Tool-Calling Failure Debugger

The problem

Tool-calling workflows fail in production due to subtle schema drifts, ambiguous descriptions, or unsafe retry logic. Developers struggle to trace why an LLM selected the wrong tool, hallucinated arguments, or ignored approval gates during complex multi-step actions.

What it does

• Analyzes tool-call traces to pinpoint where discovery, selection, or argument construction diverged from intent.
• Audits JSON schemas and tool descriptions for ambiguities that trigger routing errors or validation failures.
• Evaluates safety controls including permission scoping, idempotency keys, and manual approval gating for destructive actions.
• Detects architectural risks like tool-output injection, infinite loops, and state loss in multi-turn sequences.
• Generates a remediation plan with a 100-point safety score and specific test cases for failure recovery.

Why this beats prompting it yourself

Standard prompts often miss edge cases like partial failure recovery or tool-output injection risks. This skill applies a structured 25-step audit workflow and a specialized failure catalog to ensure no part of the tool-execution lifecycle is overlooked.

Use cases

• Debugging why an agent repeatedly calls the same tool with invalid JSON arguments.
• Reviewing MCP tool routing and permissions before deploying to a production environment.
• Auditing side-effect safety for tools that handle payments, messaging, or data deletion.
• Designing robust retry and timeout logic for tools that interact with unstable third-party APIs.

Known limitations

This tool does not execute live API calls or modify production systems. It is a diagnostic and design-time auditor, not a runtime monitoring agent or automated testing suite.