fix: preserve reasoning content signature from LiteLLM thinking models

[giulio-leone]

Problem

When streaming responses from thinking models (e.g., Gemini) via LiteLLM, the reasoning content signature is silently dropped. The LiteLLM adapter's _process_choice_content only emits reasoning text from delta.reasoning_content but never captures the signature from delta.thinking.signature.

The event loop in streaming.py already correctly handles reasoningContent.signature deltas (lines 240-248, 314-315), but the LiteLLM model provider never emits them.

Root Cause

LiteLLM provides reasoning signatures via a separate thinking attribute on stream deltas (delta.thinking.signature), distinct from delta.reasoning_content (which only contains the text). The LiteLLM model provider's _process_choice_content method only processed reasoning_content text, never checking for the thinking.signature attribute.

Fix

Added a check in _process_choice_content() for the thinking.signature attribute on each streaming delta. When a valid string signature is found, a contentBlockDelta with reasoningContent.signature is emitted so the event loop can accumulate and store it in the final content block.

An isinstance(sig, str) guard prevents false positives from unittest.mock.Mock objects that auto-create attributes.

Testing

• Added test_stream_preserves_thinking_signature that simulates a 3-chunk streaming response with reasoning text and signature
• All 42 existing tests pass with 0 errors/0 warnings

Fixes #1764

[giulio-leone]

All CI checks pass. Ready for review.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[zastrowm]

Is this something we could have an integ test for? Would help prove out that it works against a live model

[github-actions]

Assessment: Approve ✅

Clean fix that correctly preserves reasoning signatures from LiteLLM's thinking.signature attribute by emitting the expected contentBlockDelta format. The implementation follows existing patterns and the type guard (isinstance(sig, str) and sig) is a good defensive measure.

Review Details

• Implementation: Correctly mirrors the event format expected by streaming.py (lines 240-248)
• Testing: Comprehensive test covering the streaming scenario with reasoning text + signature + text response
• Coverage: Minor gap (lines 419-420) due to test not exercising content type switching during signature processing - optional improvement suggested inline

Thanks for the well-documented PR description - the root cause analysis was particularly helpful!

[giulio-leone]

Hi! Gentle ping — this PR is rebased, CI passes, and ready for review. Happy to address any feedback. Thanks!

[giulio-leone]

@zastrowm Good point — I can add an integration test that exercises the unset data type against a live model. Would you prefer it under the existing tests/ integration suite, or in a separate file? And should it target a specific provider (e.g., Bedrock)?

[zastrowm]

@zastrowm Good point — I can add an integration test that exercises the unset data type against a live model. Would you prefer it under the existing tests/ integration suite, or in a separate file? And should it target a specific provider (e.g., Bedrock)?

I was thinking more of an litellm integ test integ test that ensures that reasoning and signatures are included in responses - does that fit into an existing test and/or something we can test for?

[giulio-leone]

@zastrowm Good suggestion — I've added a litellm integration test in tests_integ/models/test_model_litellm.py that exercises reasoning and signature handling. The test verifies that reasoning content and signatures are preserved in model responses. Let me know if this covers what you had in mind.