Summary

OpenClaw's Ollama native stream handler reads chunk.message?.reasoning to accumulate thinking/reasoning tokens, but the Ollama native API (/api/chat) uses chunk.message.thinking — not reasoning. This means thinking-capable models (kimi-k2.5, glm-5, minimax-m2.5, etc.) via the native Ollama API silently discard all thinking content.

Affected file

src/ → compiled into dist/model-selection-*.js (multiple copies)

The specific code (line ~14749 in the compiled bundle):

// Current (wrong):
else if (chunk.message?.reasoning) accumulatedContent += chunk.message.reasoning;

// Should be:
else if (chunk.message?.thinking) accumulatedContent += chunk.message.thinking;

Also the non-streaming path (line ~14659):

// Current:
const text = response.message.content || response.message.reasoning || "";

// Should be:
const text = response.message.content || response.message.thinking || response.message.reasoning || "";

Steps to reproduce

1. Configure a custom provider with api: "ollama" pointing to Ollama Cloud or any Ollama instance running a thinking model
2. Add a model that uses thinking (e.g. kimi-k2.5, glm-5)
3. Do NOT set api at the model level (let it inherit the provider-level api: "ollama")
4. Send a message
5. The response arrives correctly but thinking tokens are silently dropped

Verification

Direct API call to https://ollama.com/api/chat with stream: true returns chunks like:

{"model":"kimi-k2.5","message":{"role":"assistant","content":"","thinking":"The user said..."},"done":false}

The field is thinking, not reasoning.

With stream: false:

{"message":{"role":"assistant","content":"Hi there!","thinking":"The user said hi..."}}

Impact

Thinking content is completely dropped. If a model ONLY returns content during the thinking phase and leaves content empty until the final chunk (which some models do), this can appear as an empty response.

Workaround

Patch the compiled model-selection-*.js files manually until this is fixed upstream. Note there are three copies of the file (model-selection-Zb7eBzSY.js, model-selection-ikt2OC4j.js, model-selection-CjMYMtR0.js) serving different embedded contexts — all must be patched.

References

Ollama docs confirm the field name: https://github.com/ollama/ollama/blob/main/docs/api.md