architect
Architect
Start from caller usage, sketch types and module boundaries, compare structurally different candidates, and then implement only when the user's request includes implementation.
When installed, codebase-design is authoritative for deep-module vocabulary, interface depth, and design tests. Invoke it during grounding and synthesis. Without it, apply the bundled local tests: prefer a small interface that hides meaningful policy, keep representations private, align modules with owned knowledge, reject pass-through layers, and minimize the number of files needed to trace a flow. This skill is an execution workflow, not a competing design doctrine. When installed, use safe-refactor instead for a principally behavior-preserving structural change; without it, explicitly pin behavior, compatibility constraints, and verification before restructuring.
When candidates or critics are delegated, Subagents may communicate with each other, but no agent may communicate with a person. Repeat that sentence verbatim in every child prompt. All child agents keep artifacts local and may not send, post, comment, review, or mutate external systems. If subagents are unavailable, produce two independent local sketches before comparing them.
Model and reasoning preferences are optional. Read repository-documented model configuration or .agents/agent-models.md when present. Use architect candidates for candidate models, falling back to design candidates when that role is absent. Pass these choices to arena when it is installed. Use independent judge for a delegated design judge.
Read its optional reasoning profile line and validate each explicit model/profile pair against the current host. For a missing role or an auto or inherit-parent model entry, omit both overrides. For a missing, auto, or inherit-parent profile, omit the reasoning override. Never require the file, construct model identifiers from effort labels, or use unconfirmed choices. If a configured choice is unavailable or the host rejects a model/profile pair, report the fallback and retry that seat once with both overrides omitted. If the retry fails, record the gap or use the local fallback below.
1. Ground
Run how over the relevant subsystem when installed. Otherwise trace entry points, call chains, data flow, boundaries, types, and failure paths directly. Use why when installed and existing ownership, compatibility, or historical constraints matter. Otherwise inspect read-only git history, local docs, tests, comments, and available runtime evidence, keeping inference separate from sourced intent. Read repository instructions and the smallest relevant tests and contracts.
Skip only for genuinely greenfield work with no surrounding system.