Show and tell: orchestration-skill — 长任务不该贵到必须用前沿模型 / Long tasks shouldn't cost like they need a frontier model

[pridesong]

Show and tell: orchestration-skill — 长任务不该贵到必须用前沿模型 / Long tasks shouldn't cost like they need a frontier model

orchestration-skill 是一个给 DeepSeek Harness 的 Agent Skill：让廉价模型跑出前沿模型的长任务可靠性——把可靠性从模型身上搬进结构里。

仓库：https://github.com/pridesong/orchestration-skill
发布：v0.3.0 · https://github.com/pridesong/orchestration-skill/releases/tag/v0.3.0
类型：Agent Skill（SKILL.md + 纯 Python 标准库脚本，零依赖）

核心逻辑

长任务失败不是因为模型在某一步弱，而是不可靠在长链上复利。与其花钱买更强的模型把整条链装进它的上下文，不如把长链拆成廉价模型轻松胜任的小步，让链的可靠性由机械结构承担：

• 状态在磁盘不在上下文 — 每步写 artifacts/<field>.json；断点续跑、零上下文损耗、只重跑失败的那一步
• 每个边界有机械门禁 — validate/check/compare 机械拦截幻觉与格式错误，不信任模型自检
• 多次廉价调用 > 一次昂贵调用 — 编排者 + 独立审计 + 动态审计，多脑子补足单次推理弱
• 零引导语派发（T3） — subagent prompt 是 regex 锁定的文件引用（T3FILE:v1 读取 <file> 并按内容执行）；散文无法包裹协议，模型没有漂移空间

实测证据（廉价模型全链路）

一个完整医药供应链任务（PVG→EZE 温控空运，1000kg，$11,500 硬预算）在廉价模型上全链路跑通：11 个字段（报价/海关/天气/航线/成本NPV/仪表板/审计/结论）。

降本回退对照实验展示了结构带来的差距：

	mind-reason（方案内推导）	mind-crusher（翻墙）
总成本	$15,730	$12,676
预算缺口	$4,230	$1,176（收窄 72%）
预算内可行运量	500kg（50%）	≈888kg（89%）

且当回修经由运行时 mind-decider（读审计证据——固定成本墙 vs 单价墙——再选思维）路由时，精确复现 $12,676：选择是机械可复算的判别决策，不是散文运气。

关键机制

• 模块装配：编排者是装配师不是协议设计师，从模块库挑模块、连线、设路由，模块自带协议
• op-table 由声明展开：executor 零推导，next/parallel/routing 展开为排它条件路由表；并行组写同一行 state.csv（多条件 AND）
• T3 零引导语派发：六件套协议（fill/rules/schema/data/write/forbidden），派发 prompt 被 regex 锁死
• mind-decider：复杂回退不静态绑死思维，策略师判别式运行时读审计证据选重试思维
• 两层审计 + 失败回流：静态（独立审计装配表）+ 动态（3 次同类失败 → 重新编排）
• 零依赖：纯 Python 标准库，Python 3.7+ 即可运行

文档

• SKILL.md — 协议（给 LLM，schema 驱动契约，派发 prompt pattern 锁死）
• Description.md — 人类阅读版（理念/架构/工作流）
• README.md — 卖点定位 + 实测证据
• 5 个示例：demo-task（线性）/ bad-example（负例）/ eco-analysis（研究）/ quant-adaptive（并行组）/ dsh-client（判别路由）

---

English summary: orchestration-skill is an Agent Skill for DeepSeek Harness that makes cheap models deliver frontier-model reliability on long tasks by moving reliability out of the model and into mechanical structure (disk-resident state, mechanical gates at every boundary, multi-brain audit, regex-locked zero-prose T3 dispatch). Evidence: a full pharma supply-chain task (11 fields) ran end-to-end on a cheap model; a cost-reduction rework experiment cut the budget gap by 72% ($12,676 vs $15,730), and the runtime mind-decider reproduced the result exactly. Pure Python stdlib, zero dependencies. Feedback welcome!

[zoahdev]

Strong alignment with the verification-first philosophy ? endorsing and cross-linking.

Key overlap:

• State on disk, not in context ? dsh-rule-evolve: rules/experience live in JSONL + AGENTS.md, not in the prompt (https://github.com/zoahdev/dsh-rule-evolve)
• Mechanical gates, don't trust self-checks ? dsh-plugin-doctor's preflight (tool-verify gates tools before they ship; 8 checks incl. side-effect lints)
• Bounded context ? the retrieval-efficiency contract (Glob/Grep/Read, output caps) ? https://zoahdev.github.io/dsh-tutorials/zh-15-retrieval-efficiency.html

One complement worth adding: after a long task succeeds, extract the lessons into verified rules so the NEXT long task starts smarter ? that's the extract/evolve half of dsh-rule-evolve. Happy to link your repo in the ecosystem map (dsh-ecosystem).

[hytime]

这种 superpower 早有了一套规范长任务处理了

[pridesong]

Thanks @zoahdev — the alignment is real, and your extract/evolve half is exactly the missing piece I'd love to wire in.

Quick mapping of our mechanisms to your three overlaps:

• State on disk — ours goes further: not just rules in JSONL, but every field artifact is a file (artifacts/<field>.json) plus state.csv per-step rows. The disk is the external memory; a resumed session reads the last row and continues with zero context loss.
• Mechanical gates — same philosophy. Ours are validate.py / check (output schema + routing membership + evidence-in-source) at every boundary, plus a two-layer audit (assembly audit + dynamic 3-failure re-orchestration).
• Bounded context — each dispatch carries only the current step's inputs (T3 six-piece payload), not the whole task.

On the complement you suggest (after success, extract lessons into verified rules so the NEXT task starts smarter): we currently have a failure feedback loop (failed traces flow back into templates so the engine assembles each task type better), but it's implicit. Your extract/evolve model — turning that loop into explicit audited rules with tests — is a natural next layer. I'd like to study dsh-rule-evolve's rule format and see if the revise_* routing outcomes (which already carry failure_reason + target_ref back to the field) can feed it directly.

And yes — happy to be listed in the dsh-ecosystem map. Repo: https://github.com/pridesong/orchestration-skill. If there's a PR/format for the catalog entry, point me at it and I'll submit.

[pridesong]

@hytime 完全同意 superpower 在长任务规范上是先行者，这一点没有争议。我们的定位不是替代它，而是回答一个它没覆盖的问题：长任务能不能不依赖贵模型？

差异点其实很具体（可验证，不是口号）：

1. 模型约束不同——superpower 类方案跑在旗舰模型上（上下文大、指令遵循强），靠模型能力扛长链。我们全程跑廉价模型（deepseek-v4-flash），实测两个完整长任务（pharma-scm-001 7 轮、hvac-chicago-001 11 步）18 步零失败零重试。这不是"更强"而是"更便宜"。
2. 可靠性来源不同——superpower 靠 prompt 规范 + 模型遵循；我们把可靠性搬进结构：T3 派发 prompt 被 regex 锁死（零散文滑行空间）、每字段产物过机械门禁（schema/路由/证据）、状态在磁盘不在上下文。模型只做小步，长链由 op-table + 门禁保证。
3. 公开证据——实测数据都在 README：降本回退对照 mind-crusher $12,676 vs mind-reason $15,730（缺口收窄 72%），且经运行时 mind-decider 选思维精确复现。

如果 superpower 已经有覆盖这些的方案（尤其廉价模型 + 零散文协议的组合），请指个链接，我认真对比学习；如果没有，那正是这个项目想补齐的空位。