project-vectorizer defines 4 tools with explicit schemas and descriptions. All tools have input schemas visible in the code (JSON Schema with types), descriptions present, and are READ_ONLY. However, the server exhibits several quality gaps: (1) descriptions are minimal, averaging ~25 chars, well below the 194-char baseline and insufficient to guide LLM selection; (2) parameter descriptions are sparse or missing entirely (e.g., 'file_type' in list_files has no description despite being optional); (3) no output schemas are documented, tools return JSON strings but the structure (fields, types, nesting) is not specified; (4) no error handling guidance is provided, tools catch exceptions and return bare error strings ('Error: {e}') without telling the LLM what to do next; (5) no parameter constraints are enforced (e.g., 'limit' and 'threshold' accept any value with no validation); (6) the tool chain is weak, search_code returns results with minimal context for follow-up calls. The naming is clear and verb-based (search_, get_, list_), and all tools are safely READ_ONLY, but overall the definitions lack the depth needed for confident LLM reasoning at scale.

• Expand tool descriptions to 100 - 200 characters. Example for search_code: 'Search the vectorized codebase using semantic similarity. Returns ranked code snippets with file paths and relevance scores. Use limit (1 - 100, default 10) and threshold (0 - 1, default 0.5) to filter results. Call get_file_content to read full files after identifying matches.'
• Document output schemas for all tools. Example: search_code returns {query: string, total_results: int, results: [{file_path: string, snippet: string, score: float, line_number: int}]}. This enables LLMs to plan downstream calls.
• Add parameter descriptions and constraints. Example: 'limit (int, 1 - 100, default 10): Maximum number of results to return. Higher values slow search but find more matches.' Add 'threshold (float, 0 - 1, default 0.5): Cosine similarity threshold. Results below this score are filtered out.'
• Implement error recovery hints. Instead of 'Error: {e}', return: 'Search returned 0 results. Try: (1) broadening query terms, (2) lowering threshold to 0.3, or (3) calling list_files() to explore available code.' This guides LLM reasoning.
• Enrich search_code response with chaining metadata. Include file_path, line_number, snippet, and relevance_score in each result. This lets the LLM extract exactly what it needs and pass file_path directly to get_file_content.
• Add input validation and clamp parameters. Enforce limit ∈ [1, 100], threshold ∈ [0.0, 1.0]. Return clear errors: 'Invalid limit: 1000. Must be 1 - 100.' This prevents runaway calls and LLM hallucinations.

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry