Architecture Critic.The first Lumina sub-agent.

A critique-discipline overlay routed onto Lumina when the visitor asks for an architectural review. Same Lumina voice, different shape of answer. Routing is deterministic and falls back to the default chat for anything ambiguous.

01 · When the router picks it

routeRequest(messages) inspects the most recent user message and returns either lumina (default) or architecture-critic. Trigger cases:

• Explicit prefix@architecture-critic should I store secrets in Lambda env vars?
• Bracket marker[critic] review my cross-account scanner design
• Critique verb + architecture nounwhat are the tradeoffs of this dataflow pattern? — verb 'tradeoffs' + noun 'pattern'
• Question-form critiqueis this a good approach to multi-region failover?

Everything else falls to the default Lumina path — the router's bias is deliberately toward single-agent mode (constitutional MUST: “fail back to single-agent mode”).

02 · Critique discipline

The system prompt shapes every response into the same four-step structure. Not because templates are good, but because a senior-engineer code review reads this way consistently and the visitor benefits from predictable shape.

1. 1. Strongest concern firstOpen with the highest-impact observation. What's most likely to break, scale poorly, or cost a fortune. Don't soften. No filler preamble.
2. 2. Subtler concern secondThe issue that bites in production months later — maintainability, observability, incident response, edge cases.
3. 3. Alternatives placed, not pitchedBriefly name 1-2 patterns with a different tradeoff profile so the visitor can pick.
4. 4. One specific recommendationThe single change that would most improve the design as proposed. Not a list. Not a 'depends on your priorities' cop-out.

03 · Orchestration trace

The architecture-critic's system prompt mandates a single call to the synthetic selectArchitectureCritic tool before any text response. The chat UI's existing tool-status spinner renders this as the engaging architecture-critic pill — the constitutional “visible orchestration tracing” requirement satisfied through the primitive that's already there. The tool itself does no real work; it exists only to surface the trace.

04 · Routing decisions to date

No routing decisions recorded yet. Counts land here the moment a chat turn passes through the router.

05 · Source files

• Agent definitionlib/lumina/agents/architecture-critic.tsFull system prompt, init tool, and exported identity constants.
• Routerlib/lumina/router.tsHeuristic classifier (no LLM call). Critique verbs + architecture nouns. Always falls back to default Lumina on ambiguity.
• Chat route dispatchapp/api/chat/route.tsWhere routeRequest() runs and the sub-agent prompt + synthetic init tool get composed.