Merge branch 'pr-4' into codex/merge-pr4

README.md

@@ -251,6 +251,48 @@ clawbench run --model anthropic/claude-opus-4-6 --profile profiles/frontier_opus
 clawbench diagnose profiles/frontier_opus_4_6.yaml
 ```
 
+### Running locally with small models (Ollama)
+
+A single consumer GPU running an open-weight model through
+[Ollama](https://ollama.com) is enough to develop plugin profiles, validate
+algorithmic ideas, and submit scored results — no API keys or cloud spend
+required.
+
+Profiles tested locally can still be submitted as pull requests with
+reference results. The built-in GitHub Actions workflows in this repo only
+run the test suite and deployment sync, so treat local Ollama numbers as
+contributor-side evidence unless a maintainer separately reruns them on
+other infrastructure.
+
+```bash
+# Pull a model and set your gateway token
+ollama pull gpt-oss:20b   # or llama3.1:8b, qwen3:14b, etc.
+export OPENCLAW_GATEWAY_TOKEN=<your-gateway-token>
+
+# Quick smoke test
+clawbench run --model ollama/gpt-oss:20b --task t1-fs-quick-note --runs 1
+
+# Tier-1 sweep with confidence intervals
+clawbench run --model ollama/gpt-oss:20b --tier tier1 --runs 5
+
+# Tier-2 sweep (run separately; the CLI accepts one --tier at a time)
+clawbench run --model ollama/gpt-oss:20b --tier tier2 --runs 5 --concurrency 2
+
+# Inspect the reference profile's fingerprint and historical neighbors
+clawbench diagnose profiles/local_ollama_gpt_oss.yaml
+```
+
+**Reference contributor-side results** (gpt-oss:20b, RTX 4090, Docker sandbox, network=none):
+
+| Scope | Score | CI | Completion | Trajectory | Behavior |
+|---|---|---|---|---|---|
+| Tier-1 (6 tasks × 3 runs) | 0.397 | 0.346–0.447 | 0.056 | 0.522 | 1.000 |
+
+High trajectory/behavior but low completion — the model uses tools correctly
+but writes to wrong paths or misses format constraints. This gap is where
+profile-level improvements (workspace-aware prompts, path-checking pre-flight
+calls, retry wrappers) have the most leverage.
+
 ### Docker (recommended for reproducibility)
 
 ```bash

app.py

@@ -983,7 +983,8 @@ with gr.Blocks(title="ClawBench", theme=clawbench_theme, css=CUSTOM_CSS) as demo
         gr.Markdown("### Submit a model for evaluation")
         gr.Markdown(
             "Select a preset or enter a custom model ID. Open-source models "
-            "run via HuggingFace Inference API. Proprietary models need model auth configured in the Space runtime."
+            "run via HuggingFace Inference API. You can also use locally hosted models "
+            "(for example Ollama) when your OpenClaw runtime has them configured."
         )
 
         preset_input = gr.Dropdown(
@@ -994,7 +995,7 @@ with gr.Blocks(title="ClawBench", theme=clawbench_theme, css=CUSTOM_CSS) as demo
         with gr.Row():
             model_input = gr.Textbox(
                 label="Custom Model ID (if not using preset)",
-                placeholder="e.g. huggingface/org/model-name",
+                placeholder="e.g. huggingface/org/model-name or ollama/gpt-oss:20b",
                 scale=3,
             )
             provider_input = gr.Textbox(

profiles/local_ollama_gpt_oss.yaml

@@ -0,0 +1,30 @@
+profile:
+  name: local-ollama-gpt-oss
+  base_model: ollama/gpt-oss:20b
+  notes: |
+    Reference profile metadata for local Ollama experiments. Change
+    base_model and tools_allow to match your setup, then use
+    `clawbench diagnose` to inspect the fingerprint and attach your
+    benchmark results separately in a PR.
+
+    Contributor-side baseline (gpt-oss:20b, RTX 4090, Docker sandbox,
+    network=none):
+      Tier-1+2: 0.467 (CI 0.375-0.545), T=0.666, B=0.850, C=0.202
+  plugins:
+    enabled:
+      - ollama
+      - id: memory-lancedb
+        config:
+          dimensions: 1536
+      - browser-playwright
+    slots:
+      memory: memory-lancedb
+      contextEngine: builtin
+    tools_allow:
+      - bash
+      - file_read
+      - file_edit
+      - browser_navigate
+      - browser_click
+      - memory_read
+      - memory_write