Voice mailboxes

Hand off a Claude Code conversation to the voice agent on the phone, walk while talking, get structured responses back. One open exchange = one Reminders list = one entry in ~/.claude/voice-mailboxes/.

Three roles

The vocabulary matters because the brief addresses one role and refers to the other two — getting POV wrong makes the brief unreadable to the voice agent.

The user talks to both agents; the project agent writes for the voice agent; the voice agent talks with the user out loud.

Role	Surface	POV in brief
user	Human. Types in the terminal, speaks on the phone.	Third person ("the user", "they"). Never addressed directly in writing.
project agent	Claude Code session that composes the brief.	First person ("I").
voice agent	Claude Voice on the phone. Reader of the brief.	Second person ("you"). The reader.

Lifecycle

One open exchange. Idempotent re-open from the same slug rewrites header + brief but preserves prior responses.

What the daemon owns vs the user owns

Five surfaces per exchange. Daemon-owned in green; user-written in orange; state files in amber.

Response kinds

The brief's writeback contract lists five optional title prefixes the voice agent can use when writing a response. Anything without a prefix is classified as free on drain.

Prefix	Kind	Meaning
`decision:`	decision	User committed to something on the call.
`note:`	note	Context worth keeping; no action implied.
`question:`	question	User wants the project agent to come back with an answer.
`deferred:`	deferred	Talked about, no decision yet — explicit punt to revisit later.
`done`	done	Closes the exchange. Daemon tears the list down on the next cycle.
(none)	free	Plain text. Kept as-is on drain.

Why deferred and not just question? A deferral is a conversational outcome — the user already heard the options and chose not to choose yet. Tracking it separately lets the project agent re-raise it as a discussion item next session, instead of treating it as a fresh unanswered question.

CLI surface

Command	Effect
`rbridge mailbox open --slug X --kind REMINDERS --brief -`	Idempotent. Reads brief from stdin, creates / refreshes the exchange list, mirror, state file.
`rbridge mailbox read --slug X`	JSON dump of responses (header + brief excluded). Warns on stderr if a `done` reminder is present.
`rbridge mailbox close --slug X`	Tears down: deletes list, removes mirror, drops state. Returns truthful status (list deleted / was missing / failed).
`rbridge mailbox refresh --slug X`	Re-up header + brief from on-disk brief, no other changes.
`rbridge mailbox list`	Enumerate active exchanges.

None of these require the daemon to be running — they all hit EventKit directly. The daemon only adds GC + drift correction on top.

Discoverability — silent by design

No alarms, no notifications, no "Today" due dates. The signals an open exchange leaves are:

The exchange list itself in the sidebar.
A high-priority breadcrumb reminder Voice exchange open: <slug> in the user's default list.
rbridge mailbox list for the agent itself.

This is intentional: agents may open exchanges overnight, during meetings, or while the user is AFK. Pushing alerts in those scenarios is worse than silence. The user discovers the exchange the next time they glance at Reminders.app, which on iOS happens naturally.

Disable the breadcrumb with RBRIDGE_MAILBOX_MIRROR=false.

Slug rules

Grammar: [a-z0-9][a-z0-9-]{0,47}. Kebab-case, 2-5 words.
Topic-first, not action-first: wallgen-shipping-decision, not decide-wallgen-shipping.
Skip filler verbs (figure-out-, talk-about-, discuss-).
If the conversation has a project name, lead with it (pimpelmees-shopify-cutover).
Slug collision on open is non-destructive — header + brief are rewritten, prior responses preserved.