by LofiSu
Open-source DeepSeek Super-agent style multi-agent framework with a thin harness — Router/Worker, triggers, Watchdog (not an official DeepSeek product).
# Add to your Claude Code skills
git clone https://github.com/LofiSu/dsh-super-agentGuides for using ai agents skills like dsh-super-agent.
See how dsh-super-agent compares with popular alternatives.
dsh-super-agent is an open-source ai agents skill for AI coding assistants such as Claude Code, Codex CLI, and ChatGPT, built by LofiSu. Open-source DeepSeek Super-agent style multi-agent framework with a thin harness — Router/Worker, triggers, Watchdog (not an official DeepSeek product). It has 67 GitHub stars.
dsh-super-agent's catalog security scan is still queued. You can run an instant dependency and prompt-injection check now with the "Scan for vulnerabilities" button above.
Clone the repository with "git clone https://github.com/LofiSu/dsh-super-agent" and add it to your Claude Code skills directory (see the Installation section above).
dsh-super-agent is primarily written in TypeScript. It is open-source under LofiSu on GitHub, so you can review or fork the full source.
Yes. SkillsLLM lists many other AI Agents skills you can browse and compare side by side. Open the AI Agents category from the badge at the top of this page, or use the Related Skills and comparison links further down to weigh dsh-super-agent against similar tools.
No comments yet. Be the first to share your thoughts!
Available after the catalog security scan passes.
⚠️ Third-Party Software Notice
This skill is third-party open-source software developed and hosted independently on GitHub. SkillsLLM lists this project for discovery and does not control or maintain the underlying repository.
Any security checks, ratings, or warnings displayed by SkillsLLM are automated and limited in scope. They do not constitute a security certification or guarantee that the software is safe, error-free, or free from malicious code, vulnerabilities, compromised dependencies, or prompt-injection risks.
Review the source code, permissions, dependencies, and configuration before installing or running any third-party skill. Use is at your own risk. To the maximum extent permitted by applicable law, SkillsLLM is not liable for losses arising from third-party software.
The deep catalog scan for this skill is still queued. Run an instant dependency check now instead.
English | 简体中文
An independent, MIT-licensed super-agent harness in TypeScript.
A thin async ReAct runtime at the bottom. A proactive domain engine on top: Router / Worker, Triggers, Threads, Eval, Tickets, Assignments, and a Watchdog.
Documentation · Quick start · Core features · Architecture · Safety · Modules · Demo
[!NOTE] This is not an official DeepSeek product. It is an independent implementation inspired by DeepSeek's public Super-agent article and prototype. It does not ship, wrap, or depend on DeepSeek Harness, Cordis, or any
@deepseek-ai/dsh*package.
If you use Claude Code, Codex, Cursor, or another coding agent, hand it this sentence:
Clone https://github.com/LofiSu/deepseek-super-agent, run pnpm install, then run
pnpm typecheck and pnpm lint. Read README.md and docs/install.md. Stop before
pnpm test or pnpm demo so I can supply model credentials and a writable GitHub repo.
The agent should stop with the exact commands you still need to run, rather than guessing a provider or writing to a repository on your behalf.
Requires Node.js 22+ (the store uses node:sqlite) and pnpm.
git clone https://github.com/LofiSu/deepseek-super-agent.git
cd deepseek-super-agent
pnpm install
pnpm typecheck
pnpm lint
The repository is not published to npm. package.json keeps a build script and a bin entry for packaging, but day-to-day use is pnpm ds-agent ... from the checkout.
ds-agent speaks any OpenAI-compatible API. Point it at DeepSeek directly:
export DS_AGENT_API_BASE=https://api.deepseek.com
export DS_AGENT_API_KEY=sk-...
export DS_AGENT_API_MODEL=deepseek-flash
Or set a single provider key and let the presets fill in the base URL and default model (see Models). A half-set DS_AGENT_API_BASE / DS_AGENT_API_KEY pair fails instead of falling back silently.
pnpm ds-agent init # create the SQLite store in ./.ds-agent
pnpm ds-agent chat "fix the login timeout"
pnpm ds-agent tasks # tickets and their stage
pnpm ds-agent panel # Panel · Execution
State lives in DS_AGENT_HOME, or in ./.ds-agent when unset. pnpm ds-agent serve runs the daemon and Watchdog timer over persisted state; it starts no inbound channel loop. To inspect the read model in a browser:
DS_AGENT_UI_TOKEN="$(openssl rand -hex 16)" pnpm ds-agent ui
The dashboard is read-only, binds to loopback, and requires a token of at least 16 characters. See docs/UI.md.
