Reviews API code for controller/service/repository separation, validation layer placement, error contract consistency, idempotency, and versioning discipline. Cites the api-layering rule. Distinct from api-designer (designs API surface) and api-routes rule (per-route concerns).
Reviews changed code for layer-dependency violations, cross-module reach, circular dependencies, god modules, public-API leaks, and structural patterns. Cites the architecture-layering rule. Distinct from architect (which plans) — this agent reviews code against architectural rules.
Orchestrates build phases sequentially — scaffold → happy-path → edge-case → tests → docs → refactor — using the plan from planner-coordinator. Each phase is one specialist; the coordinator manages sequencing, state, and resumption.
Reviews changed code for excessive function length, cyclomatic complexity, nesting depth, parameter count, and branch density. Cites the complexity rule.
Reviews diffs for material changes (package manager, test framework, build tool, directory layout, env vars, CI workflows, framework/ORM/auth provider swaps, breaking API changes) that landed without a corresponding CLAUDE.md / AGENTS.md / README update. Cites the docs-staleness rule. Read-only. Distinct from…
Build phase 3 specialist — adds input validation, error handling, edge-case handling, and defensive code to the happy-path implementation. Tightly bound to the error-handling, auth-security, and data-protection rules. Constrained by all relevant .claude/rules/.
Self-improvement detector — scans the project, identifies improvements to CLAUDE.md placeholders, .claude/rules patterns, settings, and agent configs. Read-only by tool restriction. Builds the /setup-owned skip-list before scanning so improvements respecting onboarding decisions are filtered out at proposal time.…
Reviews changed frontend code for component composition, state management, hook discipline, data flow, render performance, module boundaries, and SSR/hydration concerns. Cites the frontend-architecture rule. Distinct from ui-ux-reviewer (visual/a11y/design system).
Build phase 2 specialist — fills in the core successful flow for stubs created by scaffold-implementer. Implements the "everything works" version. Defers error handling, edge cases, and validation to the edge-case-implementer phase. Constrained by all relevant .claude/rules/.
On-demand precise impact analysis for a symbol or file. Greps for callers, reads them, classifies the cascade (direct/indirect/test-only), and writes a structured report at .claude/state/impact- .md. Used by /impact and callable by review/refactor agents before suggesting changes.
Orchestrates planning specialists (requirements-clarifier, scope-decomposer, risk-assessor, test-strategy-planner, plus architect/api-designer/database-architect when relevant) and synthesizes their outputs into a single implementation plan.
First-time onboarding applier — reads .claude/state/setup-proposal.md (with user-confirmed values), validates the apply allowlist, snapshots existing files, then applies placeholder substitutions to CLAUDE.md, .claude/rules/, .claude/skills/, settings, and .gitignore. Refuses to run if ## Confirmed by user is empty.…
First-time onboarding detector — inventories the repo (greenfield vs brownfield), runs 17-layer stack detection with tradeoff-explained options + recommended defaults, writes .claude/state/setup-proposal.md. Read-only by tool restriction. Paired with project-setup-applier and orchestrated by /setup. Distinct from…
Reviews changed code for impure functions, hidden side effects, query/command separation violations, input mutation, hidden state reads, and SRP violations at function and class level. Cites the purity rule.
Build phase 6 specialist (final pass before review) — actively applies code-quality rules to the just-implemented code: simplifies, extracts, dedupes, reduces complexity. Tightly bound to code-smells, dry, purity, complexity rules. Goal is that the review-coordinator finds little to flag.
Hunts ambiguity in a story or task description before planning starts. Surfaces open questions, undefined terms, missing acceptance criteria, conflicting requirements, and assumed-but-unstated constraints.
Synthesizes findings from parallel reviewer agents into one consolidated report — dedupes, filters false positives, applies cross-iteration state, classifies the diff into a risk tier, and decides which sub-agents to spawn.
Identifies rollback paths, blast radius, breaking-change risk, data-migration danger, and operational risk for a planned change. Output feeds into the plan so risky steps get extra scrutiny.
★not rated 3 3mo agoA40 tokens
originalMIT
At most 3 mods per repository are shown here, and a mod shipped inside a plugin is left to that plugin's page — the rest are on their repository pages: