troupe-mission — скилл для Claude Code: что делает и как установить
Скилл troupe-mission для Claude Code координирует параллельную работу субагентов в отдельных окнах tmux и ветках git. Сессия пишет план, отдаёт задачи ролям и проверяет код перед слиянием.
MIT ★ 140 обновлён 2026-11-13
npx skills add hitianqi/troupe --skill troupe-mission -a claude-codeЧто делает troupe-mission
Этот скилл превращает текущую сессию в оркестратора задач. Он создаёт рабочее дерево git на ветке mission/<mission> через команду troupe mission new <mission> и формирует шаблон плана по пути docs/missions/<mission>/plan.md. До утверждения плана пользователем через гейт в панели troupe роли не запускаются.
Каждая роль получает собственное окно tmux, ветку и дерево git. Сборщики работают параллельно на агентах Claude Code или Codex, если не пересекаются по файлам, а роль рецензента по умолчанию проверяет каждую готовую ветку. Сессия не пишет код сама, а отправляет инструкции через troupe send и ожидает системных ответов вместо ручного сна.
Для запуска процесса в системе должны быть установлены утилита troupe CLI и tmux. Скилл рассчитан на крупные составные изменения и не подходит для правок в одну сессию.
Какие задачи решает
- разбить объёмную задачу на изолированные роли в отдельных ветках git
- запустить независимые части разработки параллельно в окнах tmux
- согласовать архитектурный план через интерактивную панель troupe
- организовать ревью изменений выделенным агентом перед слиянием
Примеры запросов агенту
- Запусти эту задачу как mission через troupe
- Разбей рефакторинг на роли и подготовь план в docs/missions/
- Открой роли для параллельной разработки сервиса
Как установить troupe-mission
1Через Skills CLI (Vercel)
npx skills add hitianqi/troupe --skill troupe-mission -a claude-code2Вручную: положить файлы
~/.claude/skills/troupe-mission/SKILL.mdПапка скилла целиком; для одного проекта — .claude/skills/troupe-mission/
1Через Skills CLI (Vercel)
npx skills add hitianqi/troupe --skill troupe-mission -a codex2Вручную: положить файлы
.agents/skills/troupe-mission/SKILL.mdДля всех проектов — ~/.agents/skills/
1Через Skills CLI (Vercel)
npx skills add hitianqi/troupe --skill troupe-mission -a cursor2Вручную: положить файлы
.cursor/skills/troupe-mission/SKILL.mdCursor читает и .agents/skills/, и .claude/skills/
1GitHub CLI
gh skill install hitianqi/troupe skills/troupe-mission2Вручную: положить файлы
.github/skills/troupe-mission/SKILL.md1Через Skills CLI (Vercel)
npx skills add hitianqi/troupe --skill troupe-mission -a gemini-cli2Командой платформы
gemini skills install https://github.com/hitianqi/troupe --path skills/troupe-missionВручную: положить файлы
.devin/skills/troupe-mission/SKILL.mdСтарый путь .windsurf/skills/ тоже читается
Claude.ai
Упакуйте папку в zip и загрузите в Settings → Capabilities → Skills
По шагам, с командами для каждой платформы: Как установить скилл в Claude Code: пять способов
Вопросы о troupe-mission
Как установить troupe-mission в Claude Code?
Выполните команду npx skills add hitianqi/troupe --skill troupe-mission -a claude-code. Также файл скилла можно сохранить вручную по пути ~/.claude/skills/troupe-mission/SKILL.md.
Что делает скилл troupe-mission?
Он превращает текущую сессию в оркестратора, который создаёт план изменений в docs/missions/, запрашивает одобрение и распределяет задачи по отдельным веткам и окнам tmux.
Что нужно для работы troupe-mission?
Требуются установленная утилита troupe CLI, запущенный tmux и доступные агенты Claude Code или Codex для выполнения ролей.
