dsh-semble

dsh-semble

Semantic code search for DeepSeek Harness.

Ask where something lives instead of recalling what it is called, and the agent finds the file, reads it, then edits.

English | 简体中文

dsh-semble brings Semble's code index into every DSH session as two native MCP tools. grep needs a string you already know; this needs a description of behaviour — "where is the retry policy applied when a request fails", "which module decides the scoring weights". That is usually the moment an agent starts guessing file names.

Nothing to learn beyond that: after installation the tools sit next to read and grep, and the agent decides when the index is worth a call.

What you get

Tool	Hand it	Back comes
mcp__semble__search	a sentence describing the code or where it lives	ranked file:line hits with snippets
mcp__semble__find_related	one location you already know	its neighbours: other implementations, callers, tests

The bundle also registers one slash command, /semble, for the two reporting subcommands that MCP does not expose: /semble savings and /semble file-types. See Commands.

A typical exchange:

You: Where does DSH decide that a plugin is "loaded"?

Agent: calls search, gets a hit in the plugin inventory, opens it with read, finds isEntryLoaded comparing loader entry names against the package name, and answers from the source rather than from a guess.

Using the tools

• repo — required. Absolute path of the repository, or an https:// git URL. A list searches several repositories at once.
• top_k — default 5. Raise to 10–20 when the first pass misses.
• max_snippet_lines — default 10; 0 returns locations only.
• content — code (default), docs, config, or all.

The first query against a repository builds its index (slow, one-time); later queries reuse it. Hits are ranked by similarity, so a score near 0.0x usually means "nothing relevant here" rather than "the best answer".

Install

Desktop app

Settings → Plugins → Install, and paste:

dsh-semble

Pin a version with dsh-semble@2.1.0. The settings page is the only supported route for the desktop profile — the CLI refuses it because the Electron app owns that profile's node_modules and lockfile.

CLI

dsh plugin --profile web add dsh-semble

The subcommand runs pnpm add inside the profile and reconciles dsh.profile.bundles, so the bundle mounts on the next start. There is no build step, so pnpm needs no allowBuilds exemption.

From source

To track main, a fork, or an unreleased commit, install from the repository instead:

dsh plugin --profile web add github:CodeIntegrity/dsh-semble

github:owner/repo, git+https://…, https://github.com/owner/repo and .tgz URLs are accepted; these need git on PATH.

Local checkout

dsh plugin --profile web add link:/absolute/path/to/dsh-semble   # macOS / Linux
dsh plugin --profile web add link:D:\path\to\dsh-semble          # Windows

The profile then points at your working copy, so edits to cordis.patch.yml apply on the next start without reinstalling.

Prerequisites

uv must be on PATH — uvx launches Semble from PyPI:

winget install astral-sh.uv                      # Windows

brew install uv                                  # macOS
curl -LsSf https://astral.sh/uv/install.sh | sh  # macOS / Linux

First start downloads the package and the embedding model (~1–2 min, one-time, offline afterwards). Restart DSH once after installing; until then the settings row may read 待重启.

Settings

Edit the semble-mcp row in Settings → Plugins:

Field	Default	Why change it
args	--from 'semble[mcp]' semble	append --content all to index docs and config too
args (version)	floating	semble[mcp]==X.Y.Z pins a release
env	empty	HF_ENDPOINT on restricted networks, SEMBLE_CACHE_LOCATION to move the index
toolCallTimeoutMs	300000	the first call builds the index and downloads the model; lower it to fail faster
cwd	inherited	keep unset; per-call repo values should be absolute paths
command	uvx	point it at uvx.exe if only a .cmd shim is installed

Changing settings restarts the MCP connection; disk index caches survive.

Tuning

Semble reads its configuration from the environment of the process that runs it. For the MCP server, put these in the env field of the semble-mcp row. /semble spawns a separate process and inherits the host environment instead, so export a value in both places when a report must agree with search — SEMBLE_CACHE_LOCATION is the one that matters.

Variable	Default	Effect
SEMBLE_INDEX_WORKERS	4	worker processes that chunk a repository during the first index; 0 chunks in-process. Raise it on a many-core machine, lower it to stop a background search from saturating the CPU
SEMBLE_MCP_IDLE_TIMEOUT	0 (never)	seconds an idle repository index stays in the in-memory cache; with the 10-index cap, 1800 releases repositories you have moved on from
SEMBLE_MODEL_NAME	minishlab/potion-code-16M-v2	alternative Model2Vec embedding model, as a local path or a Hugging Face repository
SEMBLE_MAX_FILE_BYTES	1000000	files above this size are skipped; raise it for generated code worth indexing
SEMBLE_CLONE_TIMEOUT	60	seconds allowed for cloning an https:// repository
HF_ENDPOINT	https://huggingface.co	mirror used for the one-time model download

Excluding files

The indexer honours .gitignore, plus a .sembleignore at the repository root for files git keeps but search should skip: generated clients, fixtures, vendored trees. The syntax is gitignore's, negation included:

# .sembleignore
packages/client/generated/
test/fixtures/
*.min.js
!*.proto          # force-include a file .gitignore excludes

A pattern that only ever needs one repository should stay in that repository's .sembleignore; keep the plugin settings free of per-repository rules.

Several repositories

repo accepts a list, so one call can compare a service with the shared library it depends on:

{ "query": "where is the retry policy applied", "repo": ["D:\\work\\service", "D:\\work\\shared"] }

Git URLs work the same way and are cloned once into the cache. Each repository gets its own index; a list costs one first-time index build per repository.

