In one sentence: DeepSeek published dsh-superagent on 2026-10-31 — an MIT-licensed research prototype that turns one GitHub repository into a work queue for an agent: it listens to GitHub (optionally Feishu), opens a Ticket for each authorized request, works in its own session and directory, and answers in the original thread with a pull request. The README states plainly that it is not an official DeepSeek product, and it cannot be installed by package name.
Every row below can be checked against the repository or its docs; read on 31 October 2026. The README also says the code was adapted for publication, so it is not identical to the internal version or to the implementation described in the article Exploring Super-agent Design.
| Item | Official value | Where it says so |
|---|---|---|
| Repository | deepseek-ai/dsh-superagent-prototype — "Towards a More Autonomous Agent System" | GitHub repo page |
| Created / default branch | 2026-10-31T09:32:46Z, main | GitHub API metadata |
| Licence | MIT | LICENSE + package.json |
| Version / publishability | 0.1.0, "private": true — not an npm-installable plugin | package.json |
| Author field | DeepSeek Harness Team <[email protected]> | package.json |
| Package shape | DSH bundle: dsh.bundle.patch → ./cordis.patch.yml; a web client; entries ., ./client, ./github-watch, ./feishu-watch | package.json exports |
| Languages in the tree | TypeScript, Python, CSS | GitHub language bar |
| Docs / prompts / skills | 9 numbered docs (000–008), 14 prompt files, 2 bundled skills (dsh-superagent-github, dsh-superagent-feishu) | Repository tree |
| Runtime engines | Node ^22.19.0 || >=24.0.0; dsh ^0.2.1-alpha.2; pnpm 11.7.0 | package.json engines |
| Two runtime dependencies | proper-lockfile 4.1.2, zod 4.4.3 | package.json dependencies |
| Check command | pnpm check = typecheck + tests + build | README + package.json scripts |
| Stars at publication | 373 stars, 29 forks, 0 open issues (snapshot, moves daily) | GitHub repo page, 2026-10-31 |
"Research prototype for reference only, not an official DeepSeek product." The README adds that much of the engineering was done by coding agents, that its style and practices may be inconsistent, and that you should review and adapt it before use. It also notes that some designs discussed in the article are future directions and are not implemented in this prototype.
The order comes from docs 001 (domain model), 002 (authorization) and 003 (execution). Steps 2 and 3 are the whole point: a message being written is not enough to start work — it has to pass a gate and an assessment.
Doc 001 is explicit that the code, the tools the agents call, the panel and the other documents use these terms and no others. Learning this table before the architecture diagram saves time.
| Concept | What it is | The rule it keeps |
|---|---|---|
| Source | 一个事件流；一种可复用的实现（GitHub、飞书）加一份实例配置 | 自己持有读取位置和机器人身份；重启不重放 |
| Trigger | 一个不可变事件，带来源自己的元数据（如发起人） | 同一个 Trigger 接受两次只产生一张工单 |
| Thread | 按时间排序的 Trigger；由 Source 决定哪些归为一串 | 同一线程同一时间最多一个 Eval 在跑 |
| Channel | 线程的「写」这一侧：话题回复、评论、PR | 一个线程可以有多个 Channel，一张工单持有一组 |
| Policy | 带版本的研判规则，加上每个 Source 的部分（谁能发起、机器人名单） | 规则变了要给进行中的工单送通知，但不撤销已发的 Mandate |
| Eval | 对一批 Trigger 的判断，两层：准入闸门 + 模型研判 | 不持有状态；输入包含线程快照、所有工单和当前策略 |
| Mandate | accept 授予的东西：策略版本、范围、强度、匹配规则、发起人身份 | 不可变；范围变大＝开新工单；只有人能撤销 |
| Ticket | 一个被授权的工单位 | 只有 Eval 能创建；人停止后不再接收任何输入 |
| Assignment | 工单与一次执行器实例的绑定 | pending → running → ended，和工单状态互不跟随 |
| Value group | Possible values |
|---|---|
| Decision | accept · stop · ignore · unavailable |
| Strength | explicit · inferred |
| Assignment | pending · running · ended |
| Report | progress · waiting · ready · blocked · failed · dropped |
Doc 001 ends with a reusability table: the concepts and their rules are reusable, while the Source implementations, Policy contents, Executor and Channel implementations belong to a deployment. In practice that means four things to fill in before it works on your repository: Sources, Policy, execution environment and Channels.
