The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/run_eval.py script executes shell commands defined in the fixture field of evaluation specifications via subprocess.Popen. While intended for environment setup during evaluations, this allows arbitrary command execution within the user's environment.\n- [DYNAMIC_EXECUTION]: The evaluation runner in scripts/run_eval.py dynamically executes an external 'adapter' script specified at runtime and processes fixture setup logic from JSON-based specifications.\n- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted data including logs, transcripts, and external repository files.\n
Ingestion points: Processes external data like logs, transcripts, and upstream skill files as described in SKILL.md.\n
Boundary markers: Includes a specific 'Instruction boundary' section in SKILL.md to mitigate the risk of obeying instructions found in data.\n
Capability inventory: The skill includes powerful capabilities such as file system modifications and shell command execution via scripts/run_eval.py.\n
Sanitization: scripts/eval_contract.py performs path validation to prevent directory traversal in fixture definitions, though the command strings themselves are executed via shell.

skill-crafting