dsh-system-one-plugin

    

System Duo

@vllm-sr/dsh gives your DeepSeek Harness agent a System One: fast decision models that triage, guard, shield, mask and check every step, while System Two, vllm-sr/auto, thinks it through.

Quick start

npm install -g @deepseek-ai/dsh pnpm
dsh plugin --profile web add @vllm-sr/dsh
dsh web

On its first run the plugin fetches its own free key from System Duo, for research and evaluation use (terms, privacy), and makes vllm-sr/auto your default model. You need Node 24.9 or later; the first line installs DeepSeek Harness and pnpm, which dsh plugin uses to install plugins.

What System One does in your agent loop

Five reflexes run on their own, and two tools are there when System Two needs them. Every card below is a real one from the demo recordings: each number is System One's answer in that run.

Triage · every request Reads each request first. A bare “Fix it.” gets a question back.

Guard · before each tool call Asks before risky commands. A force-push waits for you: Dangerous, 2.00 of 3.

Shield · after each tool result Strips hidden instructions before System Two reads them: 284 characters from one README.

Privacy · after each tool result, on by default Masks personal data as [LABEL] before System Two sees it: 29 details in one file.

Done-check · when a turn ends Backs “done” with evidence, complete 0.99 and checked 1.00, or sends the turn back.

system_one · when System Two asks Typed questions over records and images: 230 tickets routed, scored and flagged in 7.0 s.

Computer use · system_one_screen Sees the screen, picks the target, checks the result, and Guard stops the risky click: “Delete This Repository” waits for you at Dangerous, 2.10 of 3.

See it live

Open app.vllm-sr.ai/#harness and press Start DeepSeek Harness: 21 real sessions recorded with this plugin open in your browser, read-only, with each System One card where it happened. Fifteen run in the Web UI, from batch triage of 230 tickets to a judge scoring subject lines; six drive a full desktop with mail, a spreadsheet, Gitea, an admin console and a web store. Then try your own in a sandbox, on a free key.

How it works

• System One, fast. The Decision Models answer typed questions with calibrated probabilities in a fraction of a second: is this command dangerous, does this file give the agent orders, is the work really done. By default vllm-sr/d3-ultra triages, scores Shield's passages and answers the system_one tool, and vllm-sr/vela3-turbo runs Guard, Privacy, the done-check, every span answer and the screen.
• System Two, slow. vllm-sr/auto plans, reasons and acts, picking the right open model for each request. The plugin makes it your default model; choose any other model in the app and System One keeps working beside it.
• One router. vLLM Semantic Router serves System One and routes System Two at app.vllm-sr.ai, and both take the one free key the plugin fetches. The router works with any model, open or closed.

Part of System Duo, built by vLLM Semantic Router with AMD.

Configuration

It works as installed. To change it, set fields on the system-one row of your profile's patch, ~/.dsh/profiles/web/cordis.patch.yml (or headless):

- id: system-one
  config:
    privacy: true      # mask personal data before System Two reads it; false turns it off
    triage: true       # each reflex and tool can be turned off
    guard: true
    shield: true
    doneCheck: true
    tool: true         # the system_one tool
    screenTool: true   # system_one_screen, for computer use
    autoKeys: true     # fetch the free key on first run
    systemOneUi: card  # how Chat shows System One: card, compact or off

Your key is kept in DeepSeek Harness's credential store and in ~/.dsh/.env, readable by you only. In CI, set VLLM_SR_API_KEY to a key from app.vllm-sr.ai; a key you set yourself is always used and never replaced.

Every option, the questions the tools take, what Chat shows and its selectors, what is sent where, and what we measured: the full reference.

FAQ

Is it free? Yes, for research and evaluation: the plugin fetches a free key, and commercial use is not permitted (terms).

What leaves my machine? What System One decides about: your request (clipped), the tool calls it reviews, the passages of tool results it screens, the end of each turn, and what System Two hands its tools. Privacy sends the results it masks to System One whole, and System Two gets the conversation, as any LLM provider does. System Duo logs requests to improve its services and train models, so read its privacy page before you use it with private code or data. Every reflex can be turned off, and the local trace never records keys.

Does it run headless or in CI? Yes: dsh plugin --profile headless add @vllm-sr/dsh, then dsh headless "…". A headless run reads its key when it starts, so its first run fetches the key and the LLM answers from the next; in CI, set VLLM_SR_API_KEY.

What if System One is unreachable? Your agent carries on. A guarded call falls back to guardOnError (allow by default) while DeepSeek Harness's own sandbox and approvals still apply, and a tool call that finds System One unavailable shows one calm Unavailable line.

Why didn't I get the newest release? pnpm, which dsh plugin add uses, installs a release only once it is a day old. To take one sooner, name its version, as @vllm-sr/dsh@<version>.

I had dsh-system-one 0.2. Run dsh plugin --profile web remove dsh-system-one, then add @vllm-sr/dsh; your keys and your system-one settings stay.

Links

System Duo · Live demo · Docs · Reference · Changelog · vLLM Semantic Router · DeepSeek Harness

License

Apache-2.0.

DeepSeek Harness is a trademark of DeepSeek. This plugin and System Duo are not DeepSeek products.