The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted email data (subjects and snippets) which are external inputs. This creates a potential surface for indirect prompt injection where malicious instructions in an email could attempt to influence the agent.
Ingestion points: Email data is fetched from the assist server via the emails and email commands in scripts/gmail-axi.mjs.
Boundary markers: The skill uses the structured TOON format to present data to the agent, providing clear boundaries between metadata and email content.
Capability inventory: The bundled CLI in scripts/gmail-axi.mjs is limited to making network requests to a specific server (defaulting to localhost) and does not have shell execution or local file system write capabilities.
Sanitization: Email snippets are truncated to 500 characters in scripts/gmail-axi.mjs before being displayed to the agent.

assist-gmail