em-workflow
v0.1.82Unified SDD + parallel-implementation + dynamic-review workflow. /em-workflow:develop drives spec → design (conditional; fully-autonomous DESIGN.md + HTML mockups + design-system/tokens.yaml — implementers never invent design, refinement happens post-implementation via /em-workflow:design) → plan (task decomposition with complexity/domain tagging) → fully-parallel implementation in git worktrees driven as a work queue (background implementer launches with up to max_parallel_implementers in flight, notification-driven slot refill, machine-written journal.jsonl + a fail-open guard-hook net (full hook inventory: em-workflow/references/implement-phase.md); each implementer merges its own task via flock-serialized merge-task.sh, resolving conflicts by parent-side adoption; before touching any file an implementer records which tests were already failing, so a post-implementation failure is attributed by set difference rather than re-investigated after the fact, and it commits that baseline plus the AC → test mapping with the observed red per criterion to test-docs/{feature}/{task_id}.tests.yaml alongside the implementation) → dynamic review selection (rule floor + orchestrator discretion): each selected perspective dispatches ONE non-Claude primary reviewer taken from the front of that perspective's registry-defined chain (Codex and, when the separately-installed vertex-review plugin is present, Vertex AI MaaS via LiteLLM), falling back to the Claude generic reviewer only when every chain entry for that perspective is unavailable; once every perspective's result is final, a single Opus evaluator subagent evaluates the whole round and the orchestrator decides the next action (auto-fix / another round / rework / completion) → bounded auto-fix → retrospect collection. Every workflow artifact (workflow.yaml, REQUIREMENTS.md/SPEC.md/DESIGN.md/IMPLEMENTATION.md/VERIFICATION.md, tasks/, feature-docs/{feature}/reviews/roundN.yaml, retrospect.yaml) lives exclusively inside a per-feature integration worktree/branch (em-workflow/{feature}/integration) — each update is committed there immediately via commit-docs.sh, so the main working tree stays clean until the final merge and a crashed/interrupted run resumes by enumerating integration branches rather than scanning the main tree. workflow.yaml-sourced shell commands are user-approved once per project (approval store outside the repo) and enforced deterministically by a bundled PreToolUse hook. Develop start (Step 0) runs a git-setup gate: it probes gitleaks availability and idempotently installs a gitleaks pre-commit hook, aborting the workflow when gitleaks is absent. License consistency is a cross-workflow constraint: create-spec records the project license (SPDX id) into workflow.yaml, create-plan checks new library choices against it (conflict → replace the library or relicense, user's choice), and the review phase adds a license perspective whenever the diff touches dependency manifests. Every phase's dialogue and file writes are separated by role: the orchestrator is the sole writer of workflow.yaml and the sole caller of AskUserQuestion, while a requirements-analyst investigates and a spec-writer authors REQUIREMENTS.md/SPEC.md for create-spec, an autonomous designer worker owns the design step, an implementation-planner proposes IMPLEMENTATION.md/tasks/VERIFICATION.md, and a rework-planner plans review/verify-triggered follow-up tasks — every worker returns a structured envelope (question packets or workflow patches) instead of writing shared state directly. A --batch flag runs develop unattended for externally-triggered headless launches: every AskUserQuestion gate resolves mechanically — a gate carried in a worker's question packet resolves per references/question-resolution.md's batch resolution sequence against references/batch-policies.yaml (spec ambiguities settled via a bounded Codex consultation loop, command approvals auto-recorded with refusal patterns still hard-failing), while a gate that never passes through a question packet (git-setup failure, feature selection, the review diff-size gate, capped auto-rework for review residuals and verify failures, keep-branch completion, and the rest) resolves per references/batch-mode.md's Non-packet gates table — batch never merges, and opens a PR only when the launch passes --pr: by default it removes the integration worktree and leaves the branch for the human to merge or PR from the main tree; failures still stop with a report. Also provides /em-workflow:review for standalone reviews, /em-workflow:retrospect for cross-feature lesson mining, /em-workflow:design for interactive post-implementation design refinement, /em-workflow:git-setup for standalone idempotent git local-config setup, and /em-workflow:gen-license for dependency-license analysis + LICENSE generation/relicensing.
