DeepSeek Harness Plugin

zilliztech/memsearch#MemSearch

Stars ★ 2915 Downloads (30d) 28,001 Category Memory Added 2026-08-24 npm @zilliz/memsearch-dsh

Shared Markdown memory for DSH and other coding agents, with automatic capture, pre-step context injection, searchable recall, and memory-to-skill self-evolution through a review panel.

Install

# from npm (prebuilt)

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

# from GitHub (first run asks for allowBuilds approval — follow the hint, retry)

dsh plugin --profile web add github:zilliztech/memsearch#path:/plugins/dsh

Any plugin you install runs third-party code with your own permissions — it can read your files, use your credentials, and reach the network, and tool approvals don’t sandbox it. GitHub-sourced plugins also run build scripts at install time — pnpm blocks those until you allow them, so an install can stop with ERR_PNPM_GIT_DEP_PREPARE_NOT_ALLOWED or ERR_PNPM_IGNORED_BUILDS; dsh prints the exact key to add under allowBuilds in your profile’s pnpm-workspace.yaml, and the install works on the next run. Allowing a build is a trust decision: only install sources you trust, and pin a commit (github:owner/repo#sha).

README

MemSearch plugin for DeepSeek Harness (DSH). It gives DSH persistent, cross-agent memory on the same .memsearch/memory/ markdown store used by the Claude Code, Codex, OpenClaw, and OpenCode plugins, backed by a Milvus hybrid search index.

capture  ── session/event turn/end ──> summarize (dsh-headless agent default, or custom-llm) ──> memory/YYYY-MM-DD.md
hint     ── agent/pre-step step 1  ──> remind the agent to use memory-recall when needed
recall   ── native memory-recall skill ──> search → expand → transcript
review   ── web UI dock panel ──> GET/POST /memsearch-dsh/* ──> list candidates / queue review / install

Prerequisites

  • Install memsearch:

    uv tool install "memsearch[onnx]"
    
  • A DSH profile (web / headless / tui) you want to attach memory to.

  • Node >= 22.19 (DSH's requirement).

  • Python 3 on the profile's PATH (python3, python, or the Windows py -3 launcher).

Install

From npm (recommended)

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

For default summaries with an API key saved on DSH's Models page, give headless jobs access to saved keys. Add or update this top-level entry in ~/.dsh/profiles/headless/cordis.patch.yml, or ask memory-config to do it.

- id: credentials
  name: '@deepseek-ai/dsh-credentials-local'
  disabled: false

From source (development)

cd /path/to/memsearch/plugins/dsh
npm install --omit=peer
dsh plugin --profile web add /path/to/memsearch/plugins/dsh

dsh plugin is a pnpm forwarder: it links the package into the profile, detects the dsh.bundle declaration in package.json, and appends it to the profile's bundle layers. cordis.patch.yml inside the package then inserts the memsearch row into the profile's plugin tree.

Replace web with your profile name (headless, tui, ...), and restart DSH for the profile (or start a new session) so the plugin mounts.

Development tests

Run npm test in plugins/dsh for the Node tests. The maintenance integration suite also requires the Python core from this checkout:

# From the repository root
uv sync --locked
cd plugins/dsh
uv run --project ../.. npm run test:integration

The integration suite invokes the registered timer and session callbacks through runtime process spawning and the real Python maintenance runner. It replaces only native LLM generation with deterministic local output; no model service, embeddings, database, or running DSH host is needed. CI runs both suites.

Manual patch insertion (no dsh plugin)

Append this row to the profile's cordis.patch.yml and make sure @zilliz/memsearch-dsh is resolvable from the profile's node_modules (for example a link: dependency):

- insert:
    - id: memsearch
      name: '@zilliz/memsearch-dsh'

Verify capture

Complete a turn, then open today's file in <project>/.memsearch/memory/ and check for a summary of what you discussed.

Configuration

Embeddings run locally with ONNX when neither global nor project MemSearch configuration exists. Existing configuration is used as-is. Configure embedding and Milvus settings through memsearch config or the memory-config skill.

The plugin is configured through the profile's cordis.patch.yml config block (patch the memsearch row you inserted). All keys are optional.

Key Type Default Meaning
captureEnabled bool true Capture completed turns into memory.
injectEnabled bool true Remind the agent to use memory-recall before each turn's first step when memory exists.
summarizeEnabled bool true Summarize turns before writing (on failure a short unavailable note is written, never a raw dump).
summarizeMode string auto Use custom-llm when [plugins.dsh.summarize] provider is set, or dsh-headless otherwise. Select either backend explicitly to pin it.

