dsh-code-intel
面向 DeepSeek Harness 的符号感知代码大纲和混合代码搜索。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add dsh-code-intel@0.1.3Code Intel 以 dsh-code-intel 发布，当前版本为 0.1.3。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Symbol-aware code outline, persistent workspace indexing, and explicit lexical or embedding-assisted search for DeepSeek Harness.
The v0.1.3 release is tested with DSH 0.1.0-rc.8 and 0.1.1-rc.1 while retaining the rc.6-compatible peer range. Prebuilt packages are distributed through GitHub Releases; npm publication is prepared but not yet live.
code_search ranks AST symbol chunks and bounded module windows, returning path:line, symbol metadata, snippets, scores, and the active retrieval mode.code_outline parses one file immediately or projects a directory from the persistent index.code-index background job instead of blocking the agent turn..dsh/code-index/ uses DSH filesystem versions for incremental rebuilds.fs/observed marks successful DSH file operations dirty; a Chokidar watcher covers local shell, IDE, and external changes.mode: lexical.DeepSeek's official API does not currently expose a public embedding endpoint, so this plugin does not invent one or silently label lexical results as semantic.
On the first code_search or directory code_outline, the tool returns an indexing result and a job id. Inspect or wait for that job with the built-in jobs tools, then call the code tool again. File outlines do not require a complete workspace index.
The cache contains derived source text and optional vectors. It stays inside the workspace, excludes .dsh from its own scan, and may be deleted at any time for a full rebuild. A retrieval-configuration fingerprint automatically invalidates vectors when the embedding endpoint or model changes.
Source traversal and reads use the mounted DSH filesystem service. SQLite and Chokidar require the provider's processPath() to be accessible from the plugin host, so the MVP targets the standard local filesystem execution world. If a remote filesystem exposes paths that are not host-accessible, cache creation fails loudly instead of indexing a different directory.
Node 22 currently labels the built-in SQLite module experimental. The on-disk cache is disposable and schema-versioned; source files remain the only source of truth.
Lexical mode scores exact query matches, query tokens, symbol names, and paths. Hybrid mode combines that score with cosine similarity over endpoint-produced vectors. A configured endpoint or credential failure fails the indexing job or search; fallback occurs only when embedding is not configured.
Credentials are references, not secret values in YAML. The plugin resolves the reference through ctx.credentials for each indexing or query operation and never stores the credential.
.dsh/code-index/ by default.The package supports DSH >=0.1.0-rc.6 <0.2.0 plugin APIs and Node.js ^22.19 || >=24.
dsh plugin --profile web add https://github.com/lonelymoon87/dsh-code-intel/releases/download/v0.1.3/dsh-code-intel-0.1.3.tgz
The release tarball is prebuilt and needs no build allowance. A pinned source install is also supported:
dsh plugin --profile web add github:lonelymoon87/dsh-code-intel#v0.1.3
The source install runs this package's prepare build. pnpm 10 and later reject it until the profile allowlists the exact package key printed by the failed command; apply that instruction and rerun the same dsh plugin add command. Replace web with headless to install into the one-shot agent profile.
To upgrade, rerun dsh plugin add with the newer release URL. To uninstall:
dsh plugin --profile web remove dsh-code-intel
Lexical mode needs no provider configuration:
- id: code-intel
name: dsh-code-intel
config:
indexDir: .dsh/code-index
include: [.ts, .tsx, .js, .jsx, .py, .go, .rs, .java]
exclude: [.dsh, .git, node_modules, dist, build, coverage, vendor]
maxFileSize: 1000000
maxChunkChars: 12000
maxResults: 20
watch: true
embedding: false
Hybrid mode uses a complete OpenAI-compatible embeddings URL:
embedding:
provider: openai-compatible
endpoint: https://embedding.example/v1/embeddings
model: your-embedding-model
credentialRef: EMBEDDING_API_KEY
batchSize: 32
indexDir and exclusions must stay workspace-relative. Only the listed, parser-backed extensions are accepted in MVP.
Tests use real temporary workspaces and SQLite databases. They cover all parser families, background indexing, lexical retrieval, direct outline, fs/observed refresh, hybrid ranking, credential absence, vector persistence, cache replacement, and invalid configuration.
dsh --dump-config.latest and next npm tags.