[Abhijeet Prasad (AbhiPrasad)]

Mistral 3.2.0 exposes sampling and probability parameters that currently disappear from Braintrust span metadata. Capture all seven parameters for chat and agents calls, and retain completion and prompt probabilities when accumulating streamed choices.

Fixes #922.

Changes

• Add logprobs, top_logprobs, prompt_logprobs, top_prompt_logprobs, min_tokens, repetition_penalty, and top_k to both metadata allowlists.
• Extend three existing VCR tests and replace only their latest recordings. Latest tests use zai-glm-5-3, which accepts the new parameters; older SDK versions keep their existing model and recordings. Non-streaming chat and agents tests check both probability fields. Streaming checks completion probabilities because the hosted model rejects prompt probabilities with streaming.
• Consolidate speech, transcription, and setup tests while preserving all 40 cases. Restore all methods and patch markers after setup cases, and remove three duplicate conversation setup recordings.

Validation

• Red: reproduced missing request metadata before the allowlist fix. With real probability recordings, removing the stream aggregation fix makes the streaming regression fail with missing output logprobs.
• Green: all supported Mistral sessions pass in replay-only mode, including auto-instrument coverage.

Mistral version	Replay result
3.2.0 (latest)	40 passed
2.10.1	40 passed
1.12.4	38 passed, 2 expected skips: speech API unavailable

Run from py/: mise exec -- nox -R -s 'test_mistral(latest)' 'test_mistral(2.10.1)' 'test_mistral(1.12.4)' -- --vcr-record=none.

The cassette usage runner completed successfully across the entire matrix and reported no unused recordings. Pre-commit, stale cassette detection, lockfile validation, and git diff --check pass. Inspected the three new recordings: they contain only successful responses and filtered authorization headers.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-13T15:36:32.752226Z	1fe41f1	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "Codex (@codex) review" or "Codex (@codex) security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.