Sometimes a software release isn't all about a major new feature, but about making a lot of incremental improvements to things that already existed. Humbug v0.35 is one of those releases.
Giving the human-in-the-loop more insights
One problem with being the human-in-the-loop for reviewing the work of an AI is to work out what the AI was trying to do, why it was trying to do it, and working out what happened afterwards.
Key to this is to offer information about tool call contents and results. v0.34 took the first steps with this, but v0.35 goes much further. Doing this highlighted a few quirkly inconsistencies in many of the tools, however. For example the response from the filesystem tool for read_file_lines and the editor tool for read_lines was different. Worse, it was slightly weird text.
This turned out to be a general problem across many of the tools. The clock would return either ISO or Unix timestamp details but the filesystem would only return Unix timestamps. The filesystem read_file operation added text above the actual file contents but the AI would have no idea what that meant if the file contents looked similar.
The resulting changes aren't perfect, but they're now dramatically less surprising to an AI that wants similar results from different tools.
Improving tool result handling
Until now we've had "Approve Tool Call" and "Reject Tool Call" as a pretty binary set of options to any tool authorization. There was also a third, undocument, option to just submit a new message and that would reject the tool call but with an explanation of why, but this was a pretty weird user experience.
In v0.35 I've added the option for the user to say "I'm unsure about this". This rejects the tool call but tells the AI the user is unsure whether to approve or reject. It does this by sending the AI a message along with the tool rejection, but that now triggers the AI to explain what it was trying to do. This turns out to be a dramatically better user experience.
Using Humbug for writing documents
One of the things I've been doing all this week is to use Humbug to help me write documents, more so than to write code. The markdown and preview capabilities make this a very powerful option. I'll be doing more to make this particular experience even better, but this showed how powerful agentic editing and agentic document reviewing can be. It also found a few bugs!
Markdown is quite a tricky format because it's not formally defined. In turn, there are a few quirky but unusual behaviours and it turned out the dmarkdown code didn't handle a couple of them correctly. Claude fixed these without much help once I gave it a test case and let it write a stand-alone test program.
These problems were pretty grating when editing markdown, but also occasionally showed up in streaming content. I'd failed to fix one of these problems in about 4 hours earlier this year. Claude fixed both in 25 minutes. Another win for Claude!
Humbug v0.35
New features:
- The filesystem AI tool now has an option to set the date/time format for
get_inforequests. This allows it to be more consistent with the clock AI tool. - Added the ability for tool results to provide pretty-printed context information. Updated all tools to offer this capability where relevant.
- Added a mindspace setting to allow terminal tabs to auto-close when the underlying shell exits (default enabled).
- Auto-hide queued user messages after they have been used for the next user prompt.
- Added a third "I'm not sure" button to the AI tool approval widget. This tells the AI the user is not actually ready to approve and wants to discuss further.
Bug fixes:
- Updated the AI tool documentation.
- Fixed a minor race condition in the message UI that could lead to queued messages becoming the active message without updating the display to show that had happened.
- The clock AI tool never used to indicate which format it would use for outputs. It now states that it defaults to ISO.
- Relax an assertion that can occur when processing an empty markdown list.
- Resolved a problem with deleting messages while the AI is streaming a response.
- Resolved a problem with deleting messages where a tool approval is visible.
- Fixed a problem in the markdown parser where we have multiple paragraphs in a list item.
- Fixed a problem with rendering horizontal rules following tables.
- Fixed a problem with the AIFPL syntax lexer/parser so it handles multi-line strings correctly.
- Fixed a problem with the Scheme syntax lexer/parser so it handles multi-line strings correctly.