Mixed quality across 9 tools. validate_content has excellent documentation (detailed wire contract, halt semantics, domain enum); scrub_pii and lens_health are minimal. Validator tier tools (lens_catches_*, lens_consistency_*) lack input validation guidance and output schema documentation. All tools have descriptions, but 5 of 9 are under 100 chars (below baseline 194). No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles. Parameter schemas present but sparse, most lack format/constraint guidance. Error handling is generic (jsonError helper) with no recovery guidance.

• Add tool annotations to all 9 tools: readOnlyHint=true for 8 tools, destructiveHint=true for lens_catches_add. Agents need to know which calls have side effects.
• Document output schemas for validator tier tools. Example: lens_catches_relevant should return {catches: [{id, type, severity, file, line}], total_count}. This enables chaining and structured result extraction.
• Expand validator tier descriptions to 100-150 chars each, explaining the specific check (e.g., 'Deterministic cross-file consistency check for numeric values, detects mismatched figures across documents'). Current descriptions are indistinguishable.
• Enhance error responses with recovery guidance. Example: 'File not found: /path/to/file. Ensure the path is absolute and the file exists.' instead of bare error messages.
• Add format/constraint guidance to file_path parameters. Example: 'Local file path (absolute or relative to CWD; must exist and be readable)' to prevent LLM hallucination of invalid paths.