The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to process external, untrusted artifacts to diagnose issues, creating a surface for indirect prompt injection.
Ingestion points: The agent is directed to read external files such as HAR traces, log dumps, and core dumps provided by users or environments in SKILL.md Phase 1.
Boundary markers: The skill lacks explicit data-instruction delimiters for these artifacts, though it does instruct the agent to "quote only the lines that carry the signal."
Capability inventory: The agent is authorized to execute complex diagnostic operations including shell script execution, network requests via curl, and browser automation using Playwright or Puppeteer.
Sanitization: The skill includes a dedicated "Redact" section that mandates replacing secrets with <REDACTED> before displaying output, which serves as a security control for data exposure but does not prevent command injection via data.
[DYNAMIC_EXECUTION]: The skill facilitates the creation and execution of dynamic code to reproduce bugs.
Patterns: Phase 1 encourages the agent to generate and run "throwaway harnesses," "differential loops," and "curl / HTTP scripts."
Evidence: The inclusion of scripts/hitl-loop.template.sh provides a baseline for generating interactive shell scripts that the agent executes and monitors for output.

diagnosing-bugs