pattern-patrol — keep eradicated problems dead

Canonical system (read first): ../common-docs/systems/intelligence/pattern-patrols/FEATURE.md Arman's target: .../pattern-patrols/VISION.md The patrol list + status: .../pattern-patrols/PATROL_REGISTRY.md This skill is the matrx-frontend mechanics. Never duplicate the registry here.

The one-paragraph model

A patrol is a named pattern Arman wants permanently eliminated (or a good pattern he wants present everywhere), backed by doctrine, a durable queue, a skill teaching the fix, a scheduled recurring run, and adversarial certification of every fix batch. The registry owns the current count. A patrol stays in ERADICATION until an independent full pass proves zero actionable backlog; only then may it enter MAINTENANCE. Detection is never the terminal job while actionable work exists.

Duty 1 — running a patrol (you were launched as one)

Follow the per-run template in CODEX_OPERATOR.md (same directory as the system doc). The repo-specific facts it needs:

• Ledger: .matrx/PATROL_SIGHTINGS.md. Permanent run history: .matrx/patrol-runs/<P#>/<run-id>.json (append only through pnpm patrol:run; latest.json is only a projection). Reports: .matrx/patrol-reports/<id>.md (create the directory on first use; one file per patrol, overwritten each run — it carries your scan baseline).
• Resume before discovering: inspect latest.json and its permanent record before a new scan. An unfinished approval, fix, certification, infrastructure retry, or delivery resumes with its exact candidate first. Never strand it, overwrite its report, or ask Arman to repeat a decision.
• Modes, not reporting tiers: ERADICATION consumes the ranked ready queue before broad discovery, uses every available parallel-agent slot (never more than ten workers), and runs multiple independently certified batches until verified zero or one hard blocker stops every remaining ready item. MAINTENANCE is allowed only after independent zero proof; its first finding switches the same run back to ERADICATION and starts repair immediately.
• Per-item routing: every verified item is repair-now, genuine-human-decision, or hard-blocked. Repair-now executes now. A decision or blocker stays open while the patrol takes the next independent repair-now item. A detector, ranked list, chip, or machinery improvement is never a completed patrol outcome while ready product work remains.
• Professional improvement standing authority: automatically fix a verified defect or weakness when one remedy is clearly superior, reuses a canonical primitive or demonstrated industry standard, preserves product behavior/contracts, and fits a bounded certified batch. The skill need not enumerate the exact callsite. Known bugs and established quality upgrades do not wait for permission.
• Decision routing: ask Arman only when legitimate alternatives materially change product behavior, policy, workflow, permissions, data meaning, destructive impact, or visual intent. Missing evidence/machinery creates a focused task or infrastructure state. If the core repair is clear and an optional enhancement is debatable, ship the core and ask only about the enhancement. Implementation uncertainty never turns into a pointless approval request.
• Human decisions are item-scoped: when Arman chooses among legitimate alternatives, apply only that decision. The resulting repair still uses a bounded Tier-M batch with normal gates and adversarial certification. Reports separate standing-authority fixes, genuine human decisions, unresolved evidence/machinery, and the batch verification/certifier verdict. It never repeats a known defect inventory as a substitute for execution.
• Hard rules, non-negotiable: never disable a check, add a suppression, touch generated files, or change how a component enters a chunk (THE FRAGMENTATION LAW — code-splitting skill before ANY such change); fixing one side must not move the other (mobile↔desktop, dark↔light); pnpm type-check before any done-claim. Commit every coherent batch immediately and push it to a remote ref within 15 minutes. After independent certification names the exact candidate SHA, integrate it through the normal fast origin/main workflow within 45 minutes. Only deployment/release stays serialized.
• Shared-checkout ownership: scheduled patrols run in the canonical shared checkout; worktrees are forbidden. Capture exact base SHA, dirty paths, and baseline diagnostics before editing. Unrelated dirty files belong to other owners: never stage, rewrite, revert, stash, or clean them. Claim disjoint repair units, re-read an owned file before editing, use path-scoped Git, and commit only owned files. Concurrent unrelated work is not patrol evidence.
• Exclusive preview lease: pnpm preview:start uses only the canonical checkout's managed server. Never stop it while another task is actively verifying. pnpm preview:status reports ownership. Read the machine-profile RSS cap from launcher status; the five-minute startup-progress cap remains fixed. The launcher never restarts automatically.
• Certification (Tier M): a second adversarial agent ("assume this batch broke something; find it") compares pre-edit and post-edit type/gate diagnostics. New batch-caused failures reject; unchanged baseline debt is loud but cannot reject. Apply FEATURE.md's risk-based visual proof: every changed file gets scoped static coverage, repeated mechanical edits get one representative surface per distinct risk class, and shared primitive/layout/ interaction/theme changes get the full relevant matrix. CERTIFIED ships; REJECTED requires a concrete batch defect and is fixed/reverted; INFRASTRUCTURE BLOCKED preserves the approved diff for retry. A broken preview is never proof that product code broke. Only one managed preview runs machine-wide; concurrent patrols queue instead of starting a second build. No independent verdict → invalid run.
• Fast integration, serialized release: push the candidate immediately. After an independent certifier records CERTIFIED for the exact candidate SHA, integrate it into origin/main through the normal shared workflow; preserve that commit as an ancestor. The machine-wide lease applies only to ./scripts/release.sh. If a newer release already contains the candidate, record that version instead of creating another bump.
• Scoping: structural novelty (new app/**/page.tsx leaves, new features/* dirs, new files matching the patrol's surface signature) + the ledger + a full pass every Nth run. NEVER scope by raw git churn.
• Loud degradation: if a patrol cannot execute its ready queue, or a required read, scan, fix, certification, gate, release, report, or memory update does not happen, follow FEATURE.md's exact Loud Degradation Contract. Begin with AUTOMATION DEGRADED — ACTION REQUIRED; when Arman must act, end with ARMAN, WE NEED YOU: <one specific next action>. Counts alone are never an adequate warning.
• Human-owned exceptions: agents may propose an exception but never clear, suppress, allowlist, or approve one. Every proposal stays an open finding, includes a production review URL, and ends with EXCEPTION APPROVAL REQUIRED plus ARMAN, WE NEED YOU. After explicit approval, record a durable approval id/reason/reference in the patrol's typed allowlist and beside the source; detectors report approved exceptions separately instead of hiding them.

