The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data which could contain malicious instructions designed to influence the agent's behavior.
Ingestion points: The skill reads deployment records from gate/latest.json, gate/summary.md, gate/runs/*.json, and external CI artifacts downloaded via the GitHub CLI. These files contain 'snippets' and 'evidence' from the codebase which can be influenced by any contributor to the repository.
Boundary markers: The instructions do not provide delimiters or specific 'ignore instructions' warnings when the agent interpolates or quotes data from these records into its report.
Capability inventory: The skill can execute a local Python script (.github/actions/gate/run.py), perform network operations via gh run download, and commit files to the repository.
Sanitization: There is no mention of sanitizing or escaping the content retrieved from the gate records before it is processed by the LLM.
[COMMAND_EXECUTION]: The skill requires the execution of several commands to perform its duties.
Overlay Discovery: Uses a shell command [ -n "$POWERSUIT_OVERLAY" ] && ls -1 "$POWERSUIT_OVERLAY/gate" 2>/dev/null || true to detect configuration overlays.
Local Script Execution: Runs python3 .github/actions/gate/run.py for both dry-runs and manual record generation. This relies on the security of the local run.py script.
GitHub CLI: Uses gh run download to fetch remote CI artifacts, which involves network communication with GitHub services.
Version Control: The agent is instructed to commit changes to latest.json and summary.md upon user request.

run-record