PR #69967: fix(ollama): forward top-level think on native /api/chat requests

• Repository: openclaw/openclaw
• Author: WZH8898
• State: closed | merged: True
• Link: https://github.com/openclaw/openclaw/pull/69967

Description (problem / solution / changelog)

Fixes #69902

Summary

This fixes the native Ollama chat path so OpenClaw correctly forwards thinking control as a top-level think field on /api/chat requests.

Root cause

Around extensions/ollama/src/stream.ts:154, think injection still depended on model.api at invocation time. In some paths, native Ollama /api/chat requests did not preserve that condition, so thinking=off could fail to send top-level think: false.

As a result, thinking-capable models such as qwen3 could silently enter reasoning mode, produce no immediate reply, and sit idle long enough to be killed by the idle watchdog.

What changed

• Moved the native Ollama transport check earlier, to wrapper construction time in extensions/ollama/src/stream.ts:177
• For native Ollama /api/chat requests:
  • thinking=off now sends top-level think: false
  • any non-off thinking level now sends top-level think: true
• Kept existing request options such as num_ctx unchanged

Tests

Added request-body regression tests that directly verify the outgoing /api/chat payload:

• extensions/ollama/src/stream-runtime.test.ts:100
  • verifies think: false is sent when thinking is off
• extensions/ollama/src/stream-runtime.test.ts:153
  • verifies think: true is sent when thinking is enabled
• both tests also verify that existing options such as num_ctx are still preserved and that options.think is not used

Docs

Updated docs/providers/ollama.md:466 to document native Ollama thinking control behavior:

• /think off and openclaw agent --thinking off send top-level think: false
• non-off thinking levels send top-level think: true

Validation

Ran:

• pnpm install
• pnpm test extensions/ollama/src/stream-runtime.test.ts
• pnpm test extensions/ollama/index.test.ts
• pnpm check:changed

All passed.

Notes

I did not run a live end-to-end validation against a local Ollama + qwen3 instance in this environment.

Validation for this PR is therefore based on:

• the adapter request body behavior
• regression tests
• existing local test/check coverage

Changed files

• CHANGELOG.md (modified, +1/-0)
• docs/providers/ollama.md (modified, +2/-0)
• extensions/ollama/src/stream-runtime.test.ts (modified, +106/-0)
• extensions/ollama/src/stream.ts (modified, +5/-8)