() throws {
+    @Test
+    func `Retry configuration presets`() {
         let aggressive = RetryConfiguration.aggressive
         #expect(aggressive.maxAttempts == 5)
         #expect(aggressive.delay == 0.5)
@@ -71,8 +70,8 @@ struct AsyncErgonomicsTests {
         #expect(conservative.delay == 2.0)
     }
 
-    @Test("Retry with cancellation - immediate success")
-    func retryWithCancellationImmediateSuccess() async throws {
+    @Test
+    func `Retry with cancellation - immediate success`() async throws {
         let result = try await retryWithCancellation(
             configuration: .init(maxAttempts: 3, delay: 0.01),
         ) {
@@ -82,8 +81,8 @@ struct AsyncErgonomicsTests {
         #expect(result == "Success")
     }
 
-    @Test("With timeout basic functionality")
-    func withTimeoutBasic() async throws {
+    @Test
+    func `With timeout basic functionality`() async throws {
         let result = try await withTimeout(0.1) {
             "Quick result"
         }
@@ -91,8 +90,8 @@ struct AsyncErgonomicsTests {
         #expect(result == "Quick result")
     }
 
-    @Test("With timeout throws on timeout")
-    func withTimeoutThrows() async throws {
+    @Test
+    func `With timeout throws on timeout`() async throws {
         do {
             _ = try await withTimeout(0.01) {
                 try await Task<Never, Never>.sleep(nanoseconds: 1_000_000_000) // 1 second
@@ -104,8 +103,8 @@ struct AsyncErgonomicsTests {
         }
     }
 
-    @Test("Async stream collect basic")
-    func asyncStreamCollectBasic() async throws {
+    @Test
+    func `Async stream collect basic`() async throws {
         let stream = AsyncThrowingStream<Int, Error> { continuation in
             continuation.yield(1)
             continuation.yield(2)
@@ -117,8 +116,8 @@ struct AsyncErgonomicsTests {
         #expect(results == [1, 2, 3])
     }
 
-    @Test("Task group with auto cancellation")
-    func taskGroupAutoCancellation() async throws {
+    @Test
+    func `Task group with auto cancellation`() async throws {
         class Flag: @unchecked Sendable {
             var cancelled = false
         }

Tests/TachikomaTests/Audio/AudioFunctionsTests.swift

@@ -3,14 +3,12 @@ import Testing
 @testable import Tachikoma
 @testable import TachikomaAudio
 
-@Suite("Audio Functions Tests")
-struct AudioFunctionsTests {
+enum AudioFunctionsTests {
     // MARK: - Basic Transcription Function Tests
 
-    @Suite("Basic Transcription Functions Tests")
     struct BasicTranscriptionFunctionsTests {
-        @Test("transcribe() convenience function works")
-        func transcribeConvenienceFunctionWorks() async throws {
+        @Test
+        func `transcribe() convenience function works`() async throws {
             try await TestHelpers.withMockProviderEnvironment {
                 try await TestHelpers.withTestConfiguration(apiKeys: ["openai": "test-key"]) { config in
                     let audioData = TestHelpers.sampleAudioData(configuration: config)
@@ -20,8 +18,8 @@ struct AudioFunctionsTests {
             }
         }
 
-        @Test("transcribe() with full model specification works")
-        func transcribeWithFullModelSpecificationWorks() async throws {
+        @Test
+        func `transcribe() with full model specification works`() async throws {
             try await TestHelpers.withMockProviderEnvironment {
                 try await TestHelpers.withTestConfiguration(apiKeys: ["openai": "test-key"]) { config in
                     let audioData = TestHelpers.sampleAudioData(configuration: config)
@@ -43,8 +41,8 @@ struct AudioFunctionsTests {
             }
         }
 
-        @Test("transcribe() from file URL works")
-        func transcribeFromFileURLWorks() async throws {
+        @Test
+        func `transcribe() from file URL works`() async throws {
             try await TestHelpers.withMockProviderEnvironment {
                 try await TestHelpers.withTestConfiguration(apiKeys: ["openai": "test-key"]) { config in
                     let audioData = TestHelpers.sampleAudioData(configuration: config)
@@ -68,8 +66,8 @@ struct AudioFunctionsTests {
             }
         }
 
