human-context-handoff

Human Context Handoff

Use this skill when useful agent work is blocked by one specific piece of context that only an enrolled human can provide and that person is not available in the current channel.

The handoff is asynchronous. A live request returns a durable request ID quickly. The host stores that ID and polls for a terminal result instead of holding one tool request open for the length of a phone call.

When to use

• A product or workflow choice has two to five bounded options.
• An operations alert needs a human preference, such as pause, escalate, or continue monitoring.
• The missing answer materially changes the next useful agent action.
• The intended recipient was previously verified, consented to calls, and is allowed by the user's current calling policy.

When not to use

• The user is available in the current conversation. Ask there first.
• The question requests a password, one-time code, credential, payment data, private key, or other secret.
• The answer would be treated as identity proof or as authorization for an irreversible, regulated, legal, medical, financial, or emergency action.
• The context cannot be reduced to one short question with bounded choices.
• The recipient declined, the request expired, or the person could not be reached. Do not call again automatically to seek a different outcome.

Required inputs

Prepare these fields before any live request:

• a short task title;
• a minimized context summary;
• one question;
• two to five choices with stable machine-readable IDs;
• the verified recipient label or identifier, never a raw number in agent output;
• what the agent will do with each answer;
• an expiry time;
• one stable idempotency key for the logical question.

If any field is missing or ambiguous, stop and ask in the current channel.

Preview first

Always show a no-call preview before a live handoff unless the user has already approved an account policy that explicitly covers this exact class of call. The preview must include: