Conversations and Transcripts

#Conversation IDs

Every foreground run belongs to a conversation. Omit conversation_id to create one, then reuse the returned ID:

first = await agent.run("Show revenue by region")
 
second = await agent.run(
    "Now compare that with the prior quarter",
    conversation_id=first.conversation_id,
)

Conversation continuity is a bounded cold projection of completed runs, not a live or resumable session. Daita projects at most eight prior runs, 40 messages, and 24,000 UTF-8 bytes. Older context falls outside the projection rather than being summarized by another model call.

#Source Pinning

A new conversation that supplies source_id pins that source:

first = await agent.run(
    "Summarize open orders",
    source_id=orders_source.id,
)
 
follow_up = await agent.run(
    "Group them by region",
    conversation_id=first.conversation_id,
)

An explicit conflicting source on a later turn is rejected. Start a new conversation to change source scope. files_only=True creates or continues a Files-only turn without granting source or MCP access.

#Inspect Runs

exists = await agent.conversation_exists(first.conversation_id)
runs = await agent.conversation_runs(first.conversation_id)
 
for item in runs:
    print(item.turn_index, item.transcript.run.message, item.result)

Each ConversationRun contains its zero-based turn index, exact Transcript, and terminal LoopExit when available.

#Per-Run Transcript

transcript = await agent.transcript(first.run_id)

The transcript contains only the canonical user, assistant, and ordered tool messages from that run. Prior conversations, memory, skills, catalog facts, and machine execution scope are model-request context, not copied transcript history.

Machine-originated job follow-ups and routine occurrences use the same loop and exact transcript records with an immutable ExecutionScope. They cannot inherit ambient authority from a conversation.

#Clear Conversations

deleted = await agent.clear_conversations()

Clearing removes ordinary terminal transcripts, learning candidate records, and review stamps. It preserves approved memory, user profile, semantic annotations, and skills. The minimum run state still required by active jobs, follow-ups, or routine occurrences remains rooted, as do artifacts referenced by durable job results or retained deliveries.

The terminal requires confirmation:

daita conversations clear atlas --yes

Copies already delivered to user-owned directories are never deleted.