Index of every page on this site: llms.txt
All skills
cat-xierluo avatar

/multi-agent-orca

@5fb69a0
Run it yourself

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。

Use this Skill: https://skilld.dev/gh/cat-xierluo/legal-skills/multi-agent-orca

Nothing lands on disk. Nothing to clean up.

Fork this Skill

Edit a local copy. It keeps the author and licence.

referenceszcode-gui-evidence-pipeline.md

≈1.7k tokens on demand. Your agent reads this file only when SKILL.md points to it.

ZCode GUI Evidence Pipeline — 合成集成反例矩阵

目的

test-zcode-gui-evidence-pipeline.py 与 zcode-gui-pipeline-fixtures.py 构成一套可移植、纯标准库的组合回归资产:用合成 SQLite 反例矩阵驱动 真实的 collector 与 adapter 子进程,端到端检测两类组合缺陷:

  1. 误判完成:把未完成/失败/冲突的会话映射成 READY_FOR_PM_REVIEW 等交付态,或让较早的成功覆盖最新的错误;
  2. 泄漏:reasoning、tool、status_reason 文本、其他会话(decoy)内容 出现在 collector/adapter 的 stdout 中。

两者此前只有各自的单元测试;本套件覆盖真实组合路径(SQLite → 只读 collector 元数据文件 → adapter),且不镜像单元测试的构造方式。

冻结生产依赖(只读消费,不修改、不复制、不提交)

角色 脚本 冻结版本
collector skills/multi-agent-orca/scripts/zcode-session-evidence.py PR250 head 554a60c1ea27155ad03ec0e2b558d70a8fbb7c1f
adapter skills/multi-agent-orca/scripts/zcode-gui-monitor-adapter.py PR253 head 9f77ef0d51cd7b240f512da5c7bfc8257490f98b
negative control 同 adapter 路径的历史 initial 版本 commit 3ed567971a7c6c41612d2cd611aef25ef9f8117f

runner 按 CLI 参数 > 环境变量 > 同目录兄弟文件 的顺序解析依赖; 三者都不可用时整个套件输出 SKIP 并以 exit 77 结束,绝不静默通过。 negative control 所需的历史 adapter 由 runner 通过 git cat-file blob 3ed5679...:<path> 自动提取到私有 workdir(不可达时仅 跳过该一项控制,不影响其余 case);也可用 --legacy-adapter <path> 显式指定,off 关闭。

用法

# 无依赖时的默认行为:SKIP, exit 77
python3 skills/multi-agent-orca/scripts/test-zcode-gui-evidence-pipeline.py

# 注入真实冻结依赖(CI 消费命令)
ZCODE_GUI_COLLECTOR=/path/to/zcode-session-evidence.py \
ZCODE_GUI_ADAPTER=/path/to/zcode-gui-monitor-adapter.py \
python3 skills/multi-agent-orca/scripts/test-zcode-gui-evidence-pipeline.py \
  --workdir /private/tmp/zcode-gui-pipeline-run

# 仅列出反例矩阵
python3 skills/multi-agent-orca/scripts/test-zcode-gui-evidence-pipeline.py --list-cases

# 单独物化一个合成库(人工检查用)
python3 skills/multi-agent-orca/scripts/zcode-gui-pipeline-fixtures.py \
  --case completed_latest_error_wins --out /tmp/case.sqlite

退出码:0 全部预期断言成立;1 存在真实失败;77 依赖缺失 (SKIP)。negative control 中旧 adapter 的预期拒绝记录为 negative control,不计入候选失败;但若旧缺陷未能复现(旧 adapter 通过了 admitted/no-turn 反例),套件按真实失败处理。

输出:stdout 仅打印有界 summary(每 case 一行 PASS/FAIL/SKIP + 汇总);逐条断言与全部进程输出写入 <workdir>/detail.log 与 <workdir>/summary.json(私有路径)。--workdir 缺省为新建临时目录。

反例矩阵(18 个具名 case)

每个 case 使用独立合成 SQLite(schema 含 collector 必需列;行内容 全部为合成值),并植入诱饵会话 sess-decoy-other-session 与 CANARY-* 令牌。统一断言:collector/adapter 输出精确 state 或精确 错误码 + exit、adapter flags 恒 false(pmAccepted/orcaSupervised/ livenessAuthoritative)、DB 运行前后 SHA-256 不变且无 journal/wal 残留文件、metadata 只携带绑定的合成 session、全部 canary/decoy 令牌 不外泄。

