The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a systematic failure capture and diagnosis workflow that involves ingesting untrusted data from the execution environment. This data could contain malicious payloads designed to subvert the agent's introspection logic or recovery actions.
Ingestion points: The Phase 1: Failure Capture section in SKILL.md instructs the agent to record error messages, stack traces, and logs from the environment.
Boundary markers: The workflow uses Markdown templates (e.g., ## Failure Capture) to delimit captured data, which provides basic structural separation but may not prevent sophisticated injection attempts.
Capability inventory: The skill assumes the agent has active capabilities to perform recovery, including file system access, branch management, and process state inspection, as noted in the Recovery Heuristics and Phase 3 sections.
Sanitization: There are no explicit instructions within the skill to sanitize, escape, or validate the content of error logs or tool outputs before they are processed by the agent during the diagnosis phase.

agent-introspection-debugging