Bug type

Behavior bug (incorrect output/state without crash)

Summary

OpenClaw sends think as a top-level parameter in Ollama /api/chat requests, but Ollama only accepts top-level think for models with native thinking support (e.g., Gemma 4, DeepSeek R1). For all other models, the top-level think field causes a 400 {"error":"\"<model>\" does not support thinking"} error.

However, sending think inside the options block (options: {think: "low"}) works correctly for ALL models and provides a reasoning token budget.

Root Cause

Two places in extensions/ollama/src/stream.ts set think at the request top level:

1. createOllamaThinkingWrapper (line ~236): Sets payloadRecord.think = think — this handles runtime thinkingLevel from agent config
2. resolveOllamaTopLevelParams (line ~355): Includes think from model.params as a top-level request param

Both should inject into options instead.

Steps to Reproduce

1. Configure any Ollama model without native thinking support (e.g., mistral, llama3.2, qwen2.5) with thinkingDefault: "low"
2. Send a request to OpenClaw gateway
3. Observe 400 error: "<model>" does not support thinking

Direct Ollama API proof

# Top-level think — FAILS for non-thinking models
curl http://localhost:11434/api/chat -d '{
  "model": "llama3.2:1b",
  "messages": [{"role":"user","content":"hi"}],
  "think": "low"
}'
# → 400: "llama3.2:1b does not support thinking"

# think inside options — WORKS for all models
curl http://localhost:11434/api/chat -d '{
  "model": "llama3.2:1b",
  "messages": [{"role":"user","content":"hi"}],
  "options": {"think": "low"}
}'
# → 200, response works normally

Expected Behavior

The think parameter should be sent inside the options block for all Ollama models (or at minimum, top-level think should only be used for models that are heuristically detected as having native thinking support).

Suggested Fix

In extensions/ollama/src/stream.ts:

1. In createOllamaThinkingWrapper, change:

   payloadRecord.think = think;

   to:

   (payloadRecord.options ??= {}).think = think;

2. In resolveOllamaTopLevelParams, remove the think handling (or move it to resolveOllamaModelOptions so it goes into options instead).

Related

• [Bug]: OpenClaw infer/web UI fails with ollama/qwen2.5:0.5b, but direct Ollama /api/chat works #67949 (original report, closed but root cause not fully fixed)
• fix: omit think parameter for Ollama models without thinking support #67958 (attempted fix, not merged — was limited to skipping think, didn't address the options placement)