Duty 2 — logging a sighting (you're on another mission)

You spot a live violation of a registered patrol. Do not fix it (unless it is literally inside the lines you're already changing — boy-scout rule). Add one line to .matrx/PATROL_SIGHTINGS.md:

- [ ] P4 | features/foo/Bar.tsx:120 | bg-white with no dark: pair on the modal shell | 2026-08-08

…and continue your mission. The patrol verifies sightings itself; yours is a hint, not a promise, so thirty seconds is the right amount of effort.

Duty 3 — nominating a new patrol (the list must grow)

When a mistake you're fixing looks like a PATTERN — same class in a third place, something Arman has ranted about, a check you wish existed — stop and tell him: the pattern in one sentence, grep-level evidence with real counts (run the greps; no vibes), proposed tier + cadence. On approval: add the registry row (or Candidate-bench line), dispatch the sweep to a subagent with a named lane (law 7 — never a spawn_task chip), and note which of the five parts already exist. Promotion patterns (presence of something good — copy-for-AI, assists, admin-map rows) qualify exactly like elimination patterns.

Anti-patterns

• Fixing sightings off-mission (derails sessions — the ledger exists so you don't).
• A patrol "improving" style beyond its registered pattern.
• Reporting a clean run as wasted effort — zero findings is the system working.
• Marking a registry status ✅ that isn't (the registry must never lie).
• Giving a polished normal-looking summary for a degraded or incomplete run.
• Rejecting or reverting valid work because an unrelated baseline gate or the preview harness failed.
• Using a noncanonical preview, leaving owned work uncommitted or unpushed, or delaying certified work behind a fictitious integration gate.
• Treating a Markdown report or automation memory as more authoritative than the permanent run record, or rewriting an earlier lifecycle event.
• Stopping after detection while any actionable repair remains, or asking Arman to approve an obvious professional improvement.
• Assigning parallel agents to summarize or rank defects instead of giving each one a disjoint repair unit with required verification.
• Treating “looks intentional” or “false positive” as approval to suppress it.
• Growing this skill with per-patrol content — that belongs in the registry row or the pattern's own skill.