-        @Test("transcribe() with timestamps")
-        func transcribeWithTimestamps() async throws {
+        @Test
+        func `transcribe() with timestamps`() async {
             await TestHelpers.withMockProviderEnvironment {
                 await TestHelpers.withTestConfiguration(apiKeys: ["openai": "test-key"]) { config in
                     let audioData = TestHelpers.sampleAudioData(configuration: config)
@@ -95,8 +93,8 @@ struct AudioFunctionsTests {
             }
         }
 
-        @Test("transcribe() with abort signal")
-        func transcribeWithAbortSignal() async throws {
+        @Test
+        func `transcribe() with abort signal`() async {
             await TestHelpers.withMockProviderEnvironment {
                 await TestHelpers.withTestConfiguration(apiKeys: ["openai": "test-key"]) { config in
                     let audioData = TestHelpers.sampleAudioData(configuration: config)
@@ -120,10 +118,9 @@ struct AudioFunctionsTests {
 
     // MARK: - Basic Speech Generation Function Tests
 
-    @Suite("Basic Speech Generation Functions Tests")
     struct BasicSpeechGenerationFunctionsTests {
-        @Test("generateSpeech() convenience function works")
-        func generateSpeechConvenienceFunctionWorks() async throws {
+        @Test
+        func `generateSpeech() convenience function works`() async throws {
             try await TestHelpers.withMockProviderEnvironment {
                 try await TestHelpers.withTestConfiguration(apiKeys: ["openai": "test-key"]) { config in
                     let audioData = try await generateSpeech("Hello, world!", configuration: config)
@@ -133,8 +130,8 @@ struct AudioFunctionsTests {
             }
         }
 
-        @Test("generateSpeech() with voice selection")
-        func generateSpeechWithVoiceSelection() async throws {
+        @Test
+        func `generateSpeech() with voice selection`() async throws {
             try await TestHelpers.withMockProviderEnvironment {
                 try await TestHelpers.withTestConfiguration(apiKeys: ["openai": "test-key"]) { config in
                     let audioData = try await generateSpeech("Hello, world!", voice: .nova, configuration: config)
@@ -144,8 +141,8 @@ struct AudioFunctionsTests {
             }
         }
 
-        @Test("generateSpeech() with full model specification works")
-        func generateSpeechWithFullModelSpecificationWorks() async throws {
+        @Test
+        func `generateSpeech() with full model specification works`() async throws {
             try await TestHelpers.withMockProviderEnvironment {
                 try await TestHelpers.withTestConfiguration(apiKeys: ["openai": "test-key"]) { config in
                     let result = try await generateSpeech(
@@ -162,8 +159,8 @@ struct AudioFunctionsTests {
             }
         }
 
-        @Test("generateSpeech() direct to file using convenience function")
-        func generateSpeechDirectToFile() async throws {
+        @Test
+        func `generateSpeech() direct to file using convenience function`() async throws {
             try await TestHelpers.withMockProviderEnvironment {
                 try await TestHelpers.withTestConfiguration(apiKeys: ["openai": "test-key"]) { config in
                     let tempDir = FileManager.default.temporaryDirectory
@@ -187,8 +184,8 @@ struct AudioFunctionsTests {
             }
         }
 
-        @Test("generateSpeech() with abort signal")
-        func generateSpeechWithAbortSignal() async throws {
+        @Test
+        func `generateSpeech() with abort signal`() async {
             await TestHelpers.withMockProviderEnvironment {
                 await TestHelpers.withTestConfiguration(apiKeys: ["openai": "test-key"]) { config in
                     let abortSignal = AbortSignal()
@@ -211,10 +208,9 @@ struct AudioFunctionsTests {
 
