fix(xai): preserve stateful Responses output-only continuations

[luvs01]

Motivation

• Prevent a regression where enabling xAI pairing caused function_call_output/custom_tool_call_output items to be rewritten into user messages when the matching function_call still lived upstream behind previous_response_id.
• Keep existing stateless/forward repair behavior (synthesizing missing outputs) while not breaking stateful xAI continuations that the upstream can resolve.

Description

• Add a conditional repair control to repairOrphanedInputItems by introducing a repairOrphanOutputs parameter so orphan-output-to-user-message conversion can be toggled, and preserve the original signature-compatible behavior when enabled (src/adapters/openai-responses/tool-output-recovery.ts).
• Compute a destination-aware decision in the Responses passthrough so stateful xAI continuations are preserved: call repairOrphanedInputItems with repairOrphanOutputs enabled only when appropriate (forward or stateless wires, or when there is no unexpanded previous_response_id) (src/adapters/openai-responses/passthrough.ts).
• Add regression tests that assert output-only continuations are preserved when the call is expected to be resolved upstream and that standalone orphan outputs are still repaired into user messages (tests/providers/xai/xai-responses-adjacency.test.ts).
• Update the compatibility documentation to clarify the distinction between adjacency and pairing and to document that pairing synthesizes placeholders only for calls present in the current input while preserving output-only continuations that rely on server-side state (structure/providers/chat-compat.md).

Testing

• Ran the focused xAI suite: bun test tests/providers/xai/xai-responses-adjacency.test.ts and related focused dangling-call suites; the focused set passed (8–25 tests depending on scope) on the repo Bun (node_modules/bun/bin/bun.exe v1.4.0).
• Ran typecheck with bun x tsc --noEmit and structure SSOT checks with bun run structure:check; both succeeded.
• Ran the focused trio: tests/providers/xai/xai-responses-adjacency.test.ts, tests/responses/responses-forward-dangling-call.test.ts, and tests/responses/responses-stateless-dangling-call-repair.test.ts; these passed end-to-end in the environment used.
• Note: a full-repo bun run test in this container showed unrelated environment/timing/process-group flakes in other suites; the changes in this PR are covered by the focused tests which passed.

---

Codex Task

---

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: Repository: luvs01/opencodex/.coderabbit.yaml

Review profile: ASSERTIVE

Plan: Advanced

Run ID: 0d0ff214-43fb-42a1-85d9-3009fa1b3f52

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[github-actions]

✅ Deterministic PR hygiene checks passed.

[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 2026-09-21T00:46:36.184027Z	4edc411	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.

[devin-ai-integration]

Fixed in 67ccd8d — dropReasoning now takes unexpandedMiss directly, restoring the replay-miss reasoning cleanup for stateful paired destinations; repairOrphanOutputs remains the independent fourth argument gating only orphan-output conversion.

[devin-ai-integration]

Fixed in 67ccd8d — dropReasoning now takes unexpandedMiss directly, restoring the replay-miss reasoning cleanup for stateful paired destinations; repairOrphanOutputs remains the independent fourth argument gating only orphan-output conversion.

[luvs01]

Consolidated into lidge-jun#5560 as a single related-function aggregate.

Source head: 67ccd8d5032685876e77cffd017ed9ce85d4b8eb. Replacement head: 2ec0cd12f5bb1567b3723dbf13555046604aa083.

Both stateful-continuation implementation 4edc411 and independent reasoning-cleanup follow-up 67ccd8d match carried 82a5f6d and aac783f by stable patch IDs, with original authors retained. The shared f700 role-fixture correction is already present in dev, with both exact provider objects carrying foldDeveloperRoleToSystem:false. All source changed-file occurrences were accounted for; the final composed adapter preserves upstream-owned output-only calls while independently stripping unrecoverable reasoning, and the new capability matrix covers those combinations. Focused evidence totals 377 distinct passing cases; types, structure, privacy and 497-page docs build passed. The broader changed run exited124 at900s with unresolved partial failures and is explicitly not green. This retires only this owner-controlled duplicate review entry; other contributors' upstream originals remain OPEN. Full suite, cross-platform CI, provider acceptance and security review remain on the draft proposal.

Closing this duplicate standalone review entry as part of the requested consolidation after verifying coverage. This is not a merge or release claim; remaining integration checks and reviews are tracked on the replacement. Original branches are retained.