The Agent Skills Directory

[COMMAND_EXECUTION]: The skill safely uses subprocess calls to interact with other components and run evaluation scripts. In scripts/rewards/execution.py, it calls a relative path to a 'memory' skill (memory/run.sh) to perform database lookups for grounding rewards. These calls use the list-based subprocess.run method, which correctly prevents shell injection vulnerabilities. Similar patterns are found in scripts/train_grpo.py for running the evaluation harness.
[EXTERNAL_DOWNLOADS]: The training pipeline is configured to download model weights and tokenizers from HuggingFace Hub, which is standard behavior for fine-tuning tasks. Remote API calls are directed to well-known or local services, including Chutes (http://localhost:4001) and Ollama (http://localhost:11434).
[CREDENTIALS_UNSAFE]: The codebase contains default development keys (e.g., 'sk-dev-proxy-123' in scripts/rewards/relevance.py). These are clearly identified as placeholders for local proxy configurations and do not represent leaked production credentials.
[SAFE]: The core training logic implements standard reinforcement learning patterns. The presence of auto-generated files in unsloth_compiled_cache is a standard artifact of the Unsloth optimization library. The watchdog script (scripts/watchdog.py) is used for legitimate task monitoring and ETA calculation.

create-intent-map