The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill attempts to programmatically extract sensitive credentials from local system files.
Evidence: The authentication detection script in SKILL.md searches ~/.git-credentials and the .env file located in the user's home directory ($HOME/.hermes/.env) to harvest the GITHUB_TOKEN.
[INDIRECT_PROMPT_INJECTION]: The skill implements an automated 'Auto-Fix Loop' that processes untrusted external data and performs sensitive repository actions.
Ingestion points: CI logs downloaded from GitHub Actions via the /actions/runs/{run_id}/logs endpoint (referenced in SKILL.md and references/ci-troubleshooting.md).
Boundary markers: None. The agent is instructed to read the raw log content directly to understand errors.
Capability inventory: The skill uses write_file, patch, git commit, and git push to apply fixes derived from the untrusted logs (SKILL.md).
Sanitization: None. There is no validation of the log content before it influences the agent's code modification decisions.
[DYNAMIC_EXECUTION]: The skill uses inline Python scripts to process JSON data from shell commands.
Evidence: Multiple instances in SKILL.md and references/ci-troubleshooting.md use python3 -c to execute dynamically provided code strings for parsing API responses.
[COMMAND_EXECUTION]: The skill executes shell commands using variables derived from git remotes and user input.
Evidence: The script extracts OWNER and REPO from git remote get-url origin using sed and cut, then interpolates these into curl commands and API endpoints.

github-pr-workflow