The Agent Skills Directory

Indirect Prompt Injection: The skill ingests data from Metro terminal logs using the BashOutput tool. These logs contain output from the running application and its dependencies. If the application processes and logs data from untrusted sources (such as external API responses or user input), there is a potential consideration where malicious instructions could be embedded in the logs to influence the agent's interpretation of errors. The skill's ability to edit files and execute shell commands based on these logs increases the potential impact, though the structured monitoring loop and model guardrails serve as mitigations.
Ingestion points: Metro terminal output is read via BashOutput during Phase A and Phase 0.5.
Boundary markers: The skill lacks explicit boundary markers or "ignore" instructions for the processed log content.
Capability inventory: The skill utilizes Edit, Write, and Bash tools, allowing for significant system and codebase changes.
Sanitization: There is no evidence of sanitization or filtering of the terminal output before it is processed by the agent.
Autonomous Command Execution: The skill uses the Bash tool to perform environment setup (e.g., mkdir), log filtering (e.g., grep), and project validation (e.g., npm run type-check). While these are standard development tasks, the autonomous execution of shell commands based on automated log analysis is a capability that warrants user awareness.
Dynamic Code Modification: To facilitate diagnostics and bug resolution, the skill autonomously modifies source files to inject console.log trace statements and apply surgical bug fixes. This involves the agent making direct changes to the application's logic, which is a powerful capability intended for debugging but requires user review of the resulting diffs.
External Dependency Management: The skill may identify missing Node.js modules and suggest or perform their installation. While the skill correctly requires user consent before executing npm install, users should verify that suggested packages are legitimate and not subject to typosquatting.

debug-app