vegafactory-setup
Bootstrap and maintain the org control room — the org, group, repo, worker and board files whose defaults every repo's dev profile layers on. Use when asked to "set up the control room", "bootstrap the org for the factory", "register this repo in the control room", "onboard a teammate", "add someone to the org", or "what are the org defaults here", and when a repo's control-room knob names a control room that does not exist yet. Not for a single repo's profile, labels, or knobs (dev-setup), not for architecture advice (dev-architect), not for writing or approving issues (dev-intake), and not for refreshing the local copy of the room (the `vegafactory sync` command).
- License 许可
- MIT
- 最近更新
- 2026-09-25
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- vegafactory-setup
- description
- Bootstrap and maintain the org control room — the org, group, repo, worker and board files whose defaults every repo's dev profile layers on. Use when asked to "set up the control room", "bootstrap the org for the factory", "register this repo in the control room", "onboard a teammate", "add someone to the org", or "what are the org defaults here", and when a repo's control-room knob names a control room that does not exist yet. Not for a single repo's profile, labels, or knobs (dev-setup), not for architecture advice (dev-architect), not for writing or approving issues (dev-intake), and not for refreshing the local copy of the room (the `vegafactory sync` command).
02 资源文件 16
- 参考文档agents/openai.yaml227 B
- 资产文件assets/control-room/boards.md.template1.3 KB
- 资产文件assets/control-room/group.md.template2.2 KB
- 资产文件assets/control-room/nodes.md.template2.1 KB
- 资产文件assets/control-room/onboarding/new-node.md.template972 B
- 资产文件assets/control-room/onboarding/new-repo.md.template818 B
- 资产文件assets/control-room/onboarding/new-teammate.md.template1.1 KB
- 资产文件assets/control-room/onboarding/worker-box.md.template14.3 KB
- 资产文件assets/control-room/org.md.template2.5 KB
- 资产文件assets/control-room/repos.md.template278 B
- 资产文件assets/control-room/stats/README.md.template1.0 KB
- 参考文档evals/evals.json9.5 KB
- 参考文档references/control-room.md14.8 KB
- 脚本tests/boards.test.ts1.7 KB
- 参考文档tests/fixtures/trigger-queries.json1.2 KB
- 脚本tests/worker-box.test.ts3.8 KB
03 SKILL.md 正文
vegafactory-setup
Act: give the org one control room, so every repo inherits the answers it would otherwise be asked for.
The control room is <org>/vegafactory-control-room, and it is seven things: org.md, groups/<g>/group.md, repos.md, nodes.md, boards.md, onboarding/ and stats/. Nothing else belongs in it. dev-setup reads it before its interview and states an inherited knob instead of asking for it. The layout, the precedence rule, the lock, the read path, and what each file may carry live in control-room; this file is the procedure.
Nearest neighbor: dev-setup owns one repo's profile — its knobs, labels, runbooks, and AGENTS.md section. This skill owns the defaults that profile layers on. When the question is "what should this repo do", that is dev-setup; when it is "what should every repo do", it is this one.
Bootstrap — the org questionnaire, then the files, then a stop
- Ask the questionnaire. Only what
org.mdholds and nothing a department owns: the org name, the goals in one paragraph, and what applies to everyone — the language, the date format (DD-MM-YYYY across the workflow), the "nothing ships without the operator's explicit instruction" stance, and the two statistics linesstats:andstats-people:. A knob a department decides — the merge style, the tests, the harness policy — is never asked here, because two departments reading one org answer would each need the other's to be wrong. - Ask which lines are locked, one at a time, and default to none. A locked line is one the org will not let a group or a repo answer differently; the comment
# lockedon the line is the whole mechanism. Propose a lock only where one answer must hold everywhere — the provider mode is the usual one — and say what a group loses by it. A lock a group will need to work around is the wrong lock, and a lockedharness-policy:must name all six stages or it is refused. - Render every template in
assets/control-room/into a working directory the operator can read before anything is pushed:org.md,groups/dev/group.md,repos.md,nodes.md,boards.md,onboarding/,stats/. Rendering is substitution only: a template is a default, and a default nobody has confirmed is still a default. - Stop before the repository exists. Creating the org repository and granting anyone access are the operator's own account actions. Name the exact commands, say what each one does, and let the operator run them. The skill positions the operator; it does not reach for the credential.
Nothing secret goes in any file — names of secrets only. A control room is readable by everyone the org onboards, so NPM_TOKEN is the entry and the value stays in the store that name points at.
Round — automation identity
The org's automated writes go out as the public GitHub App, not as a person's token. This round records the App by name and stops at every step that needs the operator's own account.
