kirei

Architecture research agent. Maps module boundaries, dependency flows, coupling issues, and structural decisions. Produces a Mermaid diagram embedded in the findings doc (renders natively on GitHub) plus written architectural findings. Advisory only — no code changes.

Use this agent to implement findings from a kirei research agent. Optimized for focused, well-scoped tasks: single-file changes, clear bug fixes, small features. For complex multi-file or architectural work use kirei-forge instead. <example> Context: kirei has produced a handoff identifying a single-file bug fix. user: "implement the fix kirei found" assistant: "Spawning kirei-build to implement the targeted fix." <commentary> Single file, clear scope — kirei-build (sonnet) is faster and cheaper than forge. </commentary> </example>

Bundle-size research agent. Measures the actual built output, identifies the heaviest contributors, finds duplicate dependencies, missing code splits, unshaken tree leaves, and bloated assets. Distinct from kirei-perf (broad runtime bottleneck mapping) — this agent owns shipped-bytes deeply. Produces a structured handoff with measurable byte savings for kirei-build or kirei-forge.

Data layer research agent. Audits database schemas, migrations, queries, and indexes for safety, correctness, and design quality. Distinct from kirei-perf — focuses on schema design, migration safety, and data-integrity, not just query speed. Produces a structured handoff for kirei-build or kirei-forge.

Debug research agent. Reproduces a specific bug, traces the failure to its root cause, and prescribes a targeted fix. May add temporary instrumentation (logs, asserts) during diagnosis — every instrumentation point is tracked and listed in the handoff so kirei-build / kirei-forge can remove it after the real fix lands.

Audit a codebase for quality and maintainability — code smells, DRY violations, god files, dead code, inconsistent conventions, and unfollowed best practices. Depth-tunable like /kirei-deps; a scout pass sizes how many kirei-refactor agents run in parallel (1 → up to 6) to match the repo. Produces one merged, dependency-ordered cleanup plan, then offers to fix the issues in order via kirei-build / kirei-forge. Use whenever a user asks to audit the codebase, find code smells, hunt technical debt, check for DRY violations, find god files or dead code, surface inconsistencies, run a code-quality / health pass, or plan a refactor — even if they don't say "kirei". Invoke with /kirei-audit; the skill asks which depth before working.

Run multiple kirei research agents in parallel against the same target and merge their findings into one combined report. Use when a question needs more than one lens (e.g., "audit this module for security AND performance AND architecture", or "audit observability AND error handling together"). Invoke with /kirei-chain followed by a task description; optionally pass --types security,perf,arch to pin the lens set. Valid lenses: security, ui, refactor, perf, arch, test, data, observability, bundle, error, general.

Audit a project's dependencies for safety. Detects the package manager (pnpm/npm/yarn/bun/poetry/uv/cargo/go/bundler), runs its audit, fetches GitHub Dependabot alerts when available, and produces a depth-tunable report — quick CVE counts, safe-bump list, or full ordered upgrade plan with risky-major handoff to /kirei migrate. Use whenever a user asks to audit deps, check for CVEs, find safe upgrades, review Dependabot alerts, run npm/pnpm/yarn/bun audit, check for outdated packages, or plan a dependency hygiene pass — even if they don't say "kirei". Invoke with /kirei-deps; the skill will ask which depth before working.

Conversational pros/cons audit for an idea, feature, or project — before any code is written. Walks the user through problem framing, value, constraints, alternatives, risks, and a recommended next step (kill / spike / build now / wait), then writes a structured decision doc to docs/discuss/. Use whenever the user says "I'm thinking about…", "should we build…", "is X a good idea", "let's talk through…", "what do you think about adding Y", "audit this idea", or wants to think through an approach before committing. Pairs with /kirei — the resulting doc can be passed back via /kirei --findings <path> if the user decides to proceed. Invoke with /kirei-discuss followed by an idea description.

Set up production-ready Sentry error + performance monitoring for a project. Detects the framework (Electron, Next.js, Vite/React SPA, Node/Nest, React Native), asks the consent + scope + region decisions, verifies the current SDK API via Ref, then spawns the kirei-sentry agent to produce a framework-specific plan and kirei-forge to implement it — with consent gating, recursive PII scrubbing, CI-only source-map upload, region handling, and the exact CI secrets. Use whenever a user wants to add Sentry, set up error tracking / crash reporting, wire production monitoring, add performance tracing, or integrate Sentry into a new or existing project — even if they don't say "kirei". Invoke with /kirei-sentry; the skill asks the key decisions before working.