Humbug v0.36 wraps up a year of Humbug development, with a lot of behind-the-scenes improvements, bug fixes, and updated GLM model support.
UI improvements
One thing that has frustrated me for a while is that Humbug was quite slow at changing the zoom level or the colour mode. In v0.36 things improve quite a lot, although I'd still like these to be quicker. At some point in the future I need to really get into profiling things, again.
Doing this also gave me a good excuse to refactor quite a few classes to have common base classes. This also let me eliminate several syntax highlighters that were essentially the same.
Another thing that has been frustrating is that keyboard shortcuts no longer worked for horizontally scrolling text (although mouse wheel updates have always worked). Several different things conspired to cause this!
The first problem was the tab activation logic was being too agressive about moving the focus to a tab when it's clicked. This was also caused by a bit of a thinko in which I'd conflated focus with spotlighting a message. They're now separated out.
The second problem was the various text edit controls were inconsistently overriding key press events.
Both of these, and a few other more minor quirks, have now been resolved.
GLM 4.7
I've been curious about GLM's capabilities for a while, but Z.ai are now offering a really interesting coding option where developers can get a coding package that is dramatically less expensive than using normal API access and that is significantly cheaper than any other equivalently powerful offering,.
As I want to try this over the next few days I updated the Z.ai backend to support both thinking/non-thinking models and to support GLM 4.7.
Z.ai support coding models via a different HTTPS endpoint to the normal API: https://api.z.ai/api/coding/paas/v4, so this does mean changing the default endpoint in the Humbug user settings.
More details are at: https://docs.z.ai/devpack/tool/others
GLM 4.7 is also available on the Ollama cloud, so added it there too.
On first glance I'm pretty impressed with GLM 4.7 as a model. It's not quite beating Claude Sonnet 4.5 for me yet, but time will tell.
Humbug v0.36
Humbug v0.36 is mainly a lot of bug fixes. The change log doesn't seem very impressive, but a lot of these are UI issues that LLMs aren't too great at fixing yet. This isn't so much because they can't code things correctly as much as the LLMs aren't yet that good at testing and evaluating UI/UX problems.
New features:
- Improved the performance of style changes (makes zoom and colour mode changes quicker).
- Added smooth scrolling for shell message navigation.
- Added support for GLM 4.7 via the Z.ai and Ollama backends.
- Updated the Z.ai backend to support thinking mode enabled/disabled.
Bug fixes:
- Tightened up parameter
message_typeschecking forread_messagesandsearchoperations in the conversation AI tool. Removed unnecessary tool-call filter. - Fixed a problem with zoom scaling of markdown text blocks.
- Fixed a markdown parsing issue affecting paragraphs inside lists.
- Improved exception handling for UI interactions.
- Fixed a problem where preview tabs needed to be clicked twice to correctly gain focus.
- Fixed problem with widget focus that meant horizontally-scrollable text could be scrolled with a mouse wheel but not with the left and right arrow keys.
Internal structure changes:
- Removed the lazy UI updates as they added a lot of complexity and more significant performance issues have been resolved.