dsh-sci 🔬

August 14, 2026 · View on GitHub

Zero-dependency scientific computing tools for DeepSeek Harness: physical-unit conversion, CODATA physical constants, and Runge–Kutta ODE/dynamical-system simulation. No core changes.

English | 简体中文

A DeepSeek Harness (DSH) pure tool plugin registering three model tools that scientists reach for daily — all implemented as zero-dependency pure functions: no network, no external software, fully unit-tested.

Tool	Purpose	Example
`units_convert`	Physical unit conversion (length/mass/time/temperature/energy/pressure/frequency/force/angle/volume/area/electric, 70+ units)	`1 eV → J`, `0 ℃ → K`, `1 kWh → J`
`physical_constants`	Look up CODATA physical constants (c, h, e, k, G, N_A, …)	`physical_constants("h")`
`ode_solve`	4th-order Runge–Kutta integration of common dynamical systems (the simulation engine)	exponential / logistic / harmonic / damped / pendulum / predator-prey / SIR / Lorenz / Van der Pol

✨ Features

Zero dependencies: unit conversion, physical constants, and the RK4 integrator are hand-written pure functions — no runtime node_modules, no network.
Safe simulation: ode_solve uses a curated model library (9 dynamical systems) with parameters, not arbitrary code execution — safe for the model to call.
Verified accuracy: the RK4 integrator is checked against analytic solutions (e^-t, cos t) in unit tests.
Zero core changes: pure ctx.tools.register, no client bundle.

📦 Install

Prereqs: DSH (dsh web works), Node ≥ 22.19, pnpm.

# From GitHub (prebuilt lib/ committed)
dsh plugin --profile web add "github:Blaczz/dsh-sci#main"

# Local directory
cd dsh-sci && npm install --legacy-peer-deps && npm run build
dsh plugin --profile web add ./dsh-sci

Restart dsh web after install, then ask the agent: "convert 5 eV to joules", "look up the Planck constant", or "simulate a damped oscillator with ω=2, ζ=0.1 out to t=20".

🔬 Usage

User: simulate the Lotka-Volterra predator-prey model, alpha=1, beta=0.1, gamma=1, delta=0.02, to t=50
agent → ode_solve(model="lotka-volterra", params={...}, tEnd=50)
       → returns the time series + final state [prey, predator]

User: how many joules is 1 electron-volt?
agent → units_convert(value=1, from="eV", to="J") → 1.602176634e-19

User: list all physical constants
agent → physical_constants() → c, h, e, k, G, N_A, … all

🛠️ Development

npm install --legacy-peer-deps
$env:DSH_NODE_MODULES = "$env:USERPROFILE\.dsh\profiles\node_modules"
npm run setup:dsh-workspace
npm run verify                          # ★ one-shot gate: clean + typecheck + test + build
npm test                                # vitest (24 tests: units / constants / ODE accuracy + manifest)

Layout

dsh-sci/
├── package.json            # dsh.bundle.patch (pure tool plugin)
├── cordis.patch.yml        # bundle patch layer
├── scripts/                # build (tsc only) / clean / setup-dsh-workspace / verify
├── src/
│   ├── units.ts            # ★ unit conversion (70+ units, dimension + factor + offset)
│   ├── constants.ts        # ★ CODATA 2018 physical constants
│   ├── ode.ts              # ★ RK4 integrator + 9-model dynamical-system library
│   └── index.ts            # host half: registers the 3 tools
└── tests/                  # units / constants / ode / manifest

🧩 Ecosystem positioning

Fills a gap: the DSH ecosystem previously had zero physics / ode / numerical / fft plugins, and the units hits were all "unit-test / unit-economics" false positives — no physical unit conversion. The existing dsh-science projects are reproducible Python/R workbenches (not numerical tools), and dsh-robotic-harness is MuJoCo robotics (a heavy dependency).
Technical route: follows the zero-dependency pure-tool pattern of omdsh-dev/dsh-tool-*, with pure functions under test.
Zero core changes: pure ctx.tools.register, no external software, no network.