- Hand the operator the creation walk in dev-setup's
references/github-app.md— the permission table, the "Any account" setting, the webhook left off, and the private-key step — rather than restating it here, so the App's contract has one home and one place to update. - Detect an existing installation by reading the org's installations endpoint and selecting the one whose
app_slugisvegafactory. Write the returned id ontoorg.md'sapp-install:line. - Leave the line as an unconfirmed placeholder when the call answers 403 or returns nothing, and say which happened — the endpoint answers organization owners only, so a 403 means "not an owner", never "no App". An unconfirmed line is a question the next run asks again.
- Never create the App, never ask for the PEM, never write a secret value into any control-room file. Generating the private key is a browser download GitHub delivers once, to whoever pressed the button; an automated session's download never reaches the operator. Record the two names,
VEGAFACTORY_APP_IDandVEGAFACTORY_APP_PRIVATE_KEY, and leave the values in GitHub org settings.
Round — the board
A board mirrors the state labels; it never drives them. This round records the board and hands the operator every command that touches a project.
- Offer the board on the operator's yes, one plain sentence: labels drive the workflow and the board follows, one way — a card dragged on the board is cosmetic until the next label change.
- Run nothing that mutates a project. Creating a board, replacing its Status field, linking a repo and switching on the built-in automations all need the
projectscope on the operator's own token. Hand them the sequence in control-room's## Boardssection verbatim, starting withgh auth refresh. - Write the
boards.mdrow fromassets/control-room/boards.md.template— board title, number, the repos mirroring onto it, and a notes cell — once the operator reports the number. - Set
board: <n>in each linked repo's.vegastack/dev.md, and leave it atnonefor a repo with no board; the workflow reads that line and does nothing when it saysnone. - Record a blocked repo rather than working around it: where the org plan's auto-add cap refuses another board workflow, the row in
boards.mdsays so with its date and the request goes back to the operator.
Seeding groups/<g>/group.md
A group file carries one default for every knob a group can decide for a repo's .vegastack/dev.md — per-repo facts such as repo:, board: and control-room: stay in the repo — so a repo that answers nothing else still gets a complete profile. Seed it from an existing repo's dev.md: confirmed knob lines transfer verbatim, including harness-policy:, because a group file uses the same parser. When no confirmed policy exists, use dev-setup's refresh-tracked harness selection and flag mapping in references/harness-facts.md; this skill does not freeze vendor model/version defaults. <!-- mirrored -->
A knob whose value differs between two repos in the group is a question, not an average: ask which one is the group's default, and let the other repo keep its hand edit — the precedence rule already protects it.
register <repo>
- Confirm which group the repo belongs to.
- Run
dev-setupin the repo. It reads the control room first, so every inherited knob is stated rather than asked, and only what no layer answers reaches the interview. - Append the repo's row to
repos.md— repo, group, board, owner. - Link the board when
boards.mdnames one for the group, and add the board-mirror workflow to the repo. - Confirm the repo's
control-room:knob names this control room and this group.
The full checklist ships as onboarding/new-repo.md, so the org can edit the procedure without editing this skill.
A machine is the third onboarding path, in two shapes. Any ordinary machine follows onboarding/new-node.md and gets one nodes.md row with worker: no, which makes statistics answerable per machine and grants nothing. A machine that works boards unattended follows onboarding/worker-box.md under two accounts, so a CI job cannot read its tokens; its row says worker: yes.
onboard <login>
Walk onboarding/new-teammate.md with the person: gh auth login at or above the group's gh-floor, the harnesses the group's harness-policy: line names, the skills install, control-room read access, and the Slack subscription through the official GitHub Slack app. Then, only on the operator's word, add their login to the group's operators: line — that line is the one place a person is recorded, and org membership alone never puts anyone on it.
The declined step
Every step here can be declined, and a declined step is recorded, not skipped: write it into org.md's ## Unconfirmed section as one line, in the same form dev-setup writes for a knob it could not confirm. The next run asks again. An unrecorded decline becomes an assumption, and an assumption in a file every repo inherits is the most expensive kind.
A control room that does not exist yet, or whose local copy is missing or stale, is not an error either: say which answers could not be inherited and let dev-setup ask them.
Routing
| Need | Read |
|---|---|
| the file tree, precedence, the lock, the read path | control-room |
| comment markers, operator identity, register line format, labels | dev-setup's references/conventions.md, installed beside this file |
| the seed text of any control-room file | assets/control-room/<file>.template |
04 同仓库更多 Skills
skills-refresh
Re-verify the dated platform and harness facts the dev skills pin, and file what changed as issues. Use when asked to "refresh the skills", "check the pinned facts", "are our platform facts still true", "run the facts sweep", "what changed in Cloudflare/Next.js/Claude Code since we last checked", when a fact is older than the 60-day window, or when a scheduled refresh run starts. Not for editing skill prose (dev-implement builds the issues this files), authoring or auditing a skill (skillify), or architecture advice from those facts (dev-architect).
beta
fixture
alpha
fixture
beta
fixture
alpha
fixture
alpha
fixture