Repository navigation
Conversation
Use the Angular notebook connection and authentication lifecycle in the React Assistant. Keep notebook paragraph rendering at the workspace composition boundary and expose generic answer rendering to the conversation panel. Cover session retirement, notebook isolation, authentication handling, and saved and streaming answer rendering with unit tests.
Mock Assistant REST and WebSocket traffic in Playwright while retaining the real notebook and authentication host. Replace the production fixture flag with test-owned wire responses and explicit streaming completion gates. Exercise saved history, streaming completion, notebook isolation, and paginated answer merging without an external LLM.
voidmatcha
left a comment
There was a problem hiding this comment.
Sorry for the delayed review. The conversation flows I checked worked against a real server, but I noticed the following:
- A conversation removed from the latest server response remains in the panel's list, and selecting it returns 404. Could we reconcile removed entries when refreshing the list, while preserving conversations created locally during the request?
- If the initial conversation list request fails, reopening the list does not retry it. Closing and reopening the panel recovers it. Could we provide a way to retry loading the list without closing the panel? This is non-blocking.
- After renaming a paragraph, its reference button retains the old title in its accessible name. The displayed number and navigation work correctly, so this is also non-blocking.
Remove conversations omitted from refreshed server lists while preserving local creations confirmed during the request. Cover empty results, concurrent creation, metadata reconciliation, and preserved drafts and runs.
Let users retry a failed conversation list request without closing the panel or losing their draft. Prevent duplicate retries while loading and cover repeated failure followed by recovery.
Center the retry button alongside the error message and wrap it below the message in narrow sidebars.
Publish updated paragraph metadata without remounting the Assistant so reference buttons use the current accessible name. Cover in-place renames and insertions while preserving the open panel and connection.
voidmatcha
left a comment
There was a problem hiding this comment.
Small question about the existing UI. Was there a reason for using an opaque background for the delete confirmation? A semi-transparent backdrop might help preserve the conversation context.
|
Yeah maybe it's little part so you can add this UX improvement in this PR! |
Use the Ant Design mask token for the Assistant delete confirmation backdrop while keeping the dialog surface opaque.
voidmatcha
left a comment
There was a problem hiding this comment.
LGTM 👍 Thanks for the updates!
I didn’t get to review the UX thoroughly at the time because of the OSSCA deadline. Revisiting it, I noticed a few things that might be worth improving.
- After deletion is requested, Cancel only closes the dialog while deletion continues. Changing the label to Close while deletion is pending could make this clearer.
- Loading conversations and No conversations yet appear at the same time while the conversation list is loading.
- Closing and reopening the panel clears an unsent draft.
- Conversations created within the same minute have identical default titles and time labels, making them difficult to distinguish. Using the first question as the default title could help.
These are non-blocking suggestions. Could you take a look and, if they make sense, address them in a separate minor PR or track them in follow-up issues?
Add inline-end padding so the delete button focus ring remains visible after cancelling confirmation.
|
Thanks for taking another look at the Assistant UX! I’ll keep the four non-blocking suggestions from your latest review out of this PR and look at them separately. Your review also prompted me to check a few other Assistant scenarios. I noticed that pressing Esc to cancel deletion correctly returned focus to the delete button, but the right edge of its focus ring was clipped. Since the fix was just two lines of CSS, I included it in b3d9d4d: the header now leaves 4px of space for the full ring. I checked it at the minimum, default, and maximum sidebar widths in both themes. |
|
I filed four follow-up issues in Jira so we can address them separately from this PR:
|
|
Following up on my earlier review, I checked a few more usage scenarios. Leaving aside the items already covered by the follow-up issues, I noticed the following. If the panel is closed and reopened while an answer is streaming, the answer can remain missing from the view even after the server finishes and saves it. I reproduced this three times against a real server. Closing and reopening the panel once more restored the answer. It would be worth checking how the panel recovers an ongoing response when reopened. I also noticed these UX improvement candidates.
For scrolling and focus behavior, it may help to compare how Databricks and other AI products handle these interactions and choose policies that fit Zeppelin. Rather than creating an issue for every item immediately, we could first discuss which improvements are worth pursuing and how to group the work. |
|
Thanks for testing this, @voidmatcha. Would you prefer to fix the streaming issue in this PR or track it as a follow-up? I’d start with 1 and 3: keep Send visible as the question grows, and keep focus in the composer after sending. If the user switches back to the notebook while waiting, the answer arriving shouldn’t pull focus back into chat. I’d take 7 next. I like Open WebUI’s approach: follow the answer while the user is at the bottom, but if they scroll up, leave them there and offer Jump to latest. I’d preserve that choice when switching conversations, too. Coming back to a thread should take you back to the passage you were reading, even if more text has arrived. I’d discuss 4 and 5 together. Seeing the full title helps you choose the right conversation from the search results; keeping the search term lets you return to those same results if you open the wrong one. For 6, would refreshing the list when you reopen it cover the case you tested, or did you need to see new conversations while the list stayed open? 2 seems worth a small follow-up of its own: if an image the user chose to load fails, show that it failed and offer a retry. |
|
Thanks, @miinhho. I agree with the grouping and priorities you suggested, including not pulling focus back into chat after the user has moved to the notebook. For the streaming recovery issue, I’d prefer to fix it in this PR if it can be addressed using the existing events and history API. If it requires broader server or connection changes, a follow-up would be reasonable. For 6, refreshing the list when reopening it would cover the case I tested. Live updates while the list stays open aren’t needed for that case. |
|
Thanks, @voidmatcha. I fixed the streaming issue in 0be38903ab using the existing events and history API. Closing the sidebar no longer drops the subscription for an in-flight reply. If you reopen it before the run finishes, the reply continues. If it finishes while the sidebar is closed, the saved answer is there when you reopen it. Thanks for clarifying item 6. I’ll keep that separate and limit it to refreshing when the conversation list is reopened. |





What is this PR for?
Replace the fixture-backed notebook Assistant flow with the existing conversation REST API and WebSocket transport. Users can create, select, and delete conversations, load saved history, and submit questions from the notebook sidebar.
The Angular host provides the notebook connection and authentication handling to the React panel. Requests and subscriptions are scoped to the open panel and current notebook session. Paragraph references in answers use the notebook’s existing selection and scrolling behavior.
What type of PR is it?
Feature
What is the Jira issue?
How should this be tested?
With Zeppelin running, set
PLAYWRIGHT_BASE_URLto the server address. If authentication is enabled, setZEPPELIN_E2E_SHIRO_INIto the server’s Shiro configuration. Fromzeppelin-web-angular, run:Verify conversation creation and submission, streamed answer completion, history restoration, conversation isolation, and loading earlier messages. The tests mock Assistant REST/WebSocket responses while using the running server for notebook and authentication traffic.
For an existing conversation, delay or fail its initial history request. The draft should remain editable, and sending should become available only after a successful history retry, without automatic submission. New local conversations and history requests after initial success should remain unaffected.
Questions: