bug: [0.5.10 & olders] Cortex server process doesn't close after Jan exit

[imtuyethan]

Jan version

0.5.10

Describe the Bug

The cortex-server.exe process occasionally remains running in the background after Jan application is closed. This is a persistent issue that has been observed across multiple versions.

Steps to Reproduce

1. Launch Jan application
2. Use Jan normally
3. Exit Jan application
4. Check Task Manager/Process List for remaining cortex-server.exe process

Expected behavior:

• cortex-server.exe should terminate automatically when Jan is closed

Actual behavior:

• cortex-server.exe sometimes continues running in background after Jan exit
• Requires manual process termination

Screenshots / Logs

What is your OS?

• MacOS
• Windows
• Linux

[louis-jan]

This is something that frontend or bundler should handle, unassigned backend folks.

[louis-jan]

To check:

1. Quick Ask isn't enabled? Since it could cause the same effect, where we don’t want to close the app when the window closes.
2. A raise from users where app should maintain on window close, only exit on force quit

[Glashkoff]

For clarity: the comment in the screenshot is mine. Windows 11 Pro x64 23H2.
Experimental Mode is disabled, so Quick Response is also off.

The issue with cortex-server.exe not closing occurred even after a clean Jan 0.5.10 installation with only the Qwen2.5-Coder-14B model. I restarted Jan, edited parameters in the Qwen2.5-Coder-14B-Instruct-Q4_K_M.yml file to investigate issues #4210 and #4209, and at some point noticed the cortex-server.exe process was still running, using about 900 MB of RAM. I had to terminate it manually and then continued exploring other issues.

Unfortunately, I haven’t been able to reproduce this behavior intentionally. I usually disable logging because I work with NDA-protected data.

What I can say is that this might be related to changing model parameters in Jan’s right sidebar and regenerating responses in a dialog. While working with models, I often adjust Temperature and Top P, triggering response regeneration. That’s all the information I have on this issue for now.