By em0 GitHub starsUpdated last week
Directory evidence
- Runtimes
- Claude Code
- Parsed components
- 21 skill or MCP entries
- Source updated
- Sep 15, 2026
- Manifest status
- Canonical path parsed
The directory validates manifest shape and source location. It does not execute the plugin or provide a security endorsement. Review the indexing methodology →
Install em-workflow for Claude Code
claude plugin marketplace add IchenDEV/agent-plugin-mkt
claude plugin marketplace update agent-plugin-marketplace
claude plugin install em-workflow@agent-plugin-marketplacePaste and run these commands in a terminal with Claude Code. They add and refresh the PluginsMP catalog, then install this plugin.
The installer fetches third-party code from the source repository shown on this page. This directory validates manifest structure and source location, but does not perform a security audit; review the manifest, components, and source before installing.
Get the source manually
git clone https://github.com/m-m-n/em-claude-pluginsClone the source repository, then follow its setup instructions to add the plugin to a compatible client. The plugin root is em-workflow/.
Plugin files
├── .claude-plugin/plugin.json├── skills/backend-impl/SKILL.md├── skills/codex-prompting/SKILL.md├── skills/contributor-consent/SKILL.md├── skills/design/SKILL.md├── skills/design-impl/SKILL.md├── skills/develop/SKILL.md├── skills/frontend-impl/SKILL.md├── skills/gen-license/SKILL.md├── skills/git-setup/SKILL.md├── skills/infra-impl/SKILL.md├── skills/plan-writing/SKILL.md├── skills/retrospect/SKILL.md├── skills/review/SKILL.md├── skills/review-architecture/SKILL.md├── skills/review-comprehensive/SKILL.md├── skills/review-license/SKILL.md├── skills/review-performance/SKILL.md├── skills/review-security/SKILL.md├── skills/review-spec/SKILL.md├── skills/tdd-testing/SKILL.md└── skills/worktree-task-workflow/SKILL.md
Included Skills21
バックエンド実装の知識(em-workflow implementer 動的注入用)。API 設計、エラーハンドリング、トランザクション境界、ユニット寄りのテスト戦略、品質チェックリストと落とし穴を提供します。タスクの skills に backend-impl が指定されたときに implementer がロードします。
Codex CLI へ送るプロンプトの構造化規則(em-workflow codex-reviewer 静的プリロード用)。GPT-5 系プロンプティング知見のエッセンス — XML タグによるブロック分類と組み合わせ規則、レビュー用途の推奨 4 ブロック(task / structured_output_contract / grounding_rules / dig_deeper_nudge)を定義します。
muse-spark contributor ティアへの同意を記録・撤回する唯一のインタラクティブな経路。LiteLLM ハーネスの可用性を確認したうえで、リポジトリの可視性・ライセンス・コントリビューター数の 3 事実を提示し、状態に応じた質問を 1 回だけ行う
デザイン詰めの手動対話コマンド(em-workflow)。実機確認後のフィードバックや実機スクリーンショットをもとに、design-system/tokens.yaml・feature-docs/{feature}/DESIGN.md・HTML モックを合意ループで更新します。コードには触れません — 実装への反映は /em-workflow:develop に委ねます。引数で feature を指定、引数なしは対象を確認します(デフォルト: システム全体)
デザイン実装の知識(em-workflow implementer 動的注入用)。UI デザイン・スタイリング・レイアウトの原則、アクセシビリティ、E2E 寄りのテスト戦略、品質チェックリストと落とし穴を提供します。タスクの skills に design-impl が指定されたときに implementer がロードします。
em-workflow の統合開発エントリポイント。SDD(spec → plan + タスク分割)から worktree 並列実装、動的レビュー、統合検証、retrospect 収集までを workflow.yaml の状態だけを根拠に自走させるステートマシン。軽い変更もタスク1個として同じフローを通します
フロントエンド実装の知識(em-workflow implementer 動的注入用)。状態管理、コンポーネント設計、非同期処理の作法、ユニット+コンポーネントテスト戦略、品質チェックリストと落とし穴を提供します。タスクの skills に frontend-impl が指定されたときに implementer がロードします。
プロジェクトの依存関係ライセンスを分析し、互換性のあるライセンスを選定して LICENSE ファイルを生成・変更する。既存 LICENSE の変更(relicense)にも使う
git リポジトリのローカル設定を冪等にセットアップする。現在の設定項目は gitleaks pre-commit hook の設置。設定済みの項目は何もしない
インフラ・設定実装の知識(em-workflow implementer 動的注入用)。CI・ビルドスクリプト・設定・環境配線の作法、検証可能性を軸にしたテスト戦略、品質チェックリストと落とし穴を提供します。タスクの skills に infra-impl が指定されたときに implementer がロードします。
実装計画・タスク分割の執筆規則(em-workflow implementation-planner 静的プリロード用)。コード禁止規則、薄い IMPLEMENTATION.md と VERIFICATION.md のテンプレート、タスク分割規則(worktree 独立性 / files 予測 / インターフェース契約)、complexity・domains の判定基準、保存前セルフチェックリストを定義します。
