Wed Oct 21

October 21, 2026

This is the Bronco dev log for Wednesday, October 21, 2026! Let's go!!! 🐎🚀🔥 This is written by Robert Austin.

A Bronco session asking to see the news on cnn.com. The session shows a navigate tool chip and Bronco's reply summarizing the CNN homepage.
  1. Namespace Workers to their Skill 🔧
  2. Reading Text From Images: OCR in Bronco 👁️
  3. Corpora Support in Bronco Knowledge Base 🗂️
  4. Larger Model Support: Design Spike 🎛️
  5. Workers: a job with a shape 👷
  6. The file tools stop talking like a shell 🐚
  7. The crop skill stops guessing 📐

1. Namespace Workers to their Skill 🔧

Two skill directories listed side by side, crop-image and crop-image-2, each shipping a worker named lasso. The two no longer collide because a worker is named by its skill: crop-image/lasso and crop-image-2/lasso.

Bronco Workers allow one Bronco session to use another, without even knowing it. The Worker acts just like a normal tool. It takes arguments, just like other Bronco tools, and it has a return value too. Both are defined by a schema. When a session calls a tool that's backed by a worker, a new session is spawned. When the session completes, the value is parsed by the schema. If it fails, the errors are provided to Bronco so it can try a few more times. Ultimately, and hopefully, the worker returns a useful answer. If not, the session that called it can try again.

Workers are defined by files, which include their skills, their tools, their instructions, and their profile. The files can be in a project's worker folder or a skill's worker folder. Workers defined in a skill are only available to sessions that are using that skill.

During development, we discovered a bug. Bronco was combining every worker from every skill into a big list when trying to find a single skill. This jumbled up workers defined in different skills and caused confusion.

Take crop-image and crop-image-2. crop-image-2 was my second attempt to make an image cropping skill. I wanted to simply duplicate the folder and try changing a few things. That's the normal Bronco workflow. But since both crop-image and crop-image-2 had several workers that still had the same name (because I didn't change the name in crop-image-2 after duplicating it), Bronco got confused and didn't run any workers. Now we've fixed the bug and when a worker is called, the one defined in the skill is used, not the one from other skills.

A Bronco session running the crop-image-2 skill: a use_skill chip, a crop-image-2/lasso worker chip, and a shell chip that cropped the test image, followed by Bronco reporting the box and the cropped file.

2. Reading Text From Images: OCR in Bronco 👁️

Diagram of how a document gets read. Read tools and the knowledge base both send a document into an auto-detect step. A clean text layer goes to markitdown; an image or a bad scan goes through the chat orchestrator to DeepSeek-OCR in the gpu worker. Both paths produce clean markdown that returns to the session or the knowledge base index. Markitdown stays the default for clean text.

OCR, or Optical Character Recognition, is a specific way that AI can read an image document. Some documents, including PDFs, contain text that's actually encoded as images. For example, a PDF file created by a scanner. Bronco's normal AI model can read these images and read the text in them, but it is not an expert transcriber, and it cannot effectively produce a text version of a long and complex multiline document on its own without taking a long time and risking errors.

I have begun groundwork on adding a new OCR specific model to Bronco. It excels at converting handwritten and scanned documents into text ready for search. Bronco auto-detects when to use the special model on its own.

3. Corpora Support in Bronco Knowledge Base 🗂️

Diagram of the knowledge base as corpora: a dashed container holds four named corpora, skills, sessions, tool results, and documents, each with its own index. A note below reads no catch-all over Bronco, one index per corpus, config file with live reload.

Bronco reads your files: it splits them into small pieces and puts each piece into a database where it can be searched by content or meaning.

Bronco used to read and search just the ~/Bronco folder. Now it searches the folders of ~/Bronco and each of your other projects.

You can also pause or disable file reading for any project, or for Bronco entirely, from the tray menu.

Each search tool points at one corpus. search_workspace searches your document corpora, the folders you pointed at for each project plus the shared ~/Bronco workspace. search_sessions searches every conversation you have had with Bronco, and search_tool_results searches the files those conversations produced. Skill search looks through the skills themselves, so Bronco can find a skill it has not loaded yet and pull it in on the spot.

4. Larger Model Support: Design Spike 🎛️

Diagram of the global model selector. A Settings panel holds one choice, Small or Large, and the GPU worker keeps one model resident, loading from the bundled dist-models or a side-load folder. If a swap fails the worker restores the previous model. A note reads one model selected, every session runs on it.