Commands

MCP exposes search and find_related; Semble's two reporting subcommands only exist in its CLI. The bundle registers /semble for them:

Command	Prints
/semble savings	tokens saved by returning snippets instead of whole files, per period and per call type
/semble file-types	the extensions Semble indexes, as JSON grouped by content type

Both run in the host process: nothing is sent to the model, no approval prompt appears, and the output is the CLI's own (colour disabled). The same uvx --from "semble[mcp]" semble installation is used, so /semble works exactly where the MCP server works.

savings reads the cache that the MCP server wrote, so it only counts searches made against the same SEMBLE_CACHE_LOCATION. To see a summary of everything, and for maintenance, stay in a terminal:

uvx --from "semble[mcp]" semble savings        # same report, from a shell
uvx --from "semble[mcp]" semble clear          # drop caches
uvx --from "semble[mcp]" semble clear index    # force a rebuild

/semble is deliberately not a search shortcut — that is what the MCP tools are for — and it has no destructive subcommands, because a slash command cannot ask for confirmation. Type /semble alone for the usage line.

Platform notes

The bundle ships no OS-specific path and no region-specific endpoint, so one installation serves all three platforms. What differs is only what you type:

Concern	Windows	macOS / Linux
Executable	uvx.exe, resolved through PATH	uvx
Repository paths	D:\work\repo	/home/me/repo
Profile home	C:\Users\\.dsh	~/.dsh
Retrieved paths	keep \ separators	POSIX /

Bundled skill

Installation also registers one routing skill, semble-code-search. The tools do not advertise when they are worth calling, and a model that reaches for grep first never learns that the index existed.

DSH loads skills lazily:

When	Enters the session	Cost
Before the first request	one catalog entry (name + capped description)	a few hundred bytes per session
Model calls skill	the full skills/semble/SKILL.md body	~2 KB, only when used
You type /semble-code-search	same body, injected into that step	manual

/semble-code-search loads instructions: the search itself still runs through the MCP tools. The real executable command is /semble (see Commands), which runs a CLI report in the host and costs no model tokens.

The body never enters the system prompt. To keep the tools and drop the skill, delete the ctx.skills.register(...) block in lib/index.js; /semble keeps working because it is registered separately.

How it works

The bundle inserts two loader rows:

1. semble-mcp-anchor — a no-op plugin loaded under this package's own name. DSH decides "loaded" by matching loader entry names against the installed package name, and the MCP client row loads under @deepseek-ai/dsh-mcp-client; without the anchor the settings row reads 未加载 / 待重启 while the tools work fine.
2. semble-mcp — the official @deepseek-ai/dsh-mcp-client row, configured to start uvx --from "semble[mcp]" semble over stdio.

Semble itself is not vendored. The plugin declares how to launch it from PyPI with a floating spec, so upstream bugfixes and new tools arrive on the next server start; only a breaking change to the CLI surface would need an edit here.

Update

Re-run the install with a newer version, or press 更新 in settings:

dsh plugin --profile web add dsh-semble@2.1.0

Source installs re-run with the new ref, e.g. dsh plugin --profile web add github:CodeIntegrity/dsh-semble#v2.1.0.

Uninstall

Settings → Plugins → 卸载, or:

dsh plugin --profile web remove dsh-semble

Installed the 1.x source build? Remove it by its old name first: dsh plugin --profile web remove @local/semble-mcp. The package was renamed in 2.0.0, and the anchor row only matches an install whose package name is dsh-semble.

The cached index and the HuggingFace model stay on disk and are reused after a reinstall.

Troubleshooting

Symptom	Cause	Fix
Row reads 待重启 after install	the host loaded its layer list before the install	restart DSH once
Tools work, row reads 未加载	cordis.patch.yml lost the anchor row	reinstall this version or newer
mcp__semble__* never appears	uvx missing, or the first download failed	run uvx --from "semble[mcp]" semble --version and read the error
First search takes minutes	one-time model download	wait, or set HF_ENDPOINT
First search fails on toolCallTimeoutMs	index build + model download outran the limit	raise toolCallTimeoutMs, or warm the cache once with a CLI search
Search looks stale after a large refactor	the disk index reuses files it believes unchanged	uvx --from "semble[mcp]" semble clear index, then search again
/semble answers "Could not start the Semble CLI"	uvx is not on the host's PATH	install uv, restart DSH
savings counts fewer calls than you made	the report reads the cache of the MCP row's env, which /semble does not inherit	set SEMBLE_CACHE_LOCATION in the host environment as well
ERR_PNPM_GIT_RESOLVE_FAILED	no network route to github.com	use a tarball URL or a mirror

Development

git clone https://github.com/CodeIntegrity/dsh-semble
dsh plugin --profile web add link:$(pwd)   # then edit and restart DSH

No build step: lib/ is plain ESM and cordis.patch.yml is read directly. The only runtime dependency is uv.

npm test runs the suite with Node's built-in runner. The spawn tests replace uvx with the Node binary, so they neither download Semble nor touch your cache.

Publishing

Releases are tag-driven by .github/workflows/publish.yml:

npm version minor          # bumps package.json and creates the vX.Y.Z tag
git push origin main --follow-tags

The workflow runs the tests, refuses to continue when the tag and package.json disagree, skips a version that is already on the registry, and publishes with npm provenance. It needs one repository secret:

Secret	What it is
NPM_TOKEN	an npm granular access token with read/write access to dsh-semble and 2FA bypass enabled for automation

Provenance requires a public repository; on a private one, drop --provenance from the publish step. A manual run of the workflow (workflow_dispatch) defaults to publishing the current package.json version; tick dry_run to only run npm pack --dry-run.

License

MIT