Huashu Agent Swarm
alchaincyf/huashu-skills
多Agent蜂群并行协作,纯git自组织,适合大型项目开发。当用户提到"蜂群模式"、"多agent"、"并行开发"、"agent swarm"时使用。
编排两个以上边界独立的本地 worker,使用 Orca Run/Task/Dispatch、独立 worktree/session 或 tmux 回退,由 PM 负责拆解、派发、巡检、429 停滞恢复、独立验收、PR 收口与临时资源清理;也用于用户明确要求“并行推进”“多个 worker”“PM 总控”“Wave Autopilot”或防止 PM…
$ npx skills add cat-xierluo/legal-skills --skill multi-agent-orchestration -a claude-codeProject install by default; add -g for ~/.claude/skills/.
$ gh skill install cat-xierluo/legal-skills multi-agent-orchestration --agent claude-codeProject scope by default; add --scope user for a personal install. Needs GitHub CLI 2.90.0 or later (public preview).
$ git clone --depth 1 https://github.com/cat-xierluo/legal-skills.git skills-src && mkdir -p .claude/skills && cp -r skills-src/skills/multi-agent-orchestration .claude/skills/multi-agent-orchestration && rm -rf skills-srcUse ~/.claude/skills/ instead of .claude/skills for a personal install. The folder must contain SKILL.md.
Claude Code skills documentation · loads skills from .claude/skills/
Install the "multi-agent-orchestration" agent skill from https://github.com/cat-xierluo/legal-skills/tree/main/skills/multi-agent-orchestration into .claude/skills/multi-agent-orchestration/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "multi-agent-orchestration", then confirm the skill loads.Claude Code copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$skill-installer install https://github.com/cat-xierluo/legal-skills/tree/main/skills/multi-agent-orchestrationType this inside Codex. $skill-installer <name> installs a curated skill from openai/skills. The installer writes to $CODEX_HOME/skills (default ~/.codex/skills). Restart Codex if the skill does not show up.
$ npx skills add cat-xierluo/legal-skills --skill multi-agent-orchestration -a codexProject install goes to .agents/skills/; add -g for ~/.codex/skills/.
$ gh skill install cat-xierluo/legal-skills multi-agent-orchestration --agent codexProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/cat-xierluo/legal-skills.git skills-src && mkdir -p .agents/skills && cp -r skills-src/skills/multi-agent-orchestration .agents/skills/multi-agent-orchestration && rm -rf skills-srcUse ~/.agents/skills/ instead of .agents/skills for a personal install.
Codex skills documentation · loads skills from .agents/skills/
Install the "multi-agent-orchestration" agent skill from https://github.com/cat-xierluo/legal-skills/tree/main/skills/multi-agent-orchestration into .agents/skills/multi-agent-orchestration/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "multi-agent-orchestration", then confirm the skill loads.Codex copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ npx skills add cat-xierluo/legal-skills --skill multi-agent-orchestration -a cursorProject install goes to .agents/skills/; add -g for ~/.cursor/skills/.
$ gh skill install cat-xierluo/legal-skills multi-agent-orchestration --agent cursorProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/cat-xierluo/legal-skills.git skills-src && mkdir -p .cursor/skills && cp -r skills-src/skills/multi-agent-orchestration .cursor/skills/multi-agent-orchestration && rm -rf skills-srcUse ~/.cursor/skills/ instead of .cursor/skills for a personal install.
Cursor skills documentation · loads skills from .cursor/skills/, .agents/skills/, .claude/skills/, .codex/skills/
Install the "multi-agent-orchestration" agent skill from https://github.com/cat-xierluo/legal-skills/tree/main/skills/multi-agent-orchestration into .cursor/skills/multi-agent-orchestration/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "multi-agent-orchestration", then confirm the skill loads.Cursor copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ gemini skills install https://github.com/cat-xierluo/legal-skills.git --path skills/multi-agent-orchestration--scope user (default) or --scope workspace; --path is the subfolder of the repo that holds the skill; --consent skips the security confirmation prompt.
$ npx skills add cat-xierluo/legal-skills --skill multi-agent-orchestration -a gemini-cliProject install goes to .agents/skills/; add -g for ~/.gemini/skills/.
$ gh skill install cat-xierluo/legal-skills multi-agent-orchestration --agent gemini-cliProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/cat-xierluo/legal-skills.git skills-src && mkdir -p .gemini/skills && cp -r skills-src/skills/multi-agent-orchestration .gemini/skills/multi-agent-orchestration && rm -rf skills-srcUse ~/.gemini/skills/ instead of .gemini/skills for a personal install, then run /skills reload.
Gemini CLI skills documentation · loads skills from .gemini/skills/, .agents/skills/
Install the "multi-agent-orchestration" agent skill from https://github.com/cat-xierluo/legal-skills/tree/main/skills/multi-agent-orchestration into .gemini/skills/multi-agent-orchestration/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "multi-agent-orchestration", then confirm the skill loads.Gemini CLI copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ gh skill install cat-xierluo/legal-skills multi-agent-orchestrationInstalls for Copilot at project scope by default; add --scope user for a personal install. Preview a skill first with gh skill preview. Needs GitHub CLI 2.90.0 or later (public preview).
$ npx skills add cat-xierluo/legal-skills --skill multi-agent-orchestration -a github-copilotProject install goes to .agents/skills/; add -g for ~/.copilot/skills/.
$ git clone --depth 1 https://github.com/cat-xierluo/legal-skills.git skills-src && mkdir -p .github/skills && cp -r skills-src/skills/multi-agent-orchestration .github/skills/multi-agent-orchestration && rm -rf skills-srcUse ~/.copilot/skills/ instead of .github/skills for a personal install. Commit .github/skills so cloud agent and code review can use it.
GitHub Copilot skills documentation · loads skills from .github/skills/, .claude/skills/, .agents/skills/
Install the "multi-agent-orchestration" agent skill from https://github.com/cat-xierluo/legal-skills/tree/main/skills/multi-agent-orchestration into .github/skills/multi-agent-orchestration/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "multi-agent-orchestration", then confirm the skill loads.GitHub Copilot copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ npx skills add cat-xierluo/legal-skills --skill multi-agent-orchestration -a opencodeOpenCode documents no install command of its own. Project install goes to .agents/skills/; add -g for ~/.config/opencode/skills/.
$ gh skill install cat-xierluo/legal-skills multi-agent-orchestration --agent opencodeProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/cat-xierluo/legal-skills.git skills-src && mkdir -p .opencode/skills && cp -r skills-src/skills/multi-agent-orchestration .opencode/skills/multi-agent-orchestration && rm -rf skills-srcUse ~/.config/opencode/skills/ instead of .opencode/skills for a personal install.
