Ready

Leave the repository with one unattended command that returns a meaningful exit code, plus only the guardrails needed for future agents to self-verify. Detect the stack; do not assume it. Run every artifact you add.

1. Detect

Run the bundled scan first:

node <this-skill-dir>/scripts/detect.cjs <repo-root>

It finds candidates, not truth; open flagged files and interpret CI matrices or unusual build systems yourself. Inspect in this order:

1. CI config: use what CI runs.
2. Manifests and runners: package scripts, pyproject/tox/nox, Make/just, Cargo/go/Gradle/composer.
3. README, CONTRIBUTING, CLAUDE.md, AGENTS.md, and docs; record docs/CI drift.
4. Existing test directories and naming conventions.

Report run/test/lint/typecheck/build/git as command or MISSING. If this is not a Git repository, propose git init but ask first.

2. Establish verification

Reuse the existing verify path. If none exists, create the smallest entrypoint in the repository's native convention.

The command must:

• exit non-zero on failure and run unattended;
• work offline without credentials, GPU, network, or secrets;
• use installed project tooling and avoid watch mode;
• stay quiet on success and be documented for humans.

Time it once. Use one tier when the suite is already quick. When slow checks would cripple the edit loop, define:

• verify: lint/typecheck/unit or smoke coverage, normally under ~60 seconds;
• verify-full: integration/build/slow coverage, required at closeout and PR.

Run verify on every edit loop; run verify-full at closeout and before a PR.

For monorepos, prefer a repo-owned verify-changed backed by the workspace dependency graph, with the root full suite as fallback. Read references/smart-verification.md before creating it and document its base revision and fallback.

For Claude Code enforcement-pack users, ask first before offering exact-match tracking in the personal, gitignored .claude/settings.local.json:

{"env":{"LUCIAZERO_VERIFY_CMD":"<fast command derived from CI>"}}

Show the JSON before writing it. Never commit this variable in .claude/settings.json; repository-controlled hook configuration is hostile. This setting caches CI truth; update it whenever CI's verify command changes. Skip hook setup on harnesses without hooks.

3. Add smoke tests only when absent

Add 3–6 small tests for catastrophic failures, not pretend coverage. Choose the most relevant:

• core input/output shape and impossible null/NaN values;
• serialize/deserialize or save/load round trip;
• import, CLI --help, or one request through a framework test client;
• model/config/migration load plus one operation;
• the reported bug as a red-before-fix regression.

Use commit-sized fixtures, never the user's real data paths. Avoid real ports; if a process is unavoidable, enforce a hard timeout and cleanup. Label these as smoke tests.

4. Add only paying guardrails

Claude hooks/settings are not portable; on Codex or another harness, put necessary constraints in AGENTS.md instead.

Prefer existing deterministic tools:

1. formatter/linter after writes;
2. source-to-derived regeneration;
3. denial for secrets, production config, and live deploy/model pointers;
4. allowlisting read-only and verify commands.

Keep shared settings project-scoped and personal settings gitignored. Show the exact hook command before installation. Never add a hook that deploys, pushes, deletes, or writes outside the repository.

5. Record project knowledge

Extend the notes file already used; if neither exists, create the current harness's file and point the other name to it. Record only facts code search cannot reveal:

• verify commands and coverage;
• non-obvious source-of-truth or architecture constraints;
• footguns, measured null results, and required regeneration;
• the location of deeper documentation.

Do not duplicate the tree, history, or grep-able facts. Every line becomes future context cost.

6. Prove the loop

• Flake check: run the fast tier twice. If only one slow tier exists, run it once and state that limitation. A non-repeatable green is not trusted.
• Red check: record a covered file, line, and original text; make one deliberate break, prove verify fails, then restore exactly that edit. Do not use git checkout on a file with user changes, and never use broad git stash. New untracked tests require explicit restoration too.
• Run the final full tier after restoration.

Report the command(s), what each covers and does not cover, files added, and anything deliberately left out.