agent-auto-kit

Kick off the AgentAutoKit multi-agent workflow for a feature or bug

Summarize AgentAutoKit run metrics into a routing scorecard

Propose (or apply) model-tier changes from the routing scorecard

Read-only architecture and design advisor. Use for design decisions, tradeoffs, and evaluating approaches before implementation.

Read-only code review specialist. Use proactively once per PR, in parallel with security-auditor, before opening the PR.

Fast read-only code explorer. Locates files, traces call sites, maps structure. Use proactively before implementing to gather context.

Deep debugging specialist for hard bugs — async/race conditions, complex types, subtle state issues. Use when implementer fails repeatedly or the bug is genuinely tricky.

Default implementation agent. Writes and edits code to satisfy a task, keeping tsc clean and tests passing.

Project conventions for AgentAutoKit-managed repos — npm, TypeScript, Vitest, Vercel. Use when writing or reviewing code in this repo.

Apply web accessibility (a11y) best practices to WCAG 2.2 AA when writing, reviewing, or fixing any UI, component, page, or form. Covers semantic HTML & landmarks, correct ARIA (name/role/value, and the "no ARIA is better than bad ARIA" rule), keyboard operability & visible focus, focus management for modals/menus, color contrast & not-color-alone, text zoom/reflow, target size, accessible forms (labels, errors, autocomplete), alt text, live regions for dynamic updates, reduced motion, and the html lang attribute. Use when asked to "make this accessible", "fix a11y", "add ARIA", "keyboard navigation", "screen reader support", "check color contrast", "fix focus", "WCAG", "add alt text", "accessible form", or when reviewing ANY interactive or content UI. Complements frontend-design (aesthetics) and responsive-design (layout). Treat findings as things to fix, not just flag.

Use when developing a feature or agent task that needs an isolated working copy — e.g. running a second dev server, testing a risky change without disturbing the main checkout, or parallel agent work. Triggers on requests to "work in a worktree", "spin up an isolated branch checkout", or run two branches side by side. Covers creating a worktree under .claude/worktrees/, deps/env/port setup, syncing with main, merging back, and cleanup.

Use when running, debugging, or authoring end-to-end user-journey Playwright specs in this repo (tests/*.spec.ts) — e.g. registration/payment flows like pwc2026-flow or team-format-flow, adding a new full-journey spec, wiring its env/seeding prerequisites, or diagnosing a failed flow run (test-results/, playwright-report/). For generic Playwright technique (locators, flakiness, POM, CI), defer to the playwright-best-practices skill.

Apply Firebase best practices when writing, reviewing, correcting, or hardening Firebase code and config — Firestore/Storage security rules, composite & single-field indexes (firestore.indexes.json), Firebase Auth, role/permission (RBAC) design, Cloud Functions, Realtime Database, and Remote Config. Use when asked to "fix/optimize Firestore indexes", "audit/tighten security rules", "standardize roles", "set up custom claims", "harden a Cloud Function", "write Realtime Database rules", "make this Firebase project production-safe", or when reviewing any .rules / indexes.json / Auth / functions / RTDB / Remote Config code. For project-specific Firestore config editing/seeding/deploys, defer to firestore-config-edit.

3 hooks across 3 events