OpenCode skills documentation · loads skills from .opencode/skills/, .claude/skills/, .agents/skills/
Install the "multi-agent-orchestration" agent skill from https://github.com/cat-xierluo/legal-skills/tree/main/skills/multi-agent-orchestration into .opencode/skills/multi-agent-orchestration/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "multi-agent-orchestration", then confirm the skill loads.OpenCode copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
multi-agent-orchestration编排两个以上边界独立的本地 worker,使用 Orca Run/Task/Dispatch、独立 worktree/session 或 tmux 回退,由 PM 负责拆解、派发、巡检、429 停滞恢复、独立验收、PR 收口与临时资源清理;也用于用户明确要求“并行推进”“多个 worker”“PM 总控”“Wave Autopilot”或防止 PM…
Multi Agent Orchestration is an agent skill from cat-xierluo/legal-skills. 编排两个以上边界独立的本地 worker,使用 Orca Run/Task/Dispatch、独立 worktree/session 或 tmux 回退,由 PM 负责拆解、派发、巡检、429 停滞恢复、独立验收、PR 收口与临时资源清理;也用于用户明确要求“并行推进”“多个 worker”“PM 总控”“Wave Autopilot”或防止 PM 直接实现逃逸。不要用于单个短任务、纯状态同步,或仅需 Git 分支、提交、PR、merge 规则的工作。
Its SKILL.md is about 6.6k tokens, which your agent loads only when the skill is triggered. The skill folder holds 289 other files, including scripts and reference files (for example `CHANGELOG.md`, `DECISIONS.md` and `TASKS.md`).
It sits in Development, covering Multi-agent orchestration, Git worktrees and Git workflow. It works with tmux and Git. The licence is MIT.
9 steps, taken from the step headings in SKILL.md.
Read from SKILL.md and the folder at commit c077fcc. It shows what the files ask for, not the result of running them.
Pre-approves nothing: there is no allowed-tools line, so your agent's usual permission prompts apply.
From allowed-tools in the SKILL.md frontmatter.
Ships 1 file in scripts/ (Shell, from the files we listed), which the agent can run.
Shell commands in SKILL.md call:
bashpython3gitghjqnodepytestjestgoFrom the folder's file list and the shell code blocks in SKILL.md.
No URLs in SKILL.md. Its commands use git and gh, which can reach the network depending on how they are called.
From URLs in SKILL.md, links to its own repository left out.
Names no API keys, tokens, secrets or passwords.
From names ending in _API_KEY, _TOKEN, _SECRET, _KEY or _PASSWORD in SKILL.md.
Multi Agent Orchestration loads about 6.6k tokens when it runs, and up to ~149k if it reads all its reference files. Until then it costs about 64 tokens; SKILL.md has 1,856 words of instructions outside code blocks.
Estimates: characters ÷ 4, the usual rule of thumb; real counts depend on the model's tokenizer. Scripts and assets cost tokens only if the agent reads them.
The automated check found no risky patterns in SKILL.md.
Automated static check — not a guarantee. Review scripts before installing. It scans the text of SKILL.md for risky patterns (piping downloads into a shell, reading credential files, hidden Unicode, destructive commands); the scripts in this folder are not scanned.
The full file from cat-xierluo/legal-skills at commit c077fcc, republished under its MIT licence (© cat-xierluo). 1,856 words, ~6,578 tokens.
.claude/skills/multi-agent-orchestration/SKILL.md (or your agent's skills folder). This skill also uses 287 other files; get the full folder from GitHub.以当前主会话作为 PM,拆解、派工、巡检、验收和收口多个本地 Agent。日常 worktree 与 terminal/session 优先由 Orca 管理;Orca 不可用、用户明确要求或需要复现兼容路径时才使用 tmux。不要把“开了终端”误写成“建立了受监管任务”。
使用本 Skill:
不要使用:
git-workflow。本 Skill 可能创建 Git/Orca worktree、分支、Session Context、终端、tmux session,以及 supervised Run/Task/Dispatch。它不自动安装依赖,也不自行扩张 push、merge、发布或外部调度授权。Orca worktree 创建固定使用 --setup skip:repo Setup 发生在本 Skill 写入 Session Context 和机械门禁之前,inherit/run 会在资源创建前以 ORCA_SETUP_REQUIRES_PRELAUNCH_AUTH_CONTRACT 拒绝;--allow-install-command 只授权门禁已就位后的 worker 阶段,不能追认 Setup。真实 provider 配置及备份不得进入 Git、日志或交付物。
交付成功后,pm-closeout.sh 默认清理一次性 worker 的远端 head、worktree 与本地分支;长期功能/集成分支及固定 worktree 必须声明 long-lived 并保留。事实未知、身份漂移或生命周期未结算时失败关闭。supervised 清理必须从 Session Context 取得精确 Run,按 worker-list --run <run> --limit 100 完整遍历 opaque cursor 后唯一定位 Dispatch;前置分页、身份或结果不可证明时,在 release、terminal、worktree 和分支 mutation 前停止。reclaimable release 后还须重新执行同一完整查询;若后置结果不可证明,release 可能已发生,但不得继续 terminal、worktree 或分支 mutation。对已经确认合并、但未走标准 closeout 的单一遗留 worker,可使用 post-merge-cleanup.sh 做严格 dry-run/execute 清理;它不替代标准 closeout,也不得用于批量扫描。
