The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the creation of AI agents that process external user input and file attachments. The provided code examples often use an approveAll permission handler, which automatically authorizes all agent actions (such as tool execution or file edits) without human review. This creates a surface where a malicious user or a poisoned file could manipulate the agent into performing harmful actions.
Ingestion points: User input through CLI prompts (readline, input()) and local file attachments via the session.send method.
Boundary markers: The examples do not demonstrate the use of specific boundary markers or 'ignore' instructions for the processed content.
Capability inventory: The agents built with this SDK can execute custom tools, interact with MCP servers, and perform file edits.
Sanitization: There are no sanitization or validation steps shown for external inputs before they are passed to the agent runtime.
[EXTERNAL_DOWNLOADS]: The skill documents the use of a bundled runtime that is downloaded from GitHub servers. Specifically, it mentions that Python wheels include a pinned version and provides a command (python -m copilot download-runtime) to manually cache this runtime. As these downloads originate from the official GitHub infrastructure for the SDK, they are standard developer operations.
[COMMAND_EXECUTION]: The skill provides instructions for running the Copilot CLI in server mode (copilot --server --port 4321) to allow the SDK to connect to a persistent agent runtime. It also demonstrates programmatically managing the lifecycle of these local processes.

copilot-sdk