Install the DeepSeek Harness Plugin¶

Prerequisites¶

Python 3.10+
Node.js 22.19 or newer, as required by DSH
A DSH profile such as web, tui, or headless
The dsh command available on PATH

1. Install MemSearch¶

The default ONNX embedding provider runs locally and does not require an API key:

uv tool install "memsearch[onnx]"

The embedding model downloads from Hugging Face the first time it is used and is cached afterward.

2. Add the Plugin to a DSH Profile¶

Install the published npm package into the profile you use:

dsh plugin --profile web add @zilliz/memsearch-dsh

Replace web with another profile name when appropriate. The command adds the package to that profile and inserts the MemSearch plugin into its bundle layers.

Restart the profile, or begin a new DSH session, so the plugin mounts.

Install from Source¶

Use a local checkout when developing the plugin:

git clone https://github.com/zilliztech/memsearch.git
dsh plugin --profile web add /absolute/path/to/memsearch/plugins/dsh

The plugin is plain ESM with no build step. Restart the profile after changing the plugin source.

3. Verify the Installation¶

Complete a normal DSH turn, then check the project memory directory:

ls .memsearch/memory/
cat .memsearch/memory/$(date +%Y-%m-%d).md

You can also ask DSH to use the registered skill:

Use memory-recall to search for our earlier database migration decision.

In a web profile, look for the compact MemSearch capsule above the composer. Expanding it shows skill candidates and the read-only .memsearch browser.

Plugin Settings¶

The plugin works without configuration. Its DSH profile settings control lifecycle behavior:

Setting	Default	Purpose
`captureEnabled`	`true`	Capture completed turns into the daily memory journal
`injectEnabled`	`true`	Search and inject relevant memory before the first model step
`summarizeEnabled`	`true`	Summarize turns before writing them
`summarizeMode`	`auto`	Use a configured API provider when present; otherwise use a one-shot DSH headless agent

To override these settings, patch the memsearch row in the profile's cordis.patch.yml:

- id: memsearch
  config:
    captureEnabled: true
    injectEnabled: true
    summarizeMode: auto

Embedding, Milvus, and provider settings continue to live in the shared MemSearch configuration. For example, to route DSH capture summarization through a configured provider:

memsearch config set llm.providers.openai.type openai
memsearch config set llm.providers.openai.model gpt-5-mini
memsearch config set llm.providers.openai.api_key env:OPENAI_API_KEY
memsearch config set plugins.dsh.summarize.provider openai

Leave plugins.dsh.summarize.provider unset to use the DSH headless-agent default. Setting a provider selects the direct custom-llm route. See How It Works for the backend behavior.

Optional Maintenance¶

These background tasks are disabled by default:

memsearch config set plugins.dsh.project_review.enabled true
memsearch config set plugins.dsh.user_profile.enabled true
memsearch config set plugins.dsh.memory_to_skill.enabled true

They maintain .memsearch/PROJECT.md, .memsearch/USER.md, and .memsearch/skill-candidates/. Skill candidates remain inert until you choose to install one.

Update¶

Update the package in the selected profile, then restart that profile:

dsh plugin --profile web update @zilliz/memsearch-dsh

Uninstall¶

dsh plugin --profile web remove @zilliz/memsearch-dsh

Removing the plugin does not delete .memsearch markdown files or the Milvus index.

Troubleshooting¶

No memory file appears: confirm the plugin mounted in the DSH session log and that captureEnabled is still true.
The recall skill is missing: restart the profile after installation and confirm @zilliz/memsearch-dsh is present in that profile.
Summarization reports an unavailable note: verify dsh is on PATH, or check the configured [plugins.dsh.summarize] provider.
The web dock is absent: the dock requires a web profile; capture, injection, and recall still work in TUI and headless profiles.
First search cannot download the model: pre-cache the ONNX model from an environment with access to Hugging Face, or configure another embedding provider.