devkit

Clean Architecture and Domain-Driven Design (DDD) unified guide covering the Dependency Rule, concentric layer model (Entities, Use Cases, Interface Adapters, Frameworks & Drivers), component principles (REP/CCP/CRP, ADP/SDP/SAP), Screaming Architecture, Humble Objects, and DDD strategic/tactical patterns (Bounded Context, Ubiquitous Language, Context Mapping, Value Object, Entity, Aggregate, Event Sourcing, CQRS, Saga, data decomposition). Use when designing application layer structure, enforcing dependency rules, separating business logic from frameworks/UI/databases, modeling complex domains with DDD patterns, or mapping Bounded Contexts to Clean Architecture layers. For code-level clean practices, use writing-clean-code. For distributed/infrastructure patterns, use cloud:architecting-infrastructure. For project foundations and team practices, use practicing-software-engineering. For writing use cases as behavioral requirements, use product:writing-use-cases instead.

Semantic Versioning 2.0.0仕様に基づくバージョン判断ガイド。MAJOR/MINOR/PATCH判定フロー、プレリリース・ビルドメタデータ、優先度比較ルール、npm/pipのバージョン範囲指定（^・~・==）、検証用正規表現を提供。 REQUIRED for all version-related decisions. Use when determining version bumps, creating releases, or managing dependencies. For commit message format (type/scope/BREAKING CHANGE), use writing-conventional-commits instead. For API versioning strategy trade-offs, use cloud:architecting-infrastructure instead.

Claude Code and Codex plugin authoring guide for Agent, Skill, Command, and Hook components. Covers the Agent Skills standard (agentskills.io), Codex skill packaging (.agents/skills, .codex-plugin manifests, agents/openai.yaml), Claude Code extensions (context: fork, when_to_use, disable-model-invocation), name/description validation, spec-freshness checks, source-to-skill conversion, multi-plugin management, version sync, agent impact analysis, cross-reference integrity, and skill-improvement loops. Use when creating, modifying, or reviewing plugin components, auditing Codex skill compatibility, splitting or adding plugins in a multi-plugin repo, or processing skill-improvement proposals. For discovering and installing existing skills, use find-skills instead. For MCP protocol/server development, use lang:developing-mcp. For converting Claude Code agents to Codex, use converting-agents-to-codex. For codebase-wide automation recommendations, use recommending-automations.

作業中の学び・エラー・ユーザー訂正・機能要望を `.learnings/` に構造化記録し継続的改善につなげるスキル。反復パターンを検出しプロジェクトメモリ（CLAUDE.md / Claude Code memory / AGENTS.md）や新スキルへ昇格する。 Use when: ユーザーが訂正した時 / 存在しない機能を求められた時 / 知識が古い・誤りと判明した時 / 反復タスクでより良い方法を発見した時 / 大きな作業前に過去の学びを振り返る時。devkit の hook（UserPromptSubmit・PostToolUse）が学び記録を自動リマインドする。 For improving sumik-claude-plugin's own skills, use authoring-plugins instead. For curating durable cross-project facts, use the Claude Code memory system / managing-claude-md.

Views the user's screen and several hours of screen-recording history via the Chronicle (Codex screen recording) feature: reads the latest per-display frame, greps OCR sidecars (*.ocr.jsonl) for recent on-screen text, inspects historical 1-minute frames across multiple displays, and consults Chronicle memories. Use when the user makes a reference to their recent work, for which it'd be helpful to see the screen. This skill MUST be used whenever you need to resolve ambiguity in a user request, where the user hasn't specified enough context to do the task. Examples include disambiguating the specific user/app/document/error the user is referring to. You must also use this skill if the user asks about any question regarding Chronicle or asks what you can see from the screen. Chronicle is passive observation only; for actively driving or automating a browser, use web:automating-browser instead.

