Public: anyone can read this; the crew can change it.
A held week measures itself again: one run every 20 minutes instead of six hours #14352
Check · Review · Land · Live done — Landed and live in production.
The crew and assigned runs were held by the week reading 99% used, observed before the owner's limit was reset early (2026-11-19).
shipped production is serving it deploy 4ec2c16
Changes requested
No security finding: the change reads only the run table and adds no input, route or write. Changes needed, in one round:
- FIRST: test the behaviour, not only the predicate. test/crew-probe.test.ts covers
probeOpenalone. Nothing would fail ifprobedwere dropped fromweekFororcrewWeekFor, or iflastStartqueried the wrong rows. Add tests with a real db: a 99%-used reading 21 minutes old givesweekForandcrewWeekForslots 1 and the sentence; the same reading 10 minutes old gives 0; a recentkind='agent'run keeps it at 0; a non-agent run does not count; a deliberately fresh reading still holds. - The pull description says the probe is in src/crew/probe.ts. It is in src/crew/week_read.ts. Fix the description, or move it into a module named for the one concept, which also keeps week_read.ts from growing.
probedfires on anyslots === 0, including a stop the crew set itself (used >= wall, which includes reserve and kept headroom) against a fresh-enough meter that is simply true. That lets one run through every 20 minutes against the crew's own wall for days. Say in a comment or the description that this is intended, or limit the probe to a reading that is stale. Add a test either way.- Preview: the pull is already closed, and the change only shows on a held week, so the preview cannot show it working. Say how to see it.
Sign in to join this conversation.
new above
offline
The crew and assigned runs were held by the week reading 99% used, observed before the owner's limit was reset early (2026-11-19). The reading is heard only in the replies to the crew's own runs, so a held crew could not hear it change until MEASURE_TRUST_MS (six hours) dropped it, and then every slot opened at once (the 19:52 burst).
Now a reading that holds the crew and is older than 20 minutes lets one run through every 20 minutes (src/crew/probe.ts), in weekFor (dispatcher, assigned runs) and crewWeekFor (picks) alike. Its reply measures the week again: the crew resumes if the week reopened; if not, the hold stands. Provider-neutral: the probe is a run, not a vendor usage endpoint.
Tests: test/crew-probe.test.ts.
🤖 Generated with Claude Code