Skill Evolution Status

Run one read-only readiness census across reports/skill-evidence/ in the current repository. This skill is a reporter, not an authorizer: the destination session's fresh Skill Evolution preflight remains final.

Arguments: none.

Hard boundaries

• Read-only repository state: never write events.jsonl or gate-status.json, run a preflight or derive command that refreshes projections, claim a review, or modify a target. Cargo may write only its ordinary build and cache artifacts while starting the compiled reporter; if the host cannot provide a writable build cache, report that the reporter did not run.
• Census only: never diagnose incidents, inspect target prose semantically, propose repairs, invoke Skill Evolution, or record evidence.
• Current repository only: never search sibling repositories or a global evidence directory.
• Canonical mechanics only: use the compiled command's shared Rust event validation, target hashing, and in-memory gate derivation; never duplicate eligibility thresholds or trust a stored projection as current.
• Safe uncertainty: never print an evolution command for queued pre-close evidence, evidence retired as untestable, self-targeting, an active timer, an owned review, an unreadable stream, or a missing target.

Workflow

1. Run the census

From the repository root, run exactly:

cargo run --locked --quiet -p skill-evidence -- skills evolution-status \
  --root . --now-epoch-milliseconds "<clock-ms>"

The authoritative census payload is the compiled reporter's stdout, beginning with # Skill Evolution Status. Cargo diagnostics on stderr are not part of the census. On a safe failure, the compiled command's lifecycle diagnostic on stderr is the authoritative failure; a Cargo failure before the reporter starts is an execution failure, not a census result.

Capture one exact census clock in epoch milliseconds and pass it as <clock-ms>; replay with the same value. The compiled command scans every evidence store, validates events.jsonl, hashes each live target, and derives current gate state without persisting it. It distinguishes the two freshness proofs explicitly:

• a captured threshold session ID is ready only in another session that exposes a different ID; a no-ID destination remains blocked and waiting does not help;
• a missing threshold session ID uses the 12-hour clock, which another session cannot bypass.

It also relays the canonical post-review provenance derived by Skill Evidence Capture. Evidence that could only have fired before a completed same-hash review is reported as deferred and receives no command; a genuinely new post-review incident can reopen the gate, with the bounded trigger cluster and provenance shown on the actionable entry.

Portability contract: run the repository's compiled skill-evidence command in a checkout containing the sibling Skill Evolution skill at its standard .claude/skills/ path and the standard reports/skill-evidence/ store. The command fails safely when that contract is absent or incompatible.

Done when the compiled command prints one reconciled census or a safe failure stating that nothing changed.

2. Relay the result and stop

Relay the authoritative payload intact. Preserve every reporter line and every emitted command without omission, duplication, reordering, or paraphrase. Host-required framing around that intact payload and normalization of its terminal newline are permitted; they are presentation, not census content. Add no substantive commentary of your own. Commands appear only under Ready to evolve and are already shaped for copy-paste into another top-level session. Deferred after review identifies the canonical queued subset of collecting stores and explicitly supplies no command, distinguishing evidence queued behind an instrument-limited close that reached no conclusion and could test nothing from evidence behind a concluded close where no threshold-supporting incident was recorded afterward; neither basis claims that the close covered the deferred evidence. Retired as untestable names stores whose open incidents left the gate because a review could not decide them, whether by an instrument-limited close or by an adjudicating close naming coverage it could not decide: real evidence that no longer drives eligibility and never will, listed separately so it is not read as an absence of evidence. Eligible but blocked gives the active blocker and all available timing, ownership, quarantine, or routing detail. Could not determine names stores that might conceal eligibility but cannot be trusted. Other closed, collecting, and current-hash-reset stores appear only in the omitted count.

Two kinds of close contribute to the standing retired set, at different scopes. Each instrument-limited close contributes only the open incidents its recorded authorization reason named at the close: friction_recurrence contributes its whole symptom cluster; material_recurrence leaves friction siblings outside the retired set; ten_use_unresolved excludes retrospective incidents from its anchor cluster; and severe contributes no retired incidents. An adjudicating close contributes the coverage it named as untestable, by name and never wider, less anything that still drives the gate on its own — a review that examined its coverage mechanism by mechanism has already said which it could not decide, and a contemporaneous severe incident it named stays in the gate regardless. For historical streams, a missing or unrecognized authorizing rule uses the prior symptom-wide reach. Report the compiled result as written; this explanation changes neither its identities nor readiness.

Do not reinterpret readiness, add commands to blocked entries, suggest early review, or continue into evolution in this session.

Done when the complete, unchanged census payload or authoritative safe failure has been delivered and no further action was taken.