Connector Action Authorization

Connector Action Authorization

Use this when the user wants Samantha to act through an external connector beyond read-only setup/backfill: sending messages, replying to email, marking messages read, labeling, spam triage, deletion, or cross-channel intervention.

Principles

• Do not treat a connected account as permission to act. Connection grants access; action policies grant behavior.
• Make the policy concrete before granting it: connector, action kind, target/account, contexts, limits, and expiration when relevant.
• Prefer narrow constraints: allowed_targets / allowed_recipients and max_per_day are safer than broad grants.
• allow_autonomous: true means heartbeat or another non-chat context may act without a live user message. Use it only when the user explicitly asks for ongoing/autonomous behavior.
• Keep destructive actions, especially delete, out of autonomous policy unless the user is exceptionally explicit and the constraints are narrow.

Flow

1. Call connector_action_policy_status to inspect existing grants.
2. If granting, summarize the proposed scope in normal language before calling grant_connector_action_policy.
3. Use requires_per_action_approval: false only when the user asks for preauthorization rather than case-by-case approval.
4. Use revoke_connector_action_policy when the user asks to remove or narrow a grant.