🐉 Priori Incantato

"Every artifact swears it works. Ask it when it last ran."

An artifact nothing invokes is not a feature, it is invisible debt — and it costs more than absence, because the roster claims coverage the repo does not have. The failure is quiet by construction: a hook wired to the wrong event returns success and its payload is dropped, a skill whose disambiguation lives somewhere the router never reads is simply never chosen, and neither state produces an error anywhere. This pass asks one question of every artifact: what causes you to run?

What to scan

• Hook scripts and the config that wires them (.claude/settings*.json and equivalents) — the pair, never one side alone
• Skills, agents, commands, output styles, and their routing-visible surfaces
• Repo-local test suites and lints, and whatever is supposed to invoke them
• The version-control index — the ground truth for what a fresh clone actually receives

Anti-patterns to find

1. Declared-vs-wired event mismatch. A hook emits one event name in its payload while the config fires it on a different event. The handler runs, the payload is discarded silently, and both sides look healthy in isolation. Only comparing the pair reveals it.
2. Orphans. A skill, agent, command, hook, lint or test that nothing invokes — no config entry, no caller, no documented route. Count them honestly and separate the ones this change created from the ones that predate it.
3. Availability mistaken for discoverability. The artifact exists, is wired, and still never runs, because nothing causes it to be reached: a router that reads only a description field while the disambiguation lives in the body; an advisory that fires after the decision it was meant to inform; a skill whose triggers no real user phrasing matches.
4. Fresh-clone unreachability. The artifact works only because of local state — an unversioned file, a locally-set executable bit, an absolute path, a machine-specific tool. Verify as a fresh clone sees it: query the VCS index, never a filesystem existence test on the machine that authored it.
5. Wired to a dead event. The trigger existed once. The event was renamed, the tool it matched was removed, or the workflow that called it was retired — and the artifact still sits there declaring a route to nowhere.
6. The finish-the-sentence test fails. The author cannot complete "this will next run when ___" with an observable event. If the sentence cannot be finished, the artifact is not wired; it is merely present.

Fix prescriptions

Bounds

• Scope: automation artifacts and the config that wires them, plus the version-control index. Not application source, not another installed plugin's artifacts.
• Cap: the census covers everything (it is a table); cap detailed findings at 15, dropped-payload and dead-event cases first.
• Done when every artifact has been paired with its wiring once. Orphans that predate this change are repo backlog, reported as a count.
• Quote once, at read time; mark anything you cannot re-confirm UNVERIFIED rather than re-reading to check yourself.
• See skills/opus-pocus/references/casting-safety.md for the full rules.

Report format

Per artifact: path · declared trigger · actual wiring · verdict — reached / dropped (runs, but its output is discarded) / orphan (nothing invokes it) / unknown (could not determine, which is not a pass). Then the census against its denominator: N artifacts examined, M reached, and how many orphans predate this change versus were created by it. SCANNED NOTHING if no automation layer was found — never "clean".