ns-autonomous — скилл для Claude Code: автономное выполнение задач
Скилл ns-autonomous для Claude Code, Codex, Cursor, Copilot и Gemini CLI, который берёт локальный план или произвольный текст и выполняет его автономно в изолированном worktree.
Apache-2.0 152 установок обновлён 2026-10-24
npx skills add nextstage-brasil/skills --skill ns-autonomous -a claude-codeЧто делает ns-autonomous
ns-autonomous работает на Claude Code, Codex, Cursor, GitHub Copilot и Gemini CLI. В оба режима — самостоятельный запуск и фаза 2 скилла ns-execution-gitlab-issue — встроен один и тот же движок: планирование глубины задачи, протокол сомнений и мультиагентная диспетчеризация.
Перед любым редактированием кода скилл создаёт изолированный worktree .worktrees/{version_san}/ с веткой work/{version_san}. Писать на основной ветке или на базовой он не будет — если создать worktree не удалось, выполнение прерывается. При двух и более рабочих единицах каждую отдаёт отдельному субагенту через ns-coder; параллельный запуск — только между единицами без зависимостей в DAG и с непересекающимися файлами. Одна единица выполняется в текущей сессии.
Внутренний цикл ревью работает через ns-reviewer и ns-judge: код считается принятым только при статусе Approved. Деструктивные сомнения в автономном режиме останавливают выполнение и выносятся в чат; в режиме фазы 2 — возвращаются вызывающему скиллу. Зависимости: ns-harness, ns-reviewer, ns-judge версии 1.5.
Какие задачи решает
- Выполнить локальный план автономно без GitLab-issue
- Распределить многошаговую задачу между параллельными субагентами
- Изолировать все изменения в отдельном worktree, не трогая основную ветку
- Прогнать цикл ревью и judge-закрытие внутри одной сессии
- Запустить фазу 2 из
ns-execution-gitlab-issueкак движок выполнения
Примеры запросов агенту
- Запусти
ns-autonomousна основеplan.mdи выполни всё автономно - Прочитай этот текст задачи и реализуй его через
ns-autonomousв изолированной ветке - Используй
ns-autonomousкак движок фазы 2 для текущего GitLab-issue
Как установить ns-autonomous
1Через Skills CLI (Vercel)
npx skills add nextstage-brasil/skills --skill ns-autonomous -a claude-code2Вручную: положить файлы
~/.claude/skills/ns-autonomous/SKILL.mdПапка скилла целиком; для одного проекта — .claude/skills/ns-autonomous/
1Через Skills CLI (Vercel)
npx skills add nextstage-brasil/skills --skill ns-autonomous -a codex2Вручную: положить файлы
.agents/skills/ns-autonomous/SKILL.mdДля всех проектов — ~/.agents/skills/
1Через Skills CLI (Vercel)
npx skills add nextstage-brasil/skills --skill ns-autonomous -a cursor2Вручную: положить файлы
.cursor/skills/ns-autonomous/SKILL.mdCursor читает и .agents/skills/, и .claude/skills/
1GitHub CLI
gh skill install nextstage-brasil/skills skills/ns-autonomous2Вручную: положить файлы
.github/skills/ns-autonomous/SKILL.md1Через Skills CLI (Vercel)
npx skills add nextstage-brasil/skills --skill ns-autonomous -a gemini-cli2Командой платформы
gemini skills install https://github.com/nextstage-brasil/skills --path skills/ns-autonomousВручную: положить файлы
.devin/skills/ns-autonomous/SKILL.mdСтарый путь .windsurf/skills/ тоже читается
Через Skills CLI (Vercel)
npx skills add nextstage-brasil/skills --skill ns-autonomous -a opencodeClaude.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 ns-autonomous
Как установить ns-autonomous в Claude Code?
Выполните npx skills add nextstage-brasil/skills --skill ns-autonomous -a claude-code, затем скилл будет доступен по пути ~/.claude/skills/ns-autonomous/SKILL.md.
Что делает ns-autonomous?
Скилл берёт локальный план или произвольный текст и выполняет его автономно: создаёт изолированный worktree, при необходимости распределяет задачи между субагентами, прогоняет ревью через ns-reviewer и закрывает сессию через ns-judge.
Чем ns-autonomous отличается от ns-coder и когда его не нужно запускать напрямую?
ns-coder предназначен для одношаговых правок; ns-autonomous берёт на себя планирование и многоагентное выполнение. Если задача приходит из GitLab-issue, запускать ns-autonomous напрямую не следует — он должен стартовать только как фаза 2 скилла ns-execution-gitlab-issue.
