You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Ship review quorum reliability: hatch approval burned by magic-close false positive; opus seats unusable (exclude/hold/expired) #243
During PR #242 ship (2026-11-06) the review-quorum Guard-B (agent-tools ci/ship/shiplib/quorum_skipci.py:341, hatch check :361) had zero usable seats and its hatch escape burned an owner approval: (1) the owner's Telegram hatch approval was consumed by a magic-close refusal triggered by the phrase 'do not close' inside the PR body — the approval should only be consumed when the ship actually proceeds, and the magic-close check should run before the approval is requested; (2) claude:claude-opus-5-5 CLI seat's only account (invntrm@gmail.com) is excluded by omp_accounts.exclude; (3) omp:anthropic/claude-opus-5-5 seat is held at 'USAGE UNKNOWN' because the alex.mex.ux account has no 5-hour window and the completion-probe path does not clear the hold (7-day usage 76% is under the 80% cap); (4) the third omp account (mextner@gmail.com) is disabled with an expired refresh token.
Why (motivation)
The quorum gate is the only remaining merge blocker for ready PRs (PR #242 was fully green 2026-11-06 and could not ship for these reasons); a one-owner approval flow that silently burns approvals erodes trust in the gate.
User impact
Ready, fully verified PRs can be blocked for days (codex quota until 2026-11-09) with no honest machine-review path and no working hatch.
Cost of inaction
Every future ship that lacks a codex/claude seat will hit the same wall and need ad-hoc human intervention.
Acceptance criteria
A granted Telegram hatch approval is never consumed by a refusal that happens after the approval (magic-close and all other refusal checks run before the approval request, or the approval is re-usable until a ship succeeds)
An omp Anthropic account under the 7-day cap becomes usable when its 5-hour window is unobservable: the probe path clears the USAGE UNKNOWN hold, or the hold has a documented manual clear
At least two live review seats exist at all times (documented rotation: exclude list, hold handling, refresh-token re-login for mextner@gmail.com)
Screenshots
(none)
Deploy
Deployed: no
Links
(none)
Skipped gates
links: Cross-repo findings for agent-tools tracked here per repo convention for ops tickets
msgref-quote: Observed directly in session logs and review seats/health output on 2026-11-06
screenshots: CLI/tooling behavior evidenced by command output, not UI
Activity