The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides instructions to download and execute a shell script from the service's official domain to install the CLI tool. Evidence: SKILL.md and references/cli-reference.md list the command curl -fsSL https://cli.inference.sh | sh.
[EXTERNAL_DOWNLOADS]: The skill references and encourages the installation of an external agent skill package from the same provider. Evidence: SKILL.md contains the instruction npx skills add belt-sh/cli to add a related dependency.
[COMMAND_EXECUTION]: The skill uses the belt CLI to perform various tasks, including system-level operations like installing shell completions. Evidence: references/cli-reference.md includes commands for writing shell completion scripts to system paths (e.g., /etc/bash_completion.d/infsh).
[DATA_EXFILTRATION]: The CLI tool has a built-in feature that automatically uploads local files to the inference.sh cloud when a local path is detected in the input arguments. Evidence: references/running-apps.md explains that any field accepting a URL also accepts a local path for automatic upload.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes external application data and search results while possessing the capability to upload local files and access the network. Ingestion points: belt app list and belt app get fetch app definitions and schemas from the remote store. Boundary markers: No delimiters or specific instructions are provided to the agent to ignore potentially malicious content. Capability inventory: The agent can write to the network and upload local files via the belt app run command. Sanitization: There is no evidence of filtering or sanitization of the data retrieved from the inference.sh service.

agent-tools