This is a VSCode extension packaged as an MCP server with a single tool. The tool is explicitly registered in package.json under 'languageModelTools' with a name, description, and input schema. However, the tool has ZERO input parameters (empty properties object), making it a reference/informational tool with no actionable parameters. The description is adequate (194 chars) but the tool itself is not composable with other tools, it only returns documentation. The server lacks ANY error handling guidance, output schema documentation, or integration with downstream tools. No evidence of pagination, authentication, or recovery patterns. The tool is read-only (no state changes), which is safe, but it provides no agency to an LLM beyond 'fetch reference docs.' This is a documentation tool, not a functional agent tool.

• Rename tool from 'sdoc_reference' to 'get_sdoc_reference' or 'fetch_sdoc_docs' to use verb_noun convention and signal the action to LLMs.
• Add input parameters to allow filtering by section (e.g., section_name enum: 'syntax', 'scopes', 'formatting', 'tables', 'lists', 'metadata') so agents can request only relevant docs without processing the entire reference.
• Document the output schema explicitly: specify whether it returns plain text, markdown, structured JSON with sections, or HTML. Include max length and any truncation behavior.
• Add error handling guidance: 'If reference file is not found, check that the SDOC package is properly installed. Available sections: syntax, formatting, scopes, meta.' This gives LLMs a recovery path.
• Consider adding a 'get_sdoc_examples' tool that returns code examples by use case (e.g., 'create_table', 'define_scope', 'add_metadata') so agents can compose learning with task execution.
• Add tool annotations (readOnlyHint: true, idempotentHint: true) in the languageModelTools config to signal to LLMs that this is safe to call repeatedly.
• If the reference is large, implement a 'section' parameter and document max response size; consider pagination guidance.