playbook: init — write this repo’s profile

Use for /pitwall:init · writes .claude/agent-loop.md (and PLAN.md when the repo has no tracker and no plan) · reads first, then asks once, then writes · never commits, pushes or edits settings files — the user reviews the result and commits it

0. Already set up?

.claude/agent-loop.md exists → playbooks/doctor.md · OK → say so and stop (/pitwall:init redo starts over) · missing / unfilled sections → fill only those, through steps 1–4

1. Read the repo (no questions yet)

Each value keeps its source for the report (file:line or command)

Read the base as the loop will see it, not the working tree: git fetch origin -q, settle base first (row below), then read every file in the table with git show origin/<base>:<path> (Git Bash on Windows rewrites origin/x:path — prefix MSYS_NO_PATHCONV=1) · the local base is behind (git rev-list --count <base>..origin/<base> > 0) → say so in the report (the first /pitwall:lap pulls it, and a value read from the stale tree would be wrong: lockfile, scripts, ports)

Value	Read from
install + check commands	lockfile → package manager (package-lock.json npm · pnpm-lock.yaml pnpm · yarn.lock yarn · bun.lock* bun), frozen install form · scripts that exist in package.json (typecheck / lint / test / build) · other stacks: Makefile, pyproject.toml, Cargo.toml, go.mod, project.godot (Godot: engine version from config/features, no install step) · no package scripts → the run: steps of .github/workflows/*.yml are the next source: a step that calls a script in the repo → that script is the command, with the env vars the step sets written as a requirement (GODOT = path to the engine binary) · a check with no command → none, never invent one
base	the branch most of the last 30 merged work PRs target — count only PRs whose head is not itself a target (a <base> → release PR is a release, not work): gh pr list --state merged --limit 30 --json headRefName,baseRefName → drop rows whose headRefName is any row’s baseRefName → most common baseRefName · compare with git symbolic-ref --short refs/remotes/origin/HEAD (strip origin/) · they differ, or no work PRs yet → it becomes a step 3 question with both as options
protected branches	gh api repos/{owner}/{repo}/branches?protected=true -q '.[].name' (403 / empty → ask)
hooks	.husky/, lefthook.yml, git config core.hooksPath
commit / PR title pattern	the repo’s own rule first: a commit-msg hook and the script it calls, commitlint.config.*, a CI step that checks messages, the commit rule in the repo rules files · none → the shape of the last 20 commit subjects · nothing consistent → <type>: <summary> (type = feat / fix / refactor / test / chore) as a default listed in the report · a hook the agent’s commits must pass always wins over the default
CI checks	.github/workflows/*.yml job names, and their run: steps when the “install + check commands” row found no commands
repo rules	CLAUDE.md, AGENTS.md, .claude/rules/, .cursor/rules/, CONTRIBUTING.md, a docs / knowledge-base index
dev server	a dev / start script + its port (script flags, framework config)
verification skill	a .claude/skills/*/SKILL.md that launches or drives the app
plan	PLAN.md, TODO.md, IMPLEMENTATION_PLAN.md, tasks.md, docs/**/plan*, .taskmaster/
tracker	tracker tools connected in this session (ClickUp MCP) · GitHub Issues: gh repo view --json hasIssuesEnabled, the labels (gh label list --json name) and open milestones (gh api "repos/{owner}/{repo}/milestones?state=open") that already exist
language	language of CLAUDE.md and of the last 20 commit subjects
gh	gh auth status — fails → stop: pitwall needs an authenticated gh

2. Draft

Copy <pitwall dir>/templates/agent-loop.md and fill every {{…}} that step 1 settled · roles in agents per task: a stronger model for design, a faster one for pattern work and the verifier, and two different models in the review panel — its variety comes from the models (a default — listed in the report)

3. Ask once

Only what step 1 could not settle, plus the rows a human must confirm — AskUserQuestion, ≤ 4 questions per call, ≤ 4 options, recommended first, in the language step 1 found:

• tracker: clickup / github / none · github → which existing labels mean agent work, in progress, review and pending (labels the repo does not have yet → name them in the report for the user to create; init never creates one), QA login, milestone: current (open milestones exist) or none
• user language · PR / commit language
• base, when step 1 found two candidates (the merged-PR target first, as recommended)
• protected branches (when the API did not answer)
• level B: dev server + port, which browser, human-only steps (wallet popup, 2FA, payment confirm, none)
• no tracker and no plan → where the plan goes: PLAN.md at the root (recommended) / another path

Never ask what a file answered; never fill a business value (prices, limits, credentials) from a guess

4. Write

• .claude/agent-loop.md from the draft + answers · a row nobody could answer → none + the reason, never left as {{…}}
• no tracker and no plan → copy <pitwall dir>/templates/plan.md to the chosen path; the profile plan section points at it
• the plan file goes into git → resolve conflicts yourself: every PR ticks a row in it, and with tracker: none also appends a log line, so two open PRs always conflict there (append-only, safe to union)
• no verification skill → verification skill: none; level B items with no browser fall to C, say so in the report
• permissions: write the list into the profile; adding it to the settings file is the user’s step
• .claude/settings.json on the base has "pitwall@pitwall": true under enabledPlugins but no pitwall key under extraKnownMarketplaces → a teammate’s clone cannot resolve the plugin: say so in the report, with the entry to add ("pitwall": { "source": { "source": "github", "repo": "aontwit/pitwall" } }) — adding it is the user’s step, like permissions
• a CI workflow triggered by pull_request without edited among its types → a stacked PR moved to base after its parent merges keeps only the checks from its old base, because GitHub starts no run on a base change: say so in the report, with the trigger to add (types: [opened, synchronize, reopened, edited], and on the job if: github.event.action != 'edited' || github.event.changes.base, so edits to the title or body start nothing) — changing the workflow is the user’s step
• git check-ignore -v .claude/agent-loop.md <plan file> matches → the file cannot be committed as is: say which .gitignore line hides it and the line that un-ignores it (e.g. !.claude/agent-loop.md) — adding it is the user’s call

5. Check + report

playbooks/doctor.md → must end in “profile OK” · report in the profile’s user language: values read (with source) · values answered · defaults taken · files written · next: commit .claude/agent-loop.md (and the plan file) to the base yourself — the loop never pushes to the base, and /pitwall:lap stops on an untracked file in the tree — then /pitwall:plan when the plan has no rows yet (or the next phase is TBD), otherwise /pitwall:queue