4-dimension parallel audit pattern for god-grade setup hygiene

For comprehensive “deep audit” requests on setup/memory files, dispatch 4 Explore subagents in parallel, each owning exactly ONE audit cluster:

  1. Format + Budget compliance — file sizes vs setup-curator format contracts, frontmatter presence, section structure
  2. References + Orphans — link integrity, phantom files (referenced but missing), orphan files (present but not indexed), hook/skill registry completeness
  3. Content currency + Stale — deprecated tool/MCP names, moved file references, version drift (e.g., v5.0.1 ↔ v5.0.5), removed component references, stale dates, TODO/FIXME/DEPRECATED markers
  4. Redundancy + Ambiguity — duplicate facts across files, conflicting statements, multiple sources of truth, restated rules with variation, vague references

Each agent returns a structured severity-tagged table (file:line / issue type / severity CRITICAL/WARN/MINOR / evidence). Plus run a parallel Bash-based live-state diff (docker ps, ls skills, claude.json mcpServers, crontab -l, systemctl list-units) — never trust documented counts.

Workflow: Dispatch all 4 agents + Bash live-state simultaneously → aggregate findings into severity-tagged table → present to user BEFORE fixing → execute fixes in dependency order (foundational refactors first, then cascades, then cosmetic) → run a final pristine sweep with 14 gates.

Evidence: This pattern was executed on 14-Apr-2026 aj-workspace setup audit and caught 12 issues in one session including: hooks-registry.md was 23% complete (1 of 13 hooks listed), Permanent MCP list was quadruple-restated across 4 files, Superpowers version drift (v5.0.1 vs v5.0.5), 3 topic files over the 400-line hard limit, a ghost file reference (st-fallback-journal.md), and a [DRIFT] entry from nightly cron sweep.

Use the subagent dispatch contract from protocols.md § Phase 3 (general-purpose type, no worktree isolation, self-contained prompts, bulletproofed via settings.local.json wildcards + watchdog hook).