Rendered from the source repository. Headings, examples, code, tables, links, and referenced images are preserved.
Design Gate
Turn "which design lens applies here?" into a small, evidence-based routing decision, then run only the selected lenses and merge their findings into one verdict. This skill is the routing authority. It does not invent a second review method or call a council.
For the workflow stage boundaries and the practical skills that follow this gate, read shared/references/workflow-stage-routing.md.
Workflow
- Identify the stage and the surfaces from the plan, Slice Contract, or changed scope. In
to-tasks, this is the one routing pass for the slice. Before implementation, use the inherited lens flags; re-route only when the slice's design surface changed. - Select lenses from the routing table. Multiple rows can match; cap at three. Keep the three most central to the change's risk. No matching row on a local change means no gate.
- Run each selected lens as a parallel read-only reviewer. Give it the plan or design problem, the files and boundaries in scope, and this gate response:
verdict,blocking_findings,advisory_findings, andrequired_changes. This response replaces a lens's standalone edit or implementation route. - Merge. Any
revisewith a concrete, load-bearing finding makes the gate verdictrevise. Cosmetic or speculative findings are advisory. - On
revise, state the required plan changes as a short numbered list. When lenses expose a real unresolved trade-off, setverdict: revise, setdecision_required, and stop. Never callmodels-consensusfrom this skill. The user may invoke it separately when more opinions are useful.
Routing table
| Change touches | Lens skill(s) |
|---|---|
| Overall shape of a new system, service, or subsystem | macro-architecture, software-design-philosophy |
| Module/service boundaries, cross-context integration, naming vs business language | domain-driven-design, macro-architecture |
| Business-logic structure inside one context (aggregates, invariants, events) | domain-driven-design |
| Code structure choices, extensibility, sprawling conditionals | design-patterns |
| Layer placement, cohesion, dependency direction, scope creep | architecture-lens |
| Module/class/API interface design — depth, information hiding, interface complexity | software-design-philosophy |
| Stored state, databases, queues, caches, migrations, retries, concurrency, external APIs | data-systems-coding-lens |
| Container/process topology of a distributed app — sidecars, ambassadors, adapters, load-balanced replicas, sharding, scatter/gather, FaaS fit, ownership election, work queues, batch workflows | distributed-systems-patterns |
| Event streams as a source of truth — adopting or migrating to event-driven microservices, Kafka/Pulsar/Kinesis topics, event schemas, choreography vs orchestration, CDC/outbox data liberation | distributed-systems-patterns (event-driven route), data-systems-coding-lens |
| The thing being built is an agent — an LLM tool-calling loop, a multi-agent pipeline, or a long-running autonomous run needing durable state, mid-run approval, or replay | agent-architecture-lens |
| React components, hooks, contexts, rendering performance, state placement | advanced-react |
| New or reworked user-facing flow, information hierarchy, or design system | ui-ux-pro-max |
| Two or more competing approaches with real trade-offs | architecture-lens |
Fallback row (only when no specific row above fits): no specific lens fits but the change is architecturally significant → run architecture-lens as a Lens 1 trade-off pass (name the 2–4 characteristics in tension, the real options, and the cost of each), and apply all three or no ADR before recording it with the canonical ADR template at architecture-lens/references/adr-template.md. Do not pair this with the rows above — it is the catch-all when a change is significant yet unroutable, not a peer lens.
No row matches and the change is not architecturally significant → the change is local; return proceed with the note "no design gate required" and do not invent a review.
Output contract
Return:
verdict:proceedorrevise.lenses_run: which skills ran and why each was selected.blocking_findings: load-bearing findings requiring plan changes (empty whenproceed).advisory_findings: non-blocking observations worth carrying into implementation.required_changes: numbered plan amendments (only whenrevise).decision_required: the unresolved trade-off and its owner, ornone.
In a Slice Contract, carry lenses_run, verdict, required_changes, decision_required, and the security flag from security-gate. Resolve blocking findings before task approval. The implementation engine receives those constraints; it does not replay every lens.
Gotchas
- Do not run every lens to be safe — the cap exists to keep the gate cheap and the findings sharp.
- Do not let lenses edit anything; the gate is read-only.
- Do not block on style opinions; only findings that change boundaries, data ownership, contracts, or correctness block.
- Do not re-run the full gate after a
revisefix — re-run only the lens that raised the blocking finding. - Do not use a consensus panel to break a tie. The gate can identify a decision; it cannot manufacture user preference.
