Dw Init

dw-init — scaffold a repo for the solo lane

Everything the loop assumes a repo has, written in one gated pass: the .ai/ state the skills read and write, the durable homes they promote into, the guardrail hooks, and a settings file whose allow-list is derived from the project rather than guessed.

What it writes

Deliberately absent: .ai/verify/ and .ai/handoffs/ — the solo lane has one thin closing pass that writes no artifact, and no one to hand off to. The root VERIFY.md is not that folder: one file saying how to drive the project, written once and read by every change, never a per-change artifact. Also absent: CLAUDE.local.md. Agent memory is tracked AGENTS.md, because a gitignored file survives neither a fresh clone nor a git worktree, and a second always-loaded file forks the corpus in two — docs/decisions/ in a scaffolded repo carries the record. Personal, cross-project notes belong in ~/.claude/CLAUDE.md. .gitignore still names CLAUDE.local.md so a stray one is never committed, and the hooks still read it when they find one, so a repo scaffolded before this keeps working untouched.

docs/agents/ ships with its README — the layer's contract, and the file declaring its per-file budget — and no topic files. The Task Router ships with rows for what the scaffold actually creates, and the topic layer grows only when dw-land promotes into it: a new topic file and its router row in the same commit, with the corpus baseline re-recorded beside them.

Templates come from ${CLAUDE_PLUGIN_ROOT}/templates/ — this lane's own payload. (${CLAUDE_PLUGIN_ROOT} is the env var Claude Code substitutes to this plugin's install dir.)

Workflow

1. Detect — never assume the stack

• Repo root (git rev-parse --show-toplevel) and default branch (git symbolic-ref --short refs/remotes/origin/HEAD, else init.defaultBranch, else main).
• Test / lint / typecheck commands from the manifests actually present — package.json scripts, Makefile, pyproject.toml, go.mod. Read the real commands; don't invent them. Keep this list — it becomes the permissions.allow entries in step 4, and a command you didn't find here must not appear there.
• Pre-commit signals for step 5: formatter and linter deps/configs, a test/typecheck script — the detection table in references/precommit.md.
• What already exists: AGENTS.md, CLAUDE.md (a real file? a symlink? to what?), a legacy CLAUDE.local.md, .claude/settings*, .gitignore, CONTEXT.md, docs/. This is rarely a greenfield tree, and step 3 must diff against reality. A repo already carrying a real CLAUDE.md with content is the case to slow down on — it becomes AGENTS.md plus a symlink, which is a rename the user has to approve at the gate, not a clobber.

2. Pick the hooks

Six are always offered because they're stack-agnostic: bash-guard (the PreToolUse dispatcher — one stdin parse, then it spawns each guard only when the command could trip it), block-dangerous-commands, block-env-access, enforce-commit-hygiene, credential-leak-guard and large-file-guard. Add the JS/TS ones only where that stack is actually present: block-non-pnpm, lint-on-edit, typecheck-on-commit (the declared typecheck before a commit that stages TS — its per-turn predecessor cost a measured 4-5 s at every turn end). On a stack with no lint or typecheck hook, offer the six alone and say the rest are stack-specific rather than silently writing nothing.

guard-plugin-canon is shape-specific, not stack-specific — offer it only where step 1 found a plugins/ directory whose entries are symlinks back into the tree. It refuses an edit aimed through one of those links and names the canon instead, so on a repo with no such layout it is a hook that can never fire.

3. HARD STOP — show what you're about to write

List every path, marked tracked or ignored, with a diff for anything that already exists. Add three things that aren't paths: the permissions.allow list derived in step 1, so what the agent may run without asking is approved rather than assumed; the optional pre-commit offer (step 5), so the one gate covers it; and, when the repo already has a real CLAUDE.md, the git mv CLAUDE.md AGENTS.md rename it needs, spelled out — that one moves a file the user wrote and must never be inferred from silence. Wait for explicit confirmation. Scaffolding mutates the repo and a wrong clobber is expensive — this gate is not optional even though the rest of the lane is light.

