Gemini 3 Flash Preview: Function call fails due to missing thought_signature

[gyuyeol]

When using the Gemini 3 Flash Preview model with picoclaw, function calls are consistently failing. The error messages indicate that a thought_signature is missing in the functionCall parts, which is stated as required for tools to work correctly and leads to degraded model performance.

This prevents the agent from properly utilizing available tools via function calling.

Error Message:

2026/02/12 15:56:24 [2026-02-12T15:56:24Z] [ERROR] agent: LLM call failed {iteration=2, error=API error: [{
  "error": {
    "code": 400,
    "message": "Function call is missing a thought_signature in functionCall parts. This is required for tools to work correctly, and missing thought_signature may lead to degraded model performance. Additional data, function call default_api:list_dir , position 10. Please refer to https://ai.google.dev/gemini-api/docs/thought-signatures for more details.",
    "status": "INVALID_ARGUMENT"
  "error": {
    "code": 400,
    "message": "Function call is missing a thought_signature in functionCall parts. This is required for tools to work correctly, and missing thought_signature may lead to degraded model performance. Additional data, function call default_api:list_dir , position 10. Please refer to https://ai.google.dev/gemini-api/docs/thought-signatures for more details.",
    "status": "INVALID_ARGUMENT"
  }
}
]}
Error: LLM call failed: API error: [{
  "error": {
    "code": 400,
    "message": "Function call is missing a thought_signature in functionCall parts. This is required for tools to work correctly, and missing thought_signature may lead to degraded model performance. Additional data, function call default_api:list_dir , position 10. Please refer to https://ai.google.dev/gemini-api/docs/thought-signatures for more details.",
    "status": "INVALID_ARGUMENT"
  }
}
]

Expected behavior:
Function calls should execute correctly without requiring manual intervention for thought_signature when using the Gemini 3 Flash Preview model, similar to how other models handle function calling.

Impact:
This issue severely limits the agent's ability to use tools and perform tasks when configured with the Gemini 3 Flash Preview model.

[SatyamDevv]

🚀 Solution found

I managed to resolve the thought_signature and tool execution errors with Gemini.

The root cause is that PicoClaw expects OpenAI API compatibility, but the default Gemini endpoint uses Google's native format. Switching to the Google standard OpenAI-compatible endpoint fixes the issue.

✅ Configuration

Update your config.json to use the /openai suffix in the api_base and ensure you are using a non-deprecated model version:

"providers": {
  "gemini": {
    "api_key": "YOUR_API_KEY",
    "api_base": "https://generativelanguage.googleapis.com/v1beta/openai"
  }
}

I verified this with the gemini-2.5-flash model and tool calls (like web_search) are now working correctly without the 400 INVALID_ARGUMENT error.

[gyuyeol]

gemini-2.5-flash is ok.
gemini-3-flash-preview is not!

[SatyamDevv]

I checked with these 2 models and they are working

"gemini-flash-lite-latest"
"gemini-2.5-flash"

[gyuyeol]

gemini-2.5-flash works fine without the suffix.

However, my report is specifically about gemini-3-flash-preview.

Could you clarify what your point is in relation to this issue?

[SatyamDevv]

Sorry brother i made a mistake in reading your issue should I delete my answer?

[gyuyeol]

No problem at all.

Thanks for clarifying.
No need to delete it.

[mrbeandev]

Update: this specific thought_signature failure has been addressed in PR #343.

PR link: #343

The fix preserves and replays thought signatures in Antigravity tool-call turns so follow-up requests no longer fail with missing thought_signature errors.

[mrbeandev]

Correction: this specific thought_signature failure has been addressed in PR #343.

PR link: #343

[nikolasdehor]

Status update: PR #343 that addressed thought_signature preservation was closed (not merged). However, PR #213 (merged) refactored providers by protocol family, which restructured how Gemini API calls are built.

The core issue is that Gemini 3 Flash Preview requires thought_signature fields in functionCall parts when replaying tool-call history. This needs to be handled in the Gemini-native provider path -- the OpenAI-compatible endpoint workaround (/v1beta/openai) avoids the issue entirely since it uses a different wire format.

For anyone still hitting this: using api_base: "https://generativelanguage.googleapis.com/v1beta/openai" in the gemini provider config routes through Google's OpenAI-compatible layer, which does not require thought_signature. This is the most reliable workaround until native Gemini protocol support properly preserves thought signatures in multi-turn tool calls.