@dreamor/dsh-memvault

Cordis plugin bridging MemVault into DeepSeek Harness (dsh).

The dsh/Cordis API surface it relies on (ctx.systemPrompt.section(), session/event payloads, Plugin.Object resolution) was verified against the actual deepseek-harness source, not secondhand docs.

What it does

• Auto-injection: keeps memory://session-inject cached and registers it as a system prompt section (ctx.systemPrompt.section()), refreshed eagerly at load and again on every turn/end. No agent tool call required.
• Auto-extraction: on turn/end with reason.kind === 'completed', calls MemVault's notify_response with both that turn's assistant text and the user's own turn text (user/message events with source.kind === 'user') automatically. Passing the user's own words matters — MemVault's extractor signal words are first-person ("我偏好"/"我喜欢"), which match a user's own statement far more reliably than an assistant's restatement of it.

If you only want the raw 20 MemVault tools exposed to the model (no auto-injection/auto-extraction), you don't need this package — dsh's own @deepseek-ai/dsh-mcp-client does the same with zero code: each upstream server gets its own Cordis plugin instance via cordis.patch.yml (transport: stdio, serverName: memvault, command: <path to memvault-proxy>).

Install

Prerequisites: a working dsh install (e.g. npx @deepseek-ai/dsh web) and the memvault-proxy binary reachable from PATH (or point binaryPath at one — see the override section below).

Build, then install into a dsh profile from the repo's dsh-plugin/ directory:

cd dsh-plugin
npm install
npm run build      # tsc --strict
npx @deepseek-ai/dsh plugin --profile <profile> add "$PWD"

dsh plugin add runs the equivalent of pnpm add and appends the package to that profile's dsh.profile.bundles automatically — no manual edit of the profile's cordis.patch.yml is needed for a default install. The plugin ships a cordis.patch.yml that registers one memvault entry with the default config (mode: spawn, embeddingProvider: native, …).

Verify the composed config, then start:

npx @deepseek-ai/dsh --profile <profile> --dump-config   # 组合后应只有一条 id: memvault 且字段正确
npx @deepseek-ai/dsh <profile> --port 0                  # 实际启动方式以你的 dsh 用法为准

At startup the plugin spawns memvault-proxy on port 3778 and probes http://127.0.0.1:3778/health until the server is ready. Transport, port and db overrides are passed as CLI flags, so an existing ~/.memvault/proxy.yaml (including upstreams and markdown_dir) is loaded but never rewritten. To confirm the injection actually lands: start a real conversation turn, then decompress the session log (zstd -d ~/.dsh/sessions/.../session.jsonl.zstd) and check that request/header.data.header.system contains the injected [MUST]/[REF] blocks.

Overriding config (local development)

To override a single field (e.g. point binaryPath at a locally built memvault-proxy), add a bare-id override patch to the profile's cordis.patch.yml and rewrite the whole config block — an override replaces the entire config, it does not merge per-field:

- id: memvault
  name: '@dreamor/dsh-memvault'
  config:
    mode: spawn
    db: '~/.memvault/data.db'
    port: 3778
    binaryPath: /absolute/path/to/memvault-proxy
    embeddingProvider: native
    agentId: dsh
    injectOnAssemble: true
    extractOnTurnEnd: true

The cordis.patch.yml this plugin ships wraps its entry in insert: because the entry doesn't exist in any earlier layer. A bare - id: patch is an override that requires the id to already exist and errors with patch: entry "memvault" not found otherwise — an override patch must rewrite the whole config block, it never merges per-field.

Config

Build

npm install
npm run build
npm test

Status

Builds clean (tsc --strict) against the real, npm-published @deepseek-ai/[email protected], @deepseek-ai/[email protected], @deepseek-ai/[email protected], and @deepseek-ai/[email protected] packages (not mocks — see src/index.smoke.test.ts, which mounts the plugin on a real Context alongside the real SystemPrompt service and asserts ctx.systemPrompt.assemble() actually contains the injected section).

Fully verified end to end against a real dsh install (npx @deepseek-ai/dsh web, v0.1.0-rc.6). Four real bugs were found and fixed in the course of the verification:

1. cordis.patch.yml insert vs. override patch semantics (a bare id patch requires the id to already exist).
2. memvault-proxy inheriting a stray OPENAI_API_KEY from dsh's own process env, plus a stale (pre-native-embedding) local binary.
3. The eager injection-cache refresh at apply() time racing memvault-proxy's own startup (fetch failed).
4. mcp-client.ts permanently caching a failed connection attempt — connected ??= (...) memoized on the first (raced, failing) call, so every later call kept re-rejecting forever, never retrying even after the proxy came up.

With all four fixed, a real conversation turn's session log (~/.dsh/sessions/.../session.jsonl.zstd, decompressed with zstd -d) shows the actual request/header.data.header.system text containing MemVault's injected [MUST]/[REF] memories, and a direct notify_response call against the running memvault-proxy moved MemVault's memory count from 8 to 10 (new preference/fact memories actually saved to Inbox).

A follow-up round extended notify_response to also extract from the user's own turn text — MemVault's extractor signal words are first-person, so a user's own statement ("我偏好用 tabs") matches far more reliably than an assistant's restatement of it ("你偏好用 tabs" used to be silently rejected; the signal-word list was also extended with second/third-person mirrors). Verified against the real running memvault-proxy: memory count went from 10 to 13 across three scenarios, with source:user/source:assistant tags on the results.

DSH 0.2 compatibility has also been checked with @deepseek-ai/dsh-*@0.2.0-rc.2 and a real [email protected] plugin add / --dump-config run; the plugin keeps its wildcard peer range so DSH 0.1 installs remain usable.