Outbound phone calls as approval-gated work

[ldbumble]

Some of the owner's work only happens on the phone: the vendor whose billing line has no email, the technician who confirms a visit by voice, the helpdesk that chases a ticket only when someone calls. Taskuary can draft, propose and send. It cannot pick up a phone.

The idea

Outbound voice as a connector, approval-gated like every other egress:

• A task's work includes a call. The agent proposes it: who to call, the number (from the connector or the thread), the three questions to ask, what counts as a result. The owner approves the brief, not each sentence.
• A voice agent places the call, navigates the IVR, waits on hold, asks the questions, and does not improvise beyond the brief. Anything outside it ("they want a PO number") comes back as a question on the task, not an answer invented on the line.
• The result is a receipt on the Timeline: transcript, outcome, recording if the owner enabled it, and the follow-up task if one was promised. The promise ledger picks it up like a reply.
• Inbound later: a number the twin answers for the owner, filing voicemail and callbacks as arrivals.

What already exists to build on

• voice.py transcribes voice notes and the prompt-box mic through several STT providers; the provider table is the shape for a telephony provider (Twilio, Vonage, or a hosted voice-agent API).
• WhatsApp and Telegram are already approval channels and doorways (remote_assistant.py), so "approve the call brief from your phone" is a line, not a feature.
• The promise ledger and the follow-up chase ladder already know what "waiting on them" means.

Open questions

• Which provider family: a raw telephony API plus our own agent loop, or a hosted voice agent we hand the brief to? The second is faster to ship and harder to audit.
• Disclosure: the agent says it is calling on the owner's behalf and is automated. Non-negotiable, and jurisdiction-dependent for recording.
• Cost and the on-hold problem: a forty-minute hold is cheap in minutes and expensive if the model is streaming the whole time. Silence detection before the model wakes.

Status: roadmap, not scheduled.