Inbox and Deliveries

#Logical Deliveries

Delivery is the sole durable representation of an autonomous outcome. The inbox is a bounded product view over those records; it is not a second message store or a copy of an artifact.

Every delivery records its producer subject, destination, conclusion, state, provenance, sensitivity, run reference when one exists, and immutable bounded artifact references. Producer finalization validates committed artifacts and the current destination, then commits the outcome and unique logical delivery atomically.

The only current destination is the originating conversation inbox. Email, webhooks, arbitrary filesystem destinations, and external delivery are not implemented.

#Python API

destinations = await agent.distribution_destinations(conversation_id)
 
items = await agent.inbox(
    conversation_id=conversation_id,
    include_acknowledged=False,
    limit=50,
)
 
inspection = await agent.inspect_delivery(delivery_id)
acknowledged = await agent.acknowledge_inbox(delivery_id)

Destination discovery accepts an optional sensitivity ceiling. Inspection returns bounded immutable safe facts for one exact delivery. Acknowledgment is idempotent and does not delete the producer result or a user-owned artifact copy.

#Terminal and CLI

Open /inbox in the Textual app or use:

daita inbox destinations atlas <conversation-id>
daita inbox list atlas --conversation-id <conversation-id>
daita inbox inspect atlas <delivery-id>
daita inbox acknowledge atlas <delivery-id>

Pass --include-acknowledged to include retained history in inbox list.

#Retention and Capacity

Unacknowledged available or blocked deliveries are not silently evicted. If the fixed per-agent delivery bound is full, producer finalization remains pending rather than dropping the outcome. Acknowledgment allows the oldest acknowledged history to be reclaimed by a later producer finalization and wakes pending job and routine delivery work.

Artifact references are resolved against the canonical artifact manifest and revalidated for identity, checksum, provenance, authorship, sensitivity, and size. This allows ordinary transcripts to be cleared while artifacts rooted by a durable job result or retained delivery remain recoverable by known ID.