This is not a plugin-sized install: it wants its own profile, its own GitHub account, and a local clone it can add worktrees to.
gh 2.48 or later; the Source CLI fields must point at gh or lark-cli wrappers, whose subcommands and output formats the host relies on.lark-cli (tested at 1.0.90 according to the guide) and a Feishu or Lark app with a bot.^0.2.1-alpha.2; the one-time block in 006 installs @deepseek-ai/[email protected] and says other versions must satisfy engines.dsh.The official route is either to let a coding agent walk you through it (recommended) or to run the block from 006 yourself. The commands below are copied from 006 and the repository root, with nothing added.
SETUP.md and set it up with you. It asks one short question at a time (Feishu or not, which repository, which account, who may ask for work), shows the full plan once, and installs only after you confirm.npm install -g @deepseek-ai/[email protected]
git clone <this repository> dsh-superagent && cd dsh-superagent
pnpm install && pnpm build
dsh superagent --from-default-profile web --dump-config > /dev/null
dsh plugin --profile superagent add "link:$PWD"auth login for each wrapper, and on the Feishu side create and publish the bot app and run config init. Afterwards, check each login with <wrapper> api user --jq .login.@<agent login> <request> on an issue in the repository; a Ticket appears on the overview within about a minute.This page does not decide for you whether to run it; it only pulls out the lines in the official docs that get skipped but do matter.
dsh-superagent-worker for Executors, dsh-superagent-router for the main Session — run with danger-full-access as the operator's OS user.PATH and git settings, which is why the guide insists on a dedicated profile (superagent) rather than the one you work in.The docs require that each one builds only on the ones before it, so a reader never meets a concept that has not been introduced. Skipping around is how people get stuck.
| Document | The question it answers |
|---|---|
| 000 Philosophy | 为什么它要「少占注意力」；对操作者的责任要求是什么 |
| 001 Domain model | 九个概念、各自守的规则、哪些概念被明确拒绝 |
| 002 Authorization | 谁能发起工作；follow-up、订阅、交回与释放怎么算 |
| 003 Execution | 一张工单从执行到交付、恢复的全过程 |
| 004 Prompts | 哪些提示词是团队可改的部分 |
| 005 Panel | 面板只显示「需要人」的东西 |
| 006 Deployment | 账号、启动检查、升级与排错 |
| 007 Architecture | 类、接口、文件与测试的索引 |
| 008 Cookbook | 怎么加一个新的 Source |
No. The first line of the README is "Research prototype for reference only, not an official DeepSeek product." It grew out of the team's day-to-day work with much of the engineering done by coding agents, and the repository asks you to treat it as research code.
No. package.json is private: true at version 0.1.0. Doc 006 clones locally, runs pnpm install && pnpm build, then dsh plugin --profile superagent add "link:$PWD" into a dedicated profile created from the Web template.
No. The README says it listens to GitHub and, optionally, Feishu. With GitHub only, limit Sources to GitHub; Feishu additionally needs lark-cli (documented at 1.0.90) and a Feishu app with a bot.
Only where a rule the team wrote allows it, and that kind of work starts in a mode that only records what the agent would have done. The stated discipline is that it acts only when asked or when a written rule allows it, and a bot can only produce an inferred candidate, never an explicit request.
The repository publishes no pricing or token budget; it says only that it uses the profile's default model unless the Eval or Execution settings name another. Cost therefore depends on your own model configuration and on how wide your Policy is — more allowed work means more spend.
It is the public prototype of that design, but the README says explicitly that the code was adapted for publication and is not identical to the internal version or to the implementation the article describes; some designs in the article are future directions that this prototype does not implement.
Every fact on this page was read from the primary sources below on 31 October 2026: the repository page, README, package.json and the numbered documents. Versions and engine requirements change — check the official repository before you start.