Four tools with READ_ONLY risk classification. Tool naming lacks action verbs (my_risk, my_guard_detail, my_findings violate verb_noun pattern). Descriptions present but generic; 'assess' is strongest at ~280 chars. Critical gap: three tools have EMPTY input schemas ({}), violating schema requirement. Only 'assess' has a proper schema with typed properties. No parameter descriptions visible for any tool. No output schemas documented. Error handling not evident in source.

• Rename tools to verb_noun pattern: 'get_risk_assessment', 'get_guard_details', 'get_findings', 'assess_command_or_server'. This signals intent to LLMs and matches production baselines (90% of A+ tools start with action verb).
• Add explicit output schemas for all tools. Document fields like risk_band, score, reasons_count, last_assessed_time for my_risk; effective_rubric, active_parsers, loaded_rule_packs for my_guard_detail; etc.
• Add parameter descriptions to 'assess' tool: clarify when to use 'command' vs 'mcp_server' mode, explain what 'args' does, document expected format for mcp_server JSON object.
• Add error handling guidance: 'If assessment fails, check that Sigil policy is loaded. If command syntax is invalid, return the specific parse error and suggest corrections.'
• Document return types and example responses in tool descriptions. E.g., 'Returns {risk_band: string, score: number, reasons: string[], last_assessed: ISO8601}' helps LLMs understand what to expect.