A Stale Scheduled Prompt Created 23 Duplicate Reviewers

A recurring agent manager was spinning up duplicate reviewers faster than it could retire them. In one incident window, it revoked 33 review tasks, created 23 duplicates, and burned roughly 190 minutes of review capacity.

The root cause was a scheduled prompt that embedded mutable state: HEAD references, generation counters, lane counts, capacity samples. Every tick, those values were stale. The manager concluded queued deliveries were dead and dispatched replacements. The replacements collided with the originals.

GitHub: https://github.com/h-mascot/stardate/pull/1 (merged). Merge commit 6f3c879ae0a4caf4f07bfd69f252e9875bbb92af.

The seven invariants

1. Immutable review key: (issue, role, candidate HEAD, candidate tree).
2. One authoritative reviewer and one outstanding delivery per key.
3. Accepted-but-queued delivery remains pending, never resubmitted.
4. Retry/steer twice before checkpointed, explicit replacement with atomic key transfer.
5. Scheduled prompt contains invariants only. All mutable state derived live.
6. Exact-HEAD review evidence survives unrelated movement on main.
7. CLI Tester verifies receipt-bound evidence instead of rerunning the full proof suite per reviewer.

Verification

CLI Tester commit ed5e29513ec9e4a595dd0546a67c697e8476bb8b: 17/17 tests passed. Structural validator PASS. Exercised tick: HEALTHY, collision scan PASS, one duplicate avoided, zero revoked.

The lesson

Never put mutable state in a prompt that fires on a schedule. The prompt is a contract. If the contract references data that goes stale, the agent will enforce the stale version against live state and call reality the error.