自己改善ループの手動判断コマンド(em-workflow)。溜まった feature-docs/*/retrospect.yaml を横断分析し、再発性のある教訓候補を抽出して帰属先(feature-docs/LESSONS.md の audience セクション / プロジェクト CLAUDE.md / プラグイン改善候補=報告のみ)に分類、既存記述との重複チェックとユーザー承認を経てプロジェクト側に追記します。プラグイン本体のファイルには書き込みません。承認なしの自動追記は絶対にしません
単体レビューのエントリポイント(em-workflow)。SDD を通さない日常レビューを吸収します。workflow.yaml 不在時は baseline(comprehensive + SPEC.md があれば spec)+ 裁量層の追加観点で動作し、選択された各観点は primary_chain 先頭の利用可能な非 Claude レビュアー(Codex、および別途 vertex-review プラグインが導入済みなら LiteLLM 経由の Vertex AI / Muse)を 1 体だけ起動、チェーン全滅時のみ Claude 汎用レビュアーがフォールバックします。全観点確定後は Opus 評価者が 1 体、ラウンドを評価します。その上で bounded auto-fix(≤ 3 ループ、--report-only でスキップ)とレビュー記録の書き出しまで行います。コミットは一切しません
アーキテクチャ観点のレビュー知識(em-workflow 動的注入用)。レイヤー違反・循環依存・神クラス・破壊的インターフェース変更・重大な SOLID 違反を検出する基準を汎用レビュアーに与えます。オーケストレーター指示以外で自発的にロードするものではありません。
包括観点のレビュー知識(em-workflow 動的注入用・常時実行 baseline)。単一観点レビュアーが見逃す横断バグ・正確性/ロジックバグ・統合と契約のドリフト・境界値・テスト齟齬を検出する基準を汎用レビュアーに与えます。オーケストレーター指示以外で自発的にロードするものではありません。
ライセンス整合観点のレビュー知識(em-workflow 動的注入用)。diff が依存マニフェストや vendored コードに触れたとき裁量層で追加され、新規依存のライセンスが project.license と両立するかを references/license-compat.md の基準で検査します。オーケストレーター指示以外で自発的にロードするものではありません。
パフォーマンス観点のレビュー知識(em-workflow 動的注入用)。N+1 クエリ・計算量爆発・メモリ非効率・ブロッキング I/O・リソースリークを検出する基準を汎用レビュアーに与えます。オーケストレーター指示以外で自発的にロードするものではありません。
セキュリティ観点のレビュー知識(em-workflow 動的注入用)。汎用レビュアーが Skill tool でロードし、インジェクション・認証認可バイパス・機密データ露出・暗号の弱点・入力検証欠如を検出する基準を得ます。レビュアーエージェントのオーケストレーター指示以外で自発的にロードするものではありません。
仕様準拠観点のレビュー知識(em-workflow 動的注入用)。SPEC.md との矛盾・必須動作の欠落・クリティカルロジックの誤実装・データ整合性違反を検出する基準を汎用レビュアーに与えます。SPEC.md 不在時はレビュアーがクリーンにスキップします。オーケストレーター指示以外で自発的にロードするものではありません。
TDD 規律(em-workflow implementer 静的プリロード用)。テストファーストの手順、受け入れ条件からテストへの翻訳方法、良いテストの性質、テストを消さない・弱めない規律、TDD が適用しにくいタスクでの判断基準を定義します。レイヤー固有のテスト戦略は各実装スキル側にあり、ここには共通規律のみを置きます。
worktree 内タスク作業の規律(em-workflow implementer 静的プリロード用)。worktree 境界、コミット規約、merge-task.sh の呼び出し方と exit code 分岐、コンフリクト時の親側採用プロトコル、workflow.yaml のコマンド文字列に対する承認ゲートを定義します。
Plugin manifests1
{
"name": "em-workflow",
"description": "Unified SDD + parallel-implementation + dynamic-review workflow. /em-workflow:develop drives spec → design (conditional; fully-autonomous DESIGN.md + HTML mockups + design-system/tokens.yaml — implementers never invent design, refinement happens post-implementation via /em-workflow:design) → plan (task decomposition with complexity/domain tagging) → fully-parallel implementation in git worktrees driven as a work queue (background implementer launches with up to max_parallel_implementers in flight, notification-driven slot refill, machine-written journal.jsonl + a fail-open guard-hook net (full hook inventory: em-workflow/references/implement-phase.md); each implementer merges its own task via flock-serialized merge-task.sh, resolving conflicts by parent-side adoption; before touching any file an implementer records which tests were already failing, so a post-implementation failure is attributed by set difference rather than re-investigated after the fact, and it commits that baseline plus the AC → test mapping with the observed red per criterion to test-docs/{feature}/{task_id}.tests.yaml alongside the implementation) → dynamic review selection (rule floor + orchestrator discretion): each selected perspective dispatches ONE non-Claude primary reviewer taken from the front of that perspective's registry-defined chain (Codex and, when the separately-installed vertex-review plugin is present, Vertex AI MaaS via LiteLLM), falling back to the Claude generic reviewer only when every chain entry for that perspective is unavailable; once