One of the headaches with AIs doing lots of tool calling is it's quite hard to work out what they're doing when they request tool approvals. I figured the best way to address this was to make them produce diffs, especially as recent updates had enabled the editor and filesystem tools to apply diffs.
Diffs
One of the headaches with AIs doing lots of tool calling is it's quite hard to work out what they're doing when they request tool approvals. I figured the best way to address this was to make them produce diffs, especially as recent updates had enabled the editor and filesystem tools to apply diffs.
The starting point was to add a diff syntax highlighter. That took about 30 minutes earlier this week.
The next part was to modify the AI tool infrastructure so the "tool call" approval UI can take additional context and display it above the approval buttons. This disappears once the approval is given or not given.
Lastly, one of the things that has frustrated me since I built the tool calling was that many of the operations in the tools generate code/text/diffs but this was only visible as part of a JSON data structure. As I was making updates to this code, I added the ability for tools to provide markdown-formatted context information that can be pretty-printed.
This feels like a big improvement. I can now get a much clearer idea what an LLM is doing before I have to agree to let it do anything.
It's also much easier to work out how an LLM is using the AIFPL tool. I've noticed that Claude Sonnet is very fond of writing small AIFPL snippets that appear to help it debate approaches to problems. This is quite curious behaviour so the new update should help me understand that more. I'll now be able to read the AIFPL code quite easily.
AI tool refactoring
While I was doing the updates to the tool system I noticed quite a lof of duplication between my various AI tools. I figured this should be refactored to simplify maintenance.
It turned out that the various tools had diverged in unexpected ways so the refactoring has put them back into a standard form. This will make it easier to extend the tooling system in the future.
Humbug v0.34
v0.34 features the changes above plus some more AI model updates. This eliminates some old ones that have been replaced by substantially better new ones.
New features:
- Added grok-4-fast and grok-4.1-fast models.
- Removed grok-3 models.
- Added diff/patch syntax highlighting.
- Added the ability to extract diffs between the current editor buffer and the previously-saved version of an editor file, and added that as a
get_diffeditor AI tool operation. - Added code to allow tool calls to register a context extraction function that can present pretty-printed context to the user in the tool call messages.
- Updated the editor AI tool diff applier to pretty-print the diff that was applied.
- Updated the AIFPL AI tool evaluate operation to pretty-print the expression being evaluated.
- Updated the filesystem AI tool to show the diff in the approval widget.
- Removed Claude Haiku 3.5 and Claude Sonnet 3.7 models.
- Added GPT 5.2 models.
Bug fixes:
- Improved took operation naming in the filesystem tool.
- Resolved problem when reloading a preview tab while find was active and after editing and saving the source file on which the preview was based.
- Resolved problem if you deleted a message and the changed the zoom level (this would duplicate the text in the input prompt box).
Internal structure changes:
- Harmonized all the AI tools to use a new standard operation-based approach.