Исходный текст
SKILL.md · MIT
--- name: troupe-mission description: Run a multi-step software change as a troupe mission, with this session as the orchestrator. You write a plan with the person, they approve it in the troupe panel, and each role then works in a real tmux window of its own (Claude Code or Codex), in its own git worktree and branch, in parallel where the files allow; a reviewer role checks each branch as it lands and you merge, then open the MR after the person's yes. Use it whenever the person wants work split across several agents or windows, says "troupe", "mission", "run this as a mission", "split this into roles", "open the roles", "dispatch", "开几个 agent 并行做", or asks to plan, approve, build, review or land a mission under docs/missions/, in Claude Code or in Codex. Not for a change one session can just make, and not for changing troupe itself. --- # Mission You are the orchestrator in the person's current agent session. The person watches the panel beside you and the roles in the `roles-<mission>` window, and can talk to any role directly. You plan, dispatch, review the results, merge, and stop at the person's gates. You do not write the roles' code. The `troupe` CLI enforces what keeps the work correct: no role opens before the plan, as it is now, is approved; one role is one worktree, one branch and one pane; nothing is typed into a window that is not ready; each role's `flags` pin its permission. When a command refuses, it says why: fix the cause, do not retry as is or work around it. Never type into a role window yourself (no `tmux send-keys`); use `troupe send`. You don't wait on a timer. Once you have dispatched, sent work or opened a gate, end your turn: when a role finishes, the person answers a gate in the panel, or the person answers a role waiting on them in its window, troupe types one line into your session (`Role api finished: done: …`, `mr-ok approved in the panel. …`, `Role api was answered in its window and went on working.`, `Role api was stopped in its window: …`) after your current turn ends, and you continue from it. It knows your session from the troupe commands you run, so run them from this session. Never `sleep` to wait. `troupe gate --no-wait` opens a gate without blocking; `troupe wait` is for when you need an answer within the same turn. How far to read, how to split, what to ask and how long your messages are is your judgment; the steps below are the order things happen in. ## 1. Brief Run `troupe open --planning` first, in the foreground (harmless if already open; never `troupe up`, which starts a second orchestrator); pass its attach line to the person when it gives one. Find out what the person wants, why, what counts as done, and what is out of scope. Read the code it touches. A question the code cannot answer becomes a decision in the plan, with your recommendation; ask now only what blocks the plan itself. The plan is shown in the panel, not in the chat. Do not lay out an approach, a role split or open decisions in the chat and ask whether to write them up, even when the person asks what you think: write them into the plan and let the panel show it. The chat carries only a blocking question, or a few lines at the gate. ## 2. Plan Run `troupe mission new <mission>` (add `--base <ref>` when the mission should not start from the checkout's current branch). It creates the mission's own worktree on branch `mission/<mission>`, prints its path, and writes the plan template to `docs/missions/<mission>/plan.md` there with `base:` already set to the mission branch: each role's worktree is cut from it at dispatch, so a role sees what you merged before it. It also binds this session's pane to the mission, so the commands below act on it without `-m`, and other missions in the same repository can run beside it. The person's checkout stays on its branch. Fill in the plan in the mission worktree, by its full path. When its output names the language the person reads, write in it as it says; if the person switches the panel's language later, troupe tells you in a line. - **A role is a window.** Name roles for what they do (`api`, `web`, `review`). Split by coupling, not by file: files that change together go to one role, and a hub file (shared types, config, a registry) is touched by one role only. Where two roles would share a file, first try to remove the sharing (split the file); only then order them with `depends_on`. Fewer, fuller roles beat many thin ones: each is a window the person watches. - **Each role says who runs it**: `agent` (`claude`, `codex`, or an adapter in `.troupe/agents/`), `model`, `effort`, and `flags`, which must pin its permission (Claude Code: `--permission-mode acceptEdits`; Codex: `-a on-request -s workspace-write`). Propose; the person decides at the gate. Propose a bypass mode only when the person asks for one. For Codex `workspace-write`, dispatch adds `--add-dir` for its Git index, shared objects and the mission's refs/reflogs. This permits commits, not writes to the main branch; roles still share the mission refs and objects. - **Each role's text** is its `## Role: <name>` section: what to do, what not to change, how to prove it works. Builders also take `text: workflow:builder`. - **One reviewer role by default**: `touches: []`, `reviews: [...]