本文へ移動
cccskills
無料GitHub で公開

factory-monitor

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.

インストール方法を見る

含まれるファイル(1)

  • SKILL.md4.4 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Factory monitor

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.

Sweep

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 automatically
  • awaiting-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 intake
  • wait-to-implement whose named blocker has since resolved → promote it
  • needs-info with an answer now in the issue comments → send back to triage

5. 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.

Filing

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.

Report

Write one unique monitor run record under docs/factory/runs/:

  • one line per finding with its issue link
  • the current queue depth by disposition
  • the review-queue depth, called out separately, because that is the number that actually constrains the factory
  • what you checked and found clean, so the absence of findings is distinguishable from the absence of checking

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.

What you never do

  • Fix anything, including one-line fixes that look obviously safe
  • File more than 10 issues in a run. If you found more, file the top 10 by severity and say how many you dropped. Never silently truncate.
  • Reopen or comment on issues a human closed

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Claim and implement one ready GitHub issue, run fail-closed gates, obtain independent verification, and open a draft pull request.

日本語の概要は準備中です。原文の説明を表示しています。

addyosmani/factory2302026年8月22日 更新

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.

日本語の概要は準備中です。原文の説明を表示しています。

addyosmani/factory2302026年8月22日 更新

Inspect factory health, stale work, CI, security advisories, and recent run evidence without implementing fixes.

日本語の概要は準備中です。原文の説明を表示しています。

addyosmani/factory2302026年8月22日 更新

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.

日本語の概要は準備中です。原文の説明を表示しています。

addyosmani/factory2302026年8月22日 更新

Turn an ambiguous factory issue into human-approved product, behavior, design, and implementation slices. Use interactively before implementation.

日本語の概要は準備中です。原文の説明を表示しています。

addyosmani/factory2302026年8月22日 更新

Show the factory control room from live GitHub labels, open PRs, recent run records, gate health, and charter limits.

日本語の概要は準備中です。原文の説明を表示しています。

addyosmani/factory2302026年8月22日 更新

addyosmani のスキルをすべて見る

このスキルの問題を報告する