MicroHEROX/dsh-unsloth-hands

Unsloth for DeepSeek Harness

dsh-unsloth-hands — give your DeepSeek Harness agent a pair of local hands.

A third-party tool plugin for DeepSeek Harness that lets the online model (your main conversation model) hand repetitive, token-cheap labor to a local Unsloth Desktop (Unsloth Studio) server — plain text work and vision work (image analysis / OCR / comparison).

The main model stays where your deployment puts it. When it decides a job is better done locally, it calls:

• unsloth_run — run one prompt on the local text model (batch rewrites, name translations, string munging, short summarization, extraction).
• unsloth_vision — send images to the local multimodal model (OCR, image analysis, multi-image comparison) with structured report templates.

The plugin is a pure client: it only connects to the Unsloth Desktop you are already running. Model selection, downloading, quantization and context settings all happen in the Unsloth app itself — the plugin never starts, owns, or stops any process, and never kills anything.

---

✨ What it does

• Two model-facing tools registered on the harness tool registry (ctx.tools), following the official dsh-tools contract (defineTool, canonical JSON values, pure render/presenters, exec.signal forwarding).
• Authenticated wire calls: every request carries Authorization: Bearer sk-unsloth-…. The key comes from the apiKey config or the UNSLOTH_API_KEY environment variable (create it in Unsloth Settings → API).
• Friendly failure modes: before each call the plugin probes /v1/models; if Unsloth Desktop is not running you get a clear, actionable error instead of a generic network failure. A wrong/missing key surfaces as AUTH with a hint.
• Text + vision wire support: non-streaming OpenAI-compatible chat-completions; images sent as the standard multimodal content array.
• Three image sources for the vision tool: local file paths, data:/http(s): URLs, or the images attached to the current conversation (read through the harness attachment service).
• Structured vision prompts — machine-verifiable report contracts: analyze (8-section report), ocr (character-exact), compare (multi-image, 5-section) — plus a fidelity rule for the online model (relay verbatim, never invent, preserve uncertainty).
• Live config: a llm-unsloth: section in the harness user-settings document overrides the plugin config without a restart.
• Safe by construction: nothing is ever spawned or killed — the plugin only talks HTTP to your Unsloth Desktop.

🚫 What it does NOT do

• Does not replace the harness LLM provider — the online model stays the main model; the local model is only reached through the two tools.
• Does not launch, configure, or stop Unsloth — you run Unsloth Desktop and load the model you want (quantization, context size, GPU settings) in its UI.
• Does not modify any DeepSeek Harness or Unsloth file; it is a pure add-on plugin.
• Does not bundle or host GGUF model files — Unsloth downloads and caches them for you.
• Does not stream responses (tool calls get the full answer in one round trip).

📋 Requirements

Item	Requirement
Node.js	≥ 20
DeepSeek Harness	installed (npx @deepseek-ai/dsh web or a source checkout), 0.1.0-rc series
Unsloth Desktop	running, with a model loaded and an API key created (Settings → API)
Model	any GGUF/safetensors model loaded in Unsloth; vision needs a multimodal model (e.g. Qwen3-VL / Gemma vision GGUFs)

📦 Install

The package is a standard harness bundle (declares dsh.bundle with its cordis.patch.yml), so the official install path works:

dsh plugin --profile <name> add dsh-unsloth-hands        # from npm registry
dsh plugin --profile <name> add github:MicroHEROX/dsh-unsloth-hands   # straight from GitHub

It can also be installed as a plain npm dependency in your harness project (the directory whose cordis.yml / cordis.patch.yml composes your deployment), then add the plugin row yourself:

npm install dsh-unsloth-hands

- insert:
    - id: unsloth-tool
      name: 'dsh-unsloth-hands'

From a source checkout of the harness, you can point the plugin row directly at a clone of this repo:

- insert:
    - id: unsloth-tool
      name: '../dsh-unsloth-hands'

Installing from GitHub? pnpm may refuse to run the package's prepare build script until you allowlist it (exact package key printed by pnpm) in your profile's pnpm-workspace.yaml:

allowBuilds:
  dsh-unsloth-hands: true

Then re-run the add. Installing from the npm registry needs no such step.

⚙️ Configure

1. Start Unsloth Desktop, load the model you want (the model hub downloads GGUFs; the loaded model is the one the tools reach).
2. Create an API key: avatar → Settings → API → Create, copy the sk-unsloth-… value (it is only shown once).
3. Add the plugin row to your profile cordis.patch.yml:

- insert:
    - id: unsloth-tool
      name: 'dsh-unsloth-hands'
      config:
        baseURL: 'http://127.0.0.1:8888'              # Unsloth's default port
        apiKey: 'sk-unsloth-xxxx...'                   # from Unsloth Settings → API

That's it. The plugin connects to whatever model is currently loaded — no model names, no config files, no launch flags. Alternatively set UNSLOTH_API_KEY in your environment instead of apiKey.

Installed via dsh plugin add? The bundle already inserts the unsloth-tool row — just override its config in your profile's cordis.patch.yml (the harness override form, no name needed):

- id: unsloth-tool
  config:
    apiKey: 'sk-unsloth-xxxx...'

Full config reference (all 10 fields with defaults): docs/api.md §1.2.

