AxonFlow SDK exposes 7 tools, all governance/policy-enforcement utilities rather than domain-specific operations. Tool names lack action verbs (mcp_check_input, mcp_check_output, check_tool_use, check_result, govern_tools, wrap_langgraph, proxy_llm_call). Descriptions are present but generic (50-150 chars). Input schemas are visible but sparse: most parameters lack type constraints, enums, or validation guidance. No output schemas documented. Error handling is implicit (PolicyViolationError mentioned but not detailed). Tools appear designed for framework integration rather than agent invocation, they wrap other tools rather than perform domain work. This is a governance SDK, not a tool server in the traditional sense.

• Rename tools to follow verb_noun pattern: mcp_check_input → validate_tool_input, mcp_check_output → validate_tool_output, check_tool_use → validate_computer_use, check_result → validate_execution_result, govern_tools → wrap_tools_with_governance, wrap_langgraph → wrap_graph_with_governance, proxy_llm_call → proxy_llm_with_governance.
• Document output schemas for all tools. E.g., mcp_check_input should return {allowed: bool, reason?: string, redacted_input?: object}. This enables LLMs to understand policy decisions and adapt.
• Convert 'statement' and 'message' parameters from JSON strings to structured objects. Accept {tool_name: string, input: object} directly instead of requiring LLMs to serialize JSON.
• Add enum constraints to 'operation' parameter: {type: 'string', enum: ['execute', 'query']}. Prevent hallucinated values.
• Enhance error descriptions with recovery guidance. E.g., 'Policy violation: PII detected in output. Redact sensitive fields or use check_result with redaction enabled.' This guides LLM retry logic.