fix: abort completion if user confirms to change the current session

[GregoMac1]

• fix manually aborting a session, it was not working
• if the user clicks on another session while a completion is being written, they have to confirm
• if the user confirms, the completion is aborted before changing the session

[fmaclen]

@GregoMac1 also, a test should have caught this bug when I switched to ollama-js.

I think we have a test for the "Stop" behavior pero I don't know if it's specifically testing the abortController behavior, can you see if we can improve the coverage?

[cloudflare-pages[bot]]

Deploying hollama with    Cloudflare Pages

Latest commit:	bfb6e58
Status:	✅  Deploy successful!
Preview URL:	https://1baa3594.hollama.pages.dev
Branch Preview URL:	https://126-output-from-multiple-ses.hollama.pages.dev

View logs

[fmaclen]

I found a bug when you stop an ongoing 'completion' and then click on another link, it asks if you are sure about leaving.

https://github.com/user-attachments/assets/a452738f-2dfa-49a8-8987-6040b3ee3c38

[GregoMac1]

@fmaclen the code is ready!

About the test, I initially didn't know how to emulate the behavior of the 'pending' requests, so I asked Claude for help and I ended up with this test case.

As far as I understand, it should work, but it doesn't. It fails because, at line 609, it expects the completion to be started, but it is not. I think it has to be something about the stream response emulated at line 591.

Do you see something wrong?

[GregoMac1]

I've changed the code and I've split the tests but I still could not make them work. I'll dig deeper tomorrow.

[fmaclen]

I've split the tests but I still could not make them work

There's a few issues with the tests:

• The shape of the streamed mocked responses should match the Ollama API
• But more importantly, Playwright doesn't appear to support mocking streamed responses.
  • There is an issue with a suggested workaround but haven't tried it to see if it works.
  • Here's another potential solution from StackOverflow: https://stackoverflow.com/questions/77030676/how-to-stub-mock-a-streamed-response-from-chatgpt
• The test can navigate away from session while a response is in progress looks incomplete to me, somewhere there we should be "accepting" the confirmation dialog.
• The test can start a new session, choose a model and stop a completion in progress is now failing, we can probably move the stop response UI assertions from that test into one of the new ones.
• Pull the latest changes from main to fix the test seed data and take screenshots for README.md.

The rest of the implementation looks good 👍

[GregoMac1]

Thanks for the feedback!

• The shape of the streamed mocked responses should match the Ollama API

About this, I followed the shape of the real responses from Ollama in the app, which look like this:

Instead of response, the replies contain message: { role: string, content: string }.

I'm still working on the rest of tests!

[fmaclen]

Instead of response, the replies contain message: { role: string, content: string }.

Sorry, you are right. I was looking at the /api/generate streamed response instead of /api/chat. My bad!

[fmaclen]

@GregoMac1 I updated this PR with the latest merges on main and fixed the conflicts: https://github.com/fmaclen/hollama/pull/156 https://github.com/fmaclen/hollama/pull/157

[fmaclen]

:tada: This PR is included in version 0.10.1 :tada:

The release is available on GitHub release

Your semantic-release bot :package::rocket: