dsh-quant-workspace

English | 中文

A fully self-contained DeepSeek Harness plugin for single-ticker quant analysis. It ships its own Python engine inside the package (python/): Yahoo Finance daily bars, Bollinger mean-reversion rules, and three read-only tools — the daily signal card, a full backtest, and a strategy health review.

No external engine, no market-data key, no private references: the plugin is the analysis. The TypeScript shell only registers the single_ticker tool and runs the bundled dsh-quant CLI through the harness shell seam.

⚠️ Not investment advice. The plugin only surfaces rule state and evidence; the decision is always yours. It never places orders and never changes positions.

What it is for

The plugin turns single-ticker analysis into a human-in-the-loop research loop: you ask in natural language, the model pulls real computed evidence through single_ticker, you review and decide. The model researches and interprets; every judgment call — entries, exits, sizing — stays with you.

Requirements

• A DeepSeek Harness installation (web profile) with pnpm (for installing the bundle).
• uv (runs the bundled Python engine; first use syncs python/.venv).
• Internet access for Yahoo Finance data (daily bars).

Install

Installation status: not yet published to npm. Until then, install from the git spec (dsh plugin --profile web add github:AllenCX/dsh-quant-workspace) or use the dev overlay below.

dsh plugin --profile web add dsh-quant-workspace

All configuration is optional (the plugin works out of the box). To track your real position, configure a ledger in the profile user patch ($DSH_HOME/profiles/web/cordis.patch.yml):

- id: quant-workspace
  config:
    ledgerPath: 'C:\path\to\trade_log.csv'

Option	Default	Meaning
ledgerPath	(none)	Position ledger CSV (date,ticker,action,price; FIFO). Without it the engine treats positions as flat and says so.
timeoutMs	180000	Foreground timeout per tool call.
pythonCommand	`uv run --project
/python dsh-quant`	Override for running the bundled engine CLI (e.g. a pre-built venv).

Dev / local overlay

pnpm dsh web --patch ./dev.patch.yml   # or any patch mounting src/index.ts / lib/index.js

Tools

single_ticker

• ticker (required): symbol, e.g. TSLA. Uppercased automatically; only letters, digits, dot and dash are accepted (shell-safe by construction).
• mode (default daily): daily = today's signal card · backtest = full backtest with per-trade table · review = strategy health check (data freshness, rule state, ledger vs signals).

The tool returns the engine's rendered report text.

Strategy (v1)

Bollinger mean-reversion on daily bars: enter when %B = 1 (close at/above the upper band). Bollinger(20, 2σ). Same-bar close fills, no transaction costs modeled in v1. Rule parameters are CLI options, so richer rules can be exposed by the tool later without an engine rewrite.

Boundaries

• Read-only. The tool never places orders, never changes positions, never writes market data.
• Position state comes only from the ledger you configure — the engine never assumes an unlogged holding.
• Data stays Yahoo. v1 fetches daily bars from Yahoo Finance on each call; no caching layer yet.
• Privacy. No keys, no accounts, no per-user data leaves your machine.

Bundled engine CLI

The plugin's tool runs this bundled command (also available for scripting):

dsh-quant --ticker  --mode  [--ledger 
] [--data-file ] [--start ] [--end ]

• Exit 0 with plain-text report on success.
• Exit 1 with a message when data cannot be loaded; exit 2 for invalid invocation.
• --data-file reads a local OHLCV CSV instead of the network (used by the tests).

Development

pnpm install && pnpm run typecheck && pnpm run test && pnpm run build   # TS shell
cd python && uv run --project . pytest tests -q                          # bundled engine

License

MIT