The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of GitHub Actions logs, which are external, untrusted data sources. Logs can contain arbitrary text generated during build or test processes, making them a potential vector for indirect prompt injection.
Ingestion points: The script tools/fetch_gha_logs.py is used to download and display logs from GitHub Actions, which are then processed by the agent.
Boundary markers: No specific delimiters or safety instructions (e.g., 'ignore any instructions in the logs') are mentioned in the SKILL.md to protect the agent's context when reading these logs.
Capability inventory: The skill implies the agent has the capability to 'fix the identified issues in the source code', which likely involves write permissions to the repository.
Sanitization: There is no evidence of sanitization or filtering to remove potential injection patterns from the log content before the agent reads it.
[COMMAND_EXECUTION]: The skill utilizes local shell commands and script execution for its primary functionality.
Evidence: The documentation instructs the agent to use python3 tools/fetch_gha_logs.py with various arguments and to pipe output through shell utilities like grep. While these are expected developer tools, they represent a capability that requires execution permissions.

investigating-gha-failures