Reponova provides 6 read-only tools for code graph analysis with consistent naming conventions (all start with 'graph_' verb prefix). Tool descriptions are present and moderately detailed (averaging 120-180 characters), exceeding the minimum but falling short of optimization for LLM selection. All tools have input schemas with proper JSON Schema structure and type definitions. However, output schemas are NOT documented in the visible code, no return type descriptions are present for any tool. Parameter descriptions are included but often lack specificity around constraints, ranges, and error recovery paths. No error handling guidance is visible (e.g., what happens when a symbol is not found in graph_impact, or when embeddings are not built for graph_similar). No tool annotations (readOnlyHint, etc.) despite all tools being read-only. Resource management is well-architected with lazy loading and hot-reloading, but this is implementation detail, not user-facing quality.

• Document output schema for each tool. Example: graph_search returns {results: [{symbol, file_path, type, relevance_score}], total_count, next_offset}. graph_impact returns {upstream_dependencies: [...], downstream_dependents: [...]}.
• Add constraint ranges to numeric parameters: top_k (1-100, default 10), context_depth (0-5, default 0), max_depth (1-10, default 3), max_text_tokens (100-16000, default 8000).
• Enhance graph_similar description: 'Find semantically similar code nodes. Requires embedding model built during graph build phase. If embeddings unavailable, returns error, ensure reponova build completes with embedding mode enabled.'
• Add readOnlyHint: true annotation to all 6 tools in MCP registration, signaling idempotence and enabling LLM optimization.
• Clarify 'type' parameter enums and defaults: 'Filter by symbol type (function|class|module|all, default: all). Omit to match all types.'
• For graph_search, add guidance: 'Use repo parameter to filter by project name. Available repos shown in graph metadata, call graph_context on any file to see applicable repos.'
• Add error handling guidance to descriptions: graph_path, 'Returns empty if no path exists between symbols'; graph_impact, 'Set include_tests=true to include test file dependencies in blast radius'; graph_similar, 'Returns error if embeddings not available during build.'
• Consider pagination for graph_search results, add next_cursor parameter and return total_count to support large result sets without context blowout.

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

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