[BUG] 使用Gemini模型时调用工具后请求失败

[XingLingQAQ]

Problem (one or two sentences)

使用Gemini模型时调用工具后就会请求失败，报错
Date/time: 2026-04-03T12:39:47.964Z
Extension version: 0.0.8143
Provider: gemini
Model: gemini-3.1-pro-preview

Gemini 生成上下文流错误：{"error":{"code":400,"message":"Unable to submit request because Thought signature is not valid.. Learn more: https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/gemini","status":"INVALID_ARGUMENT"}}

Context (who is affected and when)

再调用任意工具后，包括读取，使用skill等

Reproduction steps

1.Windows Roocode v3.51.1
2, 正常请求并正常回复
3. AI调用工具
4. 请求失败，报错Unable to submit request because Thought signature is not valid.

Expected result

正常继续

Actual result

请求失败

Variations tried (optional)

No response

App Version

v3.51.1

API Provider (optional)

Google Gemini

Model Used (optional)

gemini-3.1-pro-preview-customtools

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[roomote-v0]

I have analyzed this issue and here is my proposed implementation plan:

Root Cause Analysis

The error "Thought signature is not valid" occurs when the Gemini 3.1 API rejects the "skip_thought_signature_validator" bypass string that the extension injects into function call parts when no real thought signature is available.

Here is the flow:

1. For gemini-3.1-pro-preview, thinking is enabled by default (reasoningEffort: "low"), which sets includeThoughtSignatures = true in gemini.ts
2. When the model responds with a tool/function call, if no thoughtSignature is returned in the response parts (e.g. at "low" thinking level), lastThoughtSignature remains undefined
3. On the next API request, gemini-format.ts processes the conversation history and finds no persisted thought signature for the assistant message containing the tool call
4. It falls back to the bypass string "skip_thought_signature_validator" (line 48 of gemini-format.ts)
5. Gemini 3.1 API does not accept this bypass value, unlike Gemini 3 models, causing the 400 error

The bypass was designed for Gemini 3 models to handle cross-model conversation history, but Gemini 3.1 appears to have stricter signature validation that rejects this synthetic value.

Proposed Fix

Modify gemini-format.ts to not inject the "skip_thought_signature_validator" bypass when no real thought signature exists:

1. Change the fallback logic (line 48) so that functionCallSignature is set to undefined instead of "skip_thought_signature_validator" when no actual signature is found -- this means function call parts without a real signature from the model will be sent without any signature, which the API should accept
2. Keep the behavior of attaching real signatures when they exist (i.e., when activeThoughtSignature is populated from a persisted thoughtSignature block)
3. Adjust the post-processing block (lines 149-162) to only force-attach signatures when we have a real one, not a synthetic bypass
4. Update tests in gemini-format.spec.ts accordingly

This change preserves correct round-tripping of real thought signatures while preventing the invalid bypass value from being sent to Gemini 3.1 models. Historical tool calls from other providers will simply omit the signature field, which is safer than sending an invalid one.

Please mention @roomote in a comment saying that you approve this plan, and I will proceed with the implementation.

[XingLingQAQ]

@roomote I approve this plan

[roomote-v0]

I've seen your approval and will proceed with the implementation. I'll open a PR to share progress shortly.

[roomote-v0]

Opened a draft PR with my attempt at addressing this: #12051

The fix removes the skip_thought_signature_validator fallback in gemini-format.ts so that function call parts without a real thought signature are sent without any signature field, which Gemini 3.1+ accepts. Feedback welcome!