Server has 3 tools with complete input schemas and descriptions. Naming follows verb_noun pattern (get_, identify_, audit_). Descriptions are substantive (100-200 chars) and explain WHAT and WHEN. However, output schemas are not documented, LLMs cannot predict response structure. Error handling is minimal (generic exception messages). Parameters lack constraints (e.g., aggressive_mode has no guidance on when to enable). No tool annotations (readOnlyHint, idempotentHint). Metrics logging is privacy-aware but not exposed to the agent.

• Document output schemas for all three tools. Example: get_dpdpa_requirement returns {"results": [{"section": string, "text": string, "source": "act_2023" | "rules_2025"}], "total": number}. This enables LLMs to chain results to downstream tools.
• Add readOnlyHint: true to all three tools in the schema. This signals to agents that these tools are safe to call without side effects and can be retried.
• Replace free-form 'topic' parameter with an enum: ["consent", "erasure", "notice", "data_fiduciary", "data_principal", "retention", "breach_notification"]. Add description: 'Select a DPDPA topic. If unsure, start with "consent" or "data_fiduciary".'
• Enhance error messages with recovery guidance. Example: 'Failed to download DPDPA Act PDF. Ensure internet connectivity and try again. If persistent, the document may be temporarily unavailable, try a different topic or check https://www.meity.gov.in.'
• Document aggressive_mode tradeoff in identify_bharat_pii: 'Enable for obfuscated/embedded Aadhaar detection (slower, ~2-5x latency). Disable for standard patterns (fast, <100ms). Default: false for speed.'.