Исходный текст
SKILL.md · Apache-2.0
---
name: ns-autonomous
description: (NS) Autonomous execution engine — plans depth, resolves doubts from docs/specs, multi-agent work in an isolated worktree. Use for a local/pasted plan or "run this autonomously" without GitLab. Also Phase 2 under ns-execution-gitlab-issue (do not run standalone then). Not for single-step edits (ns-coder) or GitLab issues directly (ns-execution-gitlab-issue).
license: Apache-2.0
metadata:
author: nextstage-brasil
version: "1.5"
depends:
- ns-harness
- ns-reviewer
- ns-judge
---
# Code Autonomous
Harness-aware execution engine: planning depth, doubt resolution, multi-agent dispatch, review gate — standalone pipeline or caller's Phase 2 engine.
## Workflow mode (mandatory)
Fixed workflow — named skill handoffs or harness bridges (`../../ns-harness/references/subagent-dispatch.md` spawn gate). Review gate: `../ns-reviewer/references/review-gate-workflow.md`. **MUST** `reviewer-agent` when available (else `ns-reviewer`); then `ns-judge` in-session only if `Code Review: Approved`. No Task persona substitutes. **≥2** units → **MUST** `coder-agent` per unit (`ns-coder`); **1** unit → follow `ns-coder` in A (worktree) — **MUST NOT** extra spawn. C2 / single-unit workers = implement only; **no** per-unit review — parent owns REV at closure.
## Isolation invariant (non-negotiable)
Skill **never** writes application code on main product checkout; **never** commits `main`/`master`/base/`SOURCE_BRANCH` unless human explicitly said work in place this run.
- Standalone + Engine: isolated `WORK_BRANCH` + `WORKTREE_ROOT` before any implementation edit.
- Worktree creation fails, or CWD/branch still main checkout / base → **stop**. No "continue on main to unlock plan".
- Sandbox restrictions (`Operation not permitted`, `.cursor/` paths) **not** permission to violate invariant.
## Session boot
See `../../ns-harness/references/session-boot.md`. **Complete Session boot (blocking)** there before any other step in this skill.
### After session boot — Engine and Standalone
**Engine mode:** caller may have already completed Session boot in Phase 0 — re-read `session-boot.md` only if context is missing.
**Standalone:** you own Session boot before step 1 of the standalone pipeline.
## Routing (read first)
Entry priority table: `../../ns-harness/references/code-skill-routing.md`. Standalone triggers: `references/entry-triggers.md`.
1. **Origin is a GitLab issue** (`ISSUE_URL` or issue reference) → entry priority **1** — follow `ns-execution-gitlab-issue` end to end; this skill only runs as its Phase 2 engine (**Engine mode** below). Do not run the standalone pipeline, do not touch GitLab state, do not create a worktree.
2. **Standalone autonomous** (local plan file, pasted text/plan, ad-hoc "implement this autonomously" without an issue) → entry priority **3** — run the **standalone pipeline**: own worktree, own state, own review loop.
See `references/routing.md` for Engine-mode I/O and anti-cycle rules with `ns-execution-gitlab-issue`.
## Engine mode (invoked by `ns-execution-gitlab-issue`)
Inputs (already resolved by the caller — this skill never creates them): issue payload, `WORKTREE_ROOT`, `WORK_BRANCH`, `SOURCE_BRANCH`.
1. **Planning-depth self-decision** — evaluate the issue payload and decide single work unit vs. light `requirements.md` + `tasks/task-NNN-*.md` + `execution-plan.md` under `docs/versions/{version_san}/sdd/`. See `references/planning-decision.md`.
2. **Doubt protocol** — self-ask, docs-first lookup, self-answer non-destructive doubts, escalate destructive ones as a structured event to the caller instead of mutating GitLab state; pause dependent units until resumed. See `references/doubt-resolution.md`.
3. **Multi-agent dispatch** — parse work units. **1 unit:** follow `ns-coder` in this session inside `WORKTREE_ROOT` (defer review to caller / standalone step 5). **≥2 units:** **MUST** dispatch `coder-agent` when available (else `ns-coder`) per unit; parallel only across units with no DAG edge and disjoint file scopes; sequential otherwise. Workers implement only — **no** per-unit `reviewer-agent`. See `references/multi-agent-dispatch.md` and `../../ns-harness/references/subagent-dispatch.md`.
4. **Checkpoint commits** — one commit per completed sequential unit or per completed parallel batch, inside the worktree; the caller squashes at delivery.
5. **Fix-loop entry point** — when re-invoked after `Rejected`, run **one** coder call on the whole map **verbatim** (no unit-count split, no “or units”). Skip units already done.
