Skip to content
Conduit

skill

orca-conductor

Coordinator-side discipline for Orca multi-agent orchestration sessions: worker briefs with read-back gate, phase/step checkpoint-rewind, evidence-based interruption recovery, acceptance & safe worktr

jiwangyihao/orca-conductorcommunitylistedunreviewed

Set up

Skills are installed by placing their folder where your agent loads skills.

The download comes from github.com/jiwangyihao/orca-conductor itself, not from our server, so what you get is exactly what its author published.

Install instructions come from the README

Skills do not use an MCP config. Follow the steps in the README below.

README

synced Oct 23, 2026

orca-conductor

An Agent Skill for running a multi-agent orchestration session on top of Orca's orca orchestration CLI — the coordinator side, not the worker side.

一个用于 Agent 调度会话(coordinator) 的 Skill。它不教你 Orca CLI 的参数(那是 Orca 自带 orchestration skill 的事),它管的是调度纪律:任务指示怎么写才不会让 worker 白干、上下文怎么真正推进隔离 worktree、worker 卡住时怎么判断、什么时候绝对不该重开 Agent、产出怎么验、worktree 怎么回收而不会把自己删掉。

Why

调度会话的失败方式很少,而且高度重复:

失败模式本 skill 的对策
spec 只有"目标 + 几条约束",worker 从零重新探索、重走已否决的路12 节必填 Worker Brief + 强制 Read-back Gate
worker 起在别的 worktree,看不到你未提交的改动先 commit 取基线 SHA,--base-branch <SHA>;Orca 谱系 ≠ git 基线
因为"跑得久"就重开 Agent,丢掉全部已有进度三证据法 + 六级升级阶梯,补发优先于重开
上下文被轮询垃圾填满,忘掉 dispatch id每个 Step 一次 checkpoint → rewind,report 强制承载 id 与下一步
回收攒到最后,分不清哪个 worktree 是谁的owned 清单作为唯一删除授权 + 自我保护 + 只读巡检脚本

Install

OMP and other agents (skills CLI, recommended)

Install globally through the open skills CLI:

npx skills@latest add jiwangyihao/orca-conductor \
  --skill orca-conductor \
  --global \
  --agent universal \
  --yes

universal installs the managed copy at ~/.agents/skills/orca-conductor, which OMP discovers as a user skill. Update it later with:

npx skills@latest update --global orca-conductor --yes

Claude Code plugin marketplace

/plugin marketplace add jiwangyihao/orca-conductor
/plugin install orca-conductor@orca-conductor

Do not install both methods for the same agent; duplicate copies can make skill resolution ambiguous.

Layout

skills/orca-conductor/
├── SKILL.md                          # 总纲:不变式、Phase/Step 骨架、S1–S7 手册、反模式
├── references/
│   ├── worker-brief.md               # 任务指示契约:12 节必填 + Read-back Gate + 正反例
│   ├── context-plumbing.md           # 三条上下文通道;谱系 vs git 基线;跨机限制
│   ├── phase-step-protocol.md        # Step 转移表、checkpoint/rewind 语义、report 模板
│   ├── recovery-playbook.md          # 四类中断现场决策表、升级阶梯、补发模板
│   └── acceptance-and-reclaim.md     # 独立复跑验收、owned 清单、回收语义与自我保护
├── assets/
│   ├── worker-brief-template.md      # 复制填写,交给 worker
│   ├── progress-ledger-template.md   # worker 进度台账格式
│   └── session-kickoff-prompt.md     # 把任意会话切换成调度会话的提示词
└── scripts/
    └── orca_sweep.sh                 # 只读巡检:自身保护标记 + task 终态 + 资源对账

Core rules (TL;DR)

  1. 先落盘再派发;--spec 只做指向 brief 的指针(15–30 行)。
  2. Read-back Gate 强制:worker 先回写"已读文件 + 实际 git log -1 + brief 的 sha256 + 分步计划",收到 reply 前不得改任何业务文件。
  3. new-child 只决定 Orca 谱系,git 基线必须 --base-branch <commit SHA>(用 SHA,分支名会漂移)。
  4. 时间不是证据:禁止仅因运行时长 stop/restart worker;升级需要新增证据。
  5. 补发(send --to dispatch:<id>)优先于重开——重开等于丢掉 worker 已建立的全部上下文。
  6. 每个 Step 一次 checkpoint→rewind;rewind 会丢弃 Step 内上下文,report 必须带 id、基线 SHA、下一步。
  7. 只回收自己 owned 清单里的资源;当前 worktree 永不删(删了会终结 coordinator 自己)。

Notes

  • Worker harness 固定 OMP(--agent omp)。
  • CLI 参数、返回结构、错误语义以 orca skills get orchestration 为准,本 skill 不复述。
  • Skill 正文为中文。

License

Mozilla Public License 2.0

↑↓ navigate openesc close