claude-kit

Create conventional commits in production context with strict validation, tests, linters, and confirmation.

Surveys pending work across TODOs, planning docs, unfinished feature plans, open spec questions, in-flight git work, and open GitHub issues, then returns a value/effort table with a recommended next action. Use at session start to triage what to work on, especially after a gap when context has faded.

After a kit sync, verifies every entry in `.claude/kit-manifest.txt` is on disk and reconciles `CLAUDE.md` against `.claude/kit-tools.md` — surfacing sync gaps, drift, gaps, and redundancies. Proposes a CLAUDE.md patch the user reviews and applies manually.

Interactive spec writer for new features. Interviews the user to understand the feature (even if vague), reads the existing domain, then produces docs/spec/{feature}.md with structured TRIGRAM-NNN business rules and an optional UX draft. For new features only — not for documenting existing code.

Captures and commits visual proof screenshots for any `.tsx` / `.css` change. Generates a full preview for every component state (idle/loading/results/empty/error) in both light and dark mode, captures with Playwright via `scripts/visual-proof-capture.mjs`, and reports any console errors found. Auto-discovers project config on first run.

Pre-release validator for kit artifacts. Checks downstream-readiness of agents/skills (no kit-centric language, correct paths), script quality, and cross-component coherence. Use before releasing kit updates to downstream projects.

Audit the project for structural simplification — fewer lines, methods, and unnecessary code — without touching architecture or DDD structure. Coverage report is mandatory; refuses to proceed without one. Read-only; produces a prioritized report for human review.

One-time setup of the Tauri WebDriver E2E infrastructure. Installs npm packages, generates wdio.conf.ts from the project binary, adds test:e2e / test:e2e:ci npm scripts. Run once before the first test-writer-e2e invocation. Safe to re-run — skips steps already done.

End-of-session maintenance — audits recent work (git log, memory-file mtimes, CLAUDE.md diff) for rules earned, contradicted, or now surplus, then proposes promote / remember / trim / skip decisions for CLAUDE.md. Output-only — the user confirms each and the main agent applies. Trigger when wrapping up a session ("done for today", "/exit") or after a notable stretch of changes; not for start-of-session triage — use /whats-next for that.

Use at the start of a new task to lock in workflow (Full vs Simple), create a feature branch, and emit the session's working context. Invoked as `/start [fix|chore|test|feature|refactor]` where the scope argument pre-selects a workflow. Not for mid-session triage of in-flight work — use `whats-next` instead.

Derives or updates a domain contract (docs/contracts/{domain}-contract.md) from a validated feature spec. Upsert-aware — adds commands without overwriting. Run after spec-reviewer approves, before contract-reviewer and `/feature-planner`. Not for validating an existing contract — use `contract-reviewer` instead.

Translates a validated spec into an implementation plan (docs/plan/{feature}-plan.md) mapping TRIGRAM-NNN rules to DDD layers and the kit workflow. Use after `spec-reviewer` and `contract-reviewer` approve. Output is gated by `plan-reviewer` before any test-writer runs. Not for spec or contract authoring — use `/spec-writer` or `/contract` instead.

Triages reviewer-* findings against the (a)/(b)/(c) per-task discipline before any are applied. Reads `.review/` reports, grades each finding, emits a per-row Follow-up table, and halts for user confirmation on any (b) or (c) row. Auto-invoked at the end of every reviewer batch by `/start`; also usable standalone after ad-hoc reviewer runs. Routes (b) rows to `/techdebt` — does not replace it.

Push the current feature branch and open a GitHub pull request. Drafts PR title from the branch name and commits; drafts body from branch commits and the feature plan doc if present. Requires gh CLI. Use at the end of any feature branch workflow.

Authors and supersedes Architecture Decision Records (docs/adr/*.md). Run when a decision passes the 3-criteria gate (genuinely complex / not obvious from context / costly to reverse). After authoring or superseding, run `adr-reviewer` to validate quality before locking in. Not for tentative decisions — those stay in the spec's `## Open Questions` until ratified.

Produces a normalized tech-debt entry in chat — date-stamped, with auto-filled git context (branch + short commit). The entry is text-only output; the main agent decides whether to append it to `docs/techdebt.md` (the convention), file a GitHub issue, or surface it elsewhere per the downstream CLAUDE.md's policy. Use when a reviewer surfaces a `[DECISION]` critical or pre-existing tech-debt observation worth recording, or when the main agent notices a non-actionable smell during work.

Audit npm and Cargo dependencies for outdated versions and CVEs, with web-verified version data. Run before every release or periodically during development.