Bug Report

Summary

Every API request produces a WARN log: extra_body overwriting request payload keys: thinking. This is a framework-level issue that affects all providers (DeepSeek, Ollama, etc.), not just a single provider. The thinking parameter is being written in two separate locations within the request payload pipeline, causing a field collision on every request.

Version

• OpenClaw 2026.4.26 (be8c246)
• Node.js v25.9.0
• macOS (launchd managed gateway)

Reproduction

Always reproducible. Simply run the gateway with any model — the warning appears on every single API request.

Frequency

• ~384 occurrences/day (observed 2026-04-30, 18:56 CST)
• 271 occurrences/day (observed 2026-04-28)
• Essentially every request triggers this WARN

Impact

1. DeepSeek (V4 Pro): When thinking is enabled, this parameter collision has caused:

• Frequent API timeouts (10 consecutive timeouts observed between 04:03-04:20 UTC on 2026-04-28)
• May be related to issue [Bug]: DeepSeek v4-pro: 400 "reasoning_content must be passed back" with thinking=disabled #74374 (400 "reasoning_content must be passed back")

2. Ollama/Qwen3 (local model): This has severe consequences for local models:

• [llm-idle-timeout] ollama/qwen3:32b-q4_K_M-16k produced no reply before the idle watchdog
• Profile ollama:default timed out
• Stuck sessions with state=processing age=141s
• The root cause appears to be: OpenClaw sends "thinking" parameter, but Ollama native protocol uses "think" (top-level). When the thinking field state is unreliable due to the overwrite collision, the model may output to message.thinking instead of message.content, causing OpenClaw to see an empty reply → idle-timeout → stuck session → failover to surface_error.
• Mitigation: Setting thinking: false, think: false, and reasoning: false for all local Ollama models resolves the issue, confirming the thinking field is the root cause.

Relevant Source Location

Based on investigation, this appears to originate from createOpenAICompletionsExtraBodyWrapper in:

• src/agents/pi-embedded-runner/extra-params.ts (around line 492)
• The extra_body wrapper is applied after provider-specific wrappers, and Object.assign overwrites payload keys with a collision warning

Expected Behavior

• The thinking parameter should be written to the request payload exactly once, in a single location
• No WARN about parameter overwriting should appear on requests where thinking is explicitly disabled
• Local Ollama models should work reliably whether thinking is enabled or disabled

Workaround

For Ollama/Qwen3: explicitly disable all thinking-related parameters in the model config (thinking: false, think: false, reasoning: false).

Log Evidence

extra_body overwriting request payload keys: thinking
[llm-idle-timeout] ollama/qwen3:32b-q4_K_M-16k produced no reply before the idle watchdog
Profile ollama:default timed out. Trying next account...
stuck session: sessionId=main state=processing age=141s