The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to ingest external production traces and logs to reproduce issues. This creates a potential surface where malicious instructions embedded in log data could be interpreted by the agent during the diagnosis process.
Ingestion points: Ingestion of production request payloads and logs is described in references/feedback-loop-recipes.md under the 'Trace / log replay' section.
Boundary markers: The skill does not provide explicit instructions or delimiters to isolate ingested data from the agent's control logic.
Capability inventory: The skill is authorized to use Bash, Read, Write, Edit, Grep, and Glob tools as specified in SKILL.md frontmatter.
Sanitization: No sanitization or validation steps are defined for the ingested log or trace data.
[DYNAMIC_EXECUTION]: The skill provides a human-in-the-loop (HITL) harness script that uses the eval command to execute shell strings dynamically populated by the agent.
Evidence: The scripts/hitl-loop.template.sh file contains eval "$SETUP_LINE" and eval "$CHECK_LINE" which are intended to be filled with arbitrary shell commands generated by the agent during execution.

diagnose-loop