Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add martin1847/evolab --skill orchestrator-coregit clone --depth 1 https://github.com/martin1847/evolabWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/martin1847/evolab/orchestrator-core)<a href="https://agentmods.dev/skills/martin1847/evolab/orchestrator-core"><img src="https://agentmods.dev/badge/skills/martin1847/evolab/orchestrator-core/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/martin1847/evolab/orchestrator-core"><img src="https://agentmods.dev/badge/skills/martin1847/evolab/orchestrator-core.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00209 | $0.03926 |
| Opus 5 | $0.00105 | $0.01963 |
| Sonnet 5 | $0.00042 | $0.00785 |
| Haiku 4.5 | $0.00021 | $0.00393 |
Grade A, and why
orchestrator-core scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 9d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 142 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Orchestrator Core — 编排内核(领域无关)
这是内核,不是编排者。 它抽自两个跑通的实例——
cto-orchestration(写码团队)与yangge-ops(内容编辑部)——减去领域词后剩下的不变式。任何新领域编排者 = 本内核 + 一张薄皮。 造 CEO / 销售 / 研究编排者,别新建一套,贴皮(§10)。 复杂度守恒:内核没消灭编排的复杂度,它把复杂度从每个领域的 worker 上下文,压进了契约和合并两处(§3/§5)。
0. 内核的一句话
先编排,再工作——编排是 agent 的操作系统:任务进来先过「派 or 不派」判据(§2),亲手执行是调度的一种结果,不是默认姿态。 编排者持有判断,不持有产物;持有薄的萃取上下文,不持有脏数据;把复杂度压进契约与合并,把简单留给主 session。
分层不变式一句:Deterministic workflow outside, bounded intelligence inside, evidence at
the boundary — and intelligence periodically attacks the workflow.(确定性 workflow 在外,
受约束的智能在内,边界用证据交接——智能定期反攻 workflow。)流程、状态、验收、重试写成
确定性机制;判断力只花在编不出来的地方;层与层用证据交接、不用信任交接;机制本身会静默
腐坏,须由智能定期以异源变异攻击。
本内核的常驻 digest(6 行,可直接并入全局规则文件让每层 agent 常驻加载)见 references/resident-digest.md。
1. 铁律一:编排者不亲手做领域产物
编排者的产出是契约、调度、裁决、状态落盘,不是领域产物本身。
- CTO 皮:不写产品代码。|总编皮:不写正文。|未来皮:CEO 不亲手谈单、研究编排者不亲手读论文。
- 判据不是「能不能做」,是「做了会不会占掉编排位的稀缺性 + 污染主上下文」。你能做,恰恰是最该派的信号。
2. 铁律二:切分轴是「要不要主上下文」,不是耗时
派 or 不派,只问一句:这活需不需要贴着我此刻这摊上下文才能干?
- 不需要(自包含、可蒸馏、脏数据密集)→ 派出去,越并行越好。哪怕它只要 10 秒(读 170KB digest 找三条=该派,因为它污染主上下文)。
- 需要(要贴着主线反复迭代、共享假设)→ 留在主 session,哪怕它要一小时(贴着确认门写正文=不派,派出去 worker 拿残缺假设各干各的,合并全是冲突)。
- 「超过 N 分钟就派」是相关变量、不是判据——耗时只是「自包含」的影子。拿影子当判据会在两个边界误伤。
- 读并行,写单线(切分轴的第二维):调研/评审/取证类读操作尽管 fan-out;产物写入保持单一 writer——并行写的隐式决策会互相冲突、合并全是撞车,要并行写先按产物所有权(文件/章节/worktree)切开。
- 分形递归:本判据对派出的每层 worker 同样生效——子任务中再冒出「自包含、结论小」的块,同判据继续再派。护栏:每层 scope 只收不扩、契约逐层更窄;只用于天然树状且每层返回极小的活(per-finding 验证类),1-2 层收敛——每多一层就多一次蒸馏压缩的失真。
- CTO 皮=评审 agent 对每条 finding 派独立 verifier;总编皮=主笔对资料检索/事实核查子块再派。
3. 铁律三:派发即契约(缺一即漂移)
派出去的每个 worker 必须拿到四件,缺一它就飘:目标 / 输出格式 / 工具与来源边界 / 任务范围(含 stop-and-report)。
- 散是一句话的事,写契约是真功夫——这是编排者压进来的第一处复杂度。
- 规模按复杂度分档:简单事实 1 个 · 对比 2-4 · 开放调研 10+——防为简单任务过度 fan-out(编排的前提是任务价值撑得起协调开销;撑不起就单 agent 直接做)。
- 附理解门:worker 动手前先复述「碰哪些东西、有哪些风险、scope 是什么」。 默认路径(合同未承诺核对回合)下复述不是握手——当场纠正结构性够不着, 纠错的真实抓手在派发前评审与交付后评审;复述的价值是留痕:事后审计锚, 且经结构化通道上报(阻塞或合同门)时作固定开头 contract-as-read,编排者可快速裁决。 承诺了核对回合的合同走下一条分叉。
- 要不要等放行,看合同承诺、不看谁在场:派发合同显式承诺了开工前的核对回合 → 复述后等它; 没承诺(默认)→ 无放行可言,复述只作留痕给编排者事后核,复述完即开工。判据 worker 动手前 读合同就能算,不必推断谁在不在场——「有没有人」是环境属性,它决定不了合同欠你什么。
- 门怎么说出去,看通道:worker 能直接问人 → 当场问;够不着人(headless)→ 合同里的 STOP 门必须落 结构化通道(写文件 / 状态标记,由 runtime 投成一个终态)。两条正交:通道只决定门怎么表达,决定不了 门要不要停——靠「等交互」表达的门在够不着人时等于没有。
- CTO 皮=goal 文档模板;总编皮=选题官/主笔各自的任务契约(证据路径+脱敏红线+输出格式)。
What ships with it
3 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 9d ago First seen · 142 lines · 209 tokens per session scan A 390183c09ac1
orchestrator-core is a skill published in the GitHub repository martin1847/evolab (6 stars, last pushed today), licensed MIT. It adds 209 tokens to every session and 3,926 once invoked, about $0.0010 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
solid-principles
SOLID principles checklist with Java examples. Use when a class has too many responsibilities, an abstraction leaks, or a dependency points the wrong way, and when the user asks about Single Responsibility, Open/Closed, Liskov, Interface Segregation or Dependency Inversion. For naming, duplication and method length…
security-audit
Java security checklist covering OWASP Top 10, input validation, injection prevention, and secure coding. Works with Spring, Quarkus, Jakarta EE, and plain Java. Use when reviewing code security, before releases, or when user asks about vulnerabilities.
issue-triage
Triage and categorize GitHub issues with priority labels. Use when user says "triage issues", "check issues", "review open issues", or during regular maintenance of GitHub issue backlog.
ln-22-codebase-auditor
Audits cross-cutting code health, security, delivery, and maintainability when no specialist audit is primary. Not for a single delivery review.
ln-31-performance-optimizer
Profiles and improves a measured latency, throughput, CPU, memory, or I/O problem. Not for speculative tuning or cosmetic refactoring.
ln-51-opportunity-evaluator
Evaluates new product opportunities using demand, channels, competition, and economics. Not for committed-backlog or implementation planning.