agentsclimarketplace

Codex internal agent orchestration

Skill FairladyZ625/codex-internal-agent-orchestration

Codex internal agent orchestration for projects using Harness Anything.

Install
npx -y skills add FairladyZ625/codex-internal-agent-orchestration

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • 25 days oldThe repository was created 25 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
  • 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Coordinate complex coding and knowledge-work tasks inside a user's target repository with Codex's current main agent and in-task internal subagents. Use for multi-agent task graphs, parallel exploration, disjoint implementation slices, long-running supervision, evidence collection, adversarial review, final integration, and Harness Anything task/decision/fact workflows when the target repository has adopted Harness Anything.

SKILL.md

7.4 KB, ~2.4k tokens by cl100k_base, as published. Nobody here has run it

Codex 内部 Agent 编排

执行范围

使用当前 Codex 主 Agent,以及当前任务内通过 Codex 协作工具启动的内部 subagent。仓库本地命令、测试和构建用于执行;目标仓库接入 Harness Anything 时,用 ha 记录任务、决策、事实与证据。内部 subagent 工具不可用时,由主 Agent 本地完成或向用户说明能力限制。

先判定是否派活

依次检查:

  1. 任务成立吗? 先读原始材料或做最小测量,证伪错误前提比执行错误任务更重要。
  2. 主 Agent 已能写出确切 diff 吗? 已知改动且路径短时直接做,避免把委托固定成本包装成管理。
  3. 委托总成本更低吗? 计入 packet、等待、回读、冲突、集成和复验。
  4. 是否存在长尾价值? 大量侦察、试错、测量、上下文卸载或独立复核才值得派活。
  5. 边界能独立验收吗? 每条线必须有清晰语义面、依赖、协调边界和可观察证据。
  6. 同一路径已失败两次吗? 连续两次没有新证据就熔断,由主 Agent 接管、缩小任务或向用户升级。

不为填满并发槽派 Agent。用户指定数量时遵从上位指令,但仍使用最小写面并说明成本或冲突。

派活前建立任务图

首个派活前写出最小 orchestration brief:

  • Mission:目标、非目标、原始依据、真实使用入口。
  • Graph:任务、依赖、关键路径、可并行节点。
  • Ownership:每条线唯一 owner、预期写面、冲突/保护面、集成 owner。
  • Budget:Agent 数、并发、时间或尝试预算,以及取消、接管和升级条件。
  • Acceptance:主 Agent 将亲自回读的 diff、测试、日志、截图或真实入口。

默认按一条连贯语义线切成一个完整包,不预防性切成大量微任务;只有依赖、风险隔离或写面冲突要求时再拆细。读取可以重叠;高耦合代码保持单写者。

每个委托包必须回答五问:ContextRequestOutput FormatConstraintsCheckpoint。Context 至少给 canonical 原文、why/首位使用者、验收者视角、系统地图和真实验收入口。预期文件只是地图,不是信息围栏;Worker 可读全仓理解根因,但不得触碰正在冲突或受保护的写面。模板见 references/delegation-templates.md

选择内部角色

  • 主 Agent / CEO:目标、关键路径、任务图、用户沟通、集成、语义验收和最终裁决。
  • Explorer:只读定位、枚举风险、比较方案;提供路径与证据,不改文件。
  • Worker:完成边界清晰的实现或验证切片;只写授权范围。
  • Reviewer:独立只读挑战隐藏假设、失败模式和关键 diff;不替主 Agent 验收。

通过当前运行时的 Codex 内部协作工具管理这些角色。完整工具映射、并发规则和失败恢复见 references/runtime-channels.md,角色红线见 references/role-matrix.md

运行循环

  1. 主 Agent 先推进关键路径或当前阻塞项,再派不重叠旁路线。
  2. 派出后继续做本地工作,不为等待 Agent 空转。
  3. 用消息补充上下文或纠偏;不要为同一工作反复新建 Agent。
  4. 优先消费完成事件;只在确有运行中任务时等待或低频检查。
  5. 把 Agent 结果先视为 claim;主 Agent 回读 ground-truth 后才成为 evidence。
  6. 按依赖顺序集成并运行 fresh verification;各线分别通过不等于整体通过。
  7. 完成、需用户裁决或自主面穷尽时,回收/打断遗留 Agent 并一次性收口。

