The Agent Skills Directory

[PROMPT_INJECTION]: The skill defines a strict multi-phase debugging process. It uses directive language to ensure the agent follows a methodical approach rather than jumping to conclusions. No malicious prompt injection or safety bypass patterns were identified.
[COMMAND_EXECUTION]: The skill facilitates the execution of local development tools such as curl, git bisect, and test runners (e.g., Playwright) to create feedback loops. These are standard practices for software debugging and are scoped to the user's local environment.
[DATA_EXFILTRATION]: While the skill involves analyzing logs, network traces, and core dumps which might contain sensitive information, this data remains within the local debugging context. No patterns for exfiltrating this data to external or untrusted domains were found.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it ingests and processes untrusted data from external sources during the debugging process.
Ingestion points: Log files, network traces (SKILL.md), user-provided bug reports, and external codebase documentation.
Boundary markers: No explicit markers are defined to isolate untrusted content from the agent's instructions.
Capability inventory: The agent possesses capabilities to read files and execute local commands (e.g., shell commands, test runners) across all phases of the skill.
Sanitization: No explicit sanitization or filtering of external logs or trace data is described before analysis.
[SAFE]: The skill includes a utility script scripts/hitl-loop.template.sh which allows the agent to interact with a human user to reproduce bugs in a structured way. The script itself is safe and does not perform any malicious operations.

diagnosing-bugs