🛠 Using the tools

unsloth_run — text

param	type	required	meaning
prompt	string	yes	instruction/text sent as a user message
system	string	no	optional system instructions
temperature	number	no	sampling temperature (0–2)
max_tokens	integer	no	output cap (default maxTokens)
stop	string[]	no	stop sequences

Returns { text, reasoning?, model, usage, elapsedMs }.

unsloth_vision — images / OCR

param	type	required	meaning
mode	analyze/ocr/compare	no	built-in prompt template (default analyze)
prompt	string	no	custom instruction (overrides the template)
image_paths	string[]	no	local images (png/jpg/jpeg/webp/gif/bmp, ≤20 MB each)
image_urls	string[]	no	data:image/... or http(s):// URLs
temperature	number	no	sampling temperature (lower for OCR, ~0.2)
max_tokens	integer	no	output cap
stop	string[]	no	stop sequences

Image sources resolve in order: explicit image_paths + image_urls → the most recent image(s) attached to the conversation → clear error. compare sends 2–4 images in ONE request for joint reasoning.

Returns { text, reasoning?, model, images, usage, elapsedMs }.

Vision requires the model currently loaded in Unsloth to be multimodal. Unsloth serves one loaded model at a time — switch to a vision model in the app before calling unsloth_vision.

❓ FAQ

My main model is text-only — how do images get in?

DeepSeek's flagship chat models (and most other routes) are text-only: the harness refuses to send image messages to them (the adapter rejects them with UNSUPPORTED_CONTENT), so you cannot attach an image to the conversation. That is exactly the case unsloth_vision is built for — no harness upload is involved:

1. When you paste/drop an image in a text-only model's composer, the harness (like OpenCode and Pi) lands it as a temporary file path in your message instead of pixels.
2. The model sees that path, calls unsloth_vision with image_paths: ["<that path>"] (or an image_urls entry), and the local vision model reads the file directly.
3. You can also just tell the model a path to any image on disk.

For a main model that DOES support images, the conversation-attachment source also works automatically.

My requests get 401 Unauthorized?

Unsloth requires a valid key on every request. Create one in Settings → API (revoked keys fail with 401) and put it in apiKey or UNSLOTH_API_KEY. The health probe treats a 401 as "server is running" — the error surfaces from the tool call itself with an actionable message.

🗺 Roadmap

Possible / planned directions:

• More vision modes and prompt templates (document layouts, table extraction).
• Reading the currently loaded model from /v1/models to fill the wire model field automatically.
• Publishing to the npm registry and the dsh-plugin topic.
• Batch jobs: drive many local calls from one agent turn.

Deliberately NOT planned:

• Launching or managing the Unsloth process — the plugin stays a pure client; you own the app.
• Becoming an LLM provider adapter — the plugin stays a tool; the online model stays the main model.
• Streaming responses — tool calls get the full answer in one round trip (simpler and sufficient).
• Bundling model files or modifying DeepSeek Harness / Unsloth itself.

🗑 Uninstall

1. Remove the plugin row from your profile cordis.patch.yml (or cordis.yml):

   # delete this block
   - insert:
       - id: unsloth-tool
         name: 'dsh-unsloth-hands'
   

   Installed via dsh plugin? dsh plugin --profile <name> remove dsh-unsloth-hands removes both the dependency and its bundle layer.
2. Restart the harness (or let HMR reload if you edit the config live). The two tools (unsloth_run, unsloth_vision) are unregistered automatically — the online model no longer sees them.
3. No residue: the plugin never spawned anything, so there is nothing to stop; your Unsloth Desktop keeps running untouched. Installed via npm? npm uninstall dsh-unsloth-hands.

📌 Version & compatibility

Component	Version
This plugin	0.1.0
DeepSeek Harness	0.1.0-rc series (tested against npm @deepseek-ai/* 0.1.0-rc.8 / 0.1.1-rc.1)
Node.js	≥ 20
Unsloth Desktop	any version exposing the external API (/v1/chat/completions)

Peer dependencies (runtime): @deepseek-ai/cordis ^4.0.1, @deepseek-ai/dsh-tools/dsh-llm/dsh-session/dsh-attachment/dsh-settings/dsh-launch-environment >=0.1.0-rc.2, @deepseek-ai/schemastery ^3.18.1.

🛠 Development

npm install
npm run typecheck   # tsc --noEmit
npm test            # vitest run (46 tests: unit, tool, integration, Loader composition)
npm run build       # clean + tsc -> lib/

Tests include a REAL-composition tier (app boot → Cordis Loader → cordis.yml) per the harness testing policy, and a real-machine driver (tests/real-driver.mjs) for main / auth / not-running scenarios.

📚 Documentation

doc	content
docs/engineering.md	structure, plugin contract, commands, test tiers
docs/api.md	authoritative API reference (Config, tools, classes, error codes)
docs/glossary.md	standard terminology
docs/solutions.md	pitfalls, troubleshooting, methodology

🙏 Credits & thanks

• DeepSeek AI — the DeepSeek Harness platform this plugin plugs into, and the reference implementations (dsh-llm-deepseek, dsh-tool-todo) that define the patterns we follow.
• Unsloth — the local training/inference stack and Desktop app whose OpenAI-compatible API makes all of this possible (llama-server underneath), and its docs that guided the integration.
• Cordis — the plugin runtime that powers the harness.
• LostRuins / KoboldCpp — the sibling plugin dsh-koboldcpp-hands this project evolved from.
• The open-source models and quantizers (llama.cpp ecosystem, GGUF) that run locally on your machine.

License

MIT. Not affiliated with DeepSeek AI or Unsloth AI; dsh and unsloth are trademarks of their respective owners.