Autofix

---
description: Autonomous closed-loop quality remediation — composes /full-audit + /api-test + /browser-test + /pentest in parallel, creates issues for findings, applies minimal fixes on staging, re-verifies, loops until all four verification surfaces return zero finding. The longest-running command in the paved road. Plan/exec hybrid recommended for cost.
subtask: true
---

Run the **autofix** procedure across repos `apps/api apps/web`.

Follow the [autofix skill](../skills/autofix/SKILL.md) in full. Reuse the other skills — do NOT reimplement their procedures.

Expected behaviour per iteration:

1. Parallel phase 1 = `/full-audit` + `/api-test` + `/browser-test` + `/pentest` via Task tool (AgDR-0167 pattern #1).
2. Dedupe + triage (phase 2).
3. Issue creation via `paved-road-issue-from-finding` (phase 3).
4. Severity-ordered autonomous fixes on `staging` (phase 4) — delegate long mechanical fixes to a `pattern: cheap-exec` subagent if configured (AgDR-0167 pattern #2 / AgDR-0157 plan/exec hybrid).
5. Targeted re-verify (phase 5).
6. Loop until phase 1 returns empty.

`subtask: true` is set — runs in a fresh subagent context (AgDR-0167 pattern #3). PO-gated actions (merge to main, destructive prod, secrets rotation) MUST be escalated, never executed autonomously.

Expect multi-hour runs. Rigour over speed.