Single tool (get_library_sources) with well-structured parameters, comprehensive descriptions, and clear intent. Tool name follows verb_noun pattern. Input schema is complete with types and descriptions. However, output schema is not formally documented (returns string), error handling relies on text responses without actionable recovery guidance, and no tool annotations present. The implementation shows good practices (validation, ecosystem detection, caching) but lacks production-grade error categorization and structured output.

• Define formal output schema. When library_name is None, return {"ecosystems": {<ecosystem>: [{"identifier", "version", "is_direct"}]}, "total_count": N}. When library_name is specified, return {"matches": [{"identifier", "version", "worktree_path", "status"}], "failed": [{"identifier", "version", "error"}]}.
• Replace freetext error strings with structured responses: {"status": "error", "code": "directory_not_found|no_ecosystem_detected|no_matches", "message": "...", "recovery": "Check path exists and is directory" or "Try transitive=true to search transitive dependencies"}.
• Add tool annotations to schema: {"readOnlyHint": false, "destructiveHint": true, "idempotentHint": false} to signal that the tool clones repos (write/destructive).
• Add optional limit and offset parameters to support pagination: limit (1 - 100, default 20) and offset (default 0). Return total_dependencies_available in response.
• Remove example values from parameter descriptions. Replace 'e.g. hibernate matches org.hibernate:hibernate-core:6.4.1' with 'Performs case-insensitive substring match against full library identifiers in format group:artifact:version'.
• Document per-dependency error handling. When a source cannot be resolved for one library, return it in the "failed" array with a specific reason (e.g., 'Source repository not found for org.example:lib:1.0.0, no public repo URL available'), not a generic message.
• Add dry-run parameter (boolean, default false) to preview what would be cloned without actually cloning. Agents can validate intent before irreversible writes.

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

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