Delay when resuming a chat.

[joetifa2003]

When using Durable Agents and have durable stream re connection, when you open a chat it types out from the beginning, slowly like it is being typed out, even though the LLM already generated the response, the processes of replaying the stream (when you refresh) is like it's being typed out again, this is more of an issue with multi-turn chat model, which is supposed to be more suitable to production solution.

Flight booking example has that issue also, i thought that i implemented something wrong with my app, but you can see it clearly in flight booking app.

2026-04-21.21-34-58.mp4

---

This basically makes it practically unusable for big chats, i notice that even when i'm not using the multi-turn chat model, i'm using it for single turns and i notice it.

[joetifa2003]

Is there a pattern that resolves this? an example?

Or is it a limitation of the current version of DurableAgents with useChat?

[shtefcs]

I faced that as well, when integrated durable agents, but for me it was super slow. What I see on your video as actually fast. Probably because my chat has many more messages and tool cals.

[VaguelySerious]

@joetifa2003 Since you're using workflow's durable streams, the stream needs to be replayed. It's possible to speed this up to some degree, but for larger chat sessions, it might make sense for you to only stream the latest code, and pre-aggregate/paginate older history.

Open Agents might be a good example on how to handle this

[joetifa2003]

@VaguelySerious currently I'm doing partial strem resumption by setting initial resume index to -100 but I face another issue #1835, and currently doing a workaround to fix this by walking up the stream until I find a safe point like "step-start" or "reasoning-start" and resume from there, doesn't work great.

Also I'm using multi-turn chat model where each workflow is a single turn, not a long lived one.

And I may be wrong, but last time I checked openagents, it's not using DurableAgent.

[VaguelySerious]

@joetifa2003 I'm taking a look at #1835 to see what we can do. I understand the UX is not great currently, and we'd love to support better stream resumption natively for this case, or provide an alternative solution out of the box

[VaguelySerious]

Back to working on this, making some changes in #2082 that should help doing partial resumptions