atl rules

The deterministic collect half of rules-distill: gather the normative / imperative statements across the skill + agent corpus so the /rules-distill skill can judge which recurring principles deserve to become a core rule. The judgment — which candidate is a real principle — is the skill's; this is the CLI/Skill boundary.

This is a maintainer-side surface that runs against the monorepo's core/ + teams/. Outside the monorepo it does nothing and exits 0.

When the repo is not an ancestor of your working directory, set ATL_REPO_ROOT to name it. That is the maintainer-hub case — a checkout that clones the monorepo as a child — where an upward walk can never reach the marker, so the command skips and exits 0 while looking like it passed. The root is named rather than searched for on purpose: in that layout the same marker also matches archived clones, and a search would resolve one of them by readdir order and report drift that is real about content nobody ships.

Usage

atl rules scan            # print normative statements across the skill corpus
atl rules scan --json     # the same, machine-readable (file, line, text)
atl rules scan --record   # stamp HEAD as the last distill (after a /rules-distill sweep)

atl rules scan

Walks the skill + agent markdown in core/ and teams/ and prints every line carrying a strong normative/imperative trigger (always, never, must, don't, avoid, the grep-before-edit idiom), each with its file:line:

core/skills/drain/SKILL.md:49  Be strict — mine only what's worth never-repeating.
teams/example-team/agents/backend-agent/agent.md:88  Never expose the domain entity directly …

It over-collects on purpose — the collect step only gathers grounded candidates; the LLM decides which are real recurring principles. rules/ subtrees are skipped, because rules are the distill target, not a source.

--record stamps HEAD as the last-distilled commit (in ~/.atl/rules-distill-state.json), which resets the session-start "a distill is due" runaway-guard for ~1 day. /rules-distill calls it at the end of a sweep.

Related

• /rules-distill — the LLM half: cluster the candidates, grep existing rules, propose the recurring ones as core rules (human-confirmed).
• /rule — authors one rule you already have in mind; rules-distill discovers which rules the corpus is asking for.
• atl skills — the sibling deterministic gate (asset content-quality; this one feeds rule discovery).