sweep-issues

Sweeps the open-issue backlog.

• tier 1 — meta parents whose sub-issues are all closed. Closed on every run.
• tier 2 — issues whose referenced repository paths all no longer exist. Listed on every run; closed only under --deep, one at a time with the user's approval.

The user types at most --deep and --dry-run. Passes 2 and 3 are this skill's own calls (SI-5, SI-7) — the user never supplies a TSV path.

Procedure

Read rules/github-issues.md before SI-1 — on-demand-only, never auto-injected; it owns the close paths, the status:migrated / status:cancelled labels, and meta-parent handling.

SI-1. Run bash "$AGENTS_MAIN_ROOT/bin/sweep-issues.sh" forwarding the user's other flags verbatim (pass 1). SI-2. Print stdout verbatim. Do not summarize or filter. SI-3. Stop here when the output carries no <<<TIER2-GATE-SI3 block. SI-4. With the block: use AskUserQuestion to separate false positives (artifact not built yet) from real staleness. Write the survivors to ${WORKFLOW_PLANS_DIR:-$HOME/.workflow-plans}/<session-id>-sweep-issues-survivors.tsv. Columns: number / tokens_csv / class. Copy columns 1-2 from the gate template rows. class ∈ resolved | refactored-away | duplicate | obsolete | partial. SI-5. Re-run with --verify-candidates <survivors.tsv> --deep (pass 2) and print every EVIDENCE- line verbatim. SI-6. Present the <<<TIER2-GATE-SI5 candidates plus their evidence in one batched AskUserQuestion. Write only the approved rows to <session-id>-sweep-issues-decisions.tsv. Columns: number / action / arg / rationale. arg is the surviving issue for migrated, else -. SI-7. Emit <<WORKFLOW_ISSUE_CLOSE_VERIFIED: sweep-issues batch triage>>, re-run with --decisions <decisions.tsv> --deep (pass 3), then emit <<WORKFLOW_ISSUE_CLOSE_VERIFIED_END: sweep-issues batch triage done>>. Report every PARTIAL: line verbatim.

Rules

• Applies by default; --dry-run suppresses tier-1 closes too.
• --deep never changes the write mode, and never widens the candidate set — SI-2 scans identically with and without it. It decides only whether the tier-2 candidates reach the human gate or are dropped after being listed.
• --verify-candidates / --decisions are this skill's protocol with the script, not user flags. Never surface them in a suggestion to the user; never expect the user to pass one.
• Pass --deep on every tier-2 call (SI-5 and SI-7); omitting it exits 2.
• The flagless run is non-interactive. Never place AskUserQuestion on the default path.
• Pair --deep with a small band (e.g. --band-size 20) — SI-5 runs real tests and large bands do not finish in usable time.
• pass-1 scans every band by default; the loop runs in the script, not here. pass-2/3 have no band concept.
• --deep aggregates all-bands candidates into one gate block; total candidates scale linearly with open-issue count. To bound the gate size, use --max-bands N, not --band-size.
• Under --dry-run, SI-5 reports the test it would run instead of running it; running a repository script is an effect.
• A pass-1 run exits non-zero when any sub-step failed. Treat a non-zero exit as "this band was not swept", never as "nothing to do".
• --repo must name the working tree's own repository; pass --repo-root DIR when sweeping a repository checked out elsewhere, or the run aborts.
• Close only through the existing helpers. Never call gh issue close directly.
• The judgement axis table's SSOT is the header of bin/sweep-issues/close-batch.sh; it is not restated here.
• SI-5 evidence never rests on the issue body's own claims.
• Write both intermediate TSVs outside the repo, under ${WORKFLOW_PLANS_DIR} — in-repo writes are blocked by enforce-worktree.js.
• After a run that closed several tier-1 children of one parent, recommend /issue-reconcile: parent-body-update.sh can lose an update between concurrent closes.
• This skill sets no context: fork, unlike its sweep siblings, because its human gates need the parent conversation.