The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The SDK provides a significant interface for the agent to ingest external, potentially untrusted data that could contain malicious instructions. * Ingestion points: Data is ingested through image attachments (file paths or base64 blobs), custom tool outputs, and configuration from external servers. * Boundary markers: The instructions do not define specific delimiters or safety warnings to isolate processed data from the agent's core instructions. * Capability inventory: The skill facilitates powerful actions including subprocess execution of local binaries, management of GitHub authentication tokens, and file system modifications. * Sanitization: There is no documentation of built-in sanitization or validation for the data ingested via attachments or tool results.
[COMMAND_EXECUTION]: The SDK is designed to spawn and manage local CLI processes, which is a powerful capability that requires careful handling. * SubprocessConfig allows specifying an arbitrary cli_path and cli_args, enabling the execution of external binaries. * The SDK provides a PermissionHandler.approve_all helper that explicitly allows bypassing all interactive user confirmations for potentially dangerous operations like shell command execution.
[EXTERNAL_DOWNLOADS]: The skill documentation includes standard instructions for installing the SDK and its dependencies from package registries, such as using pip to install the current project and the opentelemetry-api package.

copilot-python-sdk