Following the theme of Humbug v46, v47 is also dominated by user experience improvements. While many of these are incremental, it does feature one new experimental idea: blueprints.
Blueprints
A few months ago I retired my Metaphor language design. It was a very interesting idea but felt too limited given the evolution of agentic search capabilities. There was one area, however, which the agentic search didn't quite do how I wanted. AGENTS.md files are great for orienting AIs to what is in a mindspace, but aren't really quite right for capturing the shared intent of humans and AIs for what the mindspace is for and how it should evolve. To fix this I've added an experimental feature: blueprints.
The idea of a blueprint is that we add a new type of file, named blueprint.md to the mindspace root folder, and potentially to sub-folders within it. These files allow both humans and AIs to define the intent.
AGENTS.md and README.md files orient AIs and humans as to what is there, blueprint.md files orient both as to how this should evolve.
This is very much experimental right now, but the results have been very positive so far. I've seen Claude start to apply discipline around dependency management and in code structure that I've not seen before.
I also updated the system prompt to tell the AIs to update both types of docs where changes are appropriate. Previously, this would have been quite ad-hoc, but now it feels much more systematic.
PDF and Microsoft Word files
One of the original design intents for Humbug was to help with non-code related projects as much as software projects. I finally took this a step further by adding support for reading the text content of PDF and Microsoft Word (.docx) files. These can either be attached to conversations or the AI can read them from the filesystem.
Now it's possible to have an AI compare its code directly against a spec, a research paper, a product requirement, etc., but with those documents in their natural form.
True to the Humbug ethos (and it's in the blueprint), this has been done without adding any new third party dependencies!
Non-baked in AI models
I need to do something more effective, but for now there's an ability to add AI models to the catalogue via a user-defined config file in [object Object]. There's an example in the docs directory.
This makes it possible to configure Humbug to use, say, custom LLMs in an enterprise environment.
Version 47
New features:
- Updated the grok model catalogue to grok 4.3, removing the now-obsolete 4.0 and 4.1 models.
- Added the ability to view files that were attached to a message in a conversation.
- Introduced the idea of [object Object] files, allowing both AIs and human users to understand the intent of a mindspace.
- Allow a [object Object] file to be used to define models not directly supported by Humbug.
- Added support for reading the text in PDF and DOCX files when attaching files to a conversation or when reading files via the filesystem AI tool.
- The message input box now floats above historical messages when scrolling, making it easier to read earlier parts of a conversation while typing.
- The alt+up/alt+down keyboard shortcuts in the conversation tabs now scroll between user messages, skipping all other types of message. This lets you scroll to the important events more quickly.
Bug fixes:
- Resolved an issue with lazy loading introduced in v46.
- Improved the tab bar rendering with a custom painter (faster and contents are positioned more accurately).
- Dragging a file from the mindspace tree that is already open in a column will now correctly open the tab in the correct location.
- Scroll wheel operations on the Windows terminal now work correctly.
- When shrinking a conversation tab's width, the message label will now elide (replace text with [object Object]) so the message does not end up too wide.
Internal structure changes:
- Attachments have been added as a specific concept inside AI conversations. This keeps the context cleaner and will allow future innovations.