{"id":"9125ff06-0940-4e04-803a-147d7b7b984e","slug":"orca-conductor","kind":"skill","name":"orca-conductor","summary":"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","owner":"jiwangyihao","repoUrl":"https://github.com/jiwangyihao/orca-conductor","license":"MPL-2.0","licenseTier":"weak-copyleft","trust":"community","stars":1,"match":null,"fewStars":true,"format":"skill","createdAt":"2026-08-10T07:39:13.000Z","addedAt":"2026-10-21T11:25:06.332Z","logoUrl":"https://avatars.githubusercontent.com/u/51940778?u=d9d5867ace910ec300b541184f53d2efbc501515&v=4","linkOnly":false,"collection":null,"skillPath":null,"categories":["developer-tools"],"tags":["ai-agents","claude-code","multi-agent","orca","orchestration"],"runtime":"binary","local":true,"cloudCapable":false,"cloudCapability":"unknown","cloudReadyOnPlatform":false,"needsApiKey":false,"reviewed":false,"syncedAt":"2026-10-23T00:01:03.649Z","sample":false,"url":"https://conduits.dev/tools/orca-conductor","language":"Shell","transports":["stdio"],"install":{},"requirements":{"env":[],"upstreamMayCharge":false},"tools":null,"readme":"# orca-conductor\n\nAn Agent Skill for running a **multi-agent orchestration session** on top of [Orca](https://onorca.dev)'s `orca orchestration` CLI — the coordinator side, not the worker side.\n\n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。\n\n## Why\n\n调度会话的失败方式很少，而且高度重复：\n\n| 失败模式 | 本 skill 的对策 |\n|---|---|\n| 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** |\n| 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** |\n| 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 |\n| 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 |\n| 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 |\n\n## Install\n\n### OMP and other agents (`skills` CLI, recommended)\n\nInstall globally through the open [`skills`](https://github.com/vercel-labs/skills) CLI:\n\n```bash\nnpx skills@latest add jiwangyihao/orca-conductor \\\n  --skill orca-conductor \\\n  --global \\\n  --agent universal \\\n  --yes\n```\n\n`universal` installs the managed copy at `~/.agents/skills/orca-conductor`, which OMP discovers as a user skill. Update it later with:\n\n```bash\nnpx skills@latest update --global orca-conductor --yes\n```\n\n### Claude Code plugin marketplace\n\n```text\n/plugin marketplace add jiwangyihao/orca-conductor\n/plugin install orca-conductor@orca-conductor\n```\n\nDo not install both methods for the same agent; duplicate copies can make skill resolution ambiguous.\n\n## Layout\n\n```\nskills/orca-conductor/\n├── SKILL.md                          # 总纲：不变式、Phase/Step 骨架、S1–S7 手册、反模式\n├── references/\n│   ├── worker-brief.md               # 任务指示契约：12 节必填 + Read-back Gate + 正反例\n│   ├── context-plumbing.md           # 三条上下文通道；谱系 vs git 基线；跨机限制\n│   ├── phase-step-protocol.md        # Step 转移表、checkpoint/rewind 语义、report 模板\n│   ├── recovery-playbook.md          # 四类中断现场决策表、升级阶梯、补发模板\n│   └── acceptance-and-reclaim.md     # 独立复跑验收、owned 清单、回收语义与自我保护\n├── assets/\n│   ├── worker-brief-template.md      # 复制填写，交给 worker\n│   ├── progress-ledger-template.md   # worker 进度台账格式\n│   └── session-kickoff-prompt.md     # 把任意会话切换成调度会话的提示词\n└── scripts/\n    └── orca_sweep.sh                 # 只读巡检：自身保护标记 + task 终态 + 资源对账\n```\n\n## Core rules (TL;DR)\n\n1. 先落盘再派发；`--spec` 只做指向 brief 的指针（15–30 行）。\n2. 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。\n3. `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）。\n4. 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。\n5. 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。\n6. 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。\n7. 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）。\n\n## Notes\n\n- Worker harness 固定 OMP（`--agent omp`）。\n- CLI 参数、返回结构、错误语义以 `orca skills get orchestration` 为准，本 skill 不复述。\n- Skill 正文为中文。\n\n## License\n\nMozilla Public License 2.0\n"}