every perspective's result is final, a single Opus evaluator subagent evaluates the whole round and the orchestrator decides the next action (auto-fix / another round / rework / completion) → bounded auto-fix → retrospect collection. Every workflow artifact (workflow.yaml, REQUIREMENTS.md/SPEC.md/DESIGN.md/IMPLEMENTATION.md/VERIFICATION.md, tasks/, feature-docs/{feature}/reviews/roundN.yaml, retrospect.yaml) lives exclusively inside a per-feature integration worktree/branch (em-workflow/{feature}/integration) — each update is committed there immediately via commit-docs.sh, so the main working tree stays clean until the final merge and a crashed/interrupted run resumes by enumerating integration branches rather than scanning the main tree. workflow.yaml-sourced shell commands are user-approved once per project (approval store outside the repo) and enforced deterministically by a bundled PreToolUse hook. Develop start (Step 0) runs a git-setup gate: it probes gitleaks availability and idempotently installs a gitleaks pre-commit hook, aborting the workflow when gitleaks is absent. License consistency is a cross-workflow constraint: create-spec records the project license (SPDX id) into workflow.yaml, create-plan checks new library choices against it (conflict → replace the library or relicense, user's choice), and the review phase adds a license perspective whenever the diff touches dependency manifests. Every phase's dialogue and file writes are separated by role: the orchestrator is the sole writer of workflow.yaml and the sole caller of AskUserQuestion, while a requirements-analyst investigates and a spec-writer authors REQUIREMENTS.md/SPEC.md for create-spec, an autonomous designer worker owns the design step, an implementation-planner proposes IMPLEMENTATION.md/tasks/VERIFICATION.md, and a rework-planner plans review/verify-triggered follow-up tasks — every worker returns a structured envelope (question packets or workflow patches) instead of writing shared state directly. A --batch flag runs develop unattended for externally-triggered headless launches: every AskUserQuestion gate resolves mechanically — a gate carried in a worker's question packet resolves per references/question-resolution.md's batch resolution sequence against references/batch-policies.yaml (spec ambiguities settled via a bounded Codex consultation loop, command approvals auto-recorded with refusal patterns still hard-failing), while a gate that never passes through a question packet (git-setup failure, feature selection, the review diff-size gate, capped auto-rework for review residuals and verify failures, keep-branch completion, and the rest) resolves per references/batch-mode.md's Non-packet gates table — batch never merges, and opens a PR only when the launch passes --pr: by default it removes the integration worktree and leaves the branch for the human to merge or PR from the main tree; failures still stop with a report. Also provides /em-workflow:review for standalone reviews, /em-workflow:retrospect for cross-feature lesson mining, /em-workflow:design for interactive post-implementation design refinement, /em-workflow:git-setup for standalone idempotent git local-config setup, and /em-workflow:gen-license for dependency-license analysis + LICENSE generation/relicensing.",
"author": {
"name": "em"
},
"version": "0.1.82"
}For maintainers
If you maintain this plugin, link to this source-backed listing from your README so users can review its manifest and indexed components.
[em-workflow on Agent Plugins Marketplace](https://pluginsmp.com/plugins/em-workflow)