@deepseek-ai/dsh-code-index

Code-index capability seam: language-aware workspace entity/import extraction contract. Consumers (arch-lens graphs, code-grounded explains, future code search) read only this contract; providers implement it per language.

Model Experience

• Model-visible inputs: none by itself — this package is a host-side contract. Providers and consumers decide what reaches a model request.
• Token effects: none. Indexing is offline rule/AST work; no LLM calls.
• KV-cache effects: none.

Role in the seam

• Service Definition (ctx.codeIndex): indexWorkspace(root) → packages, entities, imports, composition, primary language.
• Provider: @deepseek-ai/dsh-code-index-tree-sitter (tree-sitter AST extraction for TypeScript/Python/Java).
• Consumer: @deepseek-ai/dsh-arch-lens-backend (precise dependency/ER graphs and code-grounded explain prompts).

Known Limitations and Deferred Work

• Syntax-level only: cross-file semantic resolution (which class a call binds to) is derived from the import graph by consumers, never from a language server.
• LSP-backed semantic indexing is deliberately deferred (process weight, per-language servers) unless a future consumer proves it necessary.