DeepSeek Harness integration validation — 2026-10-29

This is a compatibility check of an unofficial community integration, not a benchmark of general answer quality, API cost, or latency.

Environment

  • Windows, Node.js 24.15.0, DSH 0.2.1-rc.5 from the official npm package.
  • TOMC 0.2.1 source runtime with MCP 1.30.0 and the additive DSH setup command.
  • Dedicated synthetic notebook databases and DSH home under ignored local output. No manuscript, private conversation, or production notebook was used.

Official MCP bridge

The reproducible check mounts DSH's real Cordis SystemPrompt, ToolRuntime, and @deepseek-ai/dsh-mcp-client. It starts the TOMC Python stdio server and dispatches through ToolRuntime.execute; it does not replace the bridge with a mock or a different MCP client.

Check Result
Discovery All 10 TOMC tools registered under mcp__tomc__
Write Saved the synthetic Project Lark note through DSH
Persistence A fresh Cordis context and Python process recalled the note
Isolation Unrelated notebook marker absent from recalled memory
External history 22,516 estimated input tokens → 495 estimated memory tokens under a 512-token budget; requested port retained
Status get_status completed through DSH

The deliberately repetitive synthetic history is a transport/budget fixture. Its reduction is not a representative compression rate or measured API saving. This bridge test makes no model inference request.

Machine-readable bridge evidence

To reproduce, supply the installed DSH package.json, TOMC's Python executable, and a new output directory:

node scripts/validate_deepseek_mcp.mjs /absolute/path/to/node_modules/@deepseek-ai/dsh/package.json /absolute/path/to/TOMC/.venv/bin/python /absolute/path/to/new-test-output

The script creates a synthetic SQLite database; do not point it at existing data. On Windows, use the virtual environment's Scripts/python.exe.

Web UI

The actual DSH Web profile boots with the generated TOMC overlay. This machine's browser sent an Origin without the non-default port, causing DSH to reject API calls with 403. A loopback port-80 launch restored matching authority and normal UI access. No authentication or trust check was disabled. This local workaround is not a requirement for other installations.

Live DeepSeek model test

On 2026-10-29, the actual DSH Web UI used the official DeepSeek provider with model route deepseek-flash (UI label: DeepSeek-V4.1-Flash 1017, reasoning off). This test made real model requests with a working API credential.

  1. Session A called mcp__tomc__remember_memory exactly once to save a synthetic Project Cedar notebook. The tool returned saved: true, entries: 1.
  2. A new, separate session B called mcp__tomc__recall_memory exactly once with a 512-token budget. Its prompt supplied only the notebook name and requested fields, not the answers. The model correctly reported all four recorded facts: port 4826, unchanged public API, pending retry tests, and marker cedar-7K29.
  3. A follow-up in session B produced a three-part checklist using the recalled port, API constraint, and pending tests, with no additional tool calls.

The marker is synthetic, not a real credential. The official DSH session logs confirm two distinct sessions, the two named TOMC calls, successful tool results, and the model responses. Only these synthetic prompts/results and model routes are included in the sanitized live evidence. System prompts, local paths, credentials, and unrelated runtime metadata are omitted.

The short recalled note used the raw path: 41 estimated input tokens → 41 estimated memory tokens, 0% reduction. This verifies cross-session persistence and model use of recalled constraints; it provides no API savings claim and does not test long-history answer quality. DSH's own prompt/tool overhead is separate.

DSH model saving the synthetic note

New DSH session recalling all four facts

Follow-up checklist based on recalled constraints

Repository checks

50 focused Python tests passed, covering setup generation, MCP calls, delivery limits, and existing integrations. The strict documentation build and generated Claude Code runtime consistency check passed. macOS/Linux DSH operation has not been tested in this run.

A separate, newly created Python 3.11 virtual environment also installed .[mcp] from source, generated its own overlay, and passed the same real DSH bridge check. This checks the documented fresh-install path as well as the existing development environment.

Installation · 中文安装说明