The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the Bash tool extensively to interact with git, docker, and uv for version comparison and running story-based evaluations.
[REMOTE_CODE_EXECUTION]: The skill dynamically executes code from specific git branches or tags provided via the --baseline argument to set up the control group for testing.
[CREDENTIALS_UNSAFE]: Home Assistant access tokens are extracted from stderr output and passed as plaintext command-line arguments to the ha_query.py script, exposing them to potential process monitoring.
[INDIRECT_PROMPT_INJECTION]: The skill processes session logs and story definitions which could contain adversarial prompts intended to manipulate the evaluator's scoring or logic.
Ingestion points: Story definition files in tests/uat/stories/catalog/*.yaml and session log files (session.json, session.jsonl).
Boundary markers: None mentioned; data is parsed and utilized directly by the evaluation logic.
Capability inventory: Subprocess execution via Bash and Task, Docker management, and local file system modifications.
Sanitization: No specific content sanitization or instruction-ignoring markers are used when processing the data.
[DYNAMIC_EXECUTION]: The skill executes small Python snippets at runtime using python3 -c to extract and process data from the JSON and JSONL session files.

bat-story-eval