fix: preserve Gemini thought_signature in multi-turn tool calls #718
+221
−1
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
When using Gemini models with thinking/reasoning enabled through the
AI SDK integration, the thought_signature was not preserved across
multi-turn conversations with tool calls.
The issue was that providerData was using result.providerMetadata
(overall response metadata) instead of toolCall.providerMetadata
(per-tool-call metadata containing thoughtSignature).
Changes:
- Non-streaming mode: Prioritize toolCall.providerMetadata over
result.providerMetadata for function_call outputs
- Streaming mode: Capture part.providerMetadata and include it
as providerData in function_call outputs
Reference: https://ai.google.dev/gemini-api/docs/thought-signatures
Related: openai/openai-agents-python#2137
🦋 Changeset detectedLatest commit: 130ff1c The changes in this PR will be included in the next version bump. This PR includes changesets to release 1 package
Not sure what this means? Click here to learn what changesets are. Click here if you're a maintainer who wants to add another changeset to this PR |
seratch
requested changes
Dec 2, 2025
Member
There was a problem hiding this comment.
Thanks for sending this patch. The changes should be good to go. Can you add the following changes as well?
- run
pnpm changesetand create a changeset file with patch version bump - add new unit test cases covering the scenario to prevent future regressions
- Add changeset for @openai/agents-extensions patch version bump
- Add unit tests for per-tool-call providerMetadata preservation:
- Non-streaming: prioritizes toolCall.providerMetadata
- Non-streaming: falls back to result.providerMetadata
- Streaming: captures part.providerMetadata as providerData
- Streaming: omits providerData when not present
seratch
approved these changes
Dec 2, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary
thought_signaturenot being preserved in multi-turn conversations with tool callsproviderMetadataover overall response metadata in both streaming and non-streaming modesRelated Issues
Problem
When using Gemini models with thinking/reasoning enabled (e.g.,
gemini-3.0-pro-preview) through the AI SDK integration, thethought_signaturefield was not preserved across agentloop iterations, causing subsequent API calls to fail with:
Reference: https://ai.google.dev/gemini-api/docs/thought-signatures
Root Cause
providerDatawas set toresult.providerMetadata(overall response metadata containingusageMetadata,safetyRatings, etc.) instead oftoolCall.providerMetadata(per-tool-call metadata containingthoughtSignature)providerDatawas not captured at all from tool-call partsChanges
aiSdk.ts~line 714): PrioritizetoolCall.providerMetadatawith fallback toresult.providerMetadataaiSdk.ts~line 921): Capturepart.providerMetadataasproviderDataTest Plan