4. Write

• mkdir -p .ai/work .ai/backlog .ai/archive docs/decisions docs/agents; seed .ai/work with .gitkeep (the other four get READMEs). Remove a .gitkeep that a README now supersedes — an earlier version seeded docs/decisions/.gitkeep, so a repo scaffolded then and re-run now keeps a redundant one beside the README, and the next reader cannot tell which is the convention.
• .ai/README.md — copy ${CLAUDE_PLUGIN_ROOT}/templates/work-README.md verbatim. It states the lifecycle a reader gets wrong: a CHANGE.md leaves work/ at merge — archived, never deleted.
• .ai/backlog/README.md, .ai/archive/README.md and docs/decisions/README.md — copy ${CLAUDE_PLUGIN_ROOT}/templates/backlog-README.md, ${CLAUDE_PLUGIN_ROOT}/templates/archive-README.md and ${CLAUDE_PLUGIN_ROOT}/templates/decisions-README.md verbatim. docs/agents/README.md — copy ${CLAUDE_PLUGIN_ROOT}/templates/agents-docs-README.md verbatim; the template's router row points at it, so the two arrive together or path sync fails on the first run. Existing entries in any of the three dirs are left alone — they carry real content from earlier changes. A legacy single-file .ai/BACKLOG.md, if present, is named at the gate: offer to split it into per-file entries, never clobber or silently keep it.
• CONTEXT.md — if absent, create it with a one-line purpose statement (this project's glossary; terms only, no implementation detail) and a Term budget: **90 lines / 7 KB**. line, which caps the file and holds each term to one bullet of two lines. Nothing else. If it exists, leave it alone.
• VERIFY.md — if absent, create it the same way, and write exactly this much: a one-line purpose statement, the two invariants below, then the three headings with one line under each naming what belongs there. It holds what a green suite does not — Launch (the command that starts this project, on what port, with what disposable state), Doctor (one read-only check answering "is this instance worth driving?"), Drive (the commands, each paired with the observable result it should produce). The invariants go in before any of that is filled, both learned the hard way: cleanup removes instances and scratch state but never the evidence, and kill what you started rather than killing by process name. Filling the three sections is not this skill's job, and a project with nothing to launch leaves those lines as written — that is a finished VERIFY.md, not a stub owed to anyone. A library, a docs repo or a skills catalog has no instance to drive, and an honest empty section beats an invented command every session pays for. If it exists, leave it alone.
• ${CLAUDE_PLUGIN_ROOT}/templates/settings.json → .claude/settings.json; prune the hook entries not selected, add the permissions.allow list (below), then confirm the file still parses as valid JSON.
• AGENTS.md — if absent, render ${CLAUDE_PLUGIN_ROOT}/templates/AGENTS.md, substituting {{PROJECT_NAME}} {{DEFAULT_BRANCH}} {{STACK}} {{TEST_COMMAND}} {{LINT_COMMAND}} {{TYPECHECK_COMMAND}} {{COMMIT_PATTERN}} {{COMMIT_TRAILER}} {{BOOTSTRAP_COMMAND}} {{HOOKS_INSTALLED}} {{AGENTS_CHECK_COMMAND}}. Every placeholder gets a value or the line goes — a {{…}} left in the file is read as content by the next session and evaled as a command by the hooks, which is why they carry an explicit guard against exactly these tokens.
  • The commands come from step 1 verbatim; none where the manifests had none. Write none, not a plausible guess and not _(none detected)_: the hooks read none as "skip", and a command that doesn't exist fails on every edit.
  • {{COMMIT_PATTERN}} and {{COMMIT_TRAILER}} are policy, not detected commands, so ask rather than probe: derive the pattern from the log the repo already has (git log --format=%s -30) and propose it, and write none for the trailer unless the user says every commit must carry one. Both defaults live in enforce-commit-hygiene.sh — Conventional Commits for the pattern, none for the trailer — so an honest none is always a safe answer here.
  • {{LINT_COMMAND}} is the per-file form (pnpm exec eslint --fix, ruff check --fix), since lint-on-edit appends one path to it — not the whole-project script, which would re-lint the tree on every edit. Where the project only has the whole-project form, say so at the gate and write none rather than wiring a slow hook.
