Lethe MCP server has well-structured tool definitions with clear, domain-specific descriptions (avg 150-200 chars). All 11 tools have explicit input schemas with typed parameters and descriptions. However, output schemas are not formally documented in the code, responses are inferred from return statements. Parameter descriptions are generally good but lack explicit constraints (e.g., no min/max for k, no enum for 'kind' in log()). Error handling is minimal, no recovery guidance or actionable error messages visible. Tool naming is verb-forward and clear (inscribe, recall, release, purge, supersede, pin, unpin, consolidate, blame, log). No security issues detected (no secrets in params), but no explicit permission gates or audit logging visible.

• Add formal output schema definitions to each tool. Document the structure of returned dicts (e.g., recall returns [{id: int, text: str, depth: float, score: float, similarity: float}]). Use JSON Schema or explicit type hints in docstrings.
• Convert 'kind' parameter in log() to an enum: kind: Literal['inscribe', 'recall', 'release', 'supersede', 'purge', 'pin', 'flow', 'promote']. Add min/max bounds to 'k' (e.g., 1 - 100) and 'tau_seconds' (e.g., 1 - 31536000).
• Implement error handling with actionable messages. Wrap tool calls in try-catch; return structured errors like {error: 'signing_key_not_found', message: 'Set LETHE_SIGNING_KEY env var or place key at ~/.lethe/keys/purge_signing.key', recovery: 'Configure signing key and retry'}.
• Add a dry_run parameter to purge() and purge_signed() (default false). When true, return what would be deleted without actually deleting. Alternatively, add a confirm_purge(ids, token) tool that requires a confirmation token from a prior purge_preview() call.
• Implement audit logging. Log each tool invocation (caller, tool name, params, timestamp, result) to a structured log file or event stream. Add a query_audit_log() tool so agents can verify compliance and trace actions.