Redesigned session view poorly supports non-chatbot agents.

[khoadaniel]

The redesigned session view assumes every Langfuse user is building a chatbot. We aren't, and for us this redesign has made the session view useless.

We run complex LangGraph pipelines: many nodes, branching, retries and structured LLM calls in every session. We use the session view to debug those runs. The new chat-style layout throws away the information we relied on. Here are just a few examples; the list is much longer:

• No latency metrics. We can't see which trace or node is slow anymore.
• No metadata. Model, config and routing details are gone from the view.
• No response shape. Structured outputs are forced into chat bubbles, so we can't see what our code actually receives.

A chat transcript doesn't help us debug a graph. This view is a step backwards for anyone running agents or pipelines.

Please switch back to the old session view.
Thanks.

[khoadaniel]

@bezbac FYI, could you please have a look?
Thanks.

[bezbac]

Hi @khoadaniel. Sorry to hear that you’re having a frustrating experience with the redesigned view. As mentioned in the feedback discussion, you can switch back to the old session view by turning off the feature in User Menu → Feature Preview. Alternatively this can also be turned off for your entire organization by using the "Feature Preview" org settings. See the screenshots attached below.

I understand the frustration around graph-based agents, and I’d like to better understand what an ideal session view would look like for you. Would it be helpful to offer session-level aggregation in the current trace UI, with the same tree, timeline, and graph views available in the trace view?

Org settings:

[khoadaniel]

@bezbac Thanks for the prompt response. I think you need to make the preview feature switch more explicit.

Regarding your question, I would keep the session view as is. We track sessions specifically by ids which spanned across multiple internal services. Traces are not the same for us.

If you want to introduce this new session view, consider giving it a new name and let it be a standalone feature.

Thanks.

[alibuildss]

feel this, graph runs aren't chats. what gets most of the way back:

• observations table filtered by session_id, sorted by latency desc (save it as a view), so the slow node shows up immediately
• the trace log view + ctrl+F shows full raw I/O per observation
• if you put model/route in metadata via propagate_attributes, you can slice "which route is slow" in dashboards without opening sessions

agree a toggle for the old layout would be the real fix.