The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from Gmail (email bodies, subjects, and attachments). This data is interpolated into agent reasoning for tasks like triaging, summarizing, and drafting replies. Maliciously crafted emails could attempt to influence the agent's behavior.
Ingestion points: Gmail message bodies, subjects, and attachments accessed via +read, +triage, and users.threads.get in SKILL.md.
Boundary markers: The skill instructs the agent to single-quote shell arguments and requires explicit user confirmation before sending or forwarding emails.
Capability inventory: The skill can send emails (+send, +reply), delete emails (+trash), manage labels, and update account settings like vacation responders.
Sanitization: Instructions in SKILL.md require the agent to wrap user-provided body/subject content in single quotes and escape apostrophes to prevent shell interpretation.
[COMMAND_EXECUTION]: The skill operates by having the agent construct and run shell commands via the hatch_gws_cli binary. While the instructions include a manual sanitization strategy (single-quoting arguments and escaping apostrophes), this pattern relies on the agent's ability to consistently follow complex formatting rules to avoid command injection vulnerabilities.

gmail