6. **Return to caller**: unit statuses, files changed, any open destructive doubt, and (first invocation only) a plan-based `estimate_seconds` hint — the caller applies `set_issue_estimate` only when the issue estimate is empty and the value is ≥ 60 (see `../ns-execution-gitlab-issue/references/time-tracking.md`).
## Standalone pipeline (non-GitLab origin)
Same internals as Engine mode, but this skill owns the whole run:
1. Resolve the descriptor — local plan path or pasted text. No MCP calls.
2. Infer `change_kind` (fix/feat), allocate `{version_san}`, create `docs/versions/{version_san}/sdd/` when artifacts needed.
3. **Create its own worktree**: `.worktrees/{version_san}/` + branch `work/{version_san}` from the resolved base branch, following `../../ns-harness/references/worktree-setup.md`. Path is under the **repo root**, never under `.cursor/`. On failure → abort (do not fall back to the main checkout) — see `references/standalone-pipeline.md`.
4. Planning-depth self-decision, doubt protocol (destructive doubt → chat-only gate, no GitLab actions available), multi-agent dispatch — identical logic to Engine mode.
5. **Internal review loop** — `../ns-reviewer/references/review-gate-workflow.md` (exhaustive map → one coder → `delta`; **`Approved` = 10** only): **MUST** `reviewer-agent` when available (else `ns-reviewer`) version-closure **code** mode. **Do not** run `ns-judge` until code `Approved`. Then in-session `../ns-judge/SKILL.md` (`closure`). Product files changed after judge Rejected: code review again then judge again.
6. Report per `review-gate-workflow.md` **Final report** fields plus `{version_san}`, worktree path, commit(s), and follow-ups. No GitLab board, no MR — unless `docs/context/gitlab-sync-config.md` exists and the human explicitly asked for one (out of scope for v1; standalone stays local-only otherwise).
See `references/standalone-pipeline.md` for the full flow.
## Stop and ask the human
| Condition | Action |
| ------------------------------------------------------------------ | ---------------------------------------------------- |
| Destructive doubt (Engine mode) | Return escalation event to caller — no chat pause here |
| Destructive doubt (standalone) | Stop in chat, ask, resume with the answer as context |
| Worktree conflict for the same `{run_id}` (standalone) | Stop unless explicit resume |
| `git worktree add` failed (permissions/sandbox/path) | Abort — report path attempted; never fall back to main checkout or `.cursor/` |
| Still on `main`/`master`/base branch when about to edit code | Abort — isolation missing; do not implement |
| Human explicitly said "do not create a new branch / work in place" | Only then may edit the current branch; still never invent a `.cursor/` path |
| Base branch for standalone worktree unresolved | Stop — ask once |
| Review `Blocked` or 3 rounds exhausted | Stop, report as blocked — do not fabricate success |
| Engine mode invoked with a GitLab origin but no worktree/branch given | Stop — this is a caller contract violation, not a doubt |
## Related skills
| Skill | Role |
| ---------------------- | ------------------------------------------------------------ |
| `ns-execution-gitlab-issue` | GitLab flow owner — calls this skill for Phase 2 |
| `ns-coder` | Implement via in-session (1 unit) or `coder-agent` (≥2 — **MUST** when available) |
| `ns-reviewer` | Closure code gate via `reviewer-agent` (**MUST** when available) — not per unit |
| `ns-judge` | Delivery proof in-session after `Code Review: Approved` |
| `ns-harness` | Artifact layout, worktree mechanics, discovery |
## References
| File | When |
| ------------------------------------ | --------------------------------------------------------------------- |
| `references/routing.md` | Issue vs. standalone; Engine-mode I/O and anti-cycle with `ns-execution-gitlab-issue` |
| `references/entry-triggers.md` | Priority 3 standalone entry phrases |
| `references/planning-decision.md` | Self-evaluation heuristic: single unit vs. requirements+tasks+plan |
| `references/doubt-resolution.md` | Self-ask, docs-first lookup, destructive criteria, escalation shape |
| `references/multi-agent-dispatch.md` | DAG + disjoint-scope parallel rule, subagent prompt, checkpoint commits |
| `references/standalone-pipeline.md` | Full non-GitLab flow: version alloc, own worktree, review loop |
## Forbidden
- Review substitutes (`senior-tech-lead-reviewer`, `bugbot`, `security-review`, or any non-`ns-reviewer` code gate) — harness `reviewer-agent` is allowed; see `../ns-reviewer/references/review-gate-workflow.md`
- `ns-judge` before `Code Review: Approved`
- Reporting success without both `Approved`, or explicit **blocked** state
- Skipping re-review after a `Rejected` fix before closure
© nextstage-brasil · Apache-2.0 · nextstage-brasil/skills