0.1.3

Jobs for other trainers, and two documentation changes from the fourth training test and the Linux eval run.

• train --for musubi writes the same LoRA job for kohya-ss/musubi-tuner (a dataset TOML and a train.sh), and --for onetrainer writes a OneTrainer config and concepts file. Residual site LoRAs map to musubi's --base_weights; OneTrainer cannot merge a LoRA during training, so it refuses residual mode. Formats were read from musubi-tuner v0.3.6 and OneTrainer's 2026-10-11 source; neither trainer has run them yet.
• Prompt guide: with the brand LoRA alone on Z-Image the trigger word is optional (gen --raw); the look held without it in the fourth training test. Keep the triggers when a site LoRA is loaded too: there one picture in two fell back to the base model's photographs. Naming the header text is still what stops invented site names.
• The capture, train and mockup skills say python3 on macOS and Linux, as the hub does. In the Linux eval run, 2 of 4 agents first typed python and got command not found.

0.1.2

Faster residual site training and a prompt rule from the third training test. No default changed for captions or triggers.

• train: residual site LoRAs draw 2 sample prompts every 200 steps and skip the step-0 sample, which cut about 30 minutes from a 400-step run in the second training test. Brand LoRAs keep 4 at every checkpoint. New --sample-every; --samples 0 turns sampling off.
• Prompt guide: name the header text in quotes (the header shows the name "Fishing Service"). In the third training test this stopped the model inventing a site name or drawing the trigger word in the header. A trigger that is not a word, or captions quoting the logo, made it worse, so the defaults stay.

brandwright 0.1.1

Fixes from a fresh-install dry run of the code route and from the second training test. The code route has since been used on a work computer without trouble.

• tokens no longer takes an unstyled browser-default button as the primary button, and says when the capture has no call-to-action button.
• render warns when a font fails to load (behind a login, or refusing pages opened from disk), instead of quietly drawing a fallback font.
• The code-route guide lists the real --bw-* variable names and drops fallback values, which hid wrong names.
• README: work-machine notes from a fresh-install dry run.
• pick --to no longer crashes on Windows when ComfyUI has the old LoRA loaded: it writes a step-named copy (<brand>-<lora>-s250.safetensors) and gen uses it.

brandwright 0.1.0

First release: brand and sub-site LoRAs for on-brand UI mock-ups, with a training-free reference route and a code route. Tested end to end on GOV.UK, an openly licensed site: capture, tokens, dataset, captions, both mock-up routes, and an overnight brand plus residual site LoRA training run.

• Skills: brandwright (routes, kit, data and licence rules, common myths), brandwright-capture, brandwright-train, brandwright-mockup, each evergreen with research dated 2026-10-15.
• bw.py:
  • new, doctor, import;
  • tokens: DTCG-shaped tokens.json with per-site overrides, plus brand.css with the site's own CSS variables, @font-face rules and component recipes, including the primary button;
  • prep: viewport tiles, component specimens, perceptual and same-text dedupe, caps on repeated parts, brand and site datasets, template captions;
  • caption: a local vision model through Ollama, with a cache and style-word flags; the manifest is saved after every caption;
  • review, score (palette delta-E), sheet (cells sized to the pictures' aspect);
  • help text on every argument (the wiki's Commands page is generated from it).
• Training: train writes ostris ai-toolkit JSON for Z-Image, Z-Image Turbo, Qwen-Image-2512 or FLUX.2 klein base 4B. Site LoRAs train in residual mode (the brand LoRA is merged during training), continue mode or stack mode. --run starts the job detached, and --run --wait blocks until it ends. A retrain moves the earlier run's checkpoints aside so ai-toolkit starts from step 0; --resume continues that run instead. status and pick follow up. Default steps are about 5 passes over the dataset (400 to 1,500), with a checkpoint every 100 steps and 20 kept, because the first real run peaked at step 250 and had collapsed by step 500.
• Browser (Playwright): login, where the person signs in themselves; capture, with masking, overlay hiding, same-site crawl, --cdp and --channel msedge; render for the code route.
• ComfyUI: gen with LoRA chains and automatic triggers, sweep with a contact sheet, and ref, which uses FLUX.2 klein with up to 4 reference screenshots.
• Tests:
  • 23 offline tests;
  • eval cases for every skill (triggers, decoys and actions);
  • a browser smoke test on a fictional fixture site;
  • a size budget and a public-repository check, both in CI;
  • CI on Windows, macOS and Linux with Python 3.10 and 3.14.