Claude Code Agent定義（.md）をCodex subagent定義（agents/*.toml）に変換するガイド。 フィールドマッピング・developer_instructions変換・skillsのdescription自動ロード方式・モデルtier-map（model_reasoning_effort）・config.toml/mcp_servers設定・AGENTS.mdルーティング統合・検証手順・よくある失敗パターンを網羅。 Use when converting Claude Code agents to Codex format, migrating agent definitions, or setting up Codex subagents. Triggers: "agentをcodexに変換", "codex agent変換", "agent migration"。 For Codex-side agent orchestration (Wave並列実行・max_threads運用), use orchestrating-codex instead. For authoring new Claude Code plugins/agents/skills, use authoring-plugins instead.

Converts Word, PowerPoint, Excel, OpenDocument, RTF, and CSV files to GitHub Flavored Markdown via anydoc (firecrawl/anydoc, MIT v0.1.6), run with `npx -y @firecrawl/[email protected]` (Node.js 20+ required; first run needs network access to fetch the package), with matching npm/PyPI/crates.io libraries for code integration. Use when converting office documents or spreadsheets to Markdown, especially legacy formats (.doc, .ppt, .xls, .ods) that pandoc/pdf-to-markdown cannot read. anydoc does not perform OCR — for scanned study materials, use the Apple Vision / local VLM workflow in certificate:creating-flashcards instead.

Discovers and installs agent skills from the open skills ecosystem using the Skills CLI (npx skills find / add / check / update) and the skills.sh registry and leaderboard, vetting candidates by install count and source reputation before recommending or installing them. Use when the user asks questions like "how do I do X", "find a skill for X", "is there a skill that can...", wants to search for tools, templates, or workflows, or expresses interest in extending agent capabilities with installable skills. For evaluating npm/pip/Go code libraries, use researching-libraries instead. For authoring local plugin/skill components, use authoring-plugins instead.

Operates as Tachikoma Agent (Implementation Worker) performing actual code implementation, test creation, and documentation based on task assignments from Claude Code. Covers the worker lifecycle: worktree-confined work (no worktree creation/deletion, no git add/commit/push — git writes are user-executed), serena MCP-first code editing, parallel execution criteria, tachikoma1-4 specializations (frontend/UI, backend/API, testing/QA, infrastructure/DevOps), mandatory software-security check before completion, and the completion report format. Use when receiving task assignments from Claude Code for implementation work, test creation, or documentation as a worker agent. For orchestrating multiple Tachikoma agents in parallel (team formation, planning), use orchestrating-teams instead. For Codex-side orchestration, use orchestrating-codex instead. For detailed serena MCP usage, use using-serena instead.

Manages CLAUDE.md creation, improvement, and maintenance following 8 core principles (300-line limit, avoiding /init auto-generated boilerplate, delegating style rules to linters/formatters, 3 essential elements — one-line project summary, common commands, If-X-then-Y trap rules — progressive disclosure into rules/ and docs/, priority-first placement, long-term memory operation, living documentation). Covers an interactive improvement workflow (AskUserQuestion), scope design (project CLAUDE.md, claude.local.md, global ~/.claude/CLAUDE.md, monorepo per-package files), daily/weekly/monthly maintenance cycles, plus templates and validation checklists. Use when starting new projects, writing or reviewing a CLAUDE.md, splitting an oversized CLAUDE.md, adding If-X-then-Y trap rules, or when Claude repeatedly makes the same mistakes. For skill file (SKILL.md) authoring, use authoring-plugins instead. For capturing session learnings and errors into .learnings/, use capturing-learnings instead.

Comprehensive TypeScript guide covering type system, advanced patterns, design patterns (GoF Creational/Structural/Behavioral), functional/reactive programming, framework integration (React/Angular/Vue/Express), and 83 best practice decision criteria. MUST load when working in TypeScript projects detected by tsconfig.json. Covers Generics, Conditional Types, type inference, type design, SOLID/DDD, anti-patterns, and migration tips. Includes type safety rules (any prohibition, type guards, strict mode), compiler behavior/modern tsconfig (satisfies), TS 5.0 decorators, type-safe error handling, performance/project scaling, contract evolution (Zod validation), IDE-driven workflow and common pitfalls, and TSDoc/TypeDoc docs generation. For Next.js-specific TS config/patterns, use web:developing-nextjs. For React component typing, use web:developing-react. For JavaScript language fundamentals, use web:developing-fullstack-javascript. For language-agnostic clean code, use writing-clean-code.

