Adlc init

ADLC — bootstrap a repository

Get a repository from nothing to a working Agentic Development Lifecycle: the toolkit on PATH, the .adlc/ runtime present, and the one control that is actually unbypassable wired into CI.

1. Check what is already there

There are four controls, installed by different steps. Probe each one independently and report which are present:

adlc --version                                  # 1. gate toolkit on PATH?
ls .adlc                                        # 2. repo runtime bootstrapped?
ls .github/workflows/adlc-rails-guard.yml       # 4. CI backstop wired?

Control 3 is the harness integration — check whether your own agent has the ADLC plugin loaded (skills alone are not the plugin).

Do not treat "toolkit + .adlc/ exist" as "already set up" and stop. That is the common partial state — a CLI-only install, a skills.sh adopter, or a repo initialized before CI was wired — and stopping there reports success while the native plugin and the only unbypassable control are both missing. Skip only what is genuinely present, do the rest, and call it complete when all four check out.

2. Install the toolkit

The toolkit is Node ≥ 18 and zero-dependency.

npm install -g @adlc/cli

Do not proceed if Node is older than 18. Report the version and stop; do not install or upgrade Node on the user's behalf without asking.

One command can install the toolkit and every harness integration present on the machine:

curl -fsSL https://www.agenticlifecycle.ai/install.sh | sh

macOS and Linux only. Native Windows is not supported — the toolkit passes 6 of 28 core gate suites there. Use WSL. The installer refuses to run under Git Bash / MSYS / Cygwin rather than installing a broken toolkit.

Confirm with the user before piping a remote script to a shell. If they prefer to read it first, fetch it and show it to them.

3. Create the runtime

adlc init --root /absolute/path/to/repo

This idempotently creates the committable .adlc/ runtime. Flags worth knowing:

• --harness <codex|cursor|copilot> — records harnesses.<name> in a fresh config. cursor and copilot each imply --no-codex-agents; copilot also scaffolds a .github/copilot-instructions.md block and a copilot-setup-steps.yml snippet.
• --no-codex-agents — skip the project-scoped Codex agent templates.
• --json — machine-readable output.

Writes are confined to --root. adlc init never edits user-global harness configuration implicitly.

Careful: if the repo already freezes .adlc/config.json as a rail, do not commit a newly generated one over it.

4. Wire the CI control

In-session rail enforcement is best-effort and harness-dependent. The commit-time diff gate is the real control — wire it and make it a required check.

docs/ci/rails-guard.yml lives in the ADLC source repository. It is not shipped inside @adlc/cli (that package publishes only bin/, lib/, README.md, and LICENSE) and adlc init does not create it, so a normal downstream repo has no local copy to cp. Fetch it:

Fetch it to this exact filename — the workflow protects itself by name, and any other name leaves it editable by the PRs it is meant to gate:

mkdir -p .github/workflows
test -e .github/workflows/adlc-rails-guard.yml \
  && echo "REFUSING: a workflow already exists — diff before replacing it" \
  || curl -fsSL https://raw.githubusercontent.com/voodootikigod/adlc/main/docs/ci/rails-guard.yml \
       -o .github/workflows/adlc-rails-guard.yml

The existence check is not ceremony. curl -o truncates without asking, and this destination is a protected trust root — a repo with a customized or newer workflow would have it silently replaced by whatever main serves today. If it already exists, diff the two and let the human decide.

This file is fetched from a mutable branch and will become a required CI check with repo-wide authority. Review the whole thing before committing it — not just the header. It is workflow code that will run on every pull request; main is not a fixed revision, and nothing here verifies its integrity. If the human wants a stronger guarantee, have them pin the URL to a tag or commit SHA rather than main.

Then follow its header before marking the check required. It documents a multi-stage bootstrap: the bootstrap commit merges first, the base branch needs trustedCodeownersAttested in .adlc/config.json (which adlc init deliberately does not set), and .adlc/manifest.jsonl must be absent or empty until reviewed. Marking it required early fails every subsequent PR.

Only once that ceremony is complete does marking it required in branch protection make sense. Without it, frozen rails are advisory; with it enabled too early, every PR fails.

Optionally add the maintenance cron from docs/ci/adlc-maintenance.yml.

5. Verify

adlc preflight            # baseline environment readiness
adlc ticket list          # store reachable, exit 0
adlc gate-manifest show   # evidence surface reachable

Report the three results. If adlc preflight gate-fails (exit 2), surface what it flagged rather than working around it.

6. Hand off

Tell the user which phase to start at — for most repos that is P0: author the first ticket with adlc ticket create --input <path|-> --write. The adlc phase-routing skill covers the rest of the lifecycle.

If their agent has a native ADLC plugin (Claude Code, Codex, Cursor, OpenCode, Pi, Google Antigravity, GitHub Copilot), recommend installing it — it adds in-session rail hooks, MCP tools, and prosecutor subagents that the CLI alone does not provide. See https://www.agenticlifecycle.ai/integrations.