Case 构造 关键断言
admitted_no_turn admitted、无 promoted、无 turn adapter INPUT_ACCEPTED,provider/model=unknown
promoted_no_turn promoted 但 turn 从未开始 TURN_UNKNOWN
turn_running running turn + 部分 assistant(仅 reasoning part) TURN_RUNNING,turnId 回显
completed_ready_for_review completed + 完整无错 assistant READY_FOR_PM_REVIEW;--expected-evidence-sha256(真摘要)+ provider/model 期望全部成立
completed_latest_error_wins 同 parent 两条 assistant,较早成功 seq=10,最新 error seq=20 精确选中 msg-asst-latest(errorFree=false),DELIVERY_ERROR;较早成功不覆盖
turn_cancelled 用户取消 turn TURN_CANCELLED
turn_error error turn、无 assistant TURN_ERROR
provider_missing_unknown_ok completed 但缺 providerId/modelId 保持 unknown 且不阻塞 review gate
expected_provider_mismatch 错误的 --expected-provider exit 2 PROVIDER_MISMATCH
collector_input_not_found 错误的 input 绑定查 collector exit 1 INPUT_NOT_FOUND
adapter_binding_mismatch 元数据与另一 session id 送 adapter exit 2 BINDING_MISMATCH
multi_turn_ambiguous 同一 promoted message 两条 turn_usage exit 1 TURN_AMBIGUOUS
assistant_sequence_ambiguous 两条 assistant 并列最大 sequence exit 1 ASSISTANT_AMBIGUOUS
bad_schema_missing_table 缺 part 表 exit 1 DB_SCHEMA_INVALID
bad_json_message_data assistant data 非法 JSON exit 1 JSON_MALFORMED
canary_containment status_reason/reasoning/tool part 植入 CANARY,诱饵文本 CANARY 两次运行(含 --include-final-text)均零泄漏,仍 READY_FOR_PM_REVIEW
digest_mismatch 过期期望摘要 exit 2 EVIDENCE_DIGEST_MISMATCH
legacy_nullable_negative_control 复用 admitted/no-turn 库 历史 initial adapter 3ed5679 以 exit 2 EVIDENCE_TYPE_INVALID 拒绝 collector 合法 null turn 标量(旧缺陷复现);固定 9f77 版本对同一证据返回 INPUT_ACCEPTED

安全保证

  • 合成库全部生成在 --workdir 下;runner 断言 db 路径不越出 workdir。
  • collector 以 mode=ro + PRAGMA query_only 打开数据库;套件对每个 case 断言库文件 SHA-256 与目录文件集合运行前后不变(无写入、无 journal/wal 残留)。
  • 不访问网络、不启动服务、不触碰真实用户数据库;所有进程均在 90s 超时内本地运行。
  • 泄漏扫描覆盖 collector/adapter 的 stdout 与 stderr(含 --include-final-text 变体)。

依赖

纯 Python 3.9+ 标准库(sqlite3、subprocess、hashlib、importlib), 无需安装任何第三方包。negative control 的自动提取需要本 checkout 的 git 对象库可达(浅克隆时可改用 --legacy-adapter 显式提供历史脚本, 或接受该单项 SKIP)。

Source: SKILL.md on GitHub

skilld matched fixed text patterns in SKILL.md and file names. Patterns miss obfuscated code.

skilld run checks every file with the same patterns. It asks for approval before it loads a Skill with a behavior marked Needs approval.

1 alert18d3 checks · Risk HIGH
  • Gen Agent Trust Hub18d

    This skill is a highly sophisticated multi-agent orchestration framework designed to manage and parallelize local AI worker tasks using tools like Orca, tmux, and SSH. While it incorporates an extensive array of security features—including automated installation guards, file write scope guards, and a structured authority receipt system—it fundamentally relies on broad system permissions. It performs extensive command execution, dynamic script loading, and remote execution on network nodes via SSH. Users should review the configuration and security policies (e.g., config/harness-backend-policy.json) to ensure they align with their environment's security requirements.

  • Socket18d

    2 alerts: gptAnomaly

  • Snyk18d

    Risk: LOW · No issues

Signed by skilld at 5fb69a0. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub yesterday.

Activeupdated last month
Other metadata
metadata
{
  "version": "3.1.1",
  "homepage": "https://github.com/cat-xierluo/legal-skills",
  "author": "杨卫薪律师(微信ywxlaw)"
}

README badge

README badge for cat-xierluo/legal-skills/multi-agent-orca