M
maj0rpain/orchestrator
Claude Code plugin that takes a change from plan to pull request in fresh-context sessions: a planning…
在 GitHub 查看 ↗链接已于 2026-11-12 核查 · HTTP 200 · 打开的是厂商自己的官网
- ★ 数取值
- 0 ★ · maj0rpain/orchestrator · 取自 GitHub 搜索 API · (2026-11-13 20:00:50 核对)
- 来源明细
- 从该仓库的 GitHub 页面直接读取 · 条目类型:开源项目
- 收录与活跃
- 收录于 2026-10-29 · 最近见到 2026-11-14 · 共见到 17 次 · 被 1 个来源列出
- 评分构成
- 综合分 103 · 1 个来源 · repo
常见问题(答案都来自我们的记录)
它适合谁 / 用来做什么?
我们把它归在这些标签下: 开源项目。标签来自条目自身,不是我们强加的分类。
谁维护它、还在更新吗?
仓库
maj0rpain/orchestrator;我们收录于 2026-10-29;最近见到 2026-11-14。这里没有"官方认证"或"质量评级"——我们只记录来源与时间。# orchestrator
以下摘录自该仓库的 README(片段) · 在 GitHub 读完整 README →
A Claude Code plugin that takes a change from plan to pull request: a planning
interview, a spec on a GitHub issue, tickets built test-first, and a bounded
review loop - each step in a fresh session, connected only by written handoffs.
## Why separate sessions
One long session that plans, specs, builds, and reviews carries every earlier
phase's context into the next. The reviewer already believes the
implementer's reasoning. Splitting the phases and passing only a written
handoff between them means each phase judges the work, not the story behind
it. Fresh context per phase also leaves room for the human-in-the-loop
exchanges that orch-to-spec (test seams) and the spec review depend on.
## Requirements
Claude Code, or the Junie CLI (see Install).
gh, jq, and git, with gh signed in to the repo's GitHub.
No per-repo setup: issues are labelled with the five canonical triage label
names (needs-triage, needs-info, ready-for-agent, ready-for-human,
wontfix), or with the names a docs/agents/triage-labels.md table maps
them to, when the repo has one. /orchestrator:health reports all of this
at any time, including any of those labels the repo is missing.
## Install
/plugin marketplace add maj0rpain/orchestrator
/plugin install orchestrator@orchestrator
The repo doubles as its own single-plugin marketplace, so there is no separate
marketplace repo. It installs at user scope, so it is available in every
project on that machine.
On the Junie CLI, see docs/junie/README.md.
## Quick start
1. In your repo, run /orchestrator:interview and plan the change with it
until you share an understanding.
2. It closes with one question: pick a flow, a quick implementation, a
blueprint only, or (Claude Code only) the autopilot flow.
3. For a flow, run /orchestrator:next in a fresh session (/clear) for each
phase - spec, implement, review - until the PR is ready. A quick
implementation runs to its PR on its own; a blueprint stops once its spec
and tickets are published.
## Running autopilot
The minimum steps for the autopilot flow.
orch.sh is scripts/orch.sh under the plugin's install directory, e.g.
~/.claude/plugins/cache/orchestrator/orchestrator/ /scripts/orch.sh.
1. Set the permission mode its headless sessions run in, e.g.
git config orchestrator.autopilot.permission-mode acceptEdits.
2. Start the watcher: orch.sh autopilot watch.
3. Pick "Autopilot flow" at the planning close.
4. Answer its questions on the issue, then remove the needs-human label.
To run the watcher isolated from your machine instead, use the
container.
Setup, scheduling, and the question format are in
docs/how-it-works.md.
## Four routes
Every route starts from a planning session and is chosen by a human, never by
the model. The terms are defined in GLOSSARY.md.
Flow. The full pipeline: four phases - plan, spec, implement, review -
each in a fresh session, handing off through files. The spec phase publishes
the spec issue and reviews it with you; the implement phase builds the ticket
breakdown test-first, one fresh subagent per ticket, and opens a draft PR; the
review phase runs a bounded review loop that ends with the PR ready, or on a
bounded stop that says why.
Quick implementation. The route with no phases: a linked issue -
interview, a spec on a GitHub issue, tickets built test-first, and a bounded
review loop - each step in a fresh session, connected only by written handoffs.
## Why separate sessions
One long session that plans, specs, builds, and reviews carries every earlier
phase's context into the next. The reviewer already believes the
implementer's reasoning. Splitting the phases and passing only a written
handoff between them means each phase judges the work, not the story behind
it. Fresh context per phase also leaves room for the human-in-the-loop
exchanges that orch-to-spec (test seams) and the spec review depend on.
## Requirements
Claude Code, or the Junie CLI (see Install).
gh, jq, and git, with gh signed in to the repo's GitHub.
No per-repo setup: issues are labelled with the five canonical triage label
names (needs-triage, needs-info, ready-for-agent, ready-for-human,
wontfix), or with the names a docs/agents/triage-labels.md table maps
them to, when the repo has one. /orchestrator:health reports all of this
at any time, including any of those labels the repo is missing.
## Install
/plugin marketplace add maj0rpain/orchestrator
/plugin install orchestrator@orchestrator
The repo doubles as its own single-plugin marketplace, so there is no separate
marketplace repo. It installs at user scope, so it is available in every
project on that machine.
On the Junie CLI, see docs/junie/README.md.
## Quick start
1. In your repo, run /orchestrator:interview and plan the change with it
until you share an understanding.
2. It closes with one question: pick a flow, a quick implementation, a
blueprint only, or (Claude Code only) the autopilot flow.
3. For a flow, run /orchestrator:next in a fresh session (/clear) for each
phase - spec, implement, review - until the PR is ready. A quick
implementation runs to its PR on its own; a blueprint stops once its spec
and tickets are published.
## Running autopilot
The minimum steps for the autopilot flow.
orch.sh is scripts/orch.sh under the plugin's install directory, e.g.
~/.claude/plugins/cache/orchestrator/orchestrator/ /scripts/orch.sh.
1. Set the permission mode its headless sessions run in, e.g.
git config orchestrator.autopilot.permission-mode acceptEdits.
2. Start the watcher: orch.sh autopilot watch.
3. Pick "Autopilot flow" at the planning close.
4. Answer its questions on the issue, then remove the needs-human label.
To run the watcher isolated from your machine instead, use the
container.
Setup, scheduling, and the question format are in
docs/how-it-works.md.
## Four routes
Every route starts from a planning session and is chosen by a human, never by
the model. The terms are defined in GLOSSARY.md.
Flow. The full pipeline: four phases - plan, spec, implement, review -
each in a fresh session, handing off through files. The spec phase publishes
the spec issue and reviews it with you; the implement phase builds the ticket
breakdown test-first, one fresh subagent per ticket, and opens a draft PR; the
review phase runs a bounded review loop that ends with the PR ready, or on a
bounded stop that says why.
Quick implementation. The route with no phases: a linked issue -
收录徽章(可选)
想把收录标记放进 README 就复制这行,可选,不影响收录。
[](https://anclum.com/zh/s/maj0rpain-orchestrator/)
- 来源
- github · https://github.com/maj0rpain/orchestrator
- 按领域浏览
- 开源项目
- 收录于
- 今天
同类推荐
所属领域与标签
每个标签都能点进去看同类的全部条目。数字是当前库里的真实条数。