Restore reasoning_text when replaying history to DeepSeek

[JunkaiWang-TheoPhy]

Part of #2433 (split out of #2434 for reviewability).

Problem

DeepSeek's thinking mode requires reasoning_text on every replayed reasoning item:

400 {"error":{"message":"The `reasoning_text` in the thinking mode must be passed back to the API.","type":"invalid_request_error"}}

The official Codex backend requires the opposite — reasoning items with content: [] — which is why responses_reasoning_sanitize.go migrates DeepSeek's visible reasoning into summary and empties content on the response path. That sanitisation is deliberate (it keeps a thread usable after switching to an official account), but it makes the two upstreams demand contradictory shapes for the same field: DeepSeek's own transcripts are stored summary-only and are rejected the moment they are replayed to DeepSeek.

Measured directly against https://api.deepseek.com/responses:

input	result
reasoning with content:[{reasoning_text}] + assistant message	200
reasoning with summary only, content: []	400 reasoning_text must be passed back
no reasoning item at all, but an assistant message or tool pair	400 (same)

Observed in production on this machine three times in one day (09:00, 14:35, 22:16) across unrelated threads; each failure leaves the thread unable to continue until it is reloaded.

Change

A reverse normalisation on the request path, scoped to the DeepSeek gateway: when a reasoning item has an empty content and a non-empty summary, the summary text is written back as reasoning_text. Nothing is invented — the text came from that model's own thinking — and official-bound requests keep the summary-only shape. The response-path sanitiser is untouched.

Logic lives in a new provider_gateway_reasoning.go; provider_gateway.go gains only the call site.

Verification

• go test -count=1 . -run TestProviderGatewayRestoreReasoning — 3 new cases: summary restored, existing content untouched, non-DeepSeek gateways byte-identical.
• Live: all four shapes (summary-only + assistant, no reasoning + tool pair, summary-only + tool pair, reasoning_text + assistant) return 200 after deploy, and a real two-turn session (turn 1 executes a tool, turn 2 replays that history) completes normally.

Not covered

/responses/compact and websocket paths.

Note

Independent of #2447 and #2448 — separate file, no shared functions.

[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 2026-09-15T14:38:38.119603Z	aa06bda	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 review" or "@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.