develop · coco-research — команда для Claude Code: что делает и как установить
Команда develop запускает многоагентный пайплайн разработки и исправления кода с разделением файлов между агентами и обязательной проверкой регрессий.
MIT ★ 328 обновлён 2026-10-29
.claude/commands/develop.mdЧто делает develop
Слэш-команда develop организует работу команды виртуальных специалистов в несколько слоёв. Сначала агенты уровня L1 исследуют репозиторий, зависимости и формируют карту изменений с помощью brownfield_map.py. Затем исполнители L2 берут в работу непересекающиеся группы файлов по строгим границам владения, исключая конфликты правок.
Перед передачей на ревью запускается цепочка гейтов: утилита run_gate.py фиксирует базовое состояние тестов и покрытие, а prove_red.py во временном рабочем дереве доказывает, что добавленные тесты действительно падают без внесённых правок.
Для работы пайплайна в окружении необходимы Python-скрипты по пути ~/.claude/skills/team-gate/scripts/ и режим bypassPermissions, иначе агенты не смогут редактировать файлы и прогонять проверки.
Какие задачи решает
- разделить крупную задачу на независимые файлы для разных субагентов
- доказать валидность новых тестов через откат правок во временном дереве
- провести фичу от исследования зависимостей до замеров покрытия кодовой базы
- предотвратить случайную перезапись файлов параллельными агентами L2
Примеры запросов агенту
- /team develop Добавь эндпоинт аутентификации пользователей через JWT
- /team develop Исправь гонку данных при обновлении баланса в обработчике платежей
Как установить develop
Вручную: положить файлы
.claude/commands/develop.mdВызов — /develop. Для всех проектов — ~/.claude/commands/
Вручную: положить файлы
~/.codex/prompts/develop.mdВызов — /prompts:develop. В Codex этот формат помечен устаревшим, замена — скиллы
Вручную: положить файлы
.github/prompts/develop.prompt.mdВручную: положить файлы
.gemini/commands/develop.tomlGemini хранит команды в TOML: текст запроса — в поле prompt
Вопросы о develop
Как установить develop в Claude Code?
Сохраните файл команды по пути .claude/commands/develop.md внутри вашего репозитория.
Что делает команда develop?
Она запускает многоэтапный цикл разработки, распределяя файлы по изолированным субагентам и подтверждая качество кода сквозными тестами.
Зачем скрипт запускает prove_red.py?
Утилита prove_red.py откатывает правки во временном дереве git worktree и проверяет, что новые тесты честно падают на исходном коде, защищая от ложноположительных проверок.
Исходный текст
develop.md · MIT
--- description: "Use when the user runs /team develop, or the team router picks action develop, to build or fix code through a multi-agent pipeline. Shards files across L2 agents, then reviews with captured regression evidence." --- # /team develop: Build Pipeline > Called by team.md router when action is `develop`. ## Role Selection Bias | Layer | Preferred Roles | Count | |-------|----------------|-------| | L1 | technical-analyst, security-analyst | 2 | | L2 | senior-backend-eng, senior-frontend-eng, senior-cloud-architect, qa-test-architect, senior-data-eng, sre-devops, mcp-integration, performance-eng (if perf-related) | 3-5 (domain-dependent) | | L3 | domain-accuracy, doc-quality, security-analyst (if api/backend/infrastructure), architecture-reviewer (if `.arch/index.json` exists) | 2-5 | | L4 | principal-architect | 1 | ### Domain-Specific Selection - `backend` only → senior-backend-eng + qa-test-architect + (senior-cloud-architect if `infrastructure`) - `frontend` only → senior-frontend-eng + qa-test-architect + (senior-ux-designer if `product`) - `backend` + `frontend` → senior-backend-eng + senior-frontend-eng + qa-test-architect - `data` → senior-data-eng + senior-backend-eng - `mobile` → senior-mobile-eng + qa-test-architect - `infrastructure` → senior-cloud-architect + sre-devops - `integrations` → mcp-integration + senior-backend-eng ## Stage 1: Brownfield Map ``` python3 ~/.claude/skills/team-gate/scripts/brownfield_map.py --for "<the feature or fix in words>" --repo-root . ``` exit 0: continue; read `.team-ship/BROWNFIELD-MAP.md` before reading any code. exit 1: not used; a map is not a pass or fail judgement. exit 2: UNVERIFIED; quote the gate file's summary (gates/handoff-1-map.json). No develop-pipeline aggregator exists to require this gate file; the map is the first step by instruction. ## Pipeline Customization ### Layer 1: Research L1 agents focus on the following areas. - File mapping for the feature scope - Existing patterns and conventions to follow - Dependencies and integration points - Security considerations for the feature ### Layer 2: Execution 1. Run `run_gate.py discover` then `run_gate.py run` for the baseline, before any Layer 2 agent commits. ``` python3 ~/.claude/skills/team-gate/scripts/run_gate.py --repo-root . discover python3 ~/.claude/skills/team-gate/scripts/run_gate.py --repo-root . run ``` exit 0: continue and quote each gate file's summary (gates/7-discover.json, gates/8.json), and record gates/8.json's head field now as the base for the red-green proof; that file is rewritten by the Regression Tests run. exit 1: BLOCK; for the command that stopped, quote its gate file's summary and loop back to Layer 2 execution. exit 2: UNVERIFIED; for the command that stopped, quote its gate file's summary; the suite is not confirmed passing until the gate passes or an override receipt names the gate. L2 agents receive file ownership boundaries (I10), covering the following areas. - Orchestrator divides scope into non-overlapping file groups - Each agent's prompt includes the following block. ``` YOUR FILES: [list] DO NOT TOUCH: [files owned by other agents] ``` - Agents commit atomically per feature/fix **Architecture context is advisory, not an allocation mechanism.