fix(codex): avoid orphaned Responses message IDs

[salehelsayed]

What does this PR do?

Fixes #97442.
Related to #97427, which appears to report the same GPT-5.6 Responses replay failure.

Hermes intentionally strips replayed reasoning-item IDs when store=false because some Responses-compatible endpoints otherwise try to look them up server-side. A dependent native assistant msg_* ID can nevertheless survive, producing an invalid partial item graph such as:

reasoning{encrypted_content} -> message{id: msg_*}

GPT-5.6 rejects that request because the message identity depends on the stripped rs_* reasoning identity.

This refreshed implementation is rebased onto current main and enforces the invariant at the final provider-call boundary, after request middleware and Responses preflight:

• if a native reasoning item starts an output group, dependent typed assistant message IDs are removed until a real group boundary;
• a function_call does not prematurely close the reasoning group;
• function_call_output and ordinary role messages close the group;
• if encrypted reasoning replay has been disabled for recovery, all typed assistant message IDs are omitted because their original dependency can no longer be proven;
• independent message IDs are preserved, so the existing cache optimization remains where it is demonstrably safe;
• message content, status, phase, and tool-call pairing are unchanged.

Current-main refresh

The original PR head had become conflicted after substantial upstream refactoring. On 2026-09-11 the branch was rebuilt as a single focused commit on current main (ad03f20dd61919ca2135d6904e787a94284aacaf) rather than mechanically resolving the old diff.

Current changes:

• agent/responses_replay_identity.py
  • Adds the dependency-closure guard for Responses replay identities.
• agent/turn_api_call.py
  • Applies the guard at the final provider-call boundary after middleware and preflight.
• tests/agent/test_responses_replay_identity.py
  • Covers linked reasoning/message replay, function-call grouping, tool-output boundaries, independent message IDs, recovery mode, and non-mutation of the original request.
• contributors/emails/saleh.fekry@gmail.com
  • Adds the standard contributor mapping for salehelsayed.

Type of Change

• 🐛 Bug fix (non-breaking)
• ✅ Regression tests

Validation

The refreshed head is conflict-free and GitHub reports it as mergeable against current main.

Hosted CI, Docker, and Nix workflows were triggered for the refreshed head but currently show action_required, so they require maintainer approval before they can execute for this external-contributor PR.

Checklist

• Based on current upstream main
• Focused only on the Responses replay identity bug
• Preserves independent msg_* IDs
• Adds focused regression coverage
• Does not change architecture, configuration, or user-facing behavior outside this fix
• Hosted CI completed on the refreshed head (awaiting maintainer workflow approval)

[ehz0ah]

Reviewed exact head e3a6ba38287c. The change restores the Responses replay invariant without reintroducing store-false reasoning ID lookups: linked msg_* IDs are removed whenever their rs_* identity cannot be replayed, while message content, phase, status, and tool-call pairing remain intact. The same rule is applied to conversion, preflight, normalization, stale-replay pruning, and invalid-encrypted-content recovery.

I reproduced the invalid graph on current main and confirmed that this head produces the dependency-closed graph. I also ran the focused 10-file suite through scripts/run_tests.sh: 672 passed. Ruff, compileall, and git diff --check passed. I found no correctness blockers in the submitted implementation. The branch currently conflicts with main, so it will need a rebase before merge.

[teknium1]

Thanks @salehelsayed — the work in this PR has landed on main via:

• fix(codex): GPT-5.6 Responses continuations no longer 400 on a replayed message whose reasoning id was stripped (#97427, #97442, salvage #97445) #115733 — fix(codex): GPT-5.6 Responses continuations no longer 400 on a replayed message whose reasoning id was stripped ([Bug]: GPT-5.6 Responses rejects replayed assistant message when Hermes strips its required reasoning item ID #97427, [Bug]: GPT-5.6 Responses rejects replayed message after Hermes strips its reasoning ID #97442, salvage fix(codex): avoid orphaned Responses message IDs #97445) (landed in 4d1d3d05a358)

Your contribution is credited there (cherry-picked authorship / co-author trailer or credit in the PR body; see the linked PR for what was kept and what was trimmed). Closing this one as landed / superseded so the backlog reflects reality. If something in your original diff is still missing on current main, please comment and we'll reopen or follow up.