The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill documents the use of composio run, which allows for the execution of arbitrary JavaScript and TypeScript snippets. This provides a powerful execution environment for the agent to chain tools and process data.
[INDIRECT_PROMPT_INJECTION]: The documentation for composio run and experimental_subAgent() demonstrates interpolating untrusted data (e.g., email content from GMAIL_FETCH_EMAILS) directly into LLM prompts using results.prompt(). This creates a vulnerability surface where malicious content in external data could influence the agent's behavior.
Ingestion points: Data from tool executions like GMAIL_FETCH_EMAILS, search() results, and proxy() calls (found in SKILL.md, references/power-user-examples.md).
Boundary markers: None used in the examples; strings are directly concatenated or interpolated.
Capability inventory: Access to execute, search, link, proxy, and file upload capabilities (found in SKILL.md, references/composio-dev.md).
Sanitization: No sanitization or escaping of the interpolated data is described or implemented in the examples.
[COMMAND_EXECUTION]: The skill facilitates extensive interaction with the local system via the composio CLI, including account linking, file uploads (--file), and raw API proxying, which can be used to modify state or access sensitive external resources.

composio-cli