dsh-plugin-diepi

Community DSH bundle for the constrained diepi-mcp quantitative-research adapter. It mounts the diepi-quant-research Skill and connects DSH to a separately installed MCP stdio server.

This package is orchestration glue, not the diePi engine. It supports typed, daily stock/ETF research backtests only. It contains no market data, provider connector, credential handling, arbitrary-Python tool, broker integration, or live-order capability.

This is an independent community project. It is compatible with DeepSeek Harness but is not affiliated with or endorsed by DeepSeek.

Compatibility

• DSH 0.1.0-rc.7: primary Alpha target
• DSH 0.1.0-rc.8: canary target; use a separate DSH_HOME
• Node.js ^22.19.0 || >=24.0.0
• pnpm 11.7.x, available as the bare pnpm command on PATH
• Python 3.10+
• diepi-mcp==0.1.0a1, which pins diepi==0.1.1

DSH supplies @deepseek-ai/dsh-mcp-client and @deepseek-ai/dsh-skill-filesystem; this bundle intentionally does not copy them.

Required host setup

Install the Python adapter into a dedicated, host-owned environment. Initialize a brand-new adapter home with:

diepi-mcp init-config --target "C:\host-owned\diepi\adapter-home-alpha1"

The command exclusively claims a new target, writes an .diepi-mcp-incomplete marker, creates each staged payload without replacing an existing path, and commits diepi-mcp.json last. It calls diePi's demo generator and includes no real market rows. A failure deliberately leaves an unusable incomplete target rather than deleting or taking over a racing path; choose a new directory, or have the host owner inspect it before manual cleanup. Keep the adapter home and virtual environment outside the Agent workspace.

Before DSH starts, set both variables to absolute paths:

$env:DIEPI_MCP_COMMAND = (Resolve-Path "C:\host-owned\diepi\venv\Scripts\diepi-mcp.exe").Path
$env:DIEPI_MCP_CONFIG = (Resolve-Path "C:\host-owned\diepi\adapter-home-alpha1\diepi-mcp.json").Path
& $env:DIEPI_MCP_COMMAND --check-config

The bundle validates these values during activation. Missing or relative paths fail closed; there is no fallback to a diepi-mcp executable found through PATH. MCP starts with the configuration directory as its working directory, not the Agent workspace.

Prepare the package manager used internally by dsh plugin:

corepack enable
corepack prepare [email protected] --activate
pnpm --version

If Corepack is absent from the selected Node installation, install it using its official instructions first.

Install into a real app profile

For the Web experience:

dsh plugin --profile web add [email protected]
dsh --profile web --dump-config
dsh web --no-open

For headless:

dsh plugin --profile headless add [email protected]
dsh --profile headless --dump-config
dsh --profile headless

Do not install only into an arbitrary custom profile: a newly named profile has the base bundle but no Web or headless app. Plugin membership changes apply to the next DSH process, so restart after add/remove/update.

From a source checkout, run dsh plugin --profile web add ./dsh at the repository root. The npm package lives in dsh/; a Git dependency aimed only at the repository root is not a supported installation path.

Optional external data Skill

The bundle never downloads data or installs provider code. When configured data cannot satisfy a request, the research Skill may explain how the user can manually install the official Tushare Skill. That upstream code is outside this package's trust boundary and must run under the host's own permission policy.

Acquisition writes a new staging generation. A trusted host then validates the exact market_data_v1 scope, registers an opaque dataset_id, makes the data immutable, and restarts MCP. Provider credentials stay outside prompts, MCP arguments, logs, datasets, and artifacts.

See references/tushare-data-handoff.md for the reviewed-revision reference and consent boundary.

MCP timeout and retry safety

The bundle uses a 10-minute default server tool timeout. Hosts can set DIEPI_MCP_TOOL_CALL_TIMEOUT_MS; values are clamped to 300,000 through 2,147,483,647 milliseconds, and invalid values fall back to 600,000.

start_backtest requires a submission_id matching req_[A-Za-z0-9_-]{8,64}. If a call times out or the transport fails, retry the identical request with the same ID. Reusing that ID for different arguments is an error. This prevents an unknown response from silently enqueueing a second job.

Accept a run for comparison only when all of these are true:

artifact_verified
adapter_attribution_verified
result_committed
result_status == SUCCESS
rankable

Trust boundary and disclaimer

MCP stdio runs as trusted host code outside the Agent file sandbox. The MCP executable, JSON configuration, dataset roots, state database, and results roots must not be Agent-writable. This adapter provides evidence binding, not an OS sandbox or proof of market-data truth/licensing.

Research and backtesting only. Outputs are not investment advice, a trading offer, or a promise of future performance.

Apache-2.0. See LICENSE and THIRD_PARTY_NOTICES.md.