GitLab operations via the glab CLI (GitLab official command-line tool). Comprehensive reference for all command groups: auth (login, tokens, self-managed --hostname, CI job tokens, GITLAB_TOKEN), mr (create, list, diff, approve, merge, rebase, notes), issue and incident, ci/cd pipelines and jobs (run, trace, lint, retry, artifacts, schedules, variables), repo and project management (clone, fork, mirror, members), release (create, upload assets), api (REST and GraphQL), label, milestone, snippet, runner, securefile, ssh/gpg/deploy keys, and end-to-end workflows. Use when operating on GitLab from the terminal, automating GitLab tasks, or composing glab commands for merge requests, pipelines, issues, releases, or API calls. For GitHub operations use the gh CLI and the pull-request command instead; for commit message wording use writing-conventional-commits; for code review methodology use reviewing-code.

herdr operations via the herdr CLI — a terminal-native agent multiplexer over a local unix socket. Manage workspaces/tabs/panes; split, move, resize, navigate, and run commands; read pane/agent output (visible/recent/recent-unwrapped/detection); wait for output text (`pane wait-output`) or agent status (`agent wait --until` idle/working/blocked/done/unknown); spawn agents via `pane split` + `agent start --kind <kind> --pane <id>` and coordinate them (read/prompt/send-keys/wait/attach); manage Git worktrees (`herdr worktree`) and raw terminal streams (`herdr terminal`); install lifecycle integration for authoritative status where supported (Claude Code/Codex stay screen-manifest-based). Use when running inside herdr (HERDR_ENV=1) to orchestrate agents, dev servers, tests, or log streams across panes. Guarded: stop if HERDR_ENV is not 1. For parallel Tachikoma team orchestration inside Claude Code, use orchestrating-teams instead. For generic tmux multiplexing outside herdr this skill does not apply.

Codex CLI統合スキル（基本操作・プランレビュー・Agentオーケストレーション）。 Use when using Codex for code consultation, plan review, or orchestrating Codex agents in Wave parallel execution. Triggers: "codex", "codexと相談", "codexでレビュー", "codexでオーケストレーション"、`.codex/agents/*.toml` や config.toml の [agents] セクションの検出。 tachikoma-product-manager agent に計画策定を委譲する2フェーズ方式で、Wave計画に基づき自然言語でagentを同時起動し、config.tomlのmax_threadsで並列度を宣言的に制御する。 カスタムagent定義はTOML形式（name/description/developer_instructions/sandbox_mode）。 レビュー資材（計画書・PRD・設計書・監査レポート等）はMarkdown確定後にHTML併用生成する中央変換方式を採用（producer自身ではなくCodex本体がmd-to-html.shを実行）。 For agent conversion to Codex format, use converting-agents-to-codex instead. For Claude Code teammate orchestration (parallel sub-agents via Agent + run_in_background), use orchestrating-teams instead.

Agent Teamオーケストレーション（planner-first 2フェーズ方式: planner タチコマによる要件分析・docs/plan 作成 → ドメイン別専門タチコマの並列実装）。非herdr環境の Agent（run_in_background）/ SendMessage / TaskCreate・TaskList と、HERDR_ENV=1 環境の `herdr pane split` + `agent start --kind <kind> --pane <id>` / read / prompt / wait --until によるペイン起動・進捗管理、ファイル所有権パターンによる競合防止をカバー。 Use when multiple files need parallel changes with independent concerns (2+ files, frontend+backend, UI+API+test). Triggers: 複数ファイル並列変更、マルチ関心事開発、フロントエンド+バックエンド同時変更。 HERDR_ENV=1 では operating-herdr を必ず併用し、Claude Code の split-pane を使わず herdr CLI でエージェントを起動する。 レビュー資材（計画書・PRD・設計書・監査レポート等）はMarkdown確定後にHTML併用生成する中央変換方式を採用（producer自身ではなく本体がmd-to-html.shを実行）。 For Codex CLI orchestration (config.toml max_threads, agents/*.toml), use orchestrating-codex instead. For operating as a single Tachikoma implementation worker, use implementing-as-tachikoma instead.

