feat(gemini): add thought signature persistence for conversation replay#13100
feat(gemini): add thought signature persistence for conversation replay#13100
Conversation
Implement Gemini thought signature persistence to solve the following issues: 1. Lost reasoning_content when reloading sessions 2. Error "thinking is enabled but reasoning_content is missing" on resend 3. Inability to leverage caching for performance Changes: - Extend chunk types with providerMetadata field to pass thoughtSignature - Extract thoughtSignature from AI SDK providerMetadata in AiSdkToChunkAdapter - Collect and store thoughtSignature in MessageBlock metadata during streaming - Persist thoughtSignature to message providerMetadata in AgentMessageDataSource - Restore thoughtSignature from metadata in skipGeminiThoughtSignaturePlugin Co-Authored-By: Claude Opus 4.6 <[email protected]>
Extract thoughtSignature from MainTextBlock.metadata and add it to text part's providerOptions.google.thoughtSignature when converting messages to SDK format. This ensures the thought signature is properly passed to the API when sending requests. Co-Authored-By: Claude Opus 4.6 <[email protected]>
Co-Authored-By: Claude Opus 4.6 <[email protected]>
- Remove message.providerMetadata usage for restoring thoughtSignature - Keep only OpenAI-compatible tool-call handling - text/reasoning parts now handled by messageConverter Co-Authored-By: Claude Opus 4.6 <[email protected]>
| } | ||
|
|
||
| // 0.7 Skip Gemini3 thought signature | ||
| // 0.7 Skip Gemini3 thought signature for OpenAI-compatible API |
There was a problem hiding this comment.
Note
This issue/comment/review was translated by Claude.
Check openrouter
Original Content
看下openrouterThere was a problem hiding this comment.
Note
This issue/comment/review was translated by Claude.
Tested, openrouter tool calls no longer require thought signature, no errors. It's likely the server has implemented skipping validation
Original Content
测过，openrouter工具调用现在不需要thought signature了，不会报错。估计是服务端做了跳过验证
|
Thanks for reviewing this PR! I wanted to clarify one point: Although this PR also modifies Agent Messages related processing logic (AgentMessageDataSource.ts), the Gemini thought signature currently only works when using the Google API directly. The Agent functionality uses Messages API, so the thought signature should not work with agent-related functionality at this stage. This is a known limitation - the persistence layer is in place, but the actual usage for Agent scenarios would require additional integration work in the future. |
EurFelux
changed the title
There was a problem hiding this comment.
Note
This review was translated by Claude.
📋 Code Review Report
✅ Overall Assessment: Approve to Merge
This is a high-quality PR with reasonable architectural design, clear code, implementing the Gemini thoughtSignature persistence feature.
🎯 Feature Analysis
The data flow design is reasonable:
AI SDK providerMetadata → Chunk → textCallbacks → MessageBlock.metadata → messageConverter → API Request
✅ Strengths
- Reasonable Architecture Design: Passes providerMetadata through existing chunk pipelines with minimal intrusion
- Clear Separation of Concerns: Data extraction, transmission, storage, and recovery are clearly divided
- Simplified skipGeminiThoughtSignaturePlugin: Removed duplicate text/reasoning processing logic
- All CI Passed: Basic guarantee of code quality
⚠️ Suggested Improvements (Non-blocking)
1. Type Safety Issue
File: src/renderer/src/services/StreamProcessingService.ts:86-92
Uses (data as any).providerMetadata instead of type-safe access. Suggest using type guards or safely accessing after checking chunk type:
case ChunkType.TEXT_DELTA: {
const textChunk = data as TextDeltaChunk
if (callbacks.onTextChunk) callbacks.onTextChunk(textChunk.text, textChunk.providerMetadata)
break
}2. Insufficient Test Coverage
File: src/renderer/src/aiCore/prepareParams/__tests__/message-converter.test.ts
The newly added findMainTextBlocks mock returns an empty array and does not test the thoughtSignature recovery logic. Suggest adding test cases.
