The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill attempts to automatically retrieve authentication tokens by searching sensitive local files. It includes instructions to read ~/.hermes/.env and uses regex patterns to extract plaintext credentials from the ~/.git-credentials file.
[PROMPT_INJECTION]: The skill is designed to read and interpret CI failure logs from external sources to 'auto-fix' code, creating a surface for indirect prompt injection.
Ingestion points: Failure logs are read using the GitHub CLI or downloaded as zip archives from api.github.com (documented in SKILL.md and references/ci-troubleshooting.md).
Boundary markers: No delimiters are specified to isolate external log content from the agent's primary instructions.
Capability inventory: The agent uses file-modification tools (patch, write_file) and version control commands (git commit, git push) to apply changes based on its interpretation of these logs.
Sanitization: There is no evidence of log content validation or sanitization before it is processed by the agent.
[COMMAND_EXECUTION]: The workflow relies on shell command execution for its primary logic, including the use of python3 -c to parse API responses and unzip to process downloaded log data.
[EXTERNAL_DOWNLOADS]: The skill fetches CI logs and workflow data from GitHub's official API (api.github.com).

github-pr-workflow