| 模式 | 何时选择 | 完成权威 |
|---|---|---|
| Orca supervised | 用户要求监督、等待结果、DAG、ask/reply 或 decision gate;Agent 可被 Orca 识别 | worker_done → Delivery,再由 PM 验收与 settlement |
| Orca terminal-managed | 白名单 backend 未采用 supervised,或 CLI 仅能由外部 terminal 管理 | terminal 输出 + checkpoint + 真实产物,PM 验收 |
| tmux worktree | Orca 不可用、用户指定 tmux 或兼容性回归 | checkpoint + Git/测试/产物,PM 验收 |
| tmux lightweight | 用户明确不要 worktree,或非 Git 目标且目录绝不重叠 | checkpoint + 真实产物,PM 验收 |
| 同宿主 subagent | 中小型质量敏感卡、快速迭代轮次、额度充裕时优先于独立 worker(见上方通道选择判据);无独立进程/分支开销 | 宿主决定(建议 RESULT 四段自陈) |
| 远程节点(SSH 桥) | 需要第二台机器扩并发(独立 key 池)或卸载本机 CPU/内存时,PM 经 spawn-worker-remote.sh 派到 personal config remote_nodes 声明的节点;节点侧全套门禁本机成立,worker 读节点自己的 env/key(references/25-remote-node-dispatch.md) | STATUS.json 终态 + 分支 push/PR 存在 + PM 侧 PR-fingerprint 验收(完成三证,不依赖跨机 Orca 回执) |
MiniMax 的短任务 exec 在 terminal-managed 中由启动命令读取完整输入,不等待 TUI、不再发送任务;长程任务仍使用交互 CLI。已选择 ZCode 的长程任务默认使用已配置的原生 supervised 启动桥,见 原生 Orca 合同。
backend 与模式选择、配置来源及唯一后续动作统一由 派发 profile 合同 定义。ZCode auto 缺桥配置时在资源副作用前拒绝;旧显式 native 参数保留,通用兼容入口须传 --dispatch-profile orca-generic,直连仍为 --no-orca-mode。profile 是派发计划;真实权限、开始、完成与资源状态继续来自原生回执。
同一 worker 只能有一个控制模式。terminal-managed 没有 Task/Dispatch,不得要求 worker_done;supervised 必须有 live Task/Dispatch,不得用 STATUS、UI 卡片、TUI idle、heartbeat 或 timeout 冒充完成。
PM 在任何 worker 副作用前完成:
branch_lifecycle、integration target/base、worktree、session、角色、backend/profile/model、provider slot 和资源 owner。ephemeral-worker;只有项目任务合同明确声明的长期功能/集成基线才使用 --branch-lifecycle long-lived。源分支生命周期与合并目标是两个字段,不得混同。python3 scripts/dispatch-value-gate.py --describe-policy 读取并发与待验收 PR 背压的当前默认值;项目可以收紧,探索窗口须明确授权并限期。派发合同用 capacity 绑定 PM 已盘点的共享活跃库存、证据及较小项目上限;已有 worker 与本波候选相加判断。旧合同未提供库存时只证明本波数量合法,不能据此声称全机还有槽位。机器内存 slots 是另一道门,不能替代 worker 库存或覆盖跨项目约定。--bail 1 早停;整包全量套件(全量 test/build 链)不进 worker 自验合同,全量验证单一在飞(跨项目互斥),默认归宿是 PM 收口时串行复跑。本条约束验证类别,不约束 worker 数量(见 §5 验证负载纪律)。one_wave 或 continuous:真人要求持续派发/返回后review再推进/不要每轮催促时,不默认单波。持续模式派发前核唯一监测的原PM、原任务源、正式配置及周期,读取 PM接续合同;显式continuous的Wave入口机械检查,其他模式由PM显式检查。只有总控心跳、承诺稍后设置或人工催促后恢复均不证明自动闭环。Issue 分组读取 references/12-issue-grouping.md;并发边界与真实事故读取 references/10-parallel-lessons.md。
以下四道门按阶段执行,任一非零退出均不得跳过:
templates/dispatch-value-gate.example.json 同构合同,运行 python3 scripts/dispatch-value-gate.py <spec.json>。接受 implementation、reusable_verification、绑定具名 PR/head 的 merge_gate,以及有业务目的、来源、精确产物和逐项标准的 business_artifact。研究、设计、文书和报告按业务产物验收;无明确消费的维护文档、占位调查与纯格式扩波仍拒绝。worker-value-postflight.py,工程交付核真实 diff、声明资产、验证命令和 40 位 immutable head;业务交付读取 Git 或非 Git 目录的真实文件,将合同、来源和文件指纹绑定为同一交付身份。review-acceptance-gate.py。自审、交付身份漂移、缺失绑定证据、失败验证或未清 blocker 均拒绝;业务产物还须由不同作者逐项核查内容及来源,不以文件哈希证明内容正确。acceptance-recovery.py 分类。internal_recoverable 在预算内修复并重新独立审查;external_dependency、safety_unknown 或预算耗尽才泊车。已具名 PR 的 docs-only 验收修复只能走 acceptance-repair-gate.py 的极窄 preflight/postflight 通道。需要自动记录已提交工程验证时,可按工程证据采集合同使用可选采集/check入口;它只生成原交付后门证据,继续保留独立审查与真实完成核验。
字段、命令、例外枚举、reviewer 证据预算与恢复语义统一读取 references/18-dispatch-acceptance-contracts.md;不要在项目 prompt 或别的脚本另造一套分类表。
spawn-worker.sh 从完整进程祖先链识别真实 PM harness,并对嵌套层白名单取交集。未知、冲突或不可证明的宿主失败关闭;--pm-harness 只做一致性声明,不能提权。--base-ref 只接受引用名(main、origin/main、refs/heads/x 等),不接受裸 sha——40-hex(以及 7—40 位纯十六进制且 git 解析为 commit 而非引用名)会在任何 worktree/provider/terminal 副作用前以 SPAWN_WORKER_BASE_REF_MUST_BE_REF: <值> 拒绝并退出,避免后续 pm-cleanup-worker 在 INTEGRATION_TARGET_MISMATCH(argument=main vs metadata=sha)与 PR_BASE_MISMATCH(expected=sha vs actual=main)之间死锁。字符类大小写敏感:守卫同时识别 [0-9a-fA-F],大写 sha(如 96A304DF…)与大小写混合 sha 一律按 sha 路径拒绝;分支名恰为大写 hex 形态且真实存在时(refs/heads/<v> 大小写敏感查找),仍按真实 ref 放行。config/harness-backend-policy.json 决定:支持 Claude Code、Codex、CodeBuddy、独立 Qoder CN CLI(qoder-cn)、千问办公 bundled coding CLI(qwenwork-cn)、独立 ZCode CLI(zcode-cli)、MiniMax Code CLI(minimax-code)及兼容的 ZCode app-server driver(zcode)。CodeBuddy PM 仍只可派自身,ZCode PM 的既有授权仍只含 Claude/Codex;新增 worker 支持不新增 PM 宿主。QoderWork backend 与旧别名已移除,不能把旧 bundled CLI 软链当作 Qoder CN。Hermes 宿主仍以安装路径签名识别,裸词不作证据。references/31-borrowed-existing-worktree.md,不伪造恢复 metadata。--review-repair-grant <授权来源>,且任何 config/*.local.yaml 都不可写。