Comprehensive SW engineering practices covering project foundations (fast feedback, small steps, DORA metrics), team organization (Team Topologies, 4 types), pair programming (4 patterns), developer habits (GREAT framework), IC effectiveness mindset (outcomes vs outputs, strategic prioritization), career-stage skills (junior to staff, IC vs management), cross-functional influence (PM/design, authority-free leadership), 20 anti-patterns (15 individual + 5 team-level), sustainable performance (burnout, remote work), and AI-enhanced workflows (daily AI, 90-day rollout). Use when starting projects, organizing teams, for IC effectiveness, career growth, or team AI culture.

Discipline for recognizing and cutting self-induced complexity to keep software sustainable and effective. Covers the orient-step-learn loop and a rubric for judging which option is simpler (easier to understand, fewer parts, reflects the problem); technology minimalism (dependency decision chains, minimal-and-stable framework choice, feature-as-liability restraint, incremental delivery, conservative tech adoption); personal automation and environment mastery (automate-first, day-zero deploy, terminal/shell/editor fluency, reproducible machine provisioning); collaboration for simplicity (asynchronous decoupled work, lean meetings, skill spreading, dialectical reasoning, empathy, listening to code, analogies); and data-driven simplification with error-revealing code layout. Use when code or workflows feel more complicated than warranted, when weighing a new dependency/framework/feature, when setting up developer tooling, or when untangling logic.

コードベースを解析し、Claude Code と Codex CLI 双方の自動化（hooks / subagents / skills / MCP servers / plugins）を推奨する読み取り専用スキル。検出したプラットフォーム （.claude/・CLAUDE.md / .codex/・AGENTS.md）を優先しつつ、各レコメンドに両クライアントの セットアップ手順を併記する。Use when the user asks for automation recommendations, wants to set up or optimize Claude Code or Codex for a project, asks what hooks / subagents / skills / MCP servers they should use, or mentions improving their AI coding assistant workflow. For authoring a specific plugin / skill / agent, use authoring-plugins instead. For converting an existing Claude agent to Codex, use converting-agents-to-codex.

Researches existing libraries before implementation to prevent reinventing the wheel. REQUIRED before writing any new functionality. Use when evaluating npm packages, pip packages, Go modules, or any third-party libraries to find and evaluate existing solutions. For generic web/company/people/paper search and search-tool routing, use searching-web instead.

コードレビューガイドライン（PR構成・効果的コメント・トーン原則・レビュアー/PR作成者の姿勢・チーム合意（TWA）・アンチパターン回避・CodeRabbit統合）。 Use when reviewing code, creating PRs, establishing team review agreements, or using CodeRabbit for automated review. 3原則（客観性・具体性・明確なアウトカム）、自動修正ループを含む。 For pair programming patterns and team organization, use practicing-software-engineering instead. For commit message format, use writing-conventional-commits instead. For generating PR descriptions, use the pull-request command. For GitLab MR operations via the glab CLI, use operating-gitlab instead.

Drives live Hunk diff-review sessions via the `hunk session *` CLI (Hunk is an interactive terminal diff viewer controlled through a local daemon). Inspect review focus and file/hunk structure, navigate to files/hunks/lines, reload session contents (diff/show), and add inline review comments (single `comment add` or batched `comment apply`). Use when the user has a Hunk session running — or wants to review a diff/changeset interactively — and asks you to walk them through changes, steer their view, or leave review notes. The Hunk TUI itself belongs to the user: never run interactive `hunk diff`/`hunk show` directly; only `hunk session` subcommands through the daemon. For code-review methodology (PR structure, comment tone, CodeRabbit), use reviewing-code instead. For GitLab MR operations via the glab CLI, use operating-gitlab instead. For GitHub PR descriptions, use the pull-request command instead.