pnpm test and pnpm demo use a live model; MockModel is forbidden on those paths. If DS_AGENT_API_* is unset, the test script fills a local Magpie default (http://127.0.0.1:3425/v1, key magpie, model workbuddy/deepseek-v4.1-flash), so it needs that gateway running. For a no-key fixture run, use:
pnpm test:coverage # offline fixture suite; no model key required
pnpm demo is the real-channel acceptance path. It needs gh auth and creates real GitHub issues and comments in DS_AGENT_GH_REPO:
gh auth login
export DS_AGENT_GH_REPO=your-account/your-writable-repo
pnpm demo
Read Demo (real channels) before running it.
Most agent frameworks give you a loop: messages in, tool calls out, repeat. That loop is necessary, but it is not an operating model. A personal agent also needs to notice inbound work, decide what matters, keep state across restarts, run long jobs without blocking the conversation, and recover when something gets stuck.
This repository implements that operating model in two layers:
The Router is the sequence point: it talks to people, batches triggers, and dispatches work. Workers execute in isolated sessions. Triggers drive the loop. The Watchdog lives outside the control path and nudges or fails stuck work.
The design vocabulary follows 从零打造 Personal Agent(三): the router role is the Main Agent / Main Session, the worker role is the Executor, and a Subscription is a context-only Ticket ↔ Thread relation that grants no extra authority. Deviations and corrections are recorded in docs/DOMAIN_ALIGNMENT.md.
Channels emit semi-structured Triggers into Ingress. A deterministic Fast Path applies source and noise rules; an optional model-based Slow Path (DS_AGENT_SLOW_PATH=model) refines the decision, with a deterministic rule fallback on timeout, error, or invalid output. The default is rules-only.
Each topic or pull request becomes a Thread. The Router batches triggers waiting on the same Thread, Eval separates noise from actionable work, and actionable input becomes a Ticket and an Assignment. A speculative Assignment can start before delivery, and a deliver gate drops results that are not worth sending. Restarts resume from SQLite; the demo includes a restart check.
WorkerPorts, not the Engine, so it cannot reply or decide; only the Router writes to a channel.bypass=subscription), skipping Router Eval.TEAM_MEMBER_DONE and the parent aggregates one report. Team rows persist in SQLite and are inspectable offline with ds-agent team / ds-agent teams.delegate_agent. There is no shell, the cwd is the Worker workspace, and timeout, output size, and environment are bounded. See docs/LOCAL_AGENTS.md.Short summaries stay inline. Long content is spilled out of band with an id, and tools receive typed kind:id handles first. read_object expands one hop at a time, which gives the model a Chain of Recall instead of a context dump. See docs/MODULES.md.
Long-term memory keeps hot tickets and Router decisions (1000 entries / 90 days) plus cold Thread digests (200 / 3650 days). Recall is hybrid: lexical search, one tag hop, and an optional EmbeddingPort. The default offline HashedEmbedding works without a model; DS_AGENT_SEMANTIC_RECALL=0 disables the semantic arm, and DS_AGENT_EMBEDDING_FLOOR sets its cosine floor (default 0.35, clamped to 0–1). Month-scale behavior is not yet measured, so T48 stays partial; the dated measurement protocol and its offline fixture smoke are in docs/T48_MONTH_SCALE_PROTOCOL.md.
start_job returns {status:'started', jobId} immediately and ends the turn. The harness enqueues the next turn when the job finishes as a JOB_DONE prompt, and the foreground keeps accepting new input meanwhile. This is async ReAct, not a long-blocking tool call.
For mid-task collaboration, the CLI exposes pause, intervene, resume, and collab, backed by a durable Ticket.collaboration. The harness and Watchdog honor a pause.
request_approval with an exact outward action instead of performing it. The ticket holds; the Router asks; ds-agent approve / ds-agent deny records the decision and resumes through the normal path. The model may request, never decide, and chat text is not consent. See [docs/APPROVALS.md](docs/APPROVA