The gate cannot lock itself out (#7790)
On 2026-11-01 a bug in the live gate could not be fixed through the gate, and landing stopped for about 4.7 days.
- It says so. When the last 3 merge-queue gates of the forge's own repository were all red for one cause (they share a failing test; or none names a test and they share a reason) and nothing has landed since the first of them, the flow alarms carry one sentence,
Landing has stopped: …, on home, on/boardsand in the terminal's Flow dialog. Owner-only, like every flow alarm (an agent reader andpublicOnlynever see it). "One cause" stands in for "one base", which a gate result does not keep: three unrelated broken pulls fail different tests and do not raise it.checks/gate_stopped.ts. - An owner lands one named pull past it, only while it holds. The break-glass is offered, and accepted, only while landing has stopped and only on the forge's own repository; otherwise it is refused (
not_stopped). An owner (repo admin, a person, never an agent) has three doors, all throughlandPastGateinchecks/bypass.ts:
- the pull's Files step: a "Land past the gate" form (a labelled reason field and a danger button, plain post, no script needed);
- MCP: land_past_gate { repo, number, reason, dryRun: false } (planned unless dryRun is false; the plan checks who, the stop and the reason, as the apply does);
- the terminal: L on the pull screen asks the reason, shows it back in a confirm, then lands.
The reason (8–500 characters) is written to the pull's trail first, then the pull lands, then pull.bypassed goes to the outbox; a refused merge leaves a second comment saying so. Only the checks requirement is skipped; review, conflicts and the agent rule still hold.
- Test:
test/gate-bypass.test.ts.
Not done: running the gate from a pinned earlier release rather than the tree being judged.