The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documents the ability to fetch remote configuration files and examples from external sources.
Evidence: cli.utils.fetch.fetch_from_github and cli.config.check_remote_config in references/api.md describe mechanisms to sync files from GitHub or query HTTPS URLs for YAML/JSON content.
[DYNAMIC_EXECUTION]: The skill describes a plugin and integration architecture that utilizes dynamic loading of Python modules.
Evidence: integrations.base.PluginManager.load_plugin in references/api.md implements dynamic imports by splitting a string into a module and class name to instantiate extensions.
[COMMAND_EXECUTION]: The skill provides APIs for executing commands within specific cloud environments.
Evidence: cli.cloud.modal_.run_cmd in references/api.md allows running commands inside a folder with Modal Volume integration.
[INDIRECT_PROMPT_INJECTION]: The skill handles untrusted data from external datasets and remote configurations, creating a potential surface for indirect prompt injection.
Ingestion points: datasets (SKILL.md) and check_remote_config (references/api.md).
Boundary markers: The "Agent operating procedure" in SKILL.md instructs the agent to pin down inputs, ask rather than guess values, and validate results.
Capability inventory: The skill enables model training (cli.train.do_train), inference (cli.inference.do_inference), and cloud command execution via Modal (cli.cloud.modal_.run_cmd).
Sanitization: The skill relies on standard machine learning libraries (Hugging Face Datasets, Axolotl core) for processing inputs.

axolotl