High-performance file search via the fff MCP server (fff-mcp, frecency-ranked results with a persistent in-memory index). Covers the three fff tools — grep (file CONTENTS, the default), find_files (fuzzy file NAME search), multi_grep (OR across multiple literal patterns) — plus inline constraint syntax, core search rules (search bare identifiers, avoid regex, stop after 2 greps and read), DB persistence, and when to use fff vs serena / Glob / ripgrep. Use when searching code or files in a repository and the fff MCP server is available, or when deciding which file-search tool fits. For semantic symbol-level search/edit (definitions, references, rename) use serena; for plain path-pattern listing use Glob; for Claude Code plugin MCP configuration use authoring-plugins.

Web検索統合スキル。Exa MCP（第一優先）によるカテゴリ別高度検索（7カテゴリ: Company, News, Tweet/X, People, Personal Site, Financial Report, Research Paper）・コード検索（get_code_context_exa）・URL crawling（crawling_exa）・deep research と、gemini CLI（フォールバック・scripts/web-search.sh）による複雑なクエリ検索を統合。 Use when performing web searches, researching technologies, companies, people, academic papers, news, or social media. Load Exa MCP first; fall back to gemini CLI when Exa is unavailable. MUST load as first-choice search tool. context7 for library docs, deepwiki for GitHub wiki, WebSearch built-in as last resort. For evaluating npm/pip/Go libraries before implementation, use researching-libraries instead.

REQUIRED after all code implementations. Automatically load when implementation is complete to apply secure-code review with the software-security skill. Covers input validation, secrets management, OWASP top 10 countermeasures, authentication/authorization patterns, web penetration testing (reconnaissance, attack techniques, bypass methods), and serverless security (IAM/storage/functions across AWS/GCP/Azure, 17 threat categories, supply chain attacks). Use when implementing any code, handling external input, or developing serverless applications (Lambda, Cloud Run, Azure Functions). For dynamic authorization model design (ABAC/ReBAC/Cedar), use cloud:implementing-dynamic-authorization instead. For organizational AI development security strategy (trust frameworks, governance, AI-BOM), use ai:securing-ai-development instead. For security logging patterns and compliance logging, use cloud:implementing-observability.

Project CodeGuard-based secure-by-default coding rules (23 rule files spanning injection, authentication/MFA, cryptography, secrets, authorization, sessions, cloud/Kubernetes, IaC, supply chain, MCP, mobile, logging, privacy) for writing and reviewing secure code across 25+ languages. Use when writing, reviewing, or modifying code, handling user input/credentials/cryptographic operations, or configuring cloud infrastructure/CI-CD/containers. Japanese-localized adaptation of cosai-oasis/project-codeguard (CC-BY-4.0). For organizational AI-development security strategy use ai:securing-ai-development; for dynamic authorization model design (ABAC/ReBAC/Cedar) use cloud:implementing-dynamic-authorization; for OWASP-oriented code security and penetration testing use securing-code. For architecture-level security design methodology use cloud:architecting-security instead.

REQUIRED for all feature implementations. Automatically load when writing or reviewing tests. Enforces TDD with AAA pattern, actual/expected variables, and 100% coverage goal for business logic. Covers Vitest, Jest, Playwright, AI-augmented testing. Includes the Four Pillars of good tests (regression protection, refactoring resistance, fast feedback, maintainability), three testing styles (output/state/communication-based), code classification, test anti-patterns, test pyramid, mock patterns (stub/spy/MSW), integration testing, and testable design (DI). Covers ATDD/BDD/flaky-test elimination. For RTL patterns, use web:developing-react. For Web API testing, use web:developing-web-apis. For A/B testing, use product:practicing-product-management. For full-stack test strategy (exploratory/visual/perf/a11y/mobile), use testing-strategies. For team organization, use practicing-software-engineering. For Vitest 4.x specifics (migration, projects, Browser Mode, coverage config), use web:testing-with-vitest.

