---
name: architecture-refresh
description: Thin orchestration for the recurring architecture audit→apply→re-audit loop. It holds NO architecture policy — it only sequences four predefined subagents (two auditors, two appliers), reads their convergence signal, routes each finding to the applier the auditor named, and re-calls them until an audit round is materially clean. Every judgement lives in the agents. Use to keep architecture and implementation in sync when a single pass won't finish it.
---

# Architecture Refresh — pipeline only

This skill only **calls predefined agents and manages the loop**. It carries no architecture policy —
all judgement (criteria, scoping, conformance checks, how to fix docs, how to implement code, the
safe/gated boundary, what counts as material) lives in the agents. The skill names them, reads their
signal, and routes.

The four predefined agents (spawn by `agentType`):

- `architecture-auditor` — read-only; returns findings + `ACTIONABLE FINDINGS: <n>`.
- `architecture-conformance-auditor` — read-only; returns findings, each labelled **doc-side** or
  **code-side**, + `ACTIONABLE FINDINGS: <n>`.
- `architecture-fixer` — applier for **doc-side** findings.
- `architecture-implementer` — applier for **code-side** findings.

## Pipeline

1. **Audit.** Fan out both auditors over the target, one per disjoint area, covering every unit (log any
   area you leave out). Collect each area's findings and `ACTIONABLE FINDINGS` count.
2. **Converged?** Stop only when a full audit round reports **no material findings** in any area. Never
   stop on a round count.
3. **Apply.** Per area with findings, call the applier the auditor named — doc-side → `architecture-fixer`,
   code-side → `architecture-implementer` — on **disjoint files**. Keep whatever each applier reports back
   (applied / skipped / escalated).
4. **Re-audit.** Re-run the auditors on the changed areas.
5. **Loop** 1–4 until step 2 says converged.
6. **Land** the applied changes through the repo's normal flow; pass any escalations the appliers return
   into the repo's gated backlog.

That is the whole skill. Everything else is the agents'.