Bronco includes a small and powerful model. It always will. This makes sharing skills between Bronco users trivial. Today I began experimenting with a new larger set of weights for the Bronco model: going from 4 billion parameters to 26 billion. The 4 billion parameter model was chosen to allow fitting a Virtual Machine, a Web Browser, the Knowledge Base (vector and FTS DB) and the Large Language Model all at once on a 16GB MacBook Air (a reasonable purchase even for a student.) My users are asking already for largers models so I am experimenting with supporting a second size model in Bronco. I am also experimenting with a medium sized option to replace the small 4B.

Larger model weights actually make Bronco slower on the same hardware, so going up in size doesn't always mean Bronco will perform better. Especially if you give Bronco time sensitive tasks, a smaller model can actually get work done more accurately (because it can read and react more frequently in the same time as its calculations are quicker as it has less math to do when responding.)

However small model weights, especially hybrid quantized models, can struggle with long and complex instructions while also having less rote knowledge of human culture. Furthermore, Macbook Pro and Mac Studio models can run the 26 billion parameter models quite fast. If you have any feedback, please let me know your favorite model size. My custom Bronco at home uses a 35 billion parameter set of weights and I run 128GB of RAM.

5. Workers: a job with a shape 👷

A worker is one job in its own subsession. A session loads skills, a skill is a namespace that bundles instructions, files, and workers, a worker declares its input and output schema plus the tools it needs, and a subsession runs the job with exactly those tools. A note below reads nothing is stored: a session's tools come from its use_skill calls, and skills and workers resolve by name, never a scan of everything.

Bronco Projects and Skills can define "workers". A Worker is a way for a Bronco session to run another new Bronco session in order to delegate work. Each Worker obeys a contract: it receives an exact type of input and returns an exact type of output, and it always follows its instructions. A Worker gets access to just the tools and skills it was defined with. Workers are defined by files in your Projects and Skills folders. Workers have names and also profiles, which control how Bronco generates replies, including how much Bronco generates, how much Bronco thinks, and whether Bronco's replies are random or fixed.

When you talk to Bronco, it calls tools for you. Usually these are traditional computer programs and subroutines that do things like update a file or search your files. But sometimes a tool is actually a Worker. When Bronco calls the tool, a new Bronco session is created, and it receives the instructions and tools defined by the Worker. It gets a new message including the arguments from the tool call. It does the work it is told with the arguments and calls a special tool to return the result. This allows two Bronco sessions to cooperate over a strict contract, and these Workers allow you to organize your instructions for re-use within a Skill or a Project.

6. The file tools stop talking like a shell 🐚

Diagram of the file-tool result path: a tool runs a sandbox command, the host classifies its exit code and stderr (stderr is never shown), the tool returns a typed result made of a success and a few classified failures, and the result is wrapped in a per-tool guard tag such as bronco-ls-result or bronco-ls-failure before the model reads it as data. A dashed Before strip shows the old path, where raw stdout, stderr, and exit code reached the model and an empty search counted as a failed tool. A note reads a nonzero exit is an outcome, not a failure: ripgrep exiting 1 means no matches, which is a good answer. shell stays a shell, wrapped in bronco-shell-result.

ls, glob, grep, write, edit, and shell are the tools a session uses to look at and change your files. They run in a Linux virtual machine, and they are all executed indirectly through the shell tool. shell is a general "use a computer by writing to it" command, so it can perform the other actions when given the right text. When Bronco used the tools, messages, including errors and instructions, intended for shell were being returned for the other tools. This confused Bronco.

For example, Bronco was receiving [exit_code: 1] messages after calling edit, a tool that edits a file. exit_code is a shell-related error code, so that was unexpected and incorrect.

Now each result is wrapped in its own guard snippet, a pair of tags such as <bronco-ls-result> or <bronco-edit-failure> with the data inside. The guard tells the model to read what is between the tags as data, never as instructions. The raw error text is read only inside the host, to decide which clean outcome to report, and never reaches the model. shell is the one tool that stays a shell, because that is what it is, and it carries its own guard too.

7. The crop skill stops guessing 📐

Two panels. Left, labelled lasso preview: the full photo of a model railway with numbered pixel rulers across the frame and the crop box drawn in coloured edges around a red diesel locomotive pulling sushi cars. Right, labelled cropped: the finished tight crop of the locomotive and its sushi cars.

crop-image is a skill that crops an image down to the object you name. First Bronco looked at a scaled-down version of your image, and then it provided the dimensions of a region to crop. This sorta worked, but the crops were not very pleasing. I dug into Bronco's thinking and saw that it was just guessing.

So I gave Bronco a new script that generated a version of the image that included ruler lines going up-and-down and left-and-right across the image, as well as axis number lines on the top and side. This way Bronco could determine what numbers lined up with the target region visually.

Now Bronco can crop pixel-perfect with a few iterations.