     // MARK: - Batch Operations Tests
 
-    @Suite("Batch Operations Tests")
     struct BatchOperationsTests {
-        @Test("transcribeBatch() function works")
-        func transcribeBatchFunctionWorks() async throws {
+        @Test
+        func `transcribeBatch() function works`() async throws {
             try await TestHelpers.withMockProviderEnvironment {
                 try await TestHelpers.withTestConfiguration(apiKeys: ["openai": "test-key"]) { config in
                     let audioData = TestHelpers.sampleAudioData(configuration: config)
@@ -246,8 +242,8 @@ struct AudioFunctionsTests {
             }
         }
 
-        @Test("generateSpeechBatch() function works")
-        func generateSpeechBatchFunctionWorks() async throws {
+        @Test
+        func `generateSpeechBatch() function works`() async throws {
             try await TestHelpers.withMockProviderEnvironment {
                 try await TestHelpers.withTestConfiguration(apiKeys: ["openai": "test-key"]) { config in
                     let texts = ["Hello", "World"]
@@ -267,24 +263,23 @@ struct AudioFunctionsTests {
 
     // MARK: - Utility Functions Tests
 
-    @Suite("Utility Functions Tests")
     struct UtilityFunctionsTests {
-        @Test("availableTranscriptionModels() returns models")
-        func availableTranscriptionModelsReturnsModels() {
+        @Test
+        func `availableTranscriptionModels() returns models`() {
             let models = availableTranscriptionModels()
             #expect(!models.isEmpty)
             #expect(models.contains { $0.description == TranscriptionModel.openai(.whisper1).description })
         }
 
-        @Test("availableSpeechModels() returns models")
-        func availableSpeechModelsReturnsModels() {
+        @Test
+        func `availableSpeechModels() returns models`() {
             let models = availableSpeechModels()
             #expect(!models.isEmpty)
             #expect(models.contains { $0.description == SpeechModel.openai(.tts1).description })
         }
 
-        @Test("capabilities() for transcription models")
-        func capabilitiesForTranscriptionModels() throws {
+        @Test
+        func `capabilities() for transcription models`() throws {
             let config = TestHelpers.createTestConfiguration(apiKeys: ["openai": "test-key"])
             let capabilities = try capabilities(for: TranscriptionModel.openai(.whisper1), configuration: config)
             #expect(capabilities.supportsTimestamps == true)
@@ -292,8 +287,8 @@ struct AudioFunctionsTests {
             #expect(capabilities.supportedFormats.contains(.wav))
         }
 
-        @Test("capabilities() for speech models")
-        func capabilitiesForSpeechModels() throws {
+        @Test
+        func `capabilities() for speech models`() throws {
             let config = TestHelpers.createTestConfiguration(apiKeys: ["openai": "test-key"])
             let capabilities = try capabilities(for: SpeechModel.openai(.tts1), configuration: config)
             #expect(capabilities.supportsSpeedControl == true)
@@ -304,10 +299,9 @@ struct AudioFunctionsTests {
 
     // MARK: - Error Handling Tests
 
-    @Suite("Error Handling Tests")
     struct ErrorHandlingTests {
-        @Test("transcribe() handles empty audio data")
-        func transcribeHandlesEmptyAudioData() async throws {
+        @Test
+        func `transcribe() handles empty audio data`() async {
             await TestHelpers.withTestConfiguration(apiKeys: ["openai": "test-key"]) { config in
                 let emptyAudioData = AudioData(data: Data(), format: .wav)
 
@@ -317,8 +311,8 @@ struct AudioFunctionsTests {
             }
         }
 
-        @Test("generateSpeech() handles empty text")
-        func generateSpeechHandlesEmptyText() async throws {
+        @Test
+        func `generateSpeech() handles empty text`() async {
             await TestHelpers.withTestConfiguration(apiKeys: ["openai": "test-key"]) { config in
                 await #expect(throws: TachikomaError.self) {
                     _ = try await generateSpeech("", using: .openai(.tts1), configuration: config)
@@ -326,8 +320,8 @@ struct AudioFunctionsTests {
             }
         }
 
