Clean sweep

Skill ravidsrk/orca-fleet/skills/clean-sweep

Exhaust a finite backlog of findings to zero, PR-per-finding, on Orca. Sources: an audit/adversarial-review document, the issue tracker, or verified false documentation claims. Each finding is skeptic-triaged (reproduced or refuted with evidence), fixed on an integration branch with a failing-first test, build-blind reviewed, merge-verified, and closed with a linked SHA; the loop re-enumerates until dry. Use when "clean sweep", "close every issue", "drain the backlog", "fix everything in this audit", "the README lies — verify and fix it", or a broken test that fails deterministically (N/N, cause findable — deflake-it routes those here). Not for security/perf/deps/coverage-gaps/flakes (those are harden-it / speed-it / modernize-it / prove-it / deflake-it — different convergence proofs) and not for building new work (ship-it).From its SKILL.md

Install

npx -y skills add ravidsrk/orca-fleet --skill clean-sweep

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its file declares

Copied from the file, not written here

The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.

SKILL.md

7.1 KB, as published. Nobody here has run it

clean-sweep — exhaust a finite backlog to zero, with evidence

You are the COORDINATOR of a run that closes every real item in a bounded set and leaves the repo demonstrably working. Thin loop-holder: you enumerate, dispatch the per-finding pipeline, verify against authoritative state, and keep the ledger FILE (your memory is compacted; the ledger survives). You never review, code, open PRs, or merge — every one is a dispatched worker.

Read ARCHITECTURE.md once. Composes remediate-finding, acceptance-review, build-change, compound-learn; rides merge-serialization, reviewed-sha-freshness, dispatch-lifecycle, liveness-resume, evidence-manifest, orca-dag-semantics, ledger-contract, attention-budget. Review is remediate-finding's build-blind step (acceptance-review); per-finding negative control is build-change — not a full runtime-prove pass (reserved for non-trivial feature-class findings handed to ship-it).

Two terminal outcomes

DRY — full re-enumeration finds zero items not CLOSED with evidence (no open parks).
DRY-WITH-PARKED (degraded) — set exhausted except ≥1 PARKED (needs-human, CODE_CLOSED + VERIFY_AT_SCALE, etc.). Never reported as DRY.

The source (declare it — same unit, same pipeline, source-specific enumeration)

source=audit (default): findings from a scan / adversarial-review doc. FREEZE the findings list.
source=tracker: OPEN ISSUES. Record run-start T0 FIRST; the denominator is two queries — every open issue in scope (paginated to the end; a truncated listing silently fails the run) AND every issue created/reopened since T0 any state (class externally-resolved). Re-run BOTH each loop.
source=doc-claims: falsifiable documentation claims — a false claim IS a finding (extract → verify against file:symbol or a run → correct/remove). Discover claims in docs/, README, code comments, or a user-supplied list. Generating NEW docs is not this mission (that's ship-it scoped work).

Any source: a PRIOR run's completion report over the same surface is a claims list to RE-VERIFY (liveness-resume.md inflation post-mortem), never a pre-closed set — its green-but-unverified claims enter the enumeration FIRST.

Pipeline

SELF-ORIENT → ENUMERATE (per source) → SKEPTIC-TRIAGE (reproduce-or-refute) → FREEZE
  → BOOTSTRAP integration BASE (preflight --base <BASE> --fork-point <header sha>; BASE ≠ default)
  → PER-FINDING (remediate-finding: verify-real → build-change → PR → build-blind review → merge_ready)
  → conductor LAND (merge-serialization) → CLOSE with evidence
  → re-ENUMERATE (loop until dry) → FINAL REPORT + `compound-learn` + human gates

Run the coordinator as a MANUAL loop (task-create → spawn → dispatch --inject → check --wait), not orchestration run — you want the file-ledger boolean gate under your control. Dispatch waves respect attention-budget WIP.

Convergence proof (definition of done)

A full enumeration finds ZERO items that are not (a) CLOSED with evidence (a merged, ancestry-verified PR + a test that failed pre-fix, revert-audited on a ≥10% sample; the closing comment links PR + test; ledger flags BUILD_DONE…WT_CLEAN all t) or (b) PARKED with an allowed class (refuted/duplicate closes are a one-way batch gate; needs-human / CODE_CLOSED+VERIFY_AT_SCALE name their gate/OPS ref). The final enumeration output is pasted in the ledger showing the dry state. source=tracker reconciles created/closed-mid-run issues against T0, so the count is honest. Manifest names DRY or DRY-WITH-PARKED. DRY also requires the integration TIP to pass the repo's OWN validate/test suite: the FINAL REPORT and any run-close commit land on the branch outside the per-finding gates, so a green per-finding history under a red report commit is still a red branch — verify the tip green before opening the promotion PR.

Ledger (header first, then rows)

Line 1 per liveness-resume.md: RUN · COORDINATOR · BASE · FORK_POINT · T0 · SOURCE · WIP (SOURCE = source type + enumeration digest). Phase marker + unit boolean flags per ledger-contract.md:

Gates + supervision

Batch human gate for refuted/duplicate closes (or a once-per-run recorded grant). Fix-backed closes need no extra gate — the evidence chain is authorization. BASE→default promotion is out of scope: open the promotion PR, stop. Stalls → liveness-resume WATCH; compaction → CONTEXT HANDOFF then RESUME; death → RESUME (ledger-scoped, git-verified).

Anti-patterns

Fixing without the triage repro (you'll "fix" symptoms and close real bugs unfixed). Closing from worker memory (only off verified merges). One mega-PR for many findings (merge-serialization exists). Truncated enumeration (partial denominator = a false "done"). Dropping a finding at DISPATCH: batching the frozen list into file-coherent build units can silently omit an id the denominator still carries — the wave plan is a new artifact, so assert every frozen id maps to exactly one build unit before dispatch; a finding that never reached dispatched is OPEN, not done (re-enumeration is the backstop, not the primary guard). Owning security/perf/deps/coverage-gaps/flakes — those are separate missions with different convergence proofs.

harden-it / speed-it / modernize-it / prove-it / deflake-it (specialist campaigns, distinct convergence proofs), ship-it (build new), review-it (verdict only), mission-scheduling (a bounded source stops at the promotion PR, so it schedules as a nightly unattended sweep).