  • Prune the Task Router row for anything this run didn't create. The template ships a row pointing at .claude/hooks/, and that directory only exists if step 2 selected a hook — so a scaffold that declined them all fails its own agents:check on its first run, because path sync requires every routed path to exist. Same rule for any row whose target you skipped. A row is a promise that the file is there.
  • Idempotency is per-section: if AGENTS.md already exists, leave it alone and report which of the template's sections it is missing. Never merge a rendered template into a file someone wrote.
• scripts/check-agents-docs.mjs — copy ${CLAUDE_PLUGIN_ROOT}/templates/check-agents-docs.mjs verbatim. Zero dependencies, Node built-ins only, and it finds the repo root by walking up from its own location to the nearest AGENTS.md — so scripts/ is the conventional home, not a required one. It checks the always-loaded contract — the declared budget, that no {{…}} placeholder survived, Task Router coverage and path sync, that every pnpm <script> named in AGENTS.md exists, that CLAUDE.md is a symlink — plus the size of the layers it routes to: a Ceiling: per decision record, a Topic budget: per topic file, a Term budget: over CONTEXT.md, and a word ratchet over docs/agents/ as a whole. The three declared caps are opt-in, and each switches on the one shape rule a number cannot express for its layer — no dated gotcha bullet, no term over two lines; a missing declaration is neither checked nor mentioned, because a gate that lights an existing folder red on install day is one you switch off rather than meet. The ratchet is the opt-out one: it records what a corpus already is, so it is green the day it is seeded, and a repo switches it off by deleting the file.
  • An existing scripts/check-agents-docs.mjs is left alone, with its diff against the template shown at the gate. A repo that patched the checker patched it for a reason, and the rule the AGENTS.md bullet states holds here too: never overwrite a file someone edited.
  • Seed the baseline last. --update-baseline refuses to write while any other check fails, so node scripts/check-agents-docs.mjs --update-baseline runs once every path above exists. A fresh scaffold records zero words, and the first topic file dw-land promotes pays one visible re-record — the bargain, not a bug.
  • {{AGENTS_CHECK_COMMAND}} is how AGENTS.md's own header names its enforcement. With a package.json, add "agents:check": "node scripts/check-agents-docs.mjs" to scripts and render pnpm agents:check; where the repo has an aggregate gate script (check, verify), add it to that too. Without a package.json, render the bare node scripts/check-agents-docs.mjs — and say at the gate that the checker needs Node even where the project does not.
• CLAUDE.md — ln -s AGENTS.md CLAUDE.md. A symlink, never a copy: the harnesses load CLAUDE.md, the file is AGENTS.md, and a materialized second copy is the fork this whole layout exists to prevent. If a real CLAUDE.md with content is already there, it was approved at the gate as a rename — git mv CLAUDE.md AGENTS.md, then link — and its content stays; reconcile it against the template's sections afterwards rather than overwriting.
• The selected ${CLAUDE_PLUGIN_ROOT}/templates/hooks/*.sh → .claude/hooks/, chmod +x each.
• .worktreeinclude — if absent, copy ${CLAUDE_PLUGIN_ROOT}/templates/worktreeinclude.txt verbatim. If it exists, leave it alone. Every line ships commented out, so it copies nothing until the user names a file; that is deliberate, because an uncommented guess would copy a secret nobody asked for. Tracked, and it earns its keep twice: Claude Code reads it for claude -w worktrees, worktree.sh create reads it for the loop's own. Say at the gate that it starts empty and the user should add their .env line.
• .github/PULL_REQUEST_TEMPLATE.md — if absent, mkdir -p .github and copy ${CLAUDE_PLUGIN_ROOT}/templates/PULL_REQUEST_TEMPLATE.md verbatim. If it exists, leave it alone. It ships as guidance comments only, so it renders empty until someone writes a body.
• Append ${CLAUDE_PLUGIN_ROOT}/templates/gitignore-block.txt to .gitignore between its markers. Idempotent: if the markers are already there, replace the block in place, never duplicate it.

