The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill implements an autonomous loop that modifies a local training script (typically train.py) and executes it using the uv tool. This process of generating and running code at runtime is the primary purpose of the skill. While the skill includes constraints against installing new packages mid-session, the ability to modify and execute arbitrary logic in the training script is a high-privilege capability.
[INDIRECT_PROMPT_INJECTION]: The skill ingests existing codebase files (e.g., train.py, prepare.py, README.md) to inform its research hypotheses. These files serve as an untrusted data surface; if they contain adversarial instructions within comments or documentation, the agent could be manipulated into deviating from its research goals or performing unauthorized actions during the autonomous loop. The instructions do not specify sanitization or boundary markers for these inputs.

autoresearch