dsh-code-intel

In large, multi-language codebases, agents constantly need to locate symbols, understand directory structure, or search across files. dsh-code-intel adds a symbol-aware code intelligence layer to DeepSeek Harness: the first search kicks off a cancellable background indexing job instead of blocking the current turn, and subsequent queries hit a SQLite-based persistent index with incremental refresh so every search is not a cold start.

The core tools are code_search, which ranks AST symbol chunks and bounded module windows and returns path:line, symbol metadata, snippets, scores, and the active retrieval mode, and code_outline, which parses a single file on the fly or projects an entire directory from the index. Language coverage spans TypeScript, JavaScript, Python, Go, Rust, and Java via install-script-free Tree-sitter WASM grammars. Retrieval defaults to pure lexical mode with zero external configuration; an optional OpenAI-compatible embedding endpoint can layer cosine similarity on top for hybrid ranking. When no endpoint is configured or credentials are absent, results explicitly state mode: lexical—the plugin never silently labels lexical matches as semantic.

Built for developers working multi-language projects in DSH who want persistent, workspace-local code awareness. The index lives under .dsh/code-index/, is disposable, schema-versioned, and safe to delete at any time for a full rebuild. The plugin sends no telemetry, resolves credential references without storing secret values, and makes no network calls in lexical mode. In hybrid mode, only extracted code chunks and queries are sent to the explicitly configured endpoint.