The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads a Python wheel from connect.aiscientist.tools. This domain belongs to the vendor's ToolUniverse infrastructure and is used for hosting the relay SDK. The download is secured with a SHA-256 integrity hash (3fad5eee5ecf7887a693d93ccd1aa112dc0955617a885d1fc3daded0030f9ae0) to ensure the file has not been tampered with.
[COMMAND_EXECUTION]: The instructions require running multiple shell commands, including environment setup (venv), package installation, and execution of validation scripts such as scripts/remote_validation/setup_skill_preflight.py and the tu CLI tool.
[CREDENTIALS_UNSAFE]: The skill documentation correctly advises against placing TOOLUNIVERSE_API_TOKEN or TOOLUNIVERSE_SERVICE_KEY in shell arguments or result outputs. It recommends using environment files with restrictive permissions (0600) or browser-based device authorization flows.
[INDIRECT_PROMPT_INJECTION]: The run_chrombpnet_predict operation accepts genomic sequences from potential user input. While the format is scientific data, ingesting external content creates a surface for indirect prompt injection if the output influences downstream agent logic.
Ingestion points: The run_chrombpnet_predict tool accepts a sequence argument via JSON.
Boundary markers: No explicit instructions or delimiters are provided in the example to prevent the agent from interpreting sequence data as instructions.
Capability inventory: The skill includes the ability to execute genomic predictions and share tool access over a network relay.
Sanitization: No validation or sanitization of the input sequence is detailed in the setup instructions.

setup-chrombpnet-remote-tool