Everything else — provider/model, Milvus, collection, memory dir — comes from memsearch config / environment, exactly like the other platform plugins (no per-plugin config fields):

  • Summarize provider/model → [plugins.dsh.summarize] provider / model in ~/.memsearch/config.toml (or [llm.providers.*]; see the custom-llm section below).
  • Milvus → [milvus] uri in memsearch config.
  • Collection → derived from the project path (derive-collection.sh), overridden by [milvus] collection in memsearch config or --collection passed to the memsearch CLI.
  • Memory dir → MEMSEARCH_DIR env (explicit → global scope), else <project>/.memsearch.

When upgrading from a Windows version that used a separate DSH collection, startup rebuilds the shared index from existing Markdown. The original files and old collection are retained; explicit [milvus] collection settings still apply.

Maintenance tasks (PROJECT.md / USER.md / skills)

Optional background upkeep, aligned with the other platform plugins. Each task is disabled by default; enable the ones you want in ~/.memsearch/config.toml:

[plugins.dsh.project_review]
enabled = true            # maintain .memsearch/PROJECT.md
[plugins.dsh.user_profile]
enabled = true            # maintain .memsearch/USER.md
[plugins.dsh.memory_to_skill]
enabled = true            # distill recurring workflows into skill candidates
min_occurrences = 3       # how often a workflow must recur before distilling

Common settings per task: provider (native = a one-shot DSH headless agent, default), model, min_interval_hours (default 24), input_dir, output_file. Candidates land in .memsearch/skill-candidates/ (git-tracked) and are never installed automatically — installing is a human step (see the memory-to-skill skill in the other platform plugins).

Example override layer (add this to the profile's own cordis.patch.yml):

- id: memsearch
  config:
    summarizeMode: dsh-headless   # pin the headless backend (default is auto)

Summarization modes

Two backends are available, selected by summarizeMode — the same "configured choice" the Claude Code / Codex / OpenClaw / OpenCode plugins offer (each can summarize with their own LLM or a headless agent + small model). The default (auto) matches theirs: configure a provider and you get a direct LLM call; configure nothing and you get a headless agent.

  • auto (default) — mirrors the other platform plugins:
    • if [plugins.dsh.summarize] provider is set in memsearch config (~/.memsearch/config.toml, same place the other plugins read), use custom-llm with that provider/model;
    • otherwise use dsh-headless. This means the plugin behaves like the other four: configure a provider → direct LLM; configure nothing → headless.
  • dsh-headless — boots a one-shot DSH headless agent (dsh --profile headless "<summarize task>") to write the notes, mirroring how the other plugins reuse their own agent's headless mode. The sub-agent uses the deployment's agent-default-model — the user layer of ~/.dsh/settings.yaml (the same selection the Web UI model settings write) wins over any patch, so the [plugins.dsh.summarize] provider/model do NOT apply here — change the model in DSH settings (agent-default-model: in ~/.dsh/settings.yaml, or the Web UI model picker) instead. The boot is asynchronous and fire-and-forget, so the few seconds of headless startup never block the conversation. Requires dsh on PATH or DSH_CLI set to the CLI entry. Desktop users can enable Manage dsh Command… → Install in the application menu; the Windows dsh.cmd launcher uses Desktop's bundled runtime. Restart DSH after changing its command environment. The sub-agent is booted with MEMSEARCH_DSH_SUMMARIZE=1; the plugin checks that flag and stays inert (no capture / inject / skill) inside the summarizer, so the summarizer's own session is never re-captured in a loop.
  • custom-llm — scripts/summarize.py imports memsearch's [llm.providers.*] config and calls the LLM directly. Lightweight: one python process, no DSH boot, no extra CLI dependency. Choose this when you want a specific small model (e.g. an official deepseek-v4-flash key in memsearch config) without booting an agent. Provider selection (most specific first):
    1. [plugins.dsh.summarize] provider (or the summarizeProvider CLI argument summarize.py receives from it) — looked up in [llm.providers.<name>]; a missing entry fails loudly (visible error), never a silent empty write.
    2. llm.provider when it names a configured provider or is a raw type.
    3. compact.llm_provider (deprecated) or openai as a final default.

There is no automatic fallback between modes: the backend you configure (or auto resolves) is the backend used. If it fails (missing dsh CLI, bad provider config), a short unavailable note is written with the reason — the plugin never silently switches to an LLM you did not configure.

A failed summarization writes a short unavailable note (mirroring Claude Code's behavior — memory stays clean, the transcript anchor keeps the raw content reachable for progressive disclosure), and logs a visible warning through the DSH logger.