--permission-mode auto 时,普通 Bash 交给 Claude Code 的 auto 分类器和用户 settings;编排 hook 继续拦识别出的安装命令、直接及常见 Shell 包装的受保护 Git 操作、Orca 协议和受限 tracked 删除。其他后端与 Claude Code 非 auto 模式继续使用精确 allowed_shell_commands。execution_authority.shell_policy 固定所选策略;Claude auto 不是 Shell 文件写范围或任意程序副作用的机械沙箱,PM 仍须核对真实 diff 与外部副作用。验证命令不等于安装授权;安装类命令只有精确 --allow-install-command 和可审计授权来源才可通过编排门禁。Orca repo Setup 是更早的独立阶段,默认跳过,不能复用该授权。reauthorize只用于可证明live的目标,不以ready-reset恢复失败历史。worker-start 后冻结 Dispatch 身份与 capability 摘要;发送 worker_done 前复核 live runtime/process/run。不要改写 receipt 或用 Shell allowlist 绕过完成校验;首次 ORCA_COMPLETION_AUTHORITY_INVALID 即停止并向 PM 上报。字段与手动 register 迁移见 references/13-orca-cli-worker.md §5。--mode yolo,MiniMax batch 默认 --permission full;显式较窄模式保持生效。MiniMax 长程任务保持交互 CLI,权限按 可选 CLI 合同 配置并核实,不借 batch 参数替代。宿主审批、业务范围、账号登录与编排准入是独立边界,不因 worker 模式而取消。;/&& 复合命令在每段都是安全读或安全交付命令时整体放行(git status/diff/log/show/fetch/add/commit/push/rebase、gh pr create/view、ls/grep/cat/jq/sort 等过滤器、node --version 类版本查询);重定向仅限 /dev/null 与临时目录(拒绝 .. 穿越)。仍然 fail-closed:force push(--force/-f/--force-with-lease)、push 到 main/master、远端删除(git push origin :branch)、--mirror/--tags、子 shell、输入重定向、命令替换、gh api/gh repo sync、安装类命令。identity 四件套(--git-expected-name/--git-expected-email/--git-integration-base/--git-push-remote)仍推荐用于 PR 交付任务:绑定的 safe-push 会校验从远端 PR base 到 HEAD 的完整提交链后按不可变 OID 推送,是裸 push 的强化替代而非唯一通路。git rm -- <一个 canonical repo-relative tracked file>,且该精确路径必须在 spawn 时写入不可变 allowed_write_paths;scope glob、后续 reauthorize --allow-cmd、-r/-f/--cached、多路径、目录/gitlink、pathspec、Shell 展开和复合命令均不能扩大权限。Codex/ZCode 的 prompt_only_degraded 只提供提示,不能声称机械删除保护;详见 references/02-runtime-dependencies.md §6。verification_commands 时,调用 spawn 必须同时传 --verification-contract <spec.json> --verification-task-id <ID>;无文件合同时逐条传 --verify-cmd。命令作为完整字符串原样进入 authority receipt、METADATA 与 allowed_shell_commands,不得拆开 cd <subdir> && <verify>。在 Claude auto 策略下,该列表仍是交付验收的必跑命令,不是普通 Bash 的唯一执行权限来源。--require-verification,或在项目 .claude/orchestration.config.json 设置 verification.required: true。命令解析为空、合同 task 不唯一、worker type 未声明、配置畸形、重复/空白、U+0000 或安装型命令时,必须在 terminal/Task/Dispatch/任务注入前失败。业务合同显式 acceptance_mode: content_review 时使用空命令数组并进行内容独审,不自动发现无关测试;与显式 --require-verification 冲突则拒绝。需要真实命令验证的业务使用 verifier。--verify-cmd,否则使用派发价值合同;两者互斥。都未提供时才读取项目配置,再回退根目录有界发现。Node/Make 既有发现不变;Python 只在根 manifest 与根 tests/ 同时存在时注入 python3 -m unittest discover -s tests -v。嵌套 Python/其他子项目必须在项目配置 verification.by_worker_type 显式写完整命令,不递归猜测。依赖行为读取 references/02-runtime-dependencies.md。orca skills get orca-cli
orca skills get orchestration # supervised / DAG / ask-reply 时
orca status --json以运行中 CLI 的指南和 --help 为准。spawn-worker.sh 只在当前 PROJECT_DIR 可被精确解析为同一 Orca worktree/repo 时进入 Orca;--no-orca-mode 显式走 tmux。
多 worker supervised Wave 必须先一次性写 manifest、创建一个 Run 并预建全部 Task,receipt 成功后才并行启动。不要让并发 spawn 各自创建/重绑 Run,也不要在 worker-start 注入任务后再次发送完整 prompt。
bash scripts/orca-wave-prepare.sh --manifest /tmp/wave.json --from "$PM_TERMINAL" --receipt /tmp/wave-receipt.json
WAVE_RUNTIME_ID=$(jq -er '._meta.runtimeId' /tmp/wave-receipt.json)
bash scripts/spawn-worker.sh \
--project "$PROJECT" --branch feat/worker-a --session worker-a \
--branch-lifecycle ephemeral-worker --worker-backend claude-code \
--verification-contract /tmp/dispatch-spec.json --verification-task-id TASK-A \
--command "$AGENT_COMMAND" --orca-supervised \
--orca-run-id "$RUN_ID" --orca-coordinator-handle "$COORDINATOR_HANDLE" \
--orca-runtime-id "$WAVE_RUNTIME_ID" \
--orca-task-id "$TASK_A_ID"PM_TERMINAL 必须是明确属于本轮 PM 的终端句柄,不从 UI 焦点猜测。新 Wave 把 receipt 的 _meta.runtimeId 程序化传入 --orca-runtime-id;sender、终端活性与 Run/coordinator 在创建 Worker 资源前共同核验,预建 Wave 不重复建 Task/重绑 Run。单 Worker 可在 quota/mem 通过后、新建 lease/worktree/terminal 前准备 Run;无可靠 sender 则拒绝。旧上下文缺 runtime 时只能正向重验当前绑定并明确历史连续性未验证,已有 runtime 漂移不能绕过;仍以 Orca consumer fencing 作最终判断。
完整 manifest、Terminal-managed、Dispatch 自检、cold-start 恢复、settle 与 metadata 合同读取 references/13-orca-cli-worker.md。PM 的 read/show/send/wait/reply/release/ack/settle/pr-audit/closeout 命令读取 references/14-pm-orchestrate.md。