** ``` python3 ~/.claude/skills/team-gate/scripts/arch_gate.py --repo-root . ``` exit 0: continue and quote `.arch/ARCH-GATE.json`; inline the component table from `.arch/INDEX.md` into each L2 prompt (max 25 lines) so agents know where each component's code lives. exit 1: BLOCK; a REMOVE verdict, quote the reason from `.arch/ARCH-GATE.json`. exit 2: UNVERIFIED; the index is absent (no .arch/index.json, which the script reports as not found), the pin is behind HEAD, or drift is unreconciled; omit the component-table block from each L2 prompt rather than quoting it. Do **not** derive `YOUR FILES` from the index. The index sits at C4-Container altitude while work is sharded at module altitude, and the two do not line up: in this repository a legal five-component index assigns `skills/`, 1,087 files, to a single component. A typical feature also spans three components at once. Scope-based splitting stays authoritative for ownership; the index only tells an agent which boundary it is working inside. **Toolkit integration:** - Check team:toolkit.md for "Code Implementation" entry - If Superpowers pipeline recommended → agent follows brainstorm → plan → execute pattern - If GSD active → agent reads `.planning/` context and follows phase conventions **Permission Mode:** `bypassPermissions`, agents need to create/edit files and run tests. ### Regression Tests Run after all Layer 2 agents complete, before Layer 3, per the Test Evidence Protocol (`team:evidence.md`). ``` python3 ~/.claude/skills/team-gate/scripts/run_gate.py --repo-root . discover python3 ~/.claude/skills/team-gate/scripts/run_gate.py --repo-root . run python3 ~/.claude/skills/team-gate/scripts/run_gate.py --repo-root . coverage ``` exit 0: continue and quote each gate file's summary (gates/7-discover.json, gates/8.json, gates/10.json). exit 1: BLOCK; for the command that stopped, quote its gate file's summary and loop back to Layer 2 execution. exit 2: UNVERIFIED; for `coverage`, no pytest-shaped command gives NOT_APPLICABLE, which still blocks unless an override receipt names gate "coverage"; for any other command that stopped, the suite is not confirmed passing until the gate passes or an override receipt names the gate. Provision what CI provisions. When something cannot be provisioned, record "override integration: <reason>" in your own words as the human, and run_gate's skip rule decides the rest. prove_red runs each test added since the base green at HEAD, reverts the implementation paths in a throwaway worktree and proves it goes red for the right reason; a never-red or invalid-red classification is BLOCK with the reason quoted from gates/9.json. ``` python3 ~/.claude/skills/team-gate/scripts/prove_red.py prove --base <the head recorded at the baseline step> --repo-root . python3 ~/.claude/skills/team-gate/scripts/prove_red.py recheck --repo-root . ``` exit 0: continue and quote each gate file's summary (gates/9.json, gates/9-recheck.json). exit 1: BLOCK; quote the reason from gates/9.json, or from gates/9-recheck.json for a changed or removed proved test, and loop back to Layer 2 execution. exit 2: UNVERIFIED; for the command that stopped, quote its gate file's summary; the suite is not confirmed passing until the gate passes or an override receipt names the gate. ``` python3 ~/.claude/skills/team-gate/scripts/run_gate.py --repo-root . evidence ``` exit 0: continue and quote `.team-ship/EVIDENCE.md`. exit 1: not used; evidence exits 0 or 2. exit 2: UNVERIFIED; the suite is not confirmed passing until the gate passes or an override receipt names the gate. Surface skips and failures explicitly in REVIEW-PACKAGE.md (with the skip count and any `UNVERIFIED` surface): never hide them inside a "passed" summary. ### Layer 3: Review L3 agents focus on the following areas. - Code correctness and edge cases (domain-accuracy) - Test coverage adequacy: confirmed by the Regression Tests gate above, via `EVIDENCE.md`, not by reading test descriptions - Architecture alignment with existing patterns - Security review of auth, data handling, API boundaries (security-analyst, if selected) ## GSD Integration (C4) When `.planning/` exists, the following applies. - Read ROADMAP.md for phase context - L2 agents follow GSD conventions: atomic commits, SUMMARY.md, STATE.md updates Output is compatible with `/gsd-verify-work`.
© coco-research · MIT · coco-research/coco
Похожие слэш-команды
Все слэш-команды по теме «Настройка агента: память, контекст, субагенты» →Сначала — develop от других авторов.