2025-08-20: Humbug v0.22

Improving the `delegate_ai` tool

One of the more interesting problems of the last month has been making the delegate_ai tool work better. The tool definition has gone through a number of subtle evolutions

This morning I realized I'd made a fairly obvious (in hindsight) mistake! I had assumed that an child AI instance would "just know" it had been spawned by a parent AI, but of course it didn't because the parent didn't tell it most of the time (although sometimes it did). In most cases, the child assumed it was conversing with a human. This was leading to strange problems such as the child spawning another child, or creating files assuming a human was watching for them.

Here's what the description has now evolved into:

    description=(
        "The delegate_ai tool lets you (the parent AI) delegate cognitive tasks to other specialized child AI instances "
        "by creating a focused AI conversation for a specified task. The delegated AI has access to the same tools as "
        "you and can engage in multi-turn collaboration with you if you provide it with a session_id returned from a "
        "previous delegated conversation. Each call creates a temporary UI tab that closes after the response is "
        "returned.\n"
        "Delegation is useful for many reasons, such as:\n,"
        "- To use AI capabilities from a different model, or with different settings\n"
        "- To parallelize problem solving (you must use parallel tool/function calls to do this)\n"
        "- To seek analysis from a different perspective (you must provide the perspective in the task_prompt)\n"
        "- To use a subset of your current conversation context for more focused reasoning\n"
        "- To process a lot of context/tokens where you don't need all the context in your main conversation\n"
        "- To explore multiple approaches to a problem before converging on a solution in your main conversation\n"
        "- To give you a sounding board for ideas or to brainstorm with a different AI\n"
        "Important:\n"
        "- Do not use this tool if you would simply be giving the child AI the same task and with no significant "
        "differences in context or requirements\n"
        "- The child AI does not have access to any different tools to the ones you have\n"
        "- You must use the session_id parameter to continue a previous conversation\n"
        "- You must generate the task_prompt and that can be very expensive if you need to provide a lot of "
        "context in it. If you need to provide something that is already in the filesystem, tell the "
        "delegated AI where to read it and do not provide the contents in the task_prompt. If the content is not "
        "already in the filesystem, you must provide it in the task_prompt. You must also tell the child AI that it "
        "is a child AI instance that has been tasked using the delegate_ai tool. You must also tell the child AI that "
        "if it creates any files, it must inform you (the parent AI) about them.\n"
        "- You have the option to provide your current conversation context to a new child AI instance. This is very "
        "useful if your context has a lot of information that is relevant to the task, but this can also limit the "
        "amount of new context the child AI can process."
        "Returns: the delegated AI's response to the task_prompt, or an error message if the operation fails"
    ),

I'm reminded of one of my favourite YouTube videos on the subject of precise instructions!

Humbug v0.22

I've just finished v0.22 of Humbug. As with a few of the more recent releases, this one has quite a mixed set of changes.

Model changes

Let's start with the model changes:

• Removed GPT-4.5 as that was super-expensive and is deprecated
• Added GPT-5, as that's now the main model from OpenAI (love it or hate it)
• Removed the M6R backend - it was used for local testing but no longer serves and useful purpose
• Added GPT-OSS 20b and 120b, and added support for ollama.com's Turbo instances of both (they're very fast!)
• Added Z.ai's GLM-4.5 models

UI changes

The UI got a significant update, with more due soon. The mindspace tree on the left of the screen is now split into a conversations view and a files view. The conversations view will be updated in future releases, and this new approach prevents the files view from constantly opening the conversations folder.

While I was making this change I also cleaned up the click, double-click, and drag & drop behaviours. This is based on some UX feedback and on the importance of making click and double-click do similar things. Wiki views are now available from the context menu but aren't the default, since no files would naturally default to a wiki view being their primary interaction.

delegate_ai tool changes

The delegate_ai tool has a revised description. This should make it much more effective and avoid some weird problems where one child AI would often spawn another.

Bug fixes

A lot of bugs were fixed, and performance was improved again. There were some interesting lessons learned and they're in the Notes.

Internals improvements

I modified more of the internals of the system to expose capabilities that could be used in other applications. The main change was in the handling of AI conversation transcripts, with that code now having much cleaner APIs.

I'm anticipating more of this in the future. I would like to add more of an inversion of control so backend events can trigger frontend updates rather than requiring tight coupling with the frontend.