跨 session 的重要请求使用 pm-orchestrate.sh send --message-contract:先用 worker-show 复验精确 Run/Task/Dispatch/worker,再固定 sender、业务 thread、correlation、expected action 与 evidence refs;Orca 原生 thread 承载 correlation,使无 payload 的原生 reply 仍能继承可核对的关联标识,业务 thread 保留在合同 payload。send receipt 必须绑定同一 Dispatch relay 的 Orca message ID 与完整请求摘要;成功只证明 durably_enqueued。只读巡检使用 inbox(check --peek),所有出现的顶层及 payload 身份/类型别名必须一致;结构化消息须有精确 provenance,无 payload 的原生关联消息须同时给 thread+correlation 过滤器并精确匹配 worker sender、coordinator recipient、Run 与原生 correlation thread。相关消息可见不证明执行过 reply、已消费、开始执行或完成业务。状态层级、白名单、幂等指纹和敏感载荷拒绝规则统一读取 references/14-pm-orchestrate.md,不要另造聊天层或用 terminal prompt 代替 Orca 消息。
Worker 需要 PM 回答时使用 live preamble 的 ask;timeout、cancel 或断线后只按原 message ID --resume,不得重发新问题,也不得把普通 ask 升格为 decision gate。PM 的 wait 按最多 50 条完整 FIFO Delivery 返回顺序分类 receipt,同一批在 ack 前重放;逐条完成 question reply、escalation 处置、worker_done 业务验收及 terminal ownership 后,才 ack 精确 Delivery。ack 回执若同时交付下一批,必须按 receipt 继续处理,不能把确认上一批误当作下一批已处理。Worker 在新文件前、每次 scoped test 后和 worker_done 前执行非 peek check --terminal <live worker handle>,处理整批后仅以同一 handle ack 该 Worker Delivery,并继续排空后续批次;send 成功或 inbox --peek 可见均不证明 Worker 已处理。consumer_fenced 表示 consumer generation/进程身份已被替换,dispatch_inactive 表示原 Dispatch 已 settled、stopped 或不再 active;任一出现都立即停止、不得重试 check 或发送 worker_done。强制前缀和 Shell 门禁共同拒绝以 --peek/--all/--unread 冒充处理;Worker 的 scoped ack 不能指定 coordinator handle,也不授予 reply、release、stop 或 Task mutation。
使用 templates/worker-prompt.md,至少写明:任务卡、范围、禁止项、验证命令、完成协议、branch lifecycle、integration target、资源 owner、安装授权和 Git identity。supervised 的 worker-start 是唯一任务注入器;长 prompt 可落到 WORKER_PROMPT.md,terminal 只发送短 Read 指令。
实际 Task spec 的共同前缀补齐 ask/resume、自然检查点收件、最终收件、最小实施、scoped 验证、授权文件 commit 和唯一 Session Context RESULT;review-only/no-change 不造空提交。spawn-worker.sh 在所有合法后端启动链路注入绝对 WORKER_SESSION_CONTEXT,不依赖安装/scope guard 是否启用;合法 prompt_only_degraded 且无 allow-paths 的 worker 也能定位。该值仅用于定位,不授予安装/Shell/scope 权限,不把降级冒充 hook 活跃。Worker 在自身进程中检查该值及所有非空旧 guard 绑定(SCOPE_GUARD_SESSION_ROOT、WORKER_INSTALL_AUTH_FILE 父目录)拼写一致且目录存在,全部检查成功后才采用路径;全缺失、任一相对/冲突/不可用时交 PM,不猜仓库根目录,不修改 authority。旧调用仍可由旧 guard 绑定定位。前缀不扩安装、Shell 或 push/PR 权限;辅助命令与 .venv opt-in 流程见 references/02-runtime-dependencies.md。
<worktree>/.claude/agent-sessions/<session>/
├── METADATA.json
├── STATUS.json
├── RESULT.md
├── PATCH_SUMMARY.md
└── WORKER_PROMPT.md字段与 checkpoint 读取 references/03-checkpoint-files.md。supervised 中 STATUS 只辅助观察,完成权威仍是 worker_done → Delivery。
按证据优先级巡检:
worker-show、worker-read。发现偏题、阻塞、越界或验证失败时,优先给原 worker 发送窄纠偏;需要独立审阅时另派 reviewer。运行时活性与业务进展必须分开判断:输出/cursor/CPU 前进只证明活性,文件、commit、测试和产物才证明业务进展。观察不确定时不得自动 Esc、Ctrl+C、stop、release 或按进程名批量 kill。
Wave Autopilot 只有用户明确授权并在项目任务源固定策略后才启用。L1 当前会话推进读取 references/15-wave-autopilot.md;跨会话 L2 controller 与尚未实现的 L3 scheduler 边界读取 references/16-autopilot-durability.md。不要把 session cron、Markdown 任务源或 provider lease 单独描述成持久控制器。
Worker返回后由原PM固定产物/head、不同上下文独审、同次原Task写回,再沿原问题修复或推进具名合法下一项;不能在“已派发/作者done/已交接”处结束本波责任。单卡预算/依赖泊车不停止其他独立READY。记录人工催促/恢复的来源,区分监测配置、正式自动触发与完整自动业务周期;检查配置通过不签出自动自愈,详见 接续证据等级。
跨项目检查 Orca Worker 是否因 429/usage limit 停在 idle 时,运行 scripts/orca_rate_limit_recovery.py --manifest <私有清单>;默认只读,只有显式 --execute 才对高置信 RATE_LIMIT_IDLE 通过 terminal 输入通道发送一次固定“继续”。tmux、单关键词/陈旧 tail、未分组身份和状态不确定一律不处置;WAKE_ACCEPTED 不等于额度恢复或业务继续。完整 manifest、状态机、错峰、幂等、TOCTOU 与退出码读取 references/20-orca-rate-limit-recovery.md。
Worker node OOM 识别与退避(2026-09-05 事故:多 worker 长输出使会话内 node 进程 V8 堆耗尽 FatalProcessOutOfMemory → SIGABRT,PM 周期重拉形成崩溃循环并一度触发整机强制重启)。spawn-worker.sh v2.20.0 起默认给 worker 会话注入 NODE_OPTIONS=--max-old-space-size=2048(SPAWN_WORKER_NODE_MAX_OLD_SPACE_MB=0 可关),worker 到限自身退出而非拖垮系统。PM 巡检发现 worker node OOM(退出码 134 / SIGABRT / 日志含 FatalProcessOutOfMemory / 系统崩溃报告目录(DiagnosticReports)中 node OOM 报告新增且时间吻合)时:同任务不得立即重拉,至少等下一轮巡检并全局并发 -1,且重拉前必须重跑内存预算预检(probe 每次 spawn 都现场探测、不缓存;额度不足按 PARKED_FOR_MEMORY 排队,不得绕过);同一任务连续 2 次 OOM 后停止重拉、泊车并向用户报告——这通常意味着任务本身产生超长输出(全量日志聚合、超大测试跑),需任务侧降输出或拆分,而不是更用力地重试。
