sweep

/sweep — drain the queue to a ship-ready bar, one commit per task

Run the work loop to completion. Claude scopes the queue Stop hook above to this skill's lifetime; runtimes with a persistent goal/tracker use that too. Scope: this repo's .agent/tasks/ (a task is a folder; its state is its directory — coop tasks lists and moves them).

Be agentic. Each task is taken to a ship-ready bar: built, gated, then self-reviewed from every angle it touches and iterated until clean — and only then committed, on its own. The bar is not "it compiles"; it's "I'd defend this in review from every hat and it breaks no house rule." The house rules are the spine: this repo's AGENTS.md and the worked examples in .agent/rules/. Build to them, then check the diff against them.

1. Prepare

• Set the runtime's persistent goal/tracker when it exists: do not stop until every scoped task is committed and archived or properly blocked, and both 00_todo/ and 10_in_progress/ are empty.
• Read AGENTS.md in full (the gate and the contract), .agent/tasks/README.md, and every .agent/rules/*.md (the taste KB). If coop is on PATH, run coop tasks; otherwise inspect the queue state directories directly to announce the open 00_todo/ count.

2. The loop — claim the next task, repeat until 00_todo/ and 10_in_progress/ are empty

1. Claim — if coop is on PATH, run coop tasks claim <id>; otherwise move the task folder from 00_todo/ to the existing 10_in_progress/ directory yourself. Do this first, so a parallel agent won't grab it. A task already in 10_in_progress/ is a prior attempt — resume it: read its state.md, then git status/git diff.
2. Build — wear the hats; obey AGENTS.md, match .agent/rules/ and the surrounding style exactly. /spec first if it spans more than one file (writes the task's spec.md). /verify-api before calling anything you're not certain exists.
3. Gate green — the repo's exact gate (AGENTS.md → "The gate"). No green, no review, no commit.
4. Self-review the diff from every angle it touches — correctness, security / abuse path, UX, tests (including the failure path), docs, readability — against the house rules. Fix what you find; iterate until you'd defend it.
5. Commit — one focused commit for this task; keep log.md (the what + why) current.
6. Final snapshot — after the commit, refresh state.md in 10_in_progress/: preserve the useful summary and traps, set Status to complete, and set Next action to none.
7. Done — as the final filesystem action, if coop is on PATH run coop tasks done <id>; otherwise move the task folder to the existing 99_done/ directory yourself. A finished task is moved, never deleted — leave it in 99_done/; never run coop tasks rm (pruning the archive is the human's call). Blocked instead? Fill in decision.md (question · options · recommendation), then use coop tasks block <id> when available or move the folder to the existing 50_blocked/ directory yourself.
8. Spot unrelated work? Use coop backlog add when available, or create a self-contained folder under .agent/tasks/xx_backlog/, then return to the queue — don't derail the current task.

3. Finish

• When 00_todo/ and 10_in_progress/ are empty, run a completeness pass: re-check every task you moved to 99_done/ against git log — gate green and a commit exists. Reopen anything that doesn't hold up with coop tasks claim <id> when available, or move it back to 10_in_progress/, and go again.
• Report: tasks shipped, anything parked in the backlog or 50_blocked/, gate status.