The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides the ability to launch training jobs and manage research state through W&B. The autoresearch_state_impl.py script implements functions to save and load research metadata as W&B Artifacts. The instructions explicitly require user authorization before performing cluster modifications or launching non-smoke-test jobs.
[EXTERNAL_DOWNLOADS]: The skill utilizes the uv package manager to install the wandb library and dependencies. These downloads target the official Python Package Index (PyPI) and well-known, trusted W&B infrastructure.
[REMOTE_CODE_EXECUTION]: The skill's primary purpose is to orchestrate remote training jobs via W&B Launch. It provides helpers to submit code artifacts and relaunch existing runs. This is the intended functionality of the skill and is governed by strict agent instructions to use smoke tests and adhere to user-defined compute budgets.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources, specifically W&B run metrics and local state files (AUTORESEARCH_STATE.md), to inform its research decisions.
Ingestion points: Reads run metadata via wandb.Api().runs() and local research state files.
Boundary markers: The instructions include explicit directives to confirm choices with the user and provide status reports.
Capability inventory: Includes launching remote jobs via relaunch_run and submit_code_artifact_job (via the wandb-primary skill).
Sanitization: The skill relies on natural language reasoning and user confirmation gates rather than programmatic sanitization. This surface is considered low risk given the collaborative nature of the workflow.

wandb-autoresearch