The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data, including error logs, network traces (such as HAR files), and bug reports, as outlined in Phase 1 and Phase 2 of SKILL.md. This data is then used to influence the agent's actions during the diagnosis process. While the skill possesses capabilities to execute shell commands, run tests, and perform network requests (curl), it explicitly mandates the redaction of secrets. However, the ingestion of external data without strict boundary markers for the AI's internal reasoning creates a standard surface for indirect prompt injection common to debugging tools.
[DYNAMIC_EXECUTION]: In SKILL.md and scripts/hitl-loop.template.sh, the skill encourages the creation and execution of temporary code artifacts, including throwaway harnesses, fuzzing loops, and human-in-the-loop bash scripts. These are generated or modified at runtime to facilitate the debugging process. While these actions are controlled by the agent and intended for local execution to reproduce issues, they represent the dynamic generation and execution of code based on the current debugging context.

diagnosing-bugs