物理内存预算排队(mem budget lane):spawn-worker.sh 在任何 worktree/terminal/lease/dispatch 副作用之前运行 scripts/mem_budget_probe.py,按 per-worker 预算(默认 3 GiB;SPAWN_WORKER_MEM_BUDGET_BYTES 可调,=0 显式关闭整道门)把现场可用物理内存折算成可派发额度。额度不足或探测不可读时 spawn 以专用退出码 4 拒绝,输出 SPAWN_WORKER_MEM_BUDGET_DENIED 与 可用/预算/缺口 诊断;放行则在 PM 日志留下 SPAWN_WORKER_MEM_BUDGET: available=… budget=… slots=… 账本行。PM 收到该拒绝不得忙等、不得改走手动 spawn:本轮巡检把任务记 PARKED_FOR_MEMORY 并留存 probe 输出,下一轮巡检重试;同一任务连续 3 轮额度不足则正式泊车并向用户报告(附 probe JSON)。单进程堆顶(v2.20.0)管单个 worker 的失血点,本门管总量叠加承诺,也覆盖无堆顶可依赖的非 node runtime。数据源、预算推导、压力收紧与排队状态机读取 references/22-mem-budget-lane.md。
具名任务资源 profile:确有任务测量/范围与原 PM 串行合同的受控任务,可显式设置 SPAWN_WORKER_MEMORY_TASK_PROFILE。轻 Node、既有 Codex 宿主 followup 与限定 MiniMax CLI 使用同一多信号观察门;MiniMax 整 worker 仍至少 3 GiB、单 writer、20 秒稳定窗,实际 Node 堆顶不等于 RSS 上限。没有 profile 时保留旧默认;profile 不能用预算 0 或旧快照放行。字段与独立来源/身份复核见 资源合同。
验证负载纪律(verification lane):约束验证类别,不约束 worker 数量。worker 自验默认 scoped(单 spec / 定向用例 / --bail 1);全量套件单一在飞、跨项目互斥——确需 worker 现场跑全量时,先探测既有全量测试进程(如 pgrep -fl 'vitest|pytest|jest|go test'),无法确认独占就退避等待或改 scoped。探测是尽力而为的现场信号,不建跨项目锁文件,宁可少并发不可误并发;PM 收口时的全量复跑天然串行,是全量验证的默认归宿。验证输出一律重定向日志文件、只把有界尾部(如 tail -50)带回 terminal/session——长输出无界刷屏正是会话侧 node 运行时 OOM 的喂食管。PM 巡检发现系统负载飙升且多个 worker 同时在跑验证时,纠偏为错峰排队(等在飞验证收尾再放下一批),不砍 worker 数量。本纪律与单进程堆顶、物理内存 lane 互补:堆顶管单进程失血点,mem lane 管总量承诺,本纪律管验证执行的并发类别与输出体量(2026-09-05/06 事故中三者缺最后一环:并发全量自验同时制造了负载尖峰与超长输出)。
长周期 supervised 运行使用 references/23-runtime-settlement.md 的只读证据适配器核对 runtime 身份、逻辑终态、terminal、provider lease 与 Delivery。pm-orchestrate.sh reconcile --snapshot ... --output ... 只复算已捕获证据;退出码 0 不等于 complete=true。缺证和矛盾状态必须保留未结算,由有权限的 owner 执行所列恢复动作后重新观察。
PM 依次完成:
settle。pm-closeout.sh 的 PR-first 流程。PR 唯一性、冻结 head/diff/check/review、两阶段 mutation receipt、Monorepo integration path 与结果不确定恢复统一以 references/14-pm-orchestrate.md 为准。STATUS=done(或 PR 已合并)但进程仍存活的 worker——含跨会话遗留——当轮巡检即触发收口或上报:按 owner 通道 closeout / 释放 terminal / 结算 lifecycle,无法处置时向用户报告具体滞留对象。滞留进程持续挤占物理内存派发额度(references/22-mem-budget-lane.md),不收口就直接压缩下一波可派发 worker 数。bash scripts/pm-cleanup-worker.sh \
--project "$PROJECT" --worktree "$WT" --branch feat/worker-a --session worker-a \
--branch-lifecycle ephemeral-worker --integration-target integration/feature-a \
--pr "$PR" --expected-tip "$WORKER_TIP" \
--delivery-mode remote-pr --delivery-commit "$MERGE_COMMIT"
# 手工接管时先预览,再以同一精确参数加 --execute;pm-closeout 成功路径默认自动执行。清理必须区分源分支生命周期与合并目标:
ephemeral-worker:只有 exact PR head/base、expected tip、delivery commit、干净 worktree 和 settled lifecycle 全部一致时才清理。released、或可精确结算的 reclaimable / retained external 才进入后续清理。任何缺页、游标循环、重复/缺失 Dispatch、跨 Run 行或读取失败均保留 terminal、worktree 与分支。long-lived,或源分支等于 integration target:保留远端 ref、本地 ref 与固定 worktree,输出 RETAINED_WITH_REASON reason=long-lived-branch。long-lived 降级;短 Worker 合入长期分支时只清理 Worker head,绝不清理 integration target。CLEANED、RETAINED_WITH_REASON、CLEANUP_PENDING。交付已确认后,清理失败作为独立债务继续处理,不重跑 push/merge;隐去 CLEANUP_PENDING 后声称完全闭环属于 Hard Fail。Git 生命周期与批量 stale 分支清理由 git-workflow Skill 的“分支清理”入口负责;该 Skill 会按需加载自己的清理 reference。
先区分支持与日常选路:日常 worker 池保持 Claude Code/Codex 及其既有 provider 路由。ZCode CLI、MiniMax Code CLI 是重点可选支持,CodeBuddy、Qoder CN、千问办公为按需兼容;这些 backend(含旧 zcode driver)只有用户明确指定时才派发。不得因为余额、模型能力、与 PM 同名或主力额度不足自动选择它们;也不得写入 quota tier_policy 默认链。harness-backend-policy.json 的 dispatch_selection 是 PM 选择合同,hosts 只决定调用权限,不决定优先级。
在日常池内默认优先与 PM 同宿主,只有额度、模型能力或用户明确要求时跨工具。个人偏好写入 ignored 的 config/orchestration-personal.json,项目策略写入 .claude/orchestration.config.json;个人配置只能在 harness 白名单内选择 backend。
启用 quota_aware_routing 时,派单前必须用新鲜 summary 运行 route_suggest.py;summary 缺失、过期、lane 低于判停线、provider 不健康或未映射时,quota_preflight.py 在任何副作用前拒绝。显式 override 必须携带授权来源并写入 receipt。额度只为已经通过价值门的任务选路,不能生成 quota-burn 工作。模型与 lane 判断读取 references/01-model-selection-matrix.md 和 references/17-model-capability-profile.md。summary 合同的生产方不限;zcode lane 可用 scripts/quota_summary_zcode.py 把本机 zcode-quota 监测器的真实观测合并写入 summary(只更新 zcode lane、不改写其他 lane 的 generated_at,不接触凭证),数据流与合并语义读取 references/21-zcode-quota-producer.md。ZCode CLI driver 的启动绑定、配置隔离和证据边界读取 references/24-zcode-driver-safety.md。sub2api 网关的四条积分 lane(qwenworkai / lobsterai / autoclaw / codebuddy)可用 scripts/quota_summary_sub2api.py 从网关 /ui/api/quota 聚合端点拉取合并写入(只更新这四条 lane、其余 lane 原样保留;qw 每日 100 当日过期、lobster campaign 分项临期 → lane 记录带 remaining_total / credit_items,PM 派简单批量任务前先看临期分项,把当日过期积分在过期前吃掉),数据流与 lane 定义读取 references/24-sub2api-quota-producer.md。