The permissions.allow list — derived, never invented. The template ships ask and deny only, so nothing is pre-approved and every check waits on a prompt unless a global setting happens to cover it. This is the difference between a lane that runs and one that idles, so build the list from what step 1 actually found:

• The project's own checks, exactly as detected. Match the wildcard style already used in the template's ask list — a bare entry plus an argument form, e.g. Bash(pnpm test) and Bash(pnpm test *). Never allowlist a script that isn't in the manifest: an entry for a command that doesn't exist is worse than no entry, because it reads as verified.
• The read-only git surface this lane uses — git status, git diff, git log, git rev-parse, git symbolic-ref, git branch --list, git worktree list, in the same two forms.

Write nothing that overlaps ask or deny. Don't reason about which list wins — just never add an entry that could match git commit, git push, anything in the template's ask list, or anything touching .env. Adding write or network commands here is not a speed optimisation; it removes the gate the rest of the lane is built around.

Seed .ai/backlog/ and .ai/archive/ with their READMEs and nothing else — no example entry, no TODO, nothing standing in for one. On the next read a placeholder is indistinguishable from real queued work, and a backlog you have to first decide isn't real is one you stop opening. The copied backlog-README.md states the entry shape; entries arrive later, from dw-land.

5. Optional — wire the pre-commit

Only when opted in at the gate. pnpm add -D husky lint-staged, pnpm exec husky init, then write .husky/pre-commit and .lintstagedrc.json from the shapes in references/precommit.md — globs matched to the formatter and linter step 1 actually detected, never to a tool that isn't installed. Typecheck and test lines are separate opt-ins: both run the whole project per commit. agents:check is not an opt-in — it goes in uncommented and unguarded, because it reads a handful of files rather than building the project, and the reference explains why a staged-paths guard on it leaks. On a repo that's partly wired, fill the gaps and show diffs — never overwrite blind; the re-run rules are in the reference. Worth having even solo: it catches the commits made outside a session, where no hook fires.

Where the repo has no package.json and so no husky, say what that leaves uncovered rather than inventing a gate: AGENTS.md can then drift past its budget until someone runs the checker by hand.

6. Reconcile tracking

The split is the whole point, so enforce it after writing: .ai/, docs/decisions/, CONTEXT.md, AGENTS.md, CLAUDE.md, .claude/settings.json and .claude/hooks/ (plus .husky/ when written) must not be ignored — remove any pre-existing rule that ignores them. CLAUDE.local.md and .claude/settings.local.json must be ignored: nothing writes the first any more, and the rule is what keeps a stray one from being committed.

git add the CLAUDE.md symlink and confirm git recorded it as one — git ls-files -s CLAUDE.md must show mode 120000. A repo with core.symlinks=false stores the link as a text file holding the target's name, which reads as a one-line CLAUDE.md and silently un-forks nothing; say so rather than leaving it.

7. Report

List what was written and which paths to git add. If the pre-commit was declined, say what that leaves uncovered: commits made outside a session run with no formatter and no guardrails. Name the _(…)_ placeholders left in AGENTS.md for the user to fill — the template ships the sections, only they know this project's invariants.

References

• references/precommit.md — detection signals, glob→command mapping, the .husky/pre-commit and .lintstagedrc.json shapes, and the idempotent re-run rules. Read it before step 5.

Next: dw-shape to open the first change, or dw-doctor to verify the scaffold actually fires.

$ARGUMENTS