AGENT SKILL

factory

Hand ready tickets to the factory, resolve holds, resume stopped work, deliver skipped work, or pause and resume one repository.

License 许可
MIT
最近更新
2026-11-15
Stars
★ 0
Forks
0

01 前言 Front Matter

name
factory
description
Hand ready tickets to the factory, resolve holds, resume stopped work, deliver skipped work, or pause and resume one repository.

02 资源文件 1

03 SKILL.md 正文

A prompt or brief naming this skill is its invocation.

Work in the repository the factory serves. Read CONTRIBUTING.md, the installed factory-workflow, github-repository-configuration and comments rules, and .scriptorium/config.json before acting. The guide and rules hold the requirements; this skill supplies the method. Use the configured label roles rather than assuming the shipped strings. Read the issue's authoritative contract, comments, blockers, and any existing PR and handoff before choosing a path below.

Hand a ready ticket to the factory

  1. Check the contract's accepted revision and its validator notice, parent readiness and open blockers under the guide and installed rules. Resolve missing required inputs before seeking readiness.
  2. Choose the ticket's model, weighing its triage suggestion. Prefer a stronger model for design judgment, ambiguous scope or a long cross-cutting change. Use the rules' model-label syntax and metadata; the host operator resolves an unavailable or disallowed model rather than substituting one silently.
  3. Apply the readiness policy through triage's readiness command. With an App or bot credential in a factory repository, hand the reviewed revision, suggested readiness, model and route labels, and blockers to an authorized person. Report pickup only when the factory actually records it.
  4. For a parent held under the latest-link-time fallback, wait until its latest ticket link is visible in the timeline, then remove and re-add its ready label after that timestamp. Each ticket still needs its own readiness. With complete authorized link history, use the installed actor-history rule.

Read a hold

Read the factory's run comment and current labels first. Match the reason rather than clearing a state label to see what happens:

Observation Next action
Refused readiness or parent not ready Verify the exact revision and the authorized person's readiness, then the parent's contract and link history.
Model-label error Reconcile the ticket's single model label with the chosen model; ask the host operator about resolvability or the allowlist.
Repository configuration refused Ask the host operator to verify the serving release and configuration acceptance. Keep the issue's evidence available.
Open blockers, capacity or usage drain Resolve blockers or wait for the host's gate to clear; no label reset bypasses it.
Repository or release hold Find the labelled open items and use Pause and resume below.
Running or paused Read the run comment; the factory owns the run and its automatic resumption.
A state: stop Use Resume stopped work below.
Failed Address the failure the run comment names first, then remove the failed label to request a retry subject to pickup gates.
Skip Use the directly launched path below; clearing a waiting marker does not request a factory retry.

If no comment explains a held issue, ask the host operator for its reason; a host with no cached valid configuration cannot comment on the issue. Keep credentials and secret host material out of tracker records.

Take the skip route

For work excluded by the installed rules, arrange the ready, model and skip labels together through the authorized readiness path. Launch implement directly with the contract, evidence, decisions, issue link and delivery role. The worker opens the PR and records its handoff as .repo-standards/rules/comments.md#delivery-records says. Continue as the orchestrating agent through babysit from that handoff; apply its current-head owner-approval gate when designated.

For an unanswered owner question, retain the skip marker and the question's workflow state. After resolution and any required renewed readiness, resume the worker if no PR exists, or continue the same open PR and its handoff. Live host access gates only the authorized live steps.

A live-acceptance ticket takes this route. Its worker records the ticket's live acceptance on it, as .repo-standards/rules/comments.md#disclaimers says, instead of opening a PR, and reports back. It closes only once its live acceptance is recorded on it, never on a merge; its source ticket's PR links it without a closing reference. A time-bounded host access grant belongs to the orchestrating agent that asked the maintainer for it, which records the workers using it. Workers never end it: it ends when the last of them reports back, or at its own time bound.

Resume stopped work

Answer the question through the guide's owner-question route, posting as .repo-standards/rules/comments.md#disclaimers says, and meet the installed Questions and stops rule's conditions for a changed contract. Then return the issue to state:ready, which is the only way to resume it: never clear another label for it, and never merge the stopped PR by hand. For skipped work, continue as Take the skip route says instead.

Pause and resume one repository

  1. Read labels.hold in the repository configuration and locate every open issue and PR carrying it. If the key is absent, arrange a compatible configuration with the owner before attempting a pause. Distinguish this repository hold from the release hold and the factory's per-run paused state.
  2. For an expedited change with an existing PR, put the hold on that PR. If work starts before its PR exists, put it on the change's issue first and record why the repository is paused and which change will carry it, as .repo-standards/rules/comments.md#disclaimers says.
  3. Deliver the expedited change outside factory pickup: launch a worker directly or implement as the orchestrating agent. A held repository launches no new work. An already-open factory PR can instead carry the hold and finish through its exempt merge relaunch. Runs already started finish without merging while another open item holds the repository.
  4. Once the expedited PR opens, add the hold to it before removing the hold from the issue. Verify that transfer on both items. Before merging, inspect all other open items carrying either configured hold: only the PR being merged is exempt, so a forgotten issue hold would block it too.
  5. Finish validation, applicable reviews and merge authorization through babysit. Its merge check exempts this PR's own hold, allowing the expedited change to merge. The PR's closure ends that hold automatically. To resume a deliberate pause without merging, remove the hold from the item that carries it; leave other owners' holds alone. Confirm no remaining open hold and report the resulting state. A release hold still restricts merges separately.

<!-- Seam (deferred): Scriptorium configuration reads; consumers: 6; extraction proposed in https://github.com/lutzseverino/repo-canon/issues/967. -->

Reads

Paths are relative to the repository root.

  • CONTRIBUTING.md: contract, readiness, delivery, question and credential requirements. When it is missing, use the installed rules and hand off any unresolved delivery or authorization question on the issue or PR.
  • The installed factory-workflow and github-repository-configuration rules, located with repository-standards. When they are missing, use the guide and configured roles; ask the owner before a route or merge-ownership decision that those sources do not settle.
  • .repo-standards/rules/comments.md: the installed comment rules, whose sections the posting steps name. When it is missing, post each record the steps name as one comment.
  • .scriptorium/config.json: configured label roles. When it is missing or unreadable, perform read-only diagnosis and ask the owner to establish the factory configuration before changing factory labels.

04 同仓库更多 Skills