Arch-check — architecture-analyst delegation

Layer boundary work delegates to the architecture-analyst agent so
cross-module reasoning happens in its own context.

Step 1 — Frame the scope

Identify the architectural question:

• "is this layer-clean?" → diff or files in scope
• "does this respect the import rules?" → changed imports
• "where should X live?" → proposed change + neighboring modules

If the user did not specify, ask what they're checking. The
architecture-analyst is too expensive to invoke on an undirected scan.

Step 2 — Delegate to architecture-analyst

Dispatch the architecture-analyst subagent (subagent_type='architecture-analyst'). The
analyst sources the project's layer model and invariants from
CLAUDE.md / docs/ARCHITECTURE.md / docs/CONVENTIONS.md — whichever
of those your project keeps; brief
it on anything not yet captured in the project's docs that you want
checked.

Example invariants (the harness this skill originated in):

• espalier/ imports from espalier/ only.
• tools/cc/ has zero espalier imports (standalone scripts).
• espalier/scanners/ is stdlib-only.
• Hook scripts exit 0 only (channel-XOR rule; exit 0 pairs with stdout
  JSON, exit 2 with stderr text, never both).
• Path comparisons normalize via .replace("\\\\", "/").

Your project's invariants will look different — the analyst handles
that adapter logic when it reads the project's docs.

The agent returns a structured architectural verdict — direction
violations, scanner-isolation breaches, and hook-exit-code errors are
common load-bearing failure modes.

Step 3 — Report

Pass the agent's verdict back. If it flagged a layer violation,
quote the offending import line and propose where the affected
function should move instead.