Message Format

Jaibber uses Ably for real-time communication. All messages flow through Ably channels.

Channels

Channel	Purpose
jaibber:presence	Global presence — each client enters with { userId, username, projectIds }
jaibber:project:{projectId}	Per-project channel — all members send/receive messages here

Message Types

All messages are published as channel.publish("message", payload) where payload is an AblyMessage:

interface AblyMessage {
  from: string;           // sender's userId
  fromUsername: string;    // sender's display name
  projectId: string;
  text: string;
  messageId: string;       // unique UUID
  type: MessageType;
  responseId?: string;     // links typing/chunk/response together
  agentName?: string;
  mentions?: string[];
  isAgentMessage?: boolean;
  responseDepth?: number;
  respondingChain?: string[];
  isTaskNotification?: boolean;
  attachments?: MessageAttachment[];
}

Message Types

Type	Description
message	User or agent prompt
typing	Agent is processing (shows typing indicator)
chunk	Streaming response chunk (batched every 200ms)
response	Complete response text (replaces streaming bubble)
error	Agent error

Streaming Protocol

When an agent responds to a message:

1. Publish typing — includes responseId so receivers create the correct bubble
2. Publish chunk messages — uses messageId = responseId (not a new UUID), buffered and flushed every 200ms
3. Publish response — includes complete text with isAgentMessage: true, responseDepth, respondingChain
4. Persist via REST — POST /api/projects/{id}/messages with the same id field (server skips Ably re-publish)

Conversation Context Format

When building context for an LLM prompt, the last 20 messages are formatted as:

User: first message text

Assistant (AgentName): response text

User: second message

Prepended with: "Respond ONLY to the final user message. Previous messages are for context."

Presence Data

Per-project presence includes:

{
  userId: string;
  username: string;
  isAgent: boolean;           // true if connection has an agent registered
  agentName?: string;         // e.g. "Coder"
  agentInstructions?: string; // system prompt (visible to project members)
  machineName?: string;       // e.g. "dev-server"
}

Loop Prevention

Agent-to-agent chains are limited:

• Max depth: 3 (configurable via maxResponseDepth)
• Chain tracking: respondingChain array tracks which agents have responded
• Deduplication: An agent won't respond if its name is already in the chain
• @mention required: Agents only respond when explicitly @mentioned