Bug

When using Ollama thinking-capable models (e.g. qwen3.5:4b), the Ollama API returns a valid response with both a thinking field and tool_calls. OpenClaw's Ollama adapter appears to drop the tool_calls, resulting in an empty assistant reply and auto-fallback to the next model.

Evidence

Direct Ollama API call to qwen3.5:4b returns correct tool_calls:

{
  "message": {
    "role": "assistant",
    "content": "",
    "thinking": "The user wants me to run the command...",
    "tool_calls": [{
      "id": "call_vo5f1vhv",
      "function": {
        "name": "exec",
        "arguments": {"command": "python -c \"print(2**10)\""}
      }
    }]
  }
}

Same task with granite4:7b-a1b-h (no thinking field) works correctly:

{
  "message": {
    "role": "assistant",
    "content": "",
    "tool_calls": [{
      "id": "call_5cs28dc4",
      "function": {
        "name": "exec",
        "arguments": {"command": "python -c \"print(2**10)\""}
      }
    }]
  }
}

Behavior

• granite4 (no thinking field): tool_calls parsed correctly, exec runs
• qwen3.5 (has thinking field): tool_calls dropped, empty reply, falls back to granite4

The subagent history shows qwen3.5 attempts fail silently, then retry message says "Continue where you left off. The previous model attempt failed or timed out." with granite4 as the fallback model.

Root Cause Hypothesis

The Ollama streaming adapter likely does not accumulate/parse tool_calls from chunks that also contain a thinking field. The fix in #50712 addressed sending think: false but did not address the case where thinking IS enabled and tool_calls coexist in the response.

Related Issues

• #47993 — empty visible reply when provider returns malformed thinking/tool-call output (still OPEN)
• #50712 — think:false not sent (CLOSED/fixed)

Reproduction

1. Configure OpenClaw with Ollama provider
2. Have a thinking-capable model (qwen3.5:4b) and a fallback (granite4:7b-a1b-h)
3. Spawn a subagent with model override to qwen3.5
4. Task: simple exec tool call (e.g. run a python command)
5. Observe: qwen3.5 attempt produces empty reply, fallback to granite4

Environment

• OpenClaw: latest (April 2026)
• Ollama: latest (April 2026, post tool-call-in-thinking fix)
• Models: qwen3.5:4b, granite4:7b-a1b-h
• OS: Windows 11
• GPU: RTX 3080 10GB