The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources during bug diagnosis, creating a vulnerability to indirect prompt injection.
Ingestion points: Untrusted data enters the agent context through bug reports, log files, HAR traces, and outputs from commands or the interactive hitl-loop.template.sh script.
Boundary markers: The skill lacks explicit delimiters to distinguish untrusted bug data from agent instructions, although it does instruct the agent to manually redact secrets.
Capability inventory: The skill utilizes high-privilege capabilities including shell execution, network requests via curl, and browser automation tools like Playwright and Puppeteer.
Sanitization: There is no automated sanitization or filtering applied to ingested data before processing.
[COMMAND_EXECUTION]: The skill's fundamental operation involves executing arbitrary shell commands, scripts, and test harnesses to reproduce, minimize, and verify bug fixes.
[DYNAMIC_EXECUTION]: The skill instructs the agent to create and run scripts and code harnesses, such as the hitl-loop.template.sh or minimal reproduction scenarios, to facilitate the debugging process.

diagnosing-bugs