dsh-workspace-intelligence

In LLM-driven code-agent workflows a recurring pain point looks like this: the model needs to reference a function or a class but has no idea which file it lives in. DeepSeek Harness ships with LSP tools such as goToDefinition and findReferences, yet all of them require a starting point—a file path and a cursor position—before a query can even be issued. When an agent is handed an unfamiliar codebase and only a symbol name, none of those tools can help. dsh-workspace-intelligence fills exactly that gap with a single workspace_symbol tool that lets the agent resolve a symbol by name across the entire workspace.

Under the hood the plugin speaks standard LSP workspace/symbol over stdio to one or more language servers. A single query fans out concurrently to every configured server while keeping per-server requests serial to preserve protocol invariants. Results are stably sorted—exact match, prefix match, then the rest—and de-duplicated on a composite key of name, kind, URI, and optional range, so the model always receives a clean, deterministic list. The plugin relies exclusively on DSH public interfaces (ctx.fs, ctx.subprocess, ctx.tools); it never patches the Harness runtime and never touches ctx.lsp, meaning existing goToDefinition, hover, findReferences, and other LSP features remain completely unaffected.

It is aimed at power users who already run a language server—Pyright, TypeScript Language Server, or any stdio-based server—in their DSH or Cordis profile and want to give their agent a reliable find-by-name capability. Whether the agent is completing a cross-module call, locating an interface for an integration test, or confirming which namespace a symbol belongs to during a code review, workspace_symbol removes a round of guesswork and returns a tidy set of name, kind, canonical URI, and optional range.