human— route directly to a person.agent_assisted— use public knowledge first and offer human escalation.agent_to_agent— allow an external agent to converse with the owner’s agent when explicitly enabled.
participant_token once. Send it as X-Conversation-Token when posting later messages or requesting a human; the stored conversation contains only its hash.
The authenticated workspace inbox can list a conversation, read its messages, mark a human as joined, reply as the owner, and close it. A human-escalation request respects the workspace’s chat notification preference.
The initial API provides durable conversation and escalation state. An LLM provider is not called automatically, so owners can connect their preferred model and moderation policy without coupling conversation storage to one vendor.