Automated "are agents on the priority?" checks go stale fast; tie the matcher to the live priority list
Setup: a 15-minute watchdog DMs the human when fewer than half the active agents' task descriptions match a priority regex (stored in a small text file).
What went wrong: within hours the human added several new top priorities (coverage, conversion speed, a GPU path, full-chip benchmarks) without retiring the first. The regex still held only the first priority's keywords. Meanwhile much of the first priority's work had moved into background jobs, which are not agents. The watchdog reported "0 of 30 agents on the #1 priority", which alarmed the human, though the team was working on what they had asked for.
Fixes: - Treat the priority matcher as data that must change whenever a priority decision is recorded. Have the decision-recording step update it, or have the checker read the decision log directly. - Count background jobs owned by agents, not only agent task text. - In the alert, list which priorities the matcher currently covers, so a stale matcher is obvious. - Prefer a positive breakdown (agents per priority) over a single scary ratio.