The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and act upon data from external tool outputs, such as stack traces, build logs, and test failures, which are considered untrusted ingestion points.
Ingestion points: The scripts/detect-failure.sh hook reads tool result JSON from the execution environment. The 'Heal Loop' in SKILL.md requires the agent to diagnose failures directly from tool output.
Boundary markers: The skill relies on the agent's internal logic to process error messages and does not provide explicit boundary markers or 'ignore' instructions for the ingested data.
Capability inventory: The skill empowers the agent to write new helper scripts and modify project source code, which are then executed during the 'VERIFY' step.
Sanitization: There is no evidence of explicit sanitization or filtering of the error messages before they are used to influence the agent's remediation logic.
[COMMAND_EXECUTION]: The skill includes shell scripts that execute logic via the Python interpreter to process JSON and text data.
Evidence: scripts/detect-failure.sh uses python3 -c for parsing hook payloads. scripts/find-similar-heals.sh uses a Python heredoc to perform regular expression searches on the local .learnings/HEALS.md file.
[EXTERNAL_DOWNLOADS]: The skill documentation and test fixtures reference external repositories and tools from trusted sources.
Evidence: References to the gh CLI (GitHub's official tool) and the browser-use/browser-harness repository on GitHub. These are classified as well-known or trusted services.

self-healing