The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides instructions for installing the uv package manager using a remote script from astral.sh via a piped shell command (curl -LsSf https://astral.sh/uv/install.sh | sh).
[COMMAND_EXECUTION]: The documentation details various shell commands for interacting with the Roboflow cloud API and managing local inference servers using the inference-cli tool.
[PRIVILEGE_ESCALATION]: Documentation in local-tooling.md suggests using sudo systemctl start docker as a troubleshooting step for the Docker daemon, while explicitly advising that users should be asked for confirmation before performing such operations.
[INDIRECT_PROMPT_INJECTION]: The skill has an inherent surface for indirect prompt injection as it processes and acts upon data returned from external vision models and workflows.
Ingestion points: Results from workflows_run, workflow_specs_run, and InferencePipeline callbacks (located in workflows.md and SKILL.md).
Boundary markers: No explicit delimiters or boundary instructions are provided in the documentation to separate model-generated data from agent instructions.
Capability inventory: The skill documents capabilities for filesystem writes (Path.write_bytes), network operations (InferenceHTTPClient), and shell execution via the inference-cli (located in workflows.md and local-tooling.md).
Sanitization: Documentation suggests decoding base64 image data but lacks specific guidance on sanitizing output keys before using them in filesystem paths, representing a potential path traversal risk if processing untrusted workflow definitions.

roboflow-inference