kenoxa/run-clawpatch

Drive an async Clawpatch review → classify → fix → revalidate campaign over the clawpatch CLI. Stateful and explicit — never automatic for every task; activates only on clawpatch-specific intent. Mainthread orchestrates the CLI loop directly (clawpatch holds .clawpatch/ claim/fix locks — do NOT fan findings out to parallel agents).

This skill assumes a working clawpatch install (0.4.0+) with .clawpatch/ already initialized. It makes no provider/model/feature assumptions — those live in the project's own clawpatch.config.json.

Usage — direct preflight vs /goal campaign

• Direct /run-clawpatch status (or --local-only for a bounded local run) — preflight only: clawpatch doctor + status + dry-run preview. Read-only, safe without a worktree, pauses normally for the user.
• Recommended full campaign — /goal /run-clawpatch --since <ref>. /goal installs the session Stop hook so the agent works the whole review/fix/revalidate loop autonomously without interrupting the user. Under /goal, every user-STOP gate below becomes an emit-artifact-and-halt signal (write the final report, stop — no autonomous re-launch).

Phase table

Backticked refs here are mainthread-read, not dispatched. This skill spawns no subagents — Read each phase reference with the Read tool on phase entry (lazy; do not pre-load all five). That inverts the usual backticked-ref convention (subagent-only, never mainthread-Read; see run-debug, run-curate) because there is no subagent to consume them.

| # | Phase | Action | Reference (Read on entry) | |---|-------|--------|---------------------------| | 1 | Preflight | clawpatch doctor + clawpatch status; record dirty flag + open-finding count; confirm initialized | — (inline) | | 2 | Scope | resolve scope args → a --since <ref> (or unscoped); collect changed paths | references/scope-resolution.md | | 3 | Worktree + session | isolated worktree + /use-session attach before any .clawpatch state change (skip only if --local-only) | references/worktree-session.md | | 4 | Ownership | map changed paths → manual *_manual.json features; add missing owned files or create a new feature; halt on ambiguous ownership | references/feature-ownership.md | | 5–6 | Review + finding loop | clawpatch review (scoped per feature, or bounded); then per finding: next → show → classify → fix+revalidate / triage / leave-open | references/finding-loop.md | | 7 | Final report | final revalidate --all --status open (or scoped); write the report artifact | references/final-report.md |

Direct status invocation runs Phase 1 only and stops.

Argument contract

Scope selectors (pick one; precedence --since > --days > --last-run > --all-open; default --last-run):

| Arg | Meaning | |-----|---------| | --since <ref> | surface changed since a git ref (main, a sha) | | --days <n> | surface changed in the last n days | | --last-run | since the last Clawpatch run (newest .clawpatch/reports/* mtime → ref) | | --all-open | all open findings, unscoped (triage/revalidate sweep) |

Knobs: --jobs <n> (review parallelism, default 3) · --include-dirty (explicit opt-in to uncommitted changes; else a dirty tree is a stop) · --local-only/--no-commit (preflight + local edits, no worktree/commit) · --rate-limit <n> (→ --rate-limit-per-minute / CLAWPATCH_RPM) · --feature <id> (scope to one feature).

Halt signals (stop, emit final report, do not re-launch)

• Still on main after worktree setup (campaign mode, not --local-only).
• Ambiguous feature ownership a generic repo can't resolve.
• Dirty tree without --include-dirty.
• Scope needs project-specific metadata absent in this repo.
• 5 consecutive revalidation failures.
• clawpatch missing / not initialized / provider unconfigured (preflight fails).

Anti-Patterns

• Calling any /run-* skill (no-run-cycle invariant) — compose via /goal, never an internal skill call.
• clawpatch map, init --force, or re-seeding features unsupervised — flips manual features and floods scratch-derived ones.
• Auto-pruning or status-flipping any source: manual-* feature — manual features are the source of truth; halt on ambiguity instead.
• Treating review --since as coverage proof — it is selection only; reconcile ownership explicitly.
• Marking a fix done without a green revalidate run.
• Committing on main; pushing, opening PRs, landing, merging, or deleting/cleaning the branch without an explicit request.
• Running the campaign automatically for every task — it is explicit and stateful.
• Assuming a provider or model (e.g. a specific codex/gpt-5.5 pin is project-specific, not generic).

Completion

Done when the last review yields 0 new open findings OR every remaining open item has a recorded stop reason; every changed path is owned or documented uncertain; every fix has green-run revalidation (E3: run id + exit) and every triage has evidence; the worktree branch holds commit checkpoints (feature-map changes and verified fix groups), left unpushed/unmerged for human review unless the user requested otherwise. The final report (references/final-report.md) is the DONE/STOP evidence artifact.