architecture-review

Run the architecture-review skill with a scope such as the entire codebase, a directory, or named modules. It reads actual calls, imports, shared models, and tests, then classifies each integration's strength, distance, and volatility to find unbalanced coupling.

Balanced Coupling is Vlad Khononov's model where balance equals strength XOR distance, or low volatility. High strength with high distance is only urgent when the integration is volatile; the goal is balancing coupling, not eliminating it.

Yes, the workflow launches three parallel read-only analyst subagents plus an adversarial critic. If parallel subagents cannot run, the skill stops and reports the capability failure rather than producing a shallow single-agent scan.

It writes a canonical Markdown report and a derived HTML report to docs/architecture-review/<date>/, containing at most five ranked findings with evidence citations, severity, confidence, before/after diagrams, and one top recommendation.

Avoid it when you only need a quick style check or when no verified change problem exists, since the review requires evidence of recurring change cost. It recommends direction only and does not design exact interfaces until you select a finding.