Architecture Audit — Rotational System Audit

Selects the highest-priority system from the health ledger, dispatches domain reviewers against it, and packages findings as spinoff candidates down a disposition ladder — the audit itself never edits code. Updates the health ledger's Last targeted audit clock and the atlas metadata.

Run when /workstream-start surfaces a stale targeted audit (check-arch-audit-staleness owns the verdict), or when PM intuition wants a targeted review. $ARGUMENTS names a system directly, skipping Step 1.

Announce: "I'm using /architecture-audit to audit [system name]."

Step 1: Select Rotation Target

No state/health-ledger.md and no docs/architecture/systems-index.md → stop: run /architecture-survey first. $ARGUMENTS provided → skip to Step 2 with that system. Otherwise select the highest-priority system in the ledger — selection signals and worked calibration: wiki (no engine scorer exists yet). Report: "Rotation target: [system]. Rationale: [why]."

Step 2: Review Existing Debt

Read state/debt-backlog/*.yaml (or bin/query-records --type debt --status open) for the target system; present open items to the PM before/alongside the audit — full pipeline (plan → review → execute), not inline here.

Step 2.5: Load Atlas Context

docs/architecture/systems/{target-system}.md exists → include it in the reviewer dispatch as background context. Absent → reviewer discovers from scratch.

Step 2.75: Emit Ground-Truth File-Enumeration Artifact

cartography file_index --target-root <system-dirs> → write the index field to state/audits/{run-id}/ground-truth-files.txt, commit it. Pass the path in every analysis-agent prompt; the Opus reviewer diffs any "fabricated"/"non-existent" claim against it before issuing a verdict.

Out of scope: cross-repo targets — a host↔addon boundary or any target spanning two repos. Run the audit once per repo and link the two audit records to each other, naming the boundary in each.

Step 3: Dispatch System Review

Check the target's live file count (not the atlas count — systems grow); small systems get a direct Opus domain-reviewer dispatch, larger ones need Haiku→Sonnet pre-digestion first or the reviewer overloads. Threshold, dispatch procedure, prompt sources, and multi-reviewer criteria: wiki.

Do not ask whether to dispatch — invoking this skill IS the request for the dispatch this step names; it dissolves no gate this skill's own body names.

Step 3.6: Scaffold the Canonical Audit Record

coordinator-doc-new --type audit-record --system <target> writes docs/architecture/audit-records/<YYYY-MM-DD>-<target>.md — the scaffolded skeleton is authoritative; the reviewer fills it in-place via mode: auto. Reviewer prompt must cover: populate frontmatter (grade, health_status, run_id, reviewer), write the ### Diagram (ASCII) section, fill every prose section without flattening grade-rationale/ambition-check content.

Negative-spec (hard): the three historical audit records (2026-05-28-coordinator-runtime.md et al.) are frozen; never retrofit them to the canonical schema.

Step 4: Package Findings as Spinoff Candidates (never edits code)

Disposition is EM judgment down a ladder:

Trivial, tradeoff-free, non-structural → EM dispatches an executor immediately, no PM gate. Hard guardrail: any finding touching a module boundary, interface, or cross-system surface is ineligible regardless of size — always a spinoff candidate instead.
Mid-size cluster → one bundled spinoff candidate, not N singles or one bloated plan.
Large/structural → standalone spinoff candidate, or escalate to /plan. Scaffold a DR first when it's an architectural decision: coordinator-doc-new --type decision --title "<title>" --out docs/decisions/DR-<NNN>-<slug>.md.

Spinoff PM-gate applies only to the grouped/standalone path — surface Candidate spinoff: <slug> — <topic>. Authorize? and block. No debt-backlog auto-write.

Step 6: Update Health Ledger

Update the system's row (grade, status, audit date, open-issue counts) and the Last targeted audit date only — never Last full audit. Calculate and update Next rotation target if present. Stage, do not commit (git add state/health-ledger.md) — the commit lands only after the Step 6.5 gate PASSes, so the clock never advances against a stale atlas.

Step 6.5: Update Atlas Page (Pre-Commit Gate)

No atlas page for the target → skip entirely. Otherwise pick a branch:

Branch A (refresh inline): edit the atlas body for findings, bump both last_mapped and last_attested, set grade/health_status, stage alongside the ledger.
Branch B (assert current): zero body diff, bump last_attested only, stage, and the close-out commit message must contain atlas-current-as-of: <YYYY-MM-DD>.

Run verify-arch-audit-atlas-refresh <AUDIT_DATE> <TARGET_SYSTEM> [<COMMIT_MSG_FILE>] before git commit. PASS branch=A|B → git commit -m "<subject>" -- state/health-ledger.md <atlas-page-path>. FAIL → do not commit; amend per your branch, re-stage, re-run. Anti-scope and branch mechanics: wiki.

Step 6.75: Triage Scratch Files

Large-systems path used → rm -rf tasks/scratch/weekly-architecture-audit/{run-id}/ (Haiku/Sonnet output is fully consumed by the Opus reviewer).

Step 7: Report

Precondition: the canonical audit record at docs/architecture/audit-records/<date>-<system>.md exists with populated frontmatter and a ### Diagram (ASCII) section, committed with the atlas + ledger close-out, and the Step 6.5 gate returned PASS — on FAIL, fix per that step. Chat report ≠ on-disk record (FAIL text: wiki).

## Architecture Audit Complete

**System:** [name]
**Reviewer(s):** [name] at High effort [+ second-angle reviewer, or "none — single angle sufficed"]
**Previous grade:** [X] | **New grade:** [Y]
**Findings:** N total — [X → immediate executor, Y → spinoff candidate(s) surfaced to PM, Z → escalated to /plan]
**Spinoff candidates surfaced:** [list, or "none"]
**Next rotation target:** [system]

Notes

Failure modes, cost, and relation to /architecture-survey, /review//review-code, and pipelines/weekly-architecture-audit/PIPELINE.md: wiki.

architecture-audit

SKILL.mdskills/architecture-audit/SKILL.md