Scoring was not performed

• Document return types and output field schemas for every tool. Add a 'Returns' section to each tool's description showing: {field_name: type, description}. Example: 'Returns: {files_processed: int, reduction_pct: float, normalized_source: str}'.
• Expand parameter descriptions to 50-100 chars with explicit format/constraint guidance. Change 'Path patterns to exclude' to 'Glob patterns to exclude from search (e.g. **/node_modules, **/*.test.py). Relative to project root.'
• Add explicit recovery guidance to error-prone tools. For apply_deduplication, append: 'If errors occur, use rollback_rewrite(backup_id, project_folder) to restore. Always run with dry_run=true first.'
• Clarify tool selection. Add 'Alternative tools' section to find_duplication/analyze_deduplication_candidates: 'Use find_duplication() for raw results. Use analyze_deduplication_candidates() for ranked/filtered results. Use apply_deduplication() to commit changes.'
• Document path parameter format. Add note: 'path: Absolute or relative path to file/directory. LLM will resolve relative paths from current working directory. Accepted formats: POSIX and Windows paths.'
• Add idempotency notes to stateful tools. Example for rewrite_code: 'Idempotent when dry_run=true (no side effects). When dry_run=false, repeated calls may fail if rules no longer match changed code, use backup_id to restore.'
• Remove/document the 'complexity_guided=False' reserved feature. Either implement it, remove it, or explicitly note: 'Currently unused. Future versions will enable automated complexity-guided depth selection when set to true.'