[abruption]

Problem and proposed contract

Windows opener lists, process-exit observations and owner claims provide different evidence, but the earlier evaluation did not spell out a complete request/verification boundary. This research record specifies the remaining W4/W5 conditions and the proposed owner-cooperative protocol. Refs #149 and #150.

• Keep R1 user-exit and R3 idle-unload observations diagnostic. Coincident process exit and a free lock do not identify the holder or prove normal cleanup.
• Treat an authenticated owner report as a proposed bound claim, with process/connection and duplicate-reference questions still open. It is not single-holder proof.
• Require a future R2 endpoint to bind fresh single-use requests to the exact process/thread/runtime/guard and persisted transcript baseline, atomically revalidate admission, and preserve a separate completion record. Uncertain delivery and failed post-verification do not authorize retry or force.
• Correct the scope of ordinary versus v2-child teardown, persistence failures, runtime replacement, conditional daemon termination and socket-directory validation. The earlier Windows evaluation now leaves duplicate-handle lock lifetime explicitly unverified.
• Add an unpublished source verifier for fixed Codex and Rust revisions. It requires the Rust std input as well as the Codex clone, and checks bytes/anchors/literal searches rather than semantic correctness.

The record is a proposal. Production CLI behavior, supported platforms, package contents and release settings are unchanged. The known daemon endpoint still fails the existing owner-mode policy; no shared-owner exception is adopted. #149 and #150 remain open pending native evidence and policy decisions.

Validation

• npm ci, check, lint, test (263 pass, 3 Linux-only skips) and package smoke passed.
• Source verifier: 26 upstream file hashes, 110 anchors, 13 literal searches and 11 Rust std anchors passed at Codex 83086188ae0614bd882818ff1aa23f3194962012 / Rust 59807616e1fa2540724bfbac14d7976d7e4a3860. Wrong std input returned exit 1; missing std and no arguments returned exit 2.
• Root independently checked the pinned source branches and tag mapping. No Windows host, actual Codex session or upstream executable was used.
• Two rounds of independent source review used non-interactive Claude Code 2.1.302 (claude-opus-5-5, response metadata confirmed) and Codex CLI 0.163.0 (isolated default selection; actual model unconfirmed), in parallel within each round. Both read identical frozen evidence with pinned upstream excerpts. Revised input SHA-256: e7528480d95ec4eaff62591cef95f6594e0dae222d6f0fe140de82e122729b6f. Root checked findings against source and corrected nonce/transcript binding, diagnostic route scope, pre/post-action outcomes, file-name absence, evidence limits and source wording. The reviews did not execute Windows or certify later corrections; current-head CI/reviews remain separate gates.
• A missing path name is recorded as name_absent; original-object lock release remains unknown without additional evidence. Guard release before waiting applies to diagnostic routes too. Cross-connection completion recovery is deliberately unresolved rather than assumed available.

Remaining work includes native connection/instance binding, duplicate-handle behavior, actual cleanup/transcript effects, supported storage and guard liveness, endpoint/API availability, and the final recovery policy. Source-verifier success does not resolve those gates.

At head 15b5ad822dcbeae4711c8f0df3343ed421f324ae, all 39 GitHub checks passed. Automatic Code Review and Security Review both completed for that commit with no findings. This remains an open research PR; no merge or release was performed.

[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 2026-10-18T02:37:48.204759Z	15b5ad8	PR opened
🔒 Security Review	✅ Completed 2026-10-18T02:37:54.439697Z	15b5ad8	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.