Discovering Architecture

Overview

Run a deterministic scan of the codebase, then dispatch a subagent to add narrative observations and adjudicate anything the scanner could not classify. The scanner (@unwind/core) produces a machine-readable scan-manifest.json — the ground truth for the file inventory, per-file structural symbols, the import graph, and a first-pass rebuild-layer assignment. The architecture document is derived from this manifest, so layer counts and entry points are verifiable rather than guessed.

Output: docs/unwind/architecture.md (derived) + docs/unwind/.cache/scan-manifest.json (ground truth)

Graceful fallback: if Node/pnpm or @unwind/core are unavailable, the deterministic scan is skipped and discovery falls back to the legacy pure-LLM Explore flow (Step 1-LEGACY below). The skill always functions; the scan is an enhancement, not a hard dependency.