-        @Test("functions handle missing API keys")
-        func functionsHandleMissingAPIKeys() async throws {
+        @Test
+        func `functions handle missing API keys`() async {
             guard TestHelpers.isMockAPIKey(ProcessInfo.processInfo.environment["OPENAI_API_KEY"]) else { return }
             await TestHelpers.withEmptyTestConfiguration { config in
                 let audioData = AudioData(data: Data([0x01, 0x02]), format: .wav)
@@ -345,10 +339,9 @@ struct AudioFunctionsTests {
 
     // MARK: - Integration Tests
 
-    @Suite("Integration Tests")
     struct IntegrationTests {
-        @Test("transcribe and generate speech pipeline")
-        func transcribeAndGenerateSpeechPipeline() async throws {
+        @Test
+        func `transcribe and generate speech pipeline`() async {
             await TestHelpers.withTestConfiguration(apiKeys: ["openai": "test-key"]) { config in
                 let originalAudioData = TestHelpers.sampleAudioData(configuration: config)
 
@@ -368,9 +361,9 @@ struct AudioFunctionsTests {
             }
         }
 
-        @Test("multiple provider integration")
-        func multipleProviderIntegration() async throws {
-            try await TestHelpers.withStandardTestConfiguration { config in
+        @Test
+        func `multiple provider integration`() async {
+            await TestHelpers.withStandardTestConfiguration { config in
                 let audioData = TestHelpers.sampleAudioData(configuration: config)
 
                 // Test different transcription providers

Tests/TachikomaTests/Audio/AudioProviderFactoryTests.swift

@@ -7,10 +7,10 @@ import Testing
 @testable import Tachikoma
 @testable import TachikomaAudio
 
-@Suite("Audio Provider Factories", .serialized)
+@Suite(.serialized)
 struct AudioProviderFactoryTests {
-    @Test("TranscriptionProviderFactory returns mock provider in test mode")
-    func transcriptionFactoryReturnsMockInTestMode() throws {
+    @Test
+    func `TranscriptionProviderFactory returns mock provider in test mode`() throws {
         let previousTestMode = getenv("TACHIKOMA_TEST_MODE").flatMap { String(cString: $0) }
         setenv("TACHIKOMA_TEST_MODE", "mock", 1)
         defer {
@@ -32,8 +32,8 @@ struct AudioProviderFactoryTests {
         #expect(provider is MockTranscriptionProvider)
     }
 
-    @Test("TranscriptionProviderFactory requires API key in test mode")
-    func transcriptionFactoryRequiresAPIKeyInTestMode() {
+    @Test
+    func `TranscriptionProviderFactory requires API key in test mode`() {
         let previousTestMode = getenv("TACHIKOMA_TEST_MODE").flatMap { String(cString: $0) }
         setenv("TACHIKOMA_TEST_MODE", "mock", 1)
         defer {
@@ -57,8 +57,8 @@ struct AudioProviderFactoryTests {
         }
     }
 
-    @Test("SpeechProviderFactory returns mock provider in test mode")
-    func speechFactoryReturnsMockInTestMode() throws {
+    @Test
+    func `SpeechProviderFactory returns mock provider in test mode`() throws {
         let previousTestMode = getenv("TACHIKOMA_TEST_MODE").flatMap { String(cString: $0) }
         setenv("TACHIKOMA_TEST_MODE", "mock", 1)
         defer {
@@ -80,44 +80,48 @@ struct AudioProviderFactoryTests {
         #expect(provider is MockSpeechProvider)
     }
 
-    @Test("AudioConfiguration reads configuration keys before environment")
-    func audioConfigurationPrefersExplicitConfiguration() {
-        let previousOpenAIKey = getenv("OPENAI_API_KEY").flatMap { String(cString: $0) }
-        unsetenv("OPENAI_API_KEY")
-        defer {
-            if let previousOpenAIKey {
-                setenv("OPENAI_API_KEY", previousOpenAIKey, 1)
+    @Test
+    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("AudioConfiguration falls back to environment variable")
-    func audioConfigurationFallsBackToEnvironment() {
-        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")
+    @Test
+    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")
+            }
         }
     }
 }