How to resuming a groupchat with chat history? #7021

yeah this is doable in v0.4, it just looks different from the v0.2 approach.

save_state() and load_state() are the cleanest path:

# at the end of a session, save the team state
state = await team.save_state()
# `state` is a plain dict - serialize it to json or store in your db

# later, to resume:
team2 = RoundRobinGroupChat([...])  # same agents and config
await team2.load_state(state)
result = await team2.run(task="continue from where we left off")

save_state() captures the full message history plus agent and termination condition state. when you load_state() before the next run(), the agents see the prior conversation as context.

if you only have raw logs from a database (not a save_state() snapshot) it is more manual - you would need to reconstruct the messages and inject them via the team or agent context. there is no built-in "replay from raw db logs" api in v0.4 right now, so the practical advice is to add save_state() at the end of every session from the start - much easier than reconstructing state after the fact.

the v0.2 resume docs you linked covered a similar concept but the api changed significantly in 0.4, which is probably why you could not find the equivalent.

0.2's resume GroupChat helper is not what 0.4+ AgentChat uses.

Persist TaskResult.messages (or the stream events) from the run you already have. On restore, rebuild the same SelectorGroupChat / team, then pass those messages back in as the history for the next run / on_messages call instead of starting from a bare user string.

The agents and model clients have to match. A new AssistantAgent() with the same name is not enough if tools or system text changed.

There is no "load the sqlite and continue" switch. You own the log. Serialize the message objects (text, source, tool calls) and replay them as TextMessage / UserMessage / ToolCall* types before the new task.

If you only have pretty-printed logs, you will lose tool-call ids and the next turn will look like a fresh chat.