需要账号级调度时,读取个人配置 account_routing。仅在 enabled=true、所选 backend 已获用户明确指定且包含在 backends 中时,读取本地 skill_path/SKILL.md 并依其入口取得新鲜账号与调度结果。未启用时不调用;启用但 Skill 缺失、观测失败或要求等待时,暂停该 backend。调用与停止边界见 references/29-local-account-routing-skill.md。公开 Skill 只提供调用合同,账号/卡规则及私人数据保留在本地 Skill;此调用是 PM 工作流,不能声称 spawn 已机械绑定账号。
独立 CLI 的标准启动参数、版本检查与权限边界读取 references/26-optional-cli-backends.md;千问办公的原生工具入口与 bundled coding 入口读取 references/27-qwenwork-cli-worker.md。ZCode CLI 原生 Orca supervised 启动的版本、可信环境桥、显式启用与接续合同读取 references/30-zcode-native-orca.md;显式选择 MiniMax Code 时默认要求 Orca terminal-managed;只有同时传 --no-orca-mode 才走直连 tmux,Orca 不可达或 lightweight 不得隐式回退。其余可选 backend 暂走 terminal-managed 或 tmux,未经真实生命周期验收不声明 supervised 已验证。ZCode CLI/MiniMax Code/Qoder CN/千问 bundled coding 的编排 hook 暂未集成,派发必须显式 --allow-prompt-only-install-guard "<授权来源>";prompt-only 不是机械 scope/安装保护,不自动扩大安装或 Git 权限。
系统依赖:Bash 4+、Git、jq、Python 3;PR 审计/收口需要 gh;tmux 仅回退路径需要;Orca 路径需要运行中的 Orca runtime 与版本匹配 CLI。按 backend 还需对应本地 CLI;显式任务内存 profile 另需已安装的原生 Node 与相应 CLI,缺失时拒绝并报告依赖,不自动安装。检查命令:
bash scripts/check-dependencies.sh --backend claude-code --backend codex --check-gh| 当前问题 | 读取 |
|---|---|
| 用户指定远端 ZCode GUI 作者直发文档/工程 PR(作者 safe-push + 唯一 draft PR,PM 只读核验) | references/38-zcode-gui-remote-pr.md |
| 标准快速派发(Claude Code + GLM/MiniMax + Orca,常规 worker 三步) | references/00-fast-dispatch-runbook.md |
| 模型、provider、执行模式 | references/01-model-selection-matrix.md、references/17-model-capability-profile.md |
| 依赖、checkpoint、Sentinel | references/02-runtime-dependencies.md、03-checkpoint-files.md、04-sentinel-design.md |
| 法律任务拆分、Issue 分组、并发事故 | references/05-legal-domain-patterns.md、10-parallel-lessons.md、12-issue-grouping.md |
| Agent Teams 排障、CLI backend | references/06-agent-cli-reference.md—11-agent-teams-troubleshooting.md 中对应 backend |
| Orca worker 与 PM 操作 | references/13-orca-cli-worker.md、14-pm-orchestrate.md |
| Autopilot | references/15-wave-autopilot.md、16-autopilot-durability.md |
| 派发、交付、review 与修复合同 | references/18-dispatch-acceptance-contracts.md |
| Orca Worker 429 批量巡检与错峰唤醒 | references/20-orca-rate-limit-recovery.md |
| zcode 额度 lane 的 summary 生产链路 | references/21-zcode-quota-producer.md |
| 独立 ZCode CLI、MiniMax Code、Qoder CN 与按需派发 | references/26-optional-cli-backends.md |
| 预建 long-lived 工作树首次接入、唯一 writer 与借用保留 | references/31-borrowed-existing-worktree.md |
| ZCode CLI 原生 Orca supervised、可信单次环境桥与交互接续 | references/30-zcode-native-orca.md |
| 独立 ZCode CLI 的 BigModel 套餐认证、模型切换与实测边界 | references/28-zcode-cli-bigmodel-coding-plan.md |
| 千问办公原生工具与 bundled coding CLI、账号边界 | references/27-qwenwork-cli-worker.md |
| ZCode CLI driver 的启动绑定、配置隔离与安全验证 | references/24-zcode-driver-safety.md |
| sub2api 四条积分 lane(qw/lobster/autoclaw/codebuddy)的生产链路与临期调度 | references/24-sub2api-quota-producer.md |
| 物理内存预算 lane 与派发排队 | references/22-mem-budget-lane.md |
| 远程节点 Worker 派发(SSH 桥 + 一次性 receipt + 容量/基线门) | references/25-remote-node-dispatch.md |
| 可选本地账号调度 Skill 的调用边界 | references/29-local-account-routing-skill.md |
| 只读核查 ZCode 会话/turn/最终 assistant 完成证据(GUI 任务监控方向,非实时活性权威) | references/zcode-session-evidence.md;运行 scripts/zcode-session-evidence.py(Python 3.9+ 标准库) |
| 只读GUI监控状态及一次性观察(READY仍需PM验收,无正式Orca监督) | references/zcode-gui-monitor-adapter.md、references/zcode-gui-observe.md;组合验证见 references/zcode-gui-evidence-pipeline.md |
| GUI交付三件套的校验分块传输及Git bundle消费者 | references/zcode-gui-artifacts.md |
| 修改本 Skill 后的验证 | references/19-maintainer-validation.md |
不要一次加载全部 references;只读取当前阶段与 backend 所需的文件。
出现以下任一情形,停止派发、接受、合并或清理,并保留可复查证据:
worker_done 与 settlement。WAKE_ACCEPTED 声称为额度或业务恢复。PARKED_FOR_REMOTE_CAPACITY)擅自回落本机派发;节点不可达(rc=65)回落本机必须是全新本机 spawn 重走全套门禁,不是复用远程上下文。修改本 Skill 后,按 references/19-maintainer-validation.md 运行受影响测试和完整回归。只有真实启动受支持 Agent 并观察 worker_done → Delivery → release/精确外部终端结算 → ack,才能把该 backend 的 supervised 路径标记为已验证;其他 backend 不得类推。
© cat-xierluo, MIT. Rendered from Markdown: HTML in the file is shown as text, images as links, and headings moved down two levels. Raw file
SKILL.md and 287 other files (scripts, references) in skills/multi-agent-orchestration of cat-xierluo/legal-skills.
