The Agent Skills Directory

[DYNAMIC_EXECUTION]: The module scripts/validator/main.py implements dynamic code loading using importlib.util.spec_from_file_location and spec.loader.exec_module. This is used to import model and configuration classes from generated files located in the neuron_port/ directory.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted model data and exposes an execution surface:
Ingestion points: scripts/validator/main.py and scripts/validator/enhanced_metrics.py load model configurations and tokenizers from model_path (HuggingFace repositories).
Boundary markers: Absent; the skill does not explicitly isolate or warn against instructions embedded in downloaded model files.
Capability inventory: The skill uses trust_remote_code=True in several transformers library calls (loading models, tokenizers, and configs), allowing third-party modeling code to execute on the host. It also executes shell commands for environment management.
Sanitization: Absent; model contents are trusted and executed as part of the validation pipeline.
[COMMAND_EXECUTION]: Shell scripts such as scripts/setup_autoport.sh and scripts/setup_dry_run.sh are used to manage system packages, Python virtual environments, and dependencies. These scripts include path manipulations and automated installation logic.
[EXTERNAL_DOWNLOADS]: The skill is configured to download dependencies from a non-standard AWS repository (https://pip.repos.neuron.amazonaws.com) and fetches large-scale weights and code from HuggingFace.

neuron-framework-autoport