Reference 路由矩阵

按当前命中的场景读取,不要一次性加载全部 Reference。

触发场景必读 Reference用它解决什么
准备首个派活包或 orchestration briefdelegation-templates.md五问契约、任务图、Explorer/Worker/Reviewer prompt 和收口格式
角色、权限、ownership 或写面边界不清role-matrix.md + worker-handbook.md谁负责什么、哪些边界不可越、Worker 如何开工和回传
创建、补充、等待、打断或并发管理内部 Agentruntime-channels.md协作工具映射、并发预算、共享工作区、监管与失败恢复
目标仓库启用 Harness Anything,或准备执行任何 ha 写操作harness-orchestration.mdcapabilities 驱动的 CLI 发现、原语映射、权责、证据和收口门
需要相信“已写入”“无发现”“已通过”或“与改动无关”ground-truth-discipline.md产物回读、全文/全量核验、终态判断、阳性与阴性对照
集成、review、complete 或最终交付acceptance.md功能、语义、编排、对抗、使用性验收与原始语义勾核
修改本技能skill-maintenance.md内容边界、更新流程、行为场景和反熵增规则

Harness 投影

本技能用于在用户自己的目标仓库中使用 Harness Anything,不携带 Harness Anything 源码仓库的开发流程。检测到 harness/harness.yaml 或目标仓治理明确要求时,仅在已有 task、非平凡多协作者工作、跨回合监管或承重证据需要时投影任务图。短小单步任务不因 Harness 存在自动新建 task。

每次通过 ha capabilities --json、各原语的 capabilities 与 ha <kind> --help 发现当前命令面,不背诵固定参数。主 Agent 自己也必须使用要求 Worker 使用的记录系统;排期、裁决、事实与关系不能只留在聊天里。具体协议见 references/harness-orchestration.md

主 Agent 验收

主 Agent 必须亲自完成:

  1. 功能复核:实现、测试和声明一致。
  2. 语义验收:结果符合原始目标、范围与非目标。
  3. 编排复核:任务图已收敛,无遗漏依赖、冲突写面或未回收 Agent。
  4. 对抗验证:隐藏假设和高风险异议已被挑战与裁决。
  5. 使用证明:真实入口可用,冷启动路径成立。

未验证项必须显式标注。写回执、diff、分页列表和状态快照都不是终态证据;关键结论按 references/ground-truth-discipline.md 核产物、全文、全量与对照。完整验收见 references/acceptance.md,一线纪律见 references/worker-handbook.md

维护

修改本技能前先读 references/skill-maintenance.md。正文只保留高频决策门、执行范围和生命周期;模板、Harness 细节与场景规则下沉到一层 references。

What ships with it: 12 files

28.3 KB alongside SKILL.md

agents/

Gives 0 of the 12 instructions most agent orchestration skills give in ~2.4k tokens

Counted across 742 of the 995 authors here whose files we hold, read 2026-08-07

  • Reference existing artifacts by path or URLin 53 of 742, across 25 files
  • Run the full test suite after integrating changesin 51 of 742, across 19 files
  • Dispatch one agent per independent problem domainin 50 of 742, across 17 files
  • Verify fixes do not conflictin 45 of 742, across 13 files
  • Include a suggested skills section in the documentin 45 of 742, across 17 files
  • Redact sensitive informationin 41 of 742, across 11 files
  • Save to the temporary directory of the operating systemin 39 of 742, across 10 files
  • Tailor the document to user-provided focus argumentsin 39 of 742, across 9 files
  • Spot check agent changes for systematic errorsin 34 of 742, across 7 files
  • Write a handoff document summarising the current conversationin 31 of 742, across 6 files
  • Assign each agent a specific scopein 23 of 742, across 8 files
  • Provide specific scope and clear goalin 23 of 742, across 5 files

Said here and by no other author read

  • Falsify false premises before executing tasks
  • Write a minimal orchestration brief before delegating
  • Answer five questions for each delegation packet
  • Keep highly coupled code under a single writer
  • Manage roles using current runtime collaboration tools
  • Verify ground-truth evidence after receiving claims

Grouped from the skills themselves: near-identical wordings counted once, and counted by distinct author, so one author publishing three of these counts once. Length counted with cl100k_base; the agent that loads this file may tokenize it differently.

Keep looking

Skills are one crate of 327,132. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.