Full-stack test strategy guide spanning ten quality dimensions: manual exploratory testing (equivalence/boundary/state-transition/decision-table frameworks), functional automation strategy (micro/macro test types, contract testing), continuous testing (CI etiquette, DORA), data testing (DB/cache/batch/event streams, Testcontainers), visual testing, security testing (STRIDE, threat modeling, OWASP tools), performance testing (KPIs, load patterns, JMeter/Gatling, RAIL), accessibility testing (WCAG, axe), cross-functional requirements testing (chaos/architecture/compliance), mobile testing, and AI-assisted testing (agentic test authoring, MCP-based exploration). Use when designing test strategy across the stack, choosing test types beyond unit tests, or planning exploratory/visual/performance/accessibility/mobile testing. For TDD/unit methodology (AAA, four pillars, mocks), use testing-code. For Playwright E2E implementation, use web:testing-e2e-with-playwright. For code-level security, use securing-code.

A hands-on, iterative method for transforming a legacy codebase into a domain-driven, modular target architecture. Assess maturity with the Modularity Maturity Index (MMI, 0-10: modularity, hierarchy, pattern consistency), then run four strategic steps (domain rediscovery, target architecture modeling, actual-vs-target comparison, prioritized refactoring). Covers transformation approaches (big-bang vs step-by-step vs reshaping), technical stabilization (seams, characterization tests, design by contract), fighting model anemia, sociotechnical team reshaping, and a catalog of strategic/tactical/sociotechnical refactorings with motivation and mechanics. Use when modernizing a legacy monolith, planning incremental migration, or decomposing a big ball of mud into bounded contexts. For DDD/clean-architecture layering, use applying-clean-architecture. For portfolio modernization and migration patterns, use cloud:architecting-infrastructure. For code-level refactoring, use writing-clean-code.

Token-efficient structured development via the /serena command. Covers command syntax and option flags (-q quick, -c code-focus, -d detailed analysis, -r research, -s step-by-step, -t TODO list, --summary, --focus) and keyword-based auto-selection of thinking patterns (debugging, design, implementation, optimization). Use when implementing components, APIs, system designs, tests, bug fixes, or optimizations with a structured problem-solving approach. Available for Claude Code本体 and all Agents including Tachikoma. For parallel multi-file orchestration, use orchestrating-teams instead. For Tachikoma runtime operation, use implementing-as-tachikoma instead. For choosing between serena MCP symbol search and fff file search, use searching-files-with-fff instead.

REQUIRED for all code implementations. Automatically load when writing or reviewing any code. Covers SOLID principles, MAPPER principles with 25 code smell categories, practical refactoring, software design laws, the 4 rules of simple design, 66 code smell heuristics, formatting principles, dependency injection / composition roots, and boundary management patterns. Language-agnostic clean code guide. For language-specific practices, use lang:developing-go, mastering-typescript, or lang:developing-python. For architecture-level clean design, use applying-clean-architecture. For legacy code refactoring, see CHANGEABLE-CODE.md subfile. For refactoring mindset (spikes, feature toggles, structural discovery), see REFACTORING-MINDSET.md reference. For project-, environment-, and collaboration-level simplicity beyond code (technology minimalism, personal automation, cutting self-induced complexity), use pursuing-simplicity.

Conventional Commits 1.0.0仕様に基づくコミットメッセージフォーマットガイド。type/scope/BREAKING CHANGEの判定とSemVerとの対応を提供。 REQUIRED for all commit messages and version descriptions. Use when writing git commit messages. For version bump decisions and SemVer rules, use applying-semantic-versioning instead.

Unified writing and editorial-quality guide covering natural prose, evidence-grounded revision, AI-pattern diagnosis, citation integrity, output-artifact cleanup, prose fundamentals, technical and engineering docs, academic and university writing, Japanese prose, web and UX copy, FAQs, tech blogs, README files, Zenn publishing, persuasive documents, and sensory writing. REQUIRED for all text output: apply the integrity and natural-prose pass without claiming AI detector evasion. Use when writing, reviewing, proofreading, rewriting, translating, or creating documents such as articles, reports, slides, emails, papers, theses, requirements, design docs, business documents, help content, and developer documentation. For LaTeX → studio:writing-latex. For presentations → studio:creating-slides.