@deepseek-ai/dsh-code-index-tree-sitter

tree-sitter provider for the code-index seam: offline, syntax-level TypeScript/Python/Java entity and import extraction over a workspace.

Model Experience

• Model-visible inputs: none directly — consumers decide what reaches a model request (arch-lens feeds extracted entities/imports into explain prompts).
• Token effects: none from indexing; extraction is offline AST work.
• KV-cache effects: none.

Role in the seam

• Provider of ctx.codeIndex (indexWorkspace(root) → packages with entities/imports/composition/entry files, plus manifest deps).
• Consumer: @deepseek-ai/dsh-arch-lens-backend (precise graphs and code-grounded explains).

Extraction coverage

Language	Imports	Entities	Composition	Modifiers
TypeScript	import/import type/namespace/named	class/interface/enum/type-alias/function	class → methods/fields	decorators
Python	import/from x import y	class/function	class → methods	decorators
Java	import	class/interface/enum/record	class → methods/fields/constructors	annotations

Indexing skips node_modules/dist/build/venv/target/lib and is bounded per package (400 files, 256 KiB per file).

Known Limitations and Deferred Work

• Syntax-level only; cross-file call resolution is left to consumers via the import graph. No LSP-backed semantic indexing (deliberate: process weight and per-language servers outweigh the value for graph precision).
• Python manifest dependencies are best-effort ([project].dependencies / poetry lines); Gradle dependencies are not parsed (pom.xml only).