[BUG][deep_discussion][The whole transcript becomes the next speaker message, roughly doubling context each turn]

[kyegomez]

Summary

In one_on_one_debate, each speaker's entire transcript becomes the next speaker's message. Since Agent.run returns the whole conversation by default, context roughly doubles every turn — by turn 3 the message contains turn 1 twice — and both agents also accumulate it in their own memory on top.

Where

swarms/structs/deep_discussion.py:53-61:

for i in range(max_loops):
    # Current speaker responds
    response = speaker.run(task=message, img=img)
    conversation.add(speaker.agent_name, response)

    # Swap roles
    message = response
    speaker, other = other, speaker

response is whatever the caller's agents produce, and Agent's output_type defaults to "str-all-except-first" (swarms/structs/agent.py:349) — the agent's entire conversation, not its answer. message = response then makes that blob the next prompt.

Effect

Turn 1: A answers → response ≈ A's answer.
Turn 2: B receives A's whole transcript as its task, and B's response is B's whole transcript, which now contains A's.
Turn 3: A receives that, so A's own turn-1 output comes back to it — mislabelled as the user's words, and duplicated.

Growth is roughly geometric in max_loops, and each agent's own short_memory grows alongside because run records the incoming task too.

Suggested fix

Two changes, both small:

1. Pass the answer, not the transcript:

   from swarms.structs.context_utils import agent_answer
   response = speaker.run(task=message, img=img)
   answer = agent_answer(speaker, fallback=response)
   conversation.add(speaker.agent_name, answer)
   message = answer

2. Better, deliver the shared conversation as typed turns so each speaker sees its own prior output as assistant rather than as user prose:

   from swarms.structs.context_utils import messages_for, split_last_turn
   prior, turn_task = split_last_turn(messages_for(speaker.agent_name, conversation))
   response = speaker.run(task=turn_task, messages=prior, img=img)

   (See [BUG][Multi-Agent Structures][Every structure flattens the shared conversation into one user block, defeating the Agent-side transcript fix] #2029 / [BUG][Agent.run][A caller-supplied messages= is silently discarded, so structures cannot pass typed turns] #2030.)

deep_discussion.py does get one thing right: it builds a fresh Conversation() inside the function (:39), so there is no cross-task leakage.

Found at 3e89f27b (v14.0.2).