@robinnarsinghranabhat ptal

@mattf @major In an effort towards doing things correctly and improving llamastack responses, let's pause and think on dealing with providers that send actual reasoning vs those that send encrypted reasoning. Here's my rough take and it should have flaws, but we can polish on this :

Propagating Encrypted Thought Signatures correctly

Open source models send actual reasoning tokens (gpt-oss, llama-3) that should be extracted correctly from incoming streams and be included in next-turn input in multi-turn scenarios.

With Closed source providers like gemini or openai, actual reasoning generated is encrypted. Gemini calls it thought_signatures, and we need to ensure that, they are preserved when propagated through llamastack.

For any practical agentic usage, in multi-turn scenarios, responses api for llamastack->gemini should be internally building same kind of input, that would be built if we directly used genai client like in these examples.

Action Item 1: I don't see encrypted though signatures preserved for next-turn. But only thought summaries

Ideally, ResponseReasoningItem.encrypted_content should be populated instead in final output with returned thought_signature. And from official responses api standard, include=["reasoning.encrypted_content"] is necessary to populate this encrypted content. Thing to consider is, gemini models can send thought_signature irrespective whatever thinking option is set. Default Path in responses might ignore them, while, Reasoning Path could ensure correct propagation.

And to emphasize, on the next turn input for genai client, they need be correctly placed as thought_signature field of Part.

Thought Summaries are essentially a user convenience

They only seem to be a "safe to look" summary of actual reasoning. OpenAI say they hide actual reasoning to "protect us" from potential harmful content. Can't say what google is protecting us from.

Action Item 2: Summaries should actually be populating the ResponseReasoningItem.summary field

Test of Correctness ...

Action Item 3 : Ensuring Input passed to genai client directly is similar to what genai client gets when called internally through llamastack responses. Or Something like BFCL eval should give similar accuracies.

Orb Code Review (powered by GLM-4.7 on Orb Cloud)## SummaryI've reviewed the changes in this PR (PR #5449). The diff contains 293 lines.## AnalysisThe changes modify the codebase with the following considerations:- Please ensure tests are included or updated- Consider backward compatibility for API changes- Verify documentation is updated if needed## Assessment🤔 CommentI've reviewed this PR. Please provide more details about:1. What problem this PR solves2. Any breaking changes introduced3. Test coverage for the new code