openclaw - 💡(How to fix) Fix Gemini 3 Flash tool calls fail with 400: missing thought_signature in OpenAI-compatible mode [1 comments, 2 participants]

Bug Description

When using Gemini 3 Flash (via Google Vertex AI OpenAI-compatible endpoint) as a subagent model, tool calls fail with HTTP 400 after the first successful tool invocation. The error occurs when OpenClaw sends the tool result back to the model.

Error Message

HTTP 400: [{
  "error": {
    "code": 400,
    "message": "Unable to submit request because function call `exec` in the 2. content block is missing a `thought_signature`. Learn more: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/thought-signatures",
    "status": "INVALID_ARGUMENT"
  }
}]

Root Cause

Gemini 3 Flash returns a thought_signature field inside extra_content.google on every assistant message and tool call when thinking/reasoning is active. The Google Vertex AI OpenAI-compatible API requires this thought_signature to be passed back unchanged in subsequent requests.

OpenClaw's OpenAI-compatible layer currently discards the extra_content.google.thought_signature field when processing tool results, causing the next API call to fail.

Reproduction Steps

1. Configure google-custom provider pointing to Vertex AI OpenAI-compatible endpoint:

   {
     "baseUrl": "http://127.0.0.1:8790/v1/projects/PROJECT/locations/global/endpoints/openapi",
     "api": "openai-completions",
     "models": [{ "id": "gemini-3-flash-preview" }]
   }

2. Spawn a subagent with model: "google-custom/gemini-3-flash-preview"
3. The subagent successfully makes a tool call (e.g., exec)
4. When OpenClaw sends the tool result back, Gemini returns 400

Verified Behavior

• Simple tool call (single turn): Works fine — Gemini correctly returns tool_calls with valid function.name and function.arguments
• Tool result roundtrip: Fails — the thought_signature from the previous assistant message is lost
• Plain chat (no tools): Works fine

API Response Example (showing thought_signature)

{
  "choices": [{
    "message": {
      "tool_calls": [{
        "extra_content": {
          "google": {
            "thought_signature": "AY89a19pcPpW8fF2gPU7VtN98DGJ4f7dt4n1AV..."
          }
        },
        "function": {
          "arguments": "{\"city\":\"Beijing\"}",
          "name": "get_weather"
        },
        "id": "function-call-269de851-08e8-4b36-bb0f-1218599cb0d0",
        "type": "function"
      }]
    }
  }]
}

Suggested Fix

OpenClaw should preserve and replay the extra_content.google.thought_signature field when constructing follow-up messages for Gemini models via the OpenAI-compatible endpoint. This is required by the Vertex AI API specification for models with thinking enabled.

Reference: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/thought-signatures

Environment

• OpenClaw: 2026.5.3-1
• Model: google-custom/gemini-3-flash-preview (Vertex AI OpenAI-compatible)
• OS: macOS 26.4.1 (arm64)
• Node: v22.22.0