Fix Stop finalization recovery false positives

[Wibias]

Summary

Fix the remaining Codex lifecycle Stop false-positive class observed after #393 / v1.3.5.

The new incident is a completed read-only analysis that ends with a required recommendation such as NEXT_ACTION = AUTHOR_NARROW_AUTHORITY_CORRECTION and an explicit stop condition. The assistant is finished, but the hook still injects recovery 1/3. The corrective continuation Cannot execute the selected next action... then fails to terminate recovery, producing recovery 2/3 and 3/3.

Root cause

#393 recognized a small set of terminal phrases after Stop, but two state-machine gaps remained:

1. the generic 12k hook generation budget fires before a long completed recommendation is classified as final output, so a valid final report without one of those phrases is treated as a missing tool boundary;
2. the corrective blocker recognizer did not match selected next action or an explicit Blocked: ... unauthorized disposition.

A blanket 64k allowance for every lifecycle Stop would weaken the existing hook-mode severe-recovery boundary. This PR therefore applies the larger completed-answer allowance only to an explicit finalization candidate and still runs the entire candidate through tool-intent/protocol detection.

Fix

• recognize the existing terminal/blocker dispositions, including selected next action and explicit authorization blockers;
• recognize the structured final-report form used by the incident: a Recommendation heading with a terminal NEXT_ACTION = ... recommendation;
• only those explicit finalization candidates may use the 40k/64k completed-answer character allowance;
• ordinary Stop messages retain the existing strict hook generation budget;
• finalization candidates are still fully scanned, so a newly announced tool action remains in recovery;
• terminal/finalization wording cannot bypass the 64k completed-answer hard bound;
• malformed tool-protocol hard stops are unchanged.

TDD / adversarial evidence

The first test-only commit, d002b5e6f22b976e0bef89864c79015d9be319e1, was RED on the unchanged runtime and reproduced the new final-report and corrective-blocker failures.

Additional adversarial coverage pins both fail-open directions:

• a terminal report above the completed-answer hard limit still recovers;
• a long finalization candidate that announces a new tool action still recovers;
• ordinary long repeated tool intent keeps the strict recovery behavior;
• blocker wording plus a new tool action cannot clear active recovery.

Scope

Only the Codex watchdog hook logic and its focused terminal-Stop regression file change. No version bump, release preparation, merge, tag, GitHub Release, or npm publication is included.

[coderabbitai]

Important

• 🔍 Trigger review

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

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: c770928b-fed7-49d7-96a3-3fe7ec36b1e2

---

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.}