The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to ingest and process external data such as logs, traces, and captured artifacts (HAR files, core dumps). This creates a potential surface for indirect prompt injection if these files contain malicious instructions. However, the skill explicitly mandates a redaction step to remove secrets and sensitive signals before processing, which serves as a manual sanitization boundary. The risk is considered inherent to the debugging task and is mitigated by the safety instructions.
[DYNAMIC_EXECUTION]: The workflow encourages the creation and execution of throwaway harnesses, bisection scripts, and headless browser scripts (e.g., Playwright/Puppeteer) to reproduce bugs. While this involves generating and running code at runtime, it is a standard and necessary component of the claimed diagnostic functionality. The instructions provided are for local debugging purposes and do not involve executing untrusted remote code.
[COMMAND_EXECUTION]: The skill facilitates the use of various command-line tools including curl, git bisect, and custom shell scripts (scripts/hitl-loop.template.sh) for reproduction loops. The provided bash template is benign, serving only to structure human-in-the-loop interactions and capture observations without dangerous command interpolation.
[CREDENTIALS_UNSAFE]: The skill actively prevents credential exposure by instructing the agent to redact secrets and use environment variables rather than hardcoding credentials into loops or outputs, aligning with security best practices.

diagnosing-bugs