How it works

  • Capture — listens on session/event for turn/end, renders the turn ([User] / [Assistant] / [Tool call] lines), then fire-and-forget summarizes (if enabled) and appends it to the session's own memory/YYYY-MM-DD.md with the shared anchor format <!-- session:<id> turn:<N> db:<path> -->. The project directory comes from session.header.cwd, so a long-lived web surface captures every project it hosts, not just the process's boot directory. Turns are serialized (LLM summarize calls never overlap) and captureExists dedup keeps each turn idempotent if its event replays.
  • Recall hint — on agent/pre-step at step 1, adds a brief reminder to use memory-recall when the project has memory files. The agent decides whether past context is useful and invokes the skill when needed.
  • Recall — registers a memory-recall skill (invocable through DSH's native skill tool) that performs search → expand → transcript drill-down and returns a curated summary. Its search collection is derived from the current working directory when the skill is loaded.
  • Maintenance — runs the shared maintenance runner (PROJECT.md / USER.md upkeep and memory-to-skill distillation), triggered on session/disposed plus a 6-hourly fallback timer. Each task is a due-state machine: it runs at most once per min_interval_hours (default 24h) and only when enabled in memsearch config ([plugins.dsh.project_review], [plugins.dsh.user_profile], [plugins.dsh.memory_to_skill]), mirroring the other platform plugins. The maintenance work is executed by a one-shot DSH headless agent (the same dsh --profile headless mechanism as summarization), booted with MEMSEARCH_DSH_SUMMARIZE=1 so the plugin stays inert inside it. Each native maintenance call has a ten-minute limit; the full runner allows 31 minutes for all three tasks and cleanup. Failed tasks can retry on a later trigger, and interrupted runs do not permanently block maintenance.
  • Skill review panel (web only) — a non-blocking dock strip above the composer (registered into the conversation.input.dock slot) lists skill candidates distilled into .memsearch/skill-candidates/. It is served by the plugin's browser half (client.js, declared through dsh.client in package.json) and talks to the host through two JSON routes on the DSH web server:
    • GET /memsearch-dsh/skill-candidates?sessionId=<id> — lists candidates (pending first) parsed from each candidate's meta.json.
    • POST /memsearch-dsh/skill-review with { sessionId, name, action }:
      • action: "review" queues a [memsearch] Skill candidate ... user message into the live agent's inbox (agent.inbox.append('next-turn', ...)). The agent reviews the candidate on its next turn — non-blocking, it never interrupts a running turn and never opens a blocking dialog.
      • action: "install" runs memsearch skills install <name> --path <dir> in the background (detached, unref'd) to the resolved target: the first entry of plugins.dsh.memory_to_skill.paths in memsearch config (relative entries resolve against the project dir), else the DSH default ~/.agents/skills, which the skill-filesystem provider watches and loads automatically. The project directory is resolved from the session id: a live agent's session cwd first, then the session's persisted header cwd (so merely browsing an old conversation — no live agent yet — still shows that conversation's project, not the boot directory), finally the process cwd. The panel therefore reflects the project of the session you are viewing on a multi-project web surface. The routes are registered once the DSH web server service becomes available (the memsearch plugin mounts in the base bundle layer, before the web server starts; apply retries on a 1s unref'd timer). Headless / TUI profiles have no browser: the webServer service never appears, no routes are registered, and everything else is unchanged.

Uninstall

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

Removing the dependency drops the profile-layer entry; the memory markdown files and the Milvus index are left untouched.

Development

  • The plugin is plain ESM with no build step — dsh plugin add links the checkout directly, so edits are live after a profile reload.
  • The browser half (client.js) is a prebuilt client-module bundle: it registers its factory with window.__ModuleLoader__.load({ id, factory }) (the lazy CJS table the web shell serves at /plugins/<id>/client.js) and exports the Cordis client plugin shape (inject + apply). It is checked in as-is; edit it directly and keep the __ModuleLoader__ registration format (mirror the shipped @deepseek-ai/dsh-client-ui-* lib/client.js bundles).
  • Python helpers under scripts/ are linted with the repo's ruff config and tested under plugins/dsh/tests/.
  • Keep the memory-write format byte-compatible with the other platform plugins; see the OpenCode journal writer in scripts/capture.py for the shared heading and transcript-anchor format.

Command setup

The plugin uses explicit command settings first. DSH_CLI selects the headless DSH entry, MEMSEARCH_CMD selects MemSearch, and MEMSEARCH_PYTHON selects the Python interpreter for helper scripts. Use an executable path, a quoted command with arguments, or a JSON argument array; a path containing spaces must remain one argument. An invalid explicit setting is reported rather than replaced with another installation.

Without an explicit setting, DSH is resolved from PATH, then ~/.local/share/pnpm/dsh. MemSearch uses memsearch on PATH, then ~/.local/bin/uvx, then uvx on PATH.

A missing DSH or Python command produces a dependency warning. Enable Desktop's CLI entry or configure the required command, then restart the profile. A command that starts but fails reports its execution error; check that command's profile, provider, and model configuration. The plugin does not change these settings or switch summary backends automatically. The Markdown journal retains the existing unavailable note and transcript anchor when summarization fails.

Content from the project README on GitHub ↗

Links

More in this category

View the whole category →

Community comments

Comments are public GitHub Discussions. Loading them connects to GitHub and Giscus; a GitHub account is required to post.