lodestar-setup

Write the agent-neutral config the lodestar skills need. The one file that matters is .agents/lodestar/context.md. Sibling skills that need repo facts read that file; they never read AGENTS.md for layout or commit policy.

Do not do a broad repo survey. Do not propose architectural changes (lodestar-architecture exists for that). Do not Read the full setup-state JSON into chat — pass its path to commands; use stdout projections only.

Resolve every bundled template path relative to the directory containing this SKILL.md. Paths beginning with .agents/ are output paths in the target repository.

What this skill does — and does not do

• Does: one deterministic collector pass; two user interactions

(review + permissions); drive writes from state + corrections; complete from summarize-results.

• Does not: ask which workflows the user plans to use, invent

Dependency Policy from today's import graph, or load full state into conversation.

How to talk to the user

Anything you print or ask is read by a person who is skimming.

• Ask one clear question at a time. Say what happens for each answer.
• Say what a thing is, not what it is called. Keep internal config keys

out of questions.

• Never make the user do arithmetic; give the numbers and your

recommendation.

• Never trim or postpone a warning. Never truncate consent rows.
• Point first. Bullets, not paragraphs. Blank line between blocks.
• Bold the first few words of each bullet.

Milestones

Concise chat progress — not diagnostic logs or run files:

• Opening — scope, output path, checks, mutation/commit policy
• Before long work — current step name
• At boundaries — completed count and next step
• Closing — from summarize-results only

Work through the steps in order. Before each step, load the references/ file that step names. Those files are one hop from this one — a step reference must not load another step reference.

Temp paths (before Step 0)

Create three OS-temp paths (via mktemp or runtime.tempDir) for:

• --out state JSON
• corrections JSON
• results JSON

Never commit them. Pass paths to commands; do not paste file contents into chat.

Let <setup-skill> be the directory containing this SKILL.md. Let <repo> be the target repository root.

Step 0 — Collect once

Follow references/00-confirm-scannable.md and references/01-collect-facts.md.

One discovery invocation only. Read stdout projection only. If scannable.total === 0 → stop, write nothing. If needsInput includes pkg-manager → ask once; write answers into corrections JSON (do not re-run full agent discovery).

Step 1 — Review (interaction 1)

Follow references/02-review.md. One review screen from the projection; one correction round → small corrections JSON. Skip audit-only headings when hasAudit is false.

Step 2 — Permissions (interaction 2)

Follow references/03-permissions.md. One tick list of writes outside `.agents/` only. Do not repeat layout/commands/conventions/scope from review. Omit Fallow rows when audit is absent.

Step 3 — Mutations

Honor ticks. Ask nothing. Follow in order:

• references/04-write-files.md
• references/05-cleanup.md
• references/06-fallow.md — **only when

hasAudit**; otherwise skip

• references/07-linters.md

After each op, record-result immediately (changed|skipped|failed + path + remedy). Never echo raw child stdout into chat. Never rebuild the summary from dirty git.

principles.md is never copied. Do not write CLAUDE.md or Copilot instructions. skills-only does not touch AGENTS.md. Never install over an in-range Fallow. A declined or failed install is not a setup failure.

Step 4 — Completion

Follow references/08-confirm.md. summarize-results --results <file> --state <file> only. Then setup-state.mjs cleanup --state --corrections --results (success or after recording failures).