I've been away in Dublin for a few days break and had some time to think about what might be missing from the current design. A few improvements have come out of this thinking.
Improving AI conversations
One problem that did occur to me is there's currently no way to tell if an LLM has actually responded to an HTTP request. Up to now we've just animated the border of whatever was the last message in the conversation history. This is pretty unsatisfactory because we don't know if the HTTP request is timing out, or if we're just waiting for the LLM to generate a large tool call (annoyingly they don't signal much when they're doing this).
To get round this I've added a new conversation state AI_CONNECTED that pops up a temporary message saying "AI is thinking...". This appears after a 200 OK comes back from the HTTP request. If the message is on-screen it means the connection has been made. If it's not on the screen it will mean the connection has not been made. This message disappears once any other message is added to the conversation.
The new state also adds a useful feature to the conversation transcripts. Now it's possible to see how quickly an LLM responded, and also how quickly we got to the first visible response.
Removing bookmarks - a "good idea at the time"
When bookmarking in conversations was added 6+ months ago, bookmarking solved a problem of navigating backwards and forwards in long conversations where the user needed to copy and paste code fragments. Since then we've added navigation between adjacent messages, but added the filesystem tool that lets the AI write files directly.
The bookmark code had grown from its original form to something that took over 300 lines of code, added complexity to the UI, and wasn't really used anymore. It's now gone!
Grok code-fast-1 support
Added grok-code-fast-1 support. This is a very fast and inexpensive model.
I set it a code refactoring problem in the AI delegation tool. It was very fast to propose a design, but had a few problems quite quickly:
- It confused me answering some of its questions with me asking it questions
- It threw a "grammar too complex" error when I asked for more design analysis
- It missed two key dependency injections that would be required, but that Claude Sonnet 4 (non-thinking) flagged immediately.
If they can resolve these functional issues, the speed will be amazing, but I'm not retiring Claude just yet. Too much of keeping the Humbug codebase clean relies on the sorts of insights that Claude is still better at.
File watching bug
The previous code attempted to tidy up file watching when __del__ was called on objects. In Python, however, this isn't guaranteed to be precise and we would see exceptions being raised. Now we don't as these cleanups are done precisely when a tab closes.
Conversation view duplicate bug
When duplicating files in the mindspace conversation view, the previous code didn't add file extensions correctly (a consequence of hiding them in v0.24). Now this works correctly.