@xxinchen/dsh-plugin

One command → 13 working memory tools → your agent stops re-explaining the project. Memories are governed, not hoarded: stale/conflict/redundancy checks, and nothing becomes permanent until you confirm it in the inbox.

Minta — context quality layer for DeepSeek Harness.

Installing this package composes two Cordis rows into a DSH profile:

• mcp-client-minta — official @deepseek-ai/dsh-mcp-client wired to the local Minta engine's streamable-HTTP endpoint (its 13 Community minta_* tools reach the agent).
• minta-plugin — this package as a real Cordis plugin. At mount it registers a runtime skill (minta-memory-governance, content derived from the Minta interaction guide) and a agent/session-start hook that prewarms engine health + recent memory over the Minta REST API (fail-open: an unreachable engine never breaks a session).

The memory engine itself is a separately deployed service (see Prerequisites); the plugin is a thin DSH-native surface over it. See docs/dsh-integration.md for the verified end-to-end flow.

Prerequisites

• DeepSeek Harness (npx @deepseek-ai/dsh web)
• Minta engine running locally: python minta_cli.py start (serves MCP at 127.0.0.1:18721 and API at 8772), or docker compose up -d

Install (verified 2026-08-23)

▼
bash
复制代码
dsh plugin --profile web add @xxinchen/dsh-plugin

Restart dsh web. The MCP client row is composed automatically — no manual config editing. Verify in a new session: "list all tools whose name contains minta" → the 13 mcp__minta__* Community tools.

Note: the engine must be running before the tools respond; failOnStartupError: false means a stopped engine does not break DSH sessions.

Using Minta in a session

1. Keep the engine running — python minta_cli.py start (or Docker).
2. After dsh plugin add + one restart of dsh web, start a new session and ask: "list all tools whose name contains minta" — expect the 13 mcp__minta__* tools. If they are absent, restart dsh web once more.
3. First session: call minta_login (your Minta credentials) so the tools operate with your identity.
4. Automatically at session start: the plugin checks engine health and recent memory (logged; never blocks the session). The skill minta-memory-governance is registered and available to the model.
5. Manual first answer: say "用 Minta 先把记忆找出来再回答" — or just ask about prior work; the skilled model will minta_search_context first.
6. To save something durable: tell the model to remember it — it will minta_append_inbox, and you confirm or discard in the Minta UI (http://127.0.0.1:8772, inbox review). Your memory is never edited silently.
7. Engine stopped / unreachable: sessions keep working; tools and prewarm just report unavailable (fail-open).

Minta agent preset (per-turn protocol)

For turn-level automation (DSH rc.2 has no per-step/turn-end events, so the protocol is carried by the preset persona, which the model follows every turn), install the preset shipped in this package:

▼
bash
复制代码
mkdir -p ~/.dsh/.agent-presets
cp -r <node_modules>/@xxinchen/dsh-plugin/presets/minta ~/.dsh/.agent-presets/  # or cp -r the repo's dsh-plugin/presets/minta

Then, in DSH, pick Minta in the session preset picker — or set it as the default by overriding the agent-presets row (in ~/.dsh/profiles/web/cordis.patch.yml):

▼
yaml
复制代码
- id: agent-presets
  config:
    default: minta

The preset persona adds the per-turn protocol: minta_autopilot_preflight before answering about prior work, minta_autopilot_postflight after each reply, inbox-only writes, and graceful degradation when the engine is offline.

Migrating from the manual cordis.patch.yml block

If you previously set Minta up by appending the mcp-client-minta block to ~/.dsh/profiles/web/cordis.patch.yml, remove that block before installing this plugin. The bundle patch and the manual block insert the same loader entry id mcp-client-minta; with both present the profile fails to boot with duplicate loader entry id: mcp-client-minta. Check you are clean with:

▼
bash
复制代码
grep -A6 mcp-client-minta ~/.dsh/profiles/web/cordis.patch.yml   # expect no output

(A patch file containing only [] is valid and means "no overrides".)

Manual alternative (no npm install)

Append to ~/.dsh/profiles/web/cordis.patch.yml:

▼
yaml
复制代码
- insert:
    - id: mcp-client-minta
      name: '@deepseek-ai/dsh-mcp-client'
      config:
        transport: streamable-http
        serverName: minta
        url: http://127.0.0.1:18721/mcp
        failOnStartupError: false

Roadmap and known boundaries

Shipped: MCP wiring, runtime skill registration, session-start prewarm.

Not here yet (and why):

• Automatic pre/post-flight per turn — DeepSeek Harness 0.1.1-rc.2 exposes no per-step or turn-end agent events, so per-message autopilot stays a model-facing tool call (minta_autopilot_preflight / minta_autopilot_postflight, instructed by the registered skill) rather than a hidden automatic hook.
• Prompt injection of the prewarm result — the profile-level plugin cannot safely shadow the deployment persona (the prompt registry owns that slot); prewarm currently logs engine state, and the agent reads memory on demand through the MCP tools.