The Agent Skills Directory

[PROMPT_INJECTION]: The skill is designed to process untrusted data from external sources, which constitutes an indirect prompt injection surface.
Ingestion points: In Phase 1, the agent reads network traces, event logs, and HTTP payloads. In Phase 2, it ingests error messages and output from external reproduction commands. The hitl-loop.template.sh script captures arbitrary user input for the agent to parse.
Boundary markers: The instructions do not define specific delimiters or instructions to ignore embedded commands within the ingested data.
Capability inventory: The skill makes use of shell execution (bash), network utilities (curl), version control systems (git bisect), and file system operations for logging and instrumentation.
Sanitization: No explicit sanitization or validation of the external content is prescribed before it is processed by the agent.
[COMMAND_EXECUTION]: The skill requires the agent to generate and execute shell commands and scripts to verify bugs.
Evidence: Phase 1 explicitly instructs the agent to define and run a "red-capable" command, such as a script path or test invocation. The skill also provides and encourages the use of scripts/hitl-loop.template.sh for interactive tasks.

diagnosing-bugs