3. Multiple Text Block Limitation (Already documented in PR description)
Multi-round tool calling scenarios may have multiple text blocks, and the current implementation only keeps the first thoughtSignature.
🔒 Security Review
- ✅ No hardcoded sensitive information
- ✅ No direct injection of user input
- ✅ Data is only stored in local metadata
Overall Assessment: 🟢 Approved to Merge
Suggest supplementing test cases and optimizing type safety in subsequent iterations.
### What this PR does Before this PR: - Version is 1.7.24 After this PR: - Version is 1.7.25 - Release notes updated in `electron-builder.yml` ### Why we need it and why it was done in this way Standard release workflow: collect commits since v1.7.24, generate bilingual release notes, bump version. The following tradeoffs were made: N/A The following alternatives were considered: N/A ### Breaking changes **Action Required**: The built-in filesystem MCP server now requires manual approval for write/edit/delete operations by default. ### Special notes for your reviewer **Included commits since v1.7.24:** ✨ New Features: - feat(websearch): add Querit search provider (#13050) - feat(minapp,provider): add MiniMax Agent, IMA mini apps and MiniMax Global, Z.ai providers (#13099) - feat(provider): add agent support filter for provider list (#11932) - feat(agent): add custom environment variables to agent configuration (#13357) - feat: add GPT-5.4 support (#13293) - feat(gemini): add thought signature persistence (#13100) 🐛 Bug Fixes: - fix: secure built-in filesystem MCP root handling (#13294) - fix: check underlying tool permissions for hub invoke/exec (#13282) - fix: remove approval countdown timers and add system notifications (#13281) - fix: agent tool status not stopping on abort (#13111) - fix: resolve spawn ENOENT on Windows for Code Tools (#13405) - fix: Use default assistant for topic auto-renaming (#13387) - fix(backup): defer auto backup during streaming response (#13307) - fix(ui): fix Move To submenu overflow (#13399) - fix: render xml fenced svg blocks as svg previews (#13431) - fix: correct Gemini reasoning params (#13388) - fix: improve Qwen 3.5 reasoning model detection (#13235) - fix: upgrade xAI web search to Responses API (#12812) - fix(api-server): relax chat completion validation (#13279) - fix: install or uninstall button state issues (#13114) - fix: improve update dialog behavior (#13363) - fix: auto-convert reasoning_effort to reasoningEffort (#12831) ### Checklist - [x] PR: The PR description is expressive enough and will help future contributors - [x] Code: Write code that humans can understand and Keep it simple - [x] Refactor: You have left the code cleaner than you found it (Boy Scout Rule) - [x] Upgrade: Impact of this change on upgrade flows was considered and addressed if required - [x] Documentation: A user-guide update was considered and is present (link) or not required. - [x] Self-review: I have reviewed my own code before requesting review from others ### Release note ```release-note See release notes in electron-builder.yml ``` --------- Signed-off-by: kangfenmao <[email protected]>
3 participants
What this PR does
Persist Gemini thoughtSignature to MessageBlock metadata.
Before this PR:
After this PR:
Fixes #13002
Why we need it and why it was done in this way
Main reason: Gemini models require thoughtSignature to be passed for better performance. Without persistence, users lose the performance benefits when:
The thought signature persistence enables leveraging caching for improved performance.
The following tradeoffs were made:
The following alternatives were considered:
Breaking changes
None.
Special notes for your reviewer
This PR modifies:
Known limitation: In multi-turn tool calling scenarios, there can be multiple text blocks each with its own thoughtSignature. The current implementation simplifies this by merging MainTextBlocks via getMainTextContent, thus only preserving the first block's thoughtSignature.
Note: Although this PR also modifies Agent Messages related processing logic, the Gemini thought signature currently only works when using the Google API directly. The Agent functionality uses Messages API, so the thought signature does not apply to agent-related functionality at this stage.
Checklist
This checklist is not enforcing, but it's a reminder of items that could be relevant to every PR.
Approvers are expected to review this list.
/gh-pr-review,gh pr diff, or GitHub UI) before requesting review from othersRelease note