User Story 08: Multi-Turn Conversation

Implementation-oriented user story for SDK development.

Send Follow-Up Turns Using Task.context_id and Resume Interrupted Tasks

Goal

Agent A holds a multi-turn conversation with Agent B: multiple request/reply exchanges are grouped by a shared Task.context_id so Agent B can maintain conversational continuity. When Agent B needs additional input mid-task, Agent A resumes the same task after prompting the user.

Preconditions

• Agent B is registered and discovered by Agent A.
• Agent A has subscribed to its reply topic.
• Agent B supports multi-turn conversations (maintains context across tasks sharing a Task.context_id).

End-to-End Flow — Follow-Up Turns

Turn 1: Start Conversation

1. Agent A generates Task.context_id C (UUIDv4) and Task.id T1 (UUIDv4).
2. Agent A publishes SendMessage to:
   • $a2a/v1/request/{org_id}/{unit_id}/{agent_b_id}
   • QoS: 1 (recommended)
   • MQTT v5 properties:
     • Response Topic: $a2a/v1/reply/{org_id}/{unit_id}/{agent_a_id}/{reply_suffix}
     • Correlation Data: D1 (unique request correlation bytes)
   • Optional MQTT User Property: a2a-context-id = C
   • JSON-RPC payload includes Message.task_id = T1 and Message.context_id = C
3. Agent B creates task T1 within context C, processes the request.
4. Agent B streams TaskStatusUpdateEvent / TaskArtifactUpdateEvent messages echoing D1.
5. Agent B sends terminal status (completed) echoing D1. Agent A marks T1 complete.

Turn 2: Follow-Up in Same Conversation

6. User asks a follow-up question.
7. Agent A generates new Task.id T2 (UUIDv4), reuses Task.context_id C.
8. Agent A publishes SendMessage:
   • Correlation Data: D2 (fresh)
   • JSON-RPC payload includes Message.task_id = T2 and Message.context_id = C
9. Agent B creates task T2 within context C. Agent B uses prior context from T1 to inform its response.
10. Agent B streams replies echoing D2, completes with terminal status.

End-to-End Flow — Interrupted Task (input-required)

Initial Request

1. Agent A publishes SendMessage with Task.id T3, Task.context_id C, Correlation Data D3.
2. Agent B starts processing, streams partial results echoing D3.

Input Required

3. Agent B determines it needs more information from the user.
4. Agent B publishes TaskStatusUpdateEvent with status.state = input-required and a message describing what input is needed, echoing D3.
5. Agent A receives input-required, treats it as stream-final for D3, cleans up in-flight state, and prompts the user.

User Provides Input

6. User provides the requested information.
7. Agent A publishes a new SendMessage with the same Task.id T3 and Task.context_id C, fresh Correlation Data D4.
8. Agent B continues processing task T3 with the new input, streams replies echoing D4.
9. Agent B sends terminal status (completed) echoing D4.

SDK Requirements Checklist

• Generate and persist Task.context_id (UUIDv4) for new conversations; reuse for follow-up turns.
• Include Message.context_id in JSON-RPC payload on all requests within a conversation.
• Optionally include MQTT User Property a2a-context-id for transport-level visibility.
• Detect input-required / auth-required status as stream-final for current correlation.
• Resume interrupted tasks: publish new SendMessage with same Task.id and Task.context_id, fresh Correlation Data.
• Responder SDK: provide access to prior task results within the same Task.context_id for context-aware responses.
• Responder SDK: validate Task.context_id consistency — reject if provided Task.context_id conflicts with existing task's Task.context_id.

Failure Cases to Test

• Mismatched Task.context_id: Agent A sends a follow-up for an existing Task.id with a different Task.context_id. Agent B rejects with JSON-RPC error -32602 (invalid params).
• Unknown Task.context_id: Agent A sends a new Task.id with a Task.context_id the responder hasn't seen. Agent B creates a new context (no error — the context is new).
• Expired context: Agent A sends a follow-up after Agent B's context expiration window. Agent B returns an appropriate error.
• Multiple concurrent tasks in same context: Agent A has two in-flight tasks (T4, T5) sharing the same Task.context_id C. Both complete independently; each uses its own Correlation Data.