mddl harness

Visual orchestrator for DeepSeek Harness. Drag models and tools onto a canvas, then export a real cordis.patch.yml overlay.

This is not a fork of dsh and not a child-process wrapper around npx @deepseek-ai/dsh. Upstream already ships a web profile, Cordis patches, and session/event telemetry. We compile graphs into that overlay format so the work can later mount as a dsh-plugin.

Why this stack

PDF plan	What we actually do
Next.js App Router	Vite + React SPA. DSH serves static dist; RSC is dead weight.
Fictional dsh.yaml	cordis.patch.yml (id replace + insert)
Spawn dsh run over Socket.io	Later: Cordis plugin listening to session/event / agent/*
Config inside expanded nodes	Dumb nodes + inspector (keeps XYFlow at 60fps)
xterm.js inside nodes	Docked preview only. Terminal belongs in a later host plugin.

Packages

• @mddl/graph-schema — graph IR and the shipped DSH row catalog
• @mddl/compiler — graph → Cordis patch YAML, plus overlay linting
• @mddl/studio — visual editor
• dsh-blueprint — Blueprint tab inside the DSH web client (readme)
• dsh-overlay-check — the overlay safety checks on their own, no dependencies, for anything that writes config (readme)

Run

pnpm install
pnpm dev

Studio: http://localhost:5173

pnpm lint       # biome check
pnpm format     # biome check --write
pnpm test
pnpm typecheck

CI runs lint, typecheck, test, and build on every push.

The canvas is saved to localStorage as you edit, so a reload keeps your graph. Reset in the header restores the starter graph.

Find modules in the palette by name, row id, package, or description. Press / to jump to search, and Not on canvas hides what you already placed.

Apply an exported overlay. --patch is resolved from your terminal cwd, not the studio. Running ./cordis.patch.yml from ~ looks for /Users/<you>/cordis.patch.yml.

After Export (typical macOS download):

npx @deepseek-ai/dsh web --patch "$HOME/Downloads/cordis.patch.yml"

Or use the starter overlay in this repo (from the repo root):

npx @deepseek-ai/dsh web --patch "$PWD/examples/cordis.patch.yml"

What Phase 1 does not do

Live harness execution, profile install, or a DSH client slot. "Preview telemetry" is a local animation that proves the glow/edge path. Real run state comes from session/event in a later phase.

Roadmap

Shipped in dsh-blueprint 0.5.0: reading the live loader tree, linting the config you actually booted, writing an overlay back behind a marker-delimited block, refusing to write a row that would stop the harness booting, snapshots with one-click restore, importing the running config onto the canvas, and compiling to agent presets so the canvas changes what a session actually gets. Next, in order:

1. Map session/event onto canvas telemetry, replacing the local animation.
2. English and 中文 both first class.
3. Model rows beyond the shipped DeepSeek adapters.

If a harness will not start

A row naming a package the profile cannot load stops the harness booting outright, rather than disabling one entry. dsh-blueprint refuses to write one, but if you get there by another route, two things help:

• dsh --profile <name> --dump-default-config reads the bundles while skipping the profile and home user layers, so it still answers when the overlay is the problem. Confirmed on rc.7: a row written into a profile's cordis.patch.yml shows in --dump-config and is absent from --dump-default-config.
• The DSH handbook's recovery runbook by the sandbaseai folks is the most thorough write-up of getting back from this, including which user layer owns which file.

Status

Early. DSH is itself a developer preview with breaking changes between release candidates, so the row ids and patch shape here track a moving target. Verified against @deepseek-ai/dsh 0.1.0-rc.7.

License

MIT. See LICENSE.