fix(ai): recover Codex SSE tool-choice rejections

[sj0618]

What

• Recognize the statusless Codex SSE representation of invalid_request_error when its message is the existing exact named-tool-not-in-tools rejection.
• Retry once without forced tool_choice before any output is emitted.
• Keep explicit non-400 statuses and unrelated provider error codes ineligible for this fallback.

Why

Fixes #3669.

The HTTP 400 form already uses the forced-tool-choice capability fallback, but the equivalent HTTP 200 SSE error carries a semantic provider code and no HTTP status. This remains a single contributor-authored commit rebased onto current dev.

Testing

• bun test packages/ai/test/openai-codex-responses-tool-choice.test.ts packages/ai/test/tool-choice-capability.test.ts — 47 passed, 269 assertions
• bunx biome check packages/ai/src/providers/openai-codex-responses.ts packages/ai/src/utils/tool-choice-capability.ts packages/ai/test/openai-codex-responses-tool-choice.test.ts packages/ai/test/tool-choice-capability.test.ts packages/ai/CHANGELOG.md — passed
• bun --cwd=packages/ai run check:types — passed
• git diff --check upstream/dev...HEAD — passed

GJC verdict

gajae.pr-review-verdict.v1 merge-approved sha256:988fa047450b0978a5bbd6bc4bdaa918f245ea2e reviewer:architect evidence:local-diff-review

---

• Target branch is dev
• Focused tests and package type check pass
• Tested locally
• CHANGELOG updated under ## [Unreleased]
• Verdict above matches the exact PR head

[Yeachan-Heo]

Signed: Yeachan-Heo / gajae-code AI-provider red-team

VERDICT: DEFERRED — exact-head review not admitted.

The exact PR head is still current at 1c81b55f26f8ec2a99f5c4065f5d55e6a4334cbe, and its checks are terminal-green (21 total: 15 success, 6 skipped, 0 failures, 0 non-terminal). It is not compatible with the current dev base: the PR targets b40bc271502a0802e212e5538c5d0e8436643029, current dev is 2c99a7f226635e812bc8fe8669787242029ce662, and GitHub reports mergeable=false, mergeable_state=dirty.

Per batch policy, no hostile implementation review or approval is attached while the exact head is stale/incompatible with current dev. Rebase onto current dev, rerun exact-head CI, and resubmit for independent review. No merge performed.
—
[repo owner's gaebal-gajae (clawdbot) 🦞]

[Yeachan-Heo]

REQUEST_CHANGES

Reviewed exact head 1c81b55f26f8ec2a99f5c4065f5d55e6a4334cbe against current origin/dev c1bf3be50e2cf0dfdf51b57eb43f7a99521c94b8.

This PR is stale and must be rebased onto current dev before it can be merge-ready. The PR base is b40bc271502a0802e212e5538c5d0e8436643029, not the requested current baseline. I inspected packages/ai/src/utils/tool-choice-capability.ts:23-31 and the Codex SSE error conversion. The new match is constrained to a forced tool choice plus invalid_request_error and the existing tool-choice message predicate, while nonmatching codes and HTTP 500 remain excluded; no additional credential, fallback-loop, or output-leak blocker was found in this exact diff.

Focused verification at the exact head passed: bun test packages/ai/test/openai-codex-responses-tool-choice.test.ts packages/ai/test/tool-choice-capability.test.ts (44 pass). Rebase onto c1bf3be50e2cf0dfdf51b57eb43f7a99521c94b8, rerun the clean one-shot fallback and cancellation coverage, and do not merge this stale head.

—
[repo owner's gaebal-gajae (clawdbot) 🦞]

[probepark]

Exact-head maintainer follow-up completed against current dev@54b14cb6e.

I rebased and corrected the statusless SSE fallback so it is Codex-only and accepts only an exact named-function rejection when the rejected name matches the forced name and that name exists in the final serialized tools list. The shared HTTP-400 classifier remains unchanged. Added negative coverage for managed mode, abort/output, required mode, mismatched/missing names, and second rejection/no third request.

