fix(openai-chat): preserve native Mistral thinking on replay - #7814

Summary

Mistral's thinking can appear in Codex on the first turn but disappear from the next model request. The response parser and Responses replay keep the thinking text; the Chat serializer omits it. Adding preserveReasoningContentModels alone sends reasoning_content, rather than Mistral's native thinking chunks.

Replay the prior thinking as native content chunks before assistant text for api.mistral.ai. This applies to hosted GLM and reasoning-enabled Mistral Small, Medium and Large. Explicit preserve lists, including [], retain control.

For example, the second request follows a thinking response and a tool result:

{"role":"assistant","content":[{"type":"thinking","thinking":[{"type":"text","text":"Synthetic prior thinking."}]},{"type":"text","text":"Checking."}],"tool_calls":[{"id":"call_read","type":"function","function":{"name":"read_file","arguments":"{}"}}]}

Use actual text from the existing scoped replay cache when needed, without inventing placeholders or signatures. The shared prompt projection counts the replayed thinking. Other destinations and native Chat passthrough keep their existing behavior.

This follows #7047's response-parsing fix. Mistral's multi-turn guidance calls for replaying assistant thinking chunks. The fix reconstructs thinking text; it does not restore original chunk boundaries or native signature/closed metadata, which the current parser does not retain.

Verification

Bun 1.4.4, Linux. Base: 736fa5087c0b807cc2c7fd54ebae1548e6b2d4cb (dev, 2.94.0). Published commit: 8414064de8c81dd1922a66eecaeb25e5a7de6262. The branch is two commits behind the publication-time dev tip, f94e95ce5.

• Before the fix, the 16 new continuation cases fail and all 51 existing cases pass. They cover GLM, Flash, Small and Large with streamed/buffered responses and visible/hidden thinking. Additional cases cover opt-outs, tool-free turns, missing thinking and scoped orphan recovery.
• bun scripts/test.ts --changed=736fa5087c0b807cc2c7fd54ebae1548e6b2d4cb --parallel=1 tests/adapters/openai/ tests/adapters/translator-budget.test.ts tests/adapters/translator-budget-lifetime.test.ts tests/providers/deepseek-reasoning-replay.test.ts tests/providers/zai-reasoning-replay.test.ts tests/responses/chat-native-reasoning-replay.test.ts: 950 pass across 38 files. The run used upstream/dev while it pointed to the exact base shown here.
• bun scripts/test.ts --parallel=1 tests/test-layout.test.ts tests/test-layout-tooling.test.ts tests/ci-workflows/structure-ssot.test.ts tests/ci-workflows/repo-hygiene.test.ts tests/ci-workflows/file-size-ratchet.test.ts: 105 pass across five files. These explicit checks include source-as-data guards outside the import graph.
• bun run typecheck, bun run privacy:scan, bun run structure:check, and git diff --check: pass on the published source.
• cd docs-site && bun run build: pass with the locked dependencies; 593 pages and 85,001 internal links checked.
• A separate offline capture harness passes 28 two-turn combinations through the real adapter, Responses bridge and parser, plus six settings and six other-host controls. It also reproduces the omission on unchanged installed 2.93.0, upstream main 2.93.0 and the dev base.

All upstream replies are synthetic. Used focused and import-graph checks to limit local resource use alongside other workstation work; the full suite, cross-platform CI, live model acceptance and external review remain pending. This does not establish a model-quality improvement or a fix for retry loops. Keep the PR in draft until review is complete.

Checklist

• Scope stays focused and avoids unrelated cleanup.
• Docs or release notes were updated when needed.
• Security-sensitive changes were reviewed for secrets, auth, and unsafe defaults.

Review readiness checklist

This PR stays in draft until every box below is ticked. Tick all four boxes once the requirements are met:

• Required local validation passed; commands, results, and any full-suite exception are documented.

• I pushed my PR to a recent dev commit (at most 10 behind; a maintainer may still ask for the exact tip before merge).

• I resolved all correct Codex and CodeRabbit findings.

• My PR is ready for review.

Maintainer note: the checklist was reset only by the maintainer merging dev into this branch (head ed782d9). Contributor validation applies to 8414064; exact-head CI on ed782d9 is green (27/27). Re-ticked by the maintainer for dev integration.

Summary by CodeRabbit

• New Features
  • For Mistral Chat, supplied reasoning is now replayed as native thinking content before assistant text across hosted model families. An explicit model list controls which models preserve reasoning; an empty list opts out.
• Bug Fixes
  • Prevented placeholder thinking content from appearing in Mistral responses when cached reasoning is unavailable.
• Documentation
  • Clarified how Mistral reasoning replay and model-list settings work.