`, `text: workflow:reviewer`, on a different model from the builders where the person agrees. It reviews each branch as it lands. - A change to how something looks needs a done-when criterion that the person has seen and approved it. - **The top of the plan is what the person reads at the gate**, and the panel shows each done-when item on one line. Keep the intent to a few sentences and each done-when item to one checkable line, what and not how, usually three to six of them; the how goes in the role sections. Run `troupe check` until it is clean. Commit the plan in the mission worktree (`git -C <worktree> add` and `commit`). ## 3. Gate: plan OK? Tell the person in a few plain lines what they are approving, the decisions with your recommendations, and who runs each role. Then run `troupe gate plan "<one line>"`. The panel shows the plan gate: `a` approves (and takes your recommendations), `c` sends their words back to you as `change: <words>`, which sends you back to step 2: edit, check, commit, gate again. Any later edit of `plan.md` or of a role text it references needs approval again before the next dispatch. ## 4. Build `troupe stage build`, then `troupe dispatch <role>` for every builder whose `depends_on` are done; independent roles run at once, then end your turn; each role's finish arrives as a line from troupe. A role stuck on a dialog (a permission, a question) shows in the panel as waiting on the person; tell them which window if it stays, then end your turn: troupe tells you once they answer a Claude role there (a Codex role's answer shows only in `troupe status`). `troupe status` shows every role. When a builder's result says `done`, read `.troupe/result.md` in its worktree and check `troupe status`: files changed outside its `touches` fail it (send it back, or change the plan and get it approved again). Then it goes to review. `blocked` or `failed`: read why, then answer it with `troupe send`, change the plan, or ask the person. For startup readiness or before sending follow-up work, run `troupe wait --role <role> --ready --timeout 60`; a dialog or gone pane ends the wait early so you can act on it. ## 5. Review, fix, merge `troupe stage review` once the first branch lands. Dispatch the reviewer, then give it each branch in the order they land: `troupe send review "Review role/<mission>/<role>: …"`. `troupe send` puts the role's old result aside first, so its next result is news again. - PASS: merge the role's branch onto the mission branch in the mission worktree (`git -C <worktree> merge …`; one commit per role, subject `<role>: <what it does>`), then dispatch the roles that just became ready. - Act-on findings: `troupe send <builder> "<the findings, or the review file's path>"` to the same window, on the same branch; then review again. **One fix round, then you rule**: merge it, or hold it and tell the person why. No third review. ## 6. Gate: MR OK? When every role is merged: `troupe stage mr-ok`, then `troupe gate mr-ok "<what the MR delivers; open findings with your recommendation>"`. Nothing is pushed before `approve`. A change sends you to the fix loop or the plan. ## 7. Open the MR, then gate: merged? `troupe stage mr`; push the mission branch (`git -C <worktree> push`) and open one MR against the base branch, per the project's conventions. A refused push: tell the person and stop. Then `troupe stage merged` and `troupe gate merged "<the MR link>"`, and leave every role window open: only a person merges the MR, and until they do, a change they ask for goes to the same builder's window, then is pushed to the same MR. Opening the MR does not finish the mission. ## 8. Teardown once merged On `approve` at `merged`: `troupe stage teardown -m <mission>`, then `troupe teardown -m <mission>`; only then report the cleanup result. Teardown refuses while any gate of the workflow is not approved, so it cannot run before the merge. Teardown removes mission-owned role panes, worktrees, branches and temporary briefs/results, and the mission worktree, while keeping the mission branch (the MR head) and the log. It closes this mission's panel and unbinds your pane; your session stays open. Dirty or untracked files, unmerged commits and an unpushed mission branch refuse cleanup; resolve them before retrying, or tell the person cleanup is blocked and ask before discarding them with `--force`. Repeating teardown with explicit `-m` is safe. ## Telling the person They hear from you at the gates and whenever something blocks: a refusal you cannot fix, a role waiting on them, a role held after its fix round, a plan change that needs approval again. Their language, short, plain words, your recommendation. No step names or logs.
© hitianqi · MIT · hitianqi/troupe