Evidence at reviewed commit 01c472023:

• bun test packages/ai/test/tool-choice-capability.test.ts packages/ai/test/openai-codex-responses-tool-choice.test.ts — 47 pass, 269 assertions
• bun --cwd=packages/ai run check — passed
• independent Architect — CLEAR / APPROVE

I cannot push to the contributor fork (Permission to sj0618/gajae-code.git denied). The reviewed commit is published at probepark/gajae-code:fix/issue-3669-pr3699-review:
probepark@01c472023

Please update this PR head from that commit/branch (or explicitly authorize maintainer takeover and disposition of this PR). No duplicate PR has been opened.

[Yeachan-Heo]

Census reassignment receipt: Codex SSE tool-choice transport batch reviewed for routing-contract overlap.

Exact head: 1c81b55f26f8ec2a99f5c4065f5d55e6a4334cbe. Current dev: bf37cb30d0442926ad3fb2b2c6fc26a6d3edeffc.

The changed files do not overlap MiniMax policy files, but this PR is not merge-ready: review changes are requested by Yeachan-Heo and its recorded base is stale relative to current dev. No mutation or merge was performed.

—
[repo owner's gaebal-gajae (clawdbot) 🦞]

[probepark]

Current-dev refresh of the previously reviewed corrected candidate (contributor ownership preserved; no duplicate PR opened):

• Base: 971ff98fc5c33e34de0962c3e59c18f51a65c395
• Candidate head: 3873f35b1a6499e64aec2d162c26b1329a1b756a
• Published branch: probepark/review/issue-3669-current-dev
• Focused tests: 47 passed, 269 assertions
• Changed-file Biome check: passed
• packages/ai type check: passed after isolated-worktree dependency install
• git diff --check upstream/dev...HEAD: passed

The dev advance since the earlier review overlaps only packages/ai/CHANGELOG.md, not the candidate source/test files. Adoption still requires the contributor to update this PR or an explicit maintainer takeover decision.

[probepark]

Exact current-dev corrected candidate refresh (contributor ownership preserved; no duplicate PR opened):

• Base: 971ff98fc5c33e34de0962c3e59c18f51a65c395
• Candidate head: c63c447b2ae2c608e5d8d27c541b1112b0371dcc
• Published branch: probepark/review/issue-3669-current-dev
• Focused tests: 50 passed, 287 assertions
• packages/ai check: passed
• git diff --check: passed
• Independent exact-head audit 137-G003G004ExactAudit: CLEAR / APPROVE, no findings

This candidate fixes the remaining pre-runtime hole: an initial HTTP named-tool rejection now transfers explicit downgrade state into the SSE runtime, so later same-turn provider retries cannot reintroduce tool_choice. The statusless classifier remains exact and the provider/tool fallback budgets remain independent.

Adoption still requires the contributor to update this PR from the published candidate or an explicit maintainer takeover decision.

[probepark]

Exact current-dev update after the contributor rebased PR #3699:

• Current upstream/dev: aebb0f6d119acf178a9c710f460f83e9735bb8e6
• Contributor head observed: 988fa047450b0978a5bbd6bc4bdaa918f245ea2e (OPEN/CLEAN, still CHANGES_REQUESTED)
• Corrected candidate: bb9bce6eca92dd7c51288bb459fe653cffed3d5b
• Published branch: probepark/review/issue-3669-current-dev
• Focused tests: 50 passed, 287 assertions
• packages/ai check and exact-range diff check: passed
• Independent exact-current-dev audit 138-RebasedCandidateAudit: CLEAR / APPROVE, no findings

The contributor rebase incorporates the earlier narrow classifier work but still lacks the 110-line final delta that separates provider/tool retry budgets and keeps both initial-HTTP and SSE downgrades sticky across later same-turn reopens. The published candidate contains those repairs and their three-request regressions.

No duplicate PR was opened and contributor ownership remains intact. Adoption requires updating this PR from the published candidate or an explicit maintainer takeover decision.