Claim and implement one ready GitHub issue, run fail-closed gates, obtain independent verification, and open a draft pull request.
The English summary is being prepared. Showing the original description.
Scheduled health sweep that closes the factory loop - reads CI failures, recent commits, dependency and security advisories, and live queue labels, then files issues and writes an immutable run record. Use for the nightly or weekly monitor routine.
The English summary is being prepared. Showing the original description.
See how to installReview the instructions the agent will receive before you install.
This is the stage that closes the loop. Without it a factory only processes work a human remembered to file, which means the backlog reflects attention rather than reality.
You file issues. You do not fix anything. Mixing detection and repair in one run means a bad detection becomes a bad commit before anyone sees it.
Read docs/factory/CONTRACT.md and docs/factory/CHARTER.md first.
Run each of these. Where a check is not applicable to this repo, say so rather than silently skipping it.
1. CI failures. Failed runs on the default branch since the last sweep. For each, identify the failing job and whether it is a genuine regression or a flake. A test that failed once and passed on rerun is a flake finding, not a bug finding, and it is worth tracking separately because flakes are what erode trust in the gates.
2. Gate health. Run ./.claude/scripts/gates.sh deep on the default branch. A gate
that is red on main means every downstream verdict this week was measured against a
broken baseline. This is the highest-priority finding the sweep can produce.
Treat MISCONFIGURED as a factory-blocking finding. Also report optional skips so a human
can decide whether they should become required.
3. Dependencies and advisories. New security advisories affecting direct dependencies. Group by severity. Do not file an issue per transitive dependency; that is noise that trains people to ignore the sweep.
4. Queue staleness. Query live GitHub labels and issue timestamps. Use QUEUE.md only
as supporting history:
in-progress with no new commit or run record for 2 hours → flag the deterministic
claim branch for human recovery; do not delete or take it over automaticallyawaiting-review for more than 7 days → the human review queue is the bottleneck,
and per the charter's STOP_IF the factory should be throttling intakewait-to-implement whose named blocker has since resolved → promote itneeds-info with an answer now in the issue comments → send back to triage5. Comprehension drift. Files changed by the factory more than 5 times in the last 30
days with no corresponding update to their documentation or to docs/factory/DECISIONS.md.
These are the places where the code has moved and the written understanding has not, which
is where the next surprise comes from.
6. Charter gaps. Anything triage flagged as not covered by CHARTER.md. Collect them
into a single issue for a human decision rather than one issue each.
For each genuine finding, file a GitHub issue with the factory:monitor label. Before
filing, search open issues for a duplicate; a monitor that refiles the same issue weekly
gets muted, and a muted monitor is worse than none.
Issue body:
**Detected:** <date> by factory-monitor
**Category:** ci-failure | gate-health | advisory | queue-staleness | comprehension-drift | charter-gap
**Severity:** blocks-factory | high | medium | low
**What**
<one paragraph>
**Evidence**
<log excerpt, gate line, commit range - the actual artifact, not a description of it>
**Suggested disposition**
ready-to-implement | ready-to-spec | needs-info
Use blocks-factory only for red gates on the default branch or a full review queue. Those
two conditions mean the factory should stop producing, and the severity should say so.
Write one unique monitor run record under docs/factory/runs/:
That last point matters more than it looks. A silent monitor is ambiguous between "nothing is wrong" and "the sweep did not run", and those need very different responses.
No reviews yet. Tell others what you thought.
No login needed. Share how it worked for you, ask a question, or note where it did not run.
No comments yet. Be the first to write one.
Claim and implement one ready GitHub issue, run fail-closed gates, obtain independent verification, and open a draft pull request.
The English summary is being prepared. Showing the original description.
Claim one live GitHub queue item, implement it end to end, run fail-closed gates and independent verification, then open a draft PR. Use for one ready-to-implement issue or the implementation routine.
The English summary is being prepared. Showing the original description.
Inspect factory health, stale work, CI, security advisories, and recent run evidence without implementing fixes.
The English summary is being prepared. Showing the original description.
Turn an ambiguous factory issue into human-approved product, behavior, design, and implementation slices. Use interactively before implementation.
The English summary is being prepared. Showing the original description.
Turn a ready-to-spec queue item into an approved plan through explicit human gates. Use when an item needs scope decided before code exists, when planning a feature, or before starting a migration. Runs interactively - not for unattended routines.
The English summary is being prepared. Showing the original description.
Show the factory control room from live GitHub labels, open PRs, recent run records, gate health, and charter limits.
The English summary is being prepared. Showing the original description.