Open the folder on GitHubat commit c077fcc
Multi Agent Orchestration next to the 5 skills that share the most tags, products or categories with it. Stars are the repository's; “used in” counts other GitHub owners with a copy.
| Skill | Stars | Used in | Tokens | Auto-check | Licence | Repo updated |
|---|---|---|---|---|---|---|
| Multi Agent Orchestration this skillcat-xierluo/legal-skills | 717 | — | ~6.6k | Automated safety check: Pass | MIT | |
| Huashu Agent Swarmalchaincyf/huashu-skills | 1.7k | — | ~576 | Automated safety check: Pass | MIT | |
| Project Session ManagerYeachan-Heo/oh-my-claudecode | 40k | — | ~4k | Automated safety check: Pass | MIT | |
| Polyphonyalinaqi/maggy | 707 | — | ~980 | Automated safety check: Pass | MIT | |
| Agtx Task Sweepfynnfluegge/agtx | 1.7k | — | ~1.7k | Automated safety check: Pass | Apache-2.0 | |
| Clawteamwin4r/ClawTeam-OpenClaw | 1.5k | — | ~3.1k | Automated safety check: Pass | MIT |
alchaincyf/huashu-skills
多Agent蜂群并行协作,纯git自组织,适合大型项目开发。当用户提到"蜂群模式"、"多agent"、"并行开发"、"agent swarm"时使用。
Yeachan-Heo/oh-my-claudecode
Creates isolated git worktrees, with optional tmux sessions, for PR reviews, issue fixes and feature work across projects and repositories.
alinaqi/maggy
Multi-agent orchestration with container-isolated workspaces — each agent session runs in its own Docker container with independent git branches
fynnfluegge/agtx
Breaks a conversation's results into feature-level tasks and pushes them to the agtx kanban board, where each task gets its own worktree and agent session.
win4r/ClawTeam-OpenClaw
Multi-agent swarm orchestration. An agent skill from win4r/ClawTeam-OpenClaw.
jnMetaCode/superpowers-zh
Sets up an isolated workspace before feature work or plan execution, preferring native worktree tools and falling back to git worktree, with instructions in Chinese.
cat-xierluo/legal-skills
Converts a lawyer's ordinary complaint or a described case into the Supreme People's Court's elements-style Word template, with layout checks on the result.
cat-xierluo/legal-skills
Analyzes raw lecture transcripts for verbal tics, pacing, time use and promise follow-through, with optional slide-by-slide comparison and cross-session tracking.
cat-xierluo/legal-skills
Detects and rewrites machine-sounding patterns in the body text of Chinese articles while keeping the author's facts, headings and legal terms intact.
cat-xierluo/legal-skills
Finds GitHub projects mentioned in articles or screenshots and stars them, tracks updates to your starred repos, and builds an HTML dashboard to browse them.
cat-xierluo/legal-skills
Sets up or incrementally updates AGENTS.md and CLAUDE.md for legal professionals, with a minimal safety baseline and a check that a new session loads and follows the rules.
cat-xierluo/legal-skills
Chinese-language skill that organizes a case file into a multi-role mock trial with judge, parties and clerk, producing a transcript, issue review and a to-strengthen list.
Categories
编排两个以上边界独立的本地 worker,使用 Orca Run/Task/Dispatch、独立 worktree/session 或 tmux 回退,由 PM 负责拆解、派发、巡检、429 停滞恢复、独立验收、PR 收口与临时资源清理;也用于用户明确要求“并行推进”“多个 worker”“PM 总控”“Wave Autopilot”或防止 PM…. Multi Agent Orchestration is an agent skill from cat-xierluo/legal-skills.
Multi Agent Orchestration fits situations like: tasks that involve Multi-agent orchestration; tasks that involve Git worktrees; tasks that involve Git workflow.
Run `npx skills add cat-xierluo/legal-skills --skill multi-agent-orchestration -a claude-code`. Or copy the skill folder (skills/multi-agent-orchestration in cat-xierluo/legal-skills) into .claude/skills/multi-agent-orchestration in your project. Claude Code loads it when a task matches its description.
Run `npx skills add cat-xierluo/legal-skills --skill multi-agent-orchestration -a codex`. Or copy the skill folder (skills/multi-agent-orchestration in cat-xierluo/legal-skills) into .agents/skills/multi-agent-orchestration in your project. Codex loads it when a task matches its description.
Cursor, Gemini CLI, GitHub Copilot and OpenCode also load SKILL.md folders. With the skills CLI, run `npx skills add cat-xierluo/legal-skills --skill multi-agent-orchestration -a cursor` (or -a gemini-cli, github-copilot or opencode for the others). To copy it by hand, put the folder in .cursor/skills/multi-agent-orchestration, .gemini/skills/multi-agent-orchestration, .github/skills/multi-agent-orchestration and .opencode/skills/multi-agent-orchestration in your project.
Going by SKILL.md and its folder, Multi Agent Orchestration needs a shell for the scripts in its folder and the command-line tools its instructions call (bash, python3, git, gh, jq and node). Our summary lists: Python 3; A Bash shell.
SKILL.md contains no URLs. Its commands use git and gh, which can reach the network depending on how they are called. This is read from the text; nothing was executed.
Our automated static check of SKILL.md found no risky patterns, such as piping downloads into a shell, reading credential files or hidden Unicode. It is not a guarantee. The check reads SKILL.md only: the scripts in the folder are not scanned, so read them before running anything.
Multi Agent Orchestration is published under the MIT licence (declared in SKILL.md). It allows redistribution, so the full SKILL.md is shown on this page.
About 6.6k tokens (SKILL.md is roughly 26k characters). Agents keep only the skill's name and description in context until a task matches; then they load SKILL.md in full. Its references folder adds about 143k tokens, read only when the agent opens those files.
Skills that share tags, products or a category with Multi Agent Orchestration: Huashu Agent Swarm (alchaincyf/huashu-skills, 1.7k stars), Project Session Manager (Yeachan-Heo/oh-my-claudecode, 40k stars), Polyphony (alinaqi/maggy, 707 stars) and Agtx Task Sweep (fynnfluegge/agtx, 1.7k stars). The comparison table on this page puts their stars, adoption, token cost, safety result and licence side by side.
cat-xierluo (a GitHub user) maintains it in cat-xierluo/legal-skills, which has 717 GitHub stars. The repository holds 62 skills in this directory. The repository was last updated on October 8, 2026.
Source: cat-xierluo/legal-skills on GitHub. Facts on this page come from the repository at the commit we read; the author's words are quoted as theirs.