The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes external CLI tools including git and pudu-ai to collect repository status and hardware metrics. gitIdentity in scripts/pudu-task.mjs uses execFile to run git rev-parse, git diff, and git ls-files. inventory in scripts/lib/pudu-adapter.mjs uses spawn to run pudu-ai (or a user-configured command) with arguments to gather hardware information.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from task files and model requests, which are sent to a local Ollama instance for inference. Ingestion points: Reads content from user-provided files via --task-file and --request-file in scripts/pudu-task.mjs. Boundary markers: The skill documentation notes that the runner never executes tool calls or generated commands, though it relies on host review for verification. Capability inventory: The skill can write model output to a new file in the repository using the --output flag (in scripts/lib/task-store.mjs) and communicates with a local Ollama API via HTTP. Sanitization: Uses SHA-256 hashing for telemetry data to avoid persisting raw prompt or response text in its database.
[EXTERNAL_DOWNLOADS]: The skill communicates with a local Ollama server to perform model inference and retrieve metadata. Requests are strictly validated to target loopback addresses (127.0.0.1, localhost, [::1]) to prevent data exfiltration to external cloud services.

pudu-task-telemetry