The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted external data such as log dumps and traces that may contain malicious instructions. Ingestion points: SKILL.md specifies reading logs, HAR files, and traces. Boundary markers: No explicit delimiters for untrusted data, though secret redaction is required. Capability inventory: SKILL.md allows shell command execution, file writes, and network operations. Sanitization: Prompt injection sanitization is missing.
[DYNAMIC_EXECUTION]: The skill guides the agent to generate and run arbitrary reproduction scripts and harnesses. Evidence: SKILL.md Phase 1 lists multiple methods for script generation and execution. Evidence: scripts/hitl-loop.template.sh provides a bash script template for human-interactive loops.

diagnosing-bugs