Best possible solution:

Merge this narrow Codex plugin fix after maintainer-approved checks/proof, then close #82129 as fixed by the landed PR.

Do we have a high-confidence way to reproduce the issue?

Yes. Current main source shows the accepted-turn path calls touchTurnCompletionActivity("turn:start") without arming the short completion-idle watchdog, and the linked regression provides CLI steps plus timeout logs; I did not run a live Codex/OAuth reproduction in this read-only review.

Is this the best way to solve the issue?

Yes. The PR uses the existing completion-idle watchdog at the accepted-turn boundary and preserves queued early progress handling, which is the narrow maintainable fix for the observed gap.

Acceptance criteria:

• pnpm test extensions/codex/src/app-server/run-attempt.test.ts -- --reporter=verbose
• node scripts/run-vitest.mjs extensions/codex/src/app-server/run-attempt.test.ts -t "releases the session when Codex accepts|does not treat global rate-limit notifications as turn progress|does not time out when turn progress arrives"
• pnpm check:changed
• git diff --check

What I checked:

• Current main does not arm the short watchdog at accepted-turn start: Current main sets turnTerminalIdleWatchArmed = true and calls touchTurnCompletionActivity("turn:start") after pending notification replay, but does not pass { arm: true }, so the completion-idle timer is not armed for an accepted turn with no current-turn progress. (extensions/codex/src/app-server/run-attempt.ts:1672, d91f58ee25b5)
• Existing helper contract requires explicit arming: touchTurnCompletionActivity only flips turnCompletionIdleWatchArmed when options.arm is true, matching the PR's approach of reusing the existing watchdog rather than adding a parallel timeout path. (extensions/codex/src/app-server/run-attempt.ts:1121, d91f58ee25b5)
• PR diff is narrow and targets the implicated boundary: The PR moves terminal-watch arming before pending notification replay and changes touchTurnCompletionActivity("turn:start") to touchTurnCompletionActivity("turn:start", { arm: true }), with matching tests/docs/help/changelog updates. (extensions/codex/src/app-server/run-attempt.ts:1638, 1bb4916cc073)
• Focused tests cover the timeout behavior and early progress guard: The PR switches the silent accepted-turn tests from turnTerminalIdleTimeoutMs to turnCompletionIdleTimeoutMs and adds coverage that a queued turn/started notification before turn/start returns does not trigger a false interrupt. (extensions/codex/src/app-server/run-attempt.test.ts:1916, 1bb4916cc073)
• Live PR state is protected and not a cleanup-close candidate: GitHub API shows the PR is open, non-draft, mergeable, at head 1bb4916cc0733c0f50fcc51e7c375bd42e7165dd, and labeled maintainer; protected labels require maintainer handling rather than automated stale/cleanup closure. (1bb4916cc073)
• Linked regression provides user-level reproduction evidence: The linked open bug reports OpenClaw 2026.5.12 Codex embedded tool/status turns timing out while text-only turns succeed, with CLI reproduction steps and sanitized timeout/stalled-session logs.

Likely related people:

• steipete: Current-main blame for the implicated watchdog/turn-start code points to Peter Steinberger's Codex-area commit, and the local shortlog over the touched Codex app-server, manifest, and docs paths shows Peter as the primary contributor. (role: dominant Codex area contributor; confidence: high; commits: c0fe7ab34ab8, f066dd2f31c2, 4862d349257a; files: extensions/codex/src/app-server/run-attempt.ts, extensions/codex/src/app-server/run-attempt.test.ts, extensions/codex/openclaw.plugin.json)
• duqaXxX: Recent history on the Codex plugin/docs/manifest surface includes Codex harness activation work, making this a plausible secondary routing signal for adjacent plugin behavior, though not the central timeout code. (role: adjacent Codex plugin contributor; confidence: low; commits: 69ba924b530e; files: extensions/codex/openclaw.plugin.json, docs/plugins/codex-harness.md, docs/plugins/codex-harness-reference.md)

Remaining risk / open question:

• The PR body reports mocked app-server harness and Testbox changed-check proof, but no live Codex app-server/OAuth model call after the patch; the protected maintainer review should decide whether that proof is sufficient before landing.

Codex review notes: model gpt-5.5, reasoning high; reviewed against d91f58ee25b5.