architecture

Router skill: resolve ONE topic, Read EXACTLY ONE playbook from routes:, execute it. Never load the cluster. [DOC]

When to use

Any architecture decision or design task matching a topic enum. If the request spans several (e.g. "event-driven system with caching"), pick the dominant concern, run it, then chain a second invocation — do NOT merge playbooks. [INFERENCE]

Inputs / Outputs

• In: topic (required; infer from request, ask only if genuinely ambiguous), depth (quick default | deep). [DOC]
• Out: the resolved playbook's deliverable, claims carrying Alfa-core tags ([DOC]/[CONFIG]/[CODE]/[INFERENCE]/[ASSUMPTION]; one family, one spelling — see references/verification-tags.md). [DOC]

Routing

1. Map request → exactly one topic. Ambiguous between two → ask, do not guess. [DOC]
2. Read that route's file. depth=deep → apply exhaustively, validate each step; quick → essentials only, name what you skipped. [DOC]
3. Spine: Discover → Analyze → Execute → Validate.
4. Gates: constitution v6.0.0 (enforcement), evidence tags, script-first. [CONFIG]
5. Gate enforced via assets/quality-rubric.json; pre-flight via assets/routing-checklist.md. [CONFIG]

Acceptance gate (done = all true)

• Exactly one playbook loaded; topic ∈ enum; no cross-playbook bleed. [DOC]
• Every non-obvious claim tagged, single family, consistent spelling. [DOC]
• Trade-offs stated with the rejected option and why. [INFERENCE]
• [ASSUMPTION] claims each paired with a verification step. [DOC]

Anti-patterns

• Loading 2+ playbooks "to be safe" — defeats the router. [INFERENCE]
• Guessing topic instead of asking when truly ambiguous. [INFERENCE]
• Recommending a pattern with no named trade-off or rejected alternative. [INFERENCE]
• Inventing prices, or marking green/done without the gate above. [DOC]