Sweep

/sweep — The Full-System Audit Fan-Out

The pattern behind every high-yield mega-audit (21-agent data audits, FAB-style architecture maps, exit dossiers): decompose into dimensions → independent parallel auditors → adversarial verification → graded synthesis → every finding becomes a mechanism. One agent can't hold a system; a swept fan-out can. This skill is the repeatable harness.

Siblings (if your workflow has them): a per-object forensic review skill goes deep on ONE planning object against its claims; a cleanup skill is fix-first reconciliation. /sweep is breadth-first discovery across a whole system — it FEEDS both.

Invocation

/sweep data stack --grade            # full-stack audit, scored vs best-in-class
/sweep ui/dashboard                  # product-surface sweep
/sweep docs/specs vs src --drift     # spec-vs-reality sweep (mass /drift)
/sweep security                      # one dimension, full depth

Phase 0 — Scope & dimension slate (inline, before any fan-out)

1. Fix the boundary: what's IN the sweep (paths, schemas, surfaces) and what's explicitly OUT. An unbounded sweep never converges.
2. Pick dimensions — default slate, prune/extend per scope:
   • correctness (does it do what it claims) · spec-drift (mass /drift)
   • security/permissions · performance/scale · data quality
   • test coverage (real tests, not shells) · architecture/placement (/altitude violations, duplicated owners) · honesty (claimed-done vs actually-done — the completion-audit dimension) · UX/product (if user-facing)
3. Declare the sweep's own gates (/gate): effort ceiling, and a finding budget ("if >N criticals in dimension X, stop sweeping and start fixing").

Phase 1 — Fan-out (parallel, one auditor per dimension)

Dispatch independent subagents/workflow stages, one per dimension. Each auditor's prompt MUST carry a handoff contract:

• the dimension's question, the scope boundary, where to look first
• evidence format: every finding = one line of file:line / query / output proof — no vibes, no "seems like"
• severity grammar: Critical (wrong results / data loss / security) · High (will bite soon) · Medium (debt) · Low (polish)
• return findings as structured list, not prose

Auditors are blind to each other — convergent findings from independent auditors are the strongest signal a sweep produces. Note them explicitly.

Phase 2 — Adversarial verification (the step lazy sweeps skip)

Raw findings lie. Before synthesis:

1. Dedup across dimensions (same root cause surfaces in many coats — use /altitude to name the shared layer).
2. Every Critical/High gets a /refute pass by a verifier that did NOT find it: reproduce it or kill it. Verdicts: CONFIRMED / REFUTED / PLAUSIBLE.
3. Only CONFIRMED findings may use the word "broken" in the report. PLAUSIBLE ships in an appendix, clearly labeled.

Phase 3 — Graded synthesis

One report, led by the number that matters:

SWEEP: <scope> | <n> dimensions | X confirmed (C/H/M/L: a/b/c/d) | grade: B-

• If --grade: score each dimension 1–5 against the best-in-class benchmark for its domain, WITH evidence per score. The grade's job is honesty, not motivation — a B- that's real beats an A that isn't.
• Convergent findings and systemic patterns first (three dimensions hitting the same subsystem = an /altitude problem, not three bugs).
• The "what's NOT broken" section is mandatory — a sweep that only lists problems can't be used to decide what's safe to build on.

Phase 4 — Findings become mechanisms (or the sweep was theater)

Every confirmed finding lands as exactly one of:

• patch (fixed in the sweep's follow-up PRs, smallest first)
• ticket (filed with the evidence line attached, severity mapped)
• gate (a tripwire/CI check so the class can't recur — preferred for anything that recurred)
• accepted risk (/verdict log, named human on the acceptance)

Close by logging the sweep itself: /verdict log SWEEP: <scope> | grade | top systemic finding | date. The next sweep of the same scope starts by diffing against this line — grades that don't move are the real report.

Rules

• Breadth-first, fix-later. Auditors that stop to fix lose coverage; the single exception is a Critical actively corrupting data — stop the sweep, raise it immediately.
• Cap the fan-out to what the finding budget can absorb. 21 agents producing 400 findings nobody triages is worse than 6 producing 40 that all land.
• Sweeps are periodic, not heroic: the value compounds when grade N+1 is compared against grade N.

Composes with

• /drift (the spec-drift dimension), /refute (Phase 2), /altitude (dedup + systemic naming), /gate (sweep budgets + recurrence tripwires), /verdict (the sweep ledger line). If your workflow has a per-object deep-review or planning-review step, a sweep is its strongest input.