The Agent Skills Directory

[PROMPT_INJECTION]: The skill classifies and displays unread emails, including sender information and subjects, which are sourced from external entities. This creates an indirect prompt injection surface where a malicious sender could include instructions in the email subject or name to manipulate the agent's triage behavior.\n
Ingestion points: _cmd_classify in src/inbox_triage/cli.py reads Gmail search results from stdin.\n
Boundary markers: None. Email metadata is directly interpolated into the markdown summary.\n
Capability inventory: subprocess.run in src/inbox_triage/cli.py is used to execute gog commands for modifying email threads.\n
Sanitization: None detected for email subject or sender fields.\n- [COMMAND_EXECUTION]: The archive command in src/inbox_triage/cli.py executes the gog utility using subprocess.run.\n
Evidence: The code constructs a command string and uses cmd.split() to pass it as an argument list. While shell=True is not enabled, the skill relies on the external gog tool to perform actions on the user's Gmail account.

inbox-triage