← Back to Whittl
WHITTL DEV LOG
Last updated: October 2026
What just shipped, how the plan changed along the way, and what's next. Plans change once a release is in real use, so this page also says which ideas shipped, which are still coming, and which were dropped.
v2.6.1 followed v2.6.0 the same day with two fixes. In the installed version, a project with tests could open a second Whittl window and pause the AI for up to a minute; your tests now run with Whittl's own Python. And when the AI answered a question in the chat, for example with a list of ideas, and you said "do all of them", it could not see its own answer. Now it can.
v2.6.0 does two big things. You can now extend Whittl itself with plugins, and every model on every backend gets the same tools: it reads what it needs, makes targeted edits, runs your app and your tests, and fixes what it broke. Most of the rest came from using Whittl on real projects every day and fixing whatever got in the way.
- Plugins. A folder with a manifest and a Python file can add a tool the AI calls, a tab, a status bar readout, a dialog, a menu item, a skill, or code that runs at set moments. Whittl can write one for you, checks every plugin before it runs, and keeps new ones off until you switch them on. Whittl's own Assets, Archives, Terminal and Help tabs are plugins now.
- One agent for every model. What a model gets depends on what it can do, not on who makes it. Models that can't call tools reliably make the same edits with short scripts, and the AI can open an app it built, click through it and show you a picture of it.
- Your work is safe. Edits made in another editor are kept, a turn that fails partway keeps its edits and is never regenerated over them, and a Stop puts back the last version of a file that compiled.
- Pictures in the chat. Several images per message, thumbnails before you send, the full picture on a click, and they survive a restart.
- Know what you're paying. A running cost readout, a context gauge, an optional Usage tab, and an OpenRouter provider choice that refuses reduced-precision copies of a model.
- New models and skills. Claude Opus 5.5, Sonnet 5.5 and Fable 5.1, Gemini 3.1 Pro, and new built-in skills for drawing, visual styles, games, tables, files, settings, serial ports and 3D, with every rule checked on the real library first.
- 8,050 automated tests at v2.6.1, up from 3,765 at v2.5.0.
Full v2.6.1 and v2.6.0 release notes →
- v2.6.0 (October 2026). Plugins, and one agent for every model.
- v2.5.0 (September 2026). Web research, auto-learned skills, the test gate, and a long run of reliability fixes from real projects.
- v2.4.1 (May 2026). Fixed an auto-fix rule that kept breaking correct WebEngine imports.
- v2.4.0 (April 2026). The Whittl Layer, the in-app updater, the question tool, Claude Code skill paths, and the documentation site.
- v2.3.0 (April 2026). Screenshot to app, Agent Mode, and a complete visual redesign.
All release notes →
Back in April this page promised a different v2.5 and v2.6: a shared rule library, a review pass, custom agents, and a ladder that would start on a cheap model and escalate to a better one. Using Whittl every day pointed somewhere else. Models kept getting stuck with no way to look things up, so v2.5 added web research. The same mistakes kept coming back, so every auto-fix became a lesson every model sees. And the biggest gap turned out to be the agent loop itself: the same model did noticeably better in other coding agents than it did in Whittl, so v2.6 rebuilt that loop for every backend.
Here's where April's ideas ended up:
- Shipped as planned: the Whittl Layer, the in-app updater, the question tool and Claude Code skill paths, all in v2.4.
- Shipped in another form: Learn-From-Autofix became auto-learned skills in v2.5. Design to Desktop became screenshot to app, which now takes several images, plus an AI that can look at the app it built.
- Still ideas: Whittl Commons (sharing learned rules between installs; auto-learned skills are the local half), Custom Agents, and Consensus Mode.
- Dropped: the Uplift Ladder. Whittl does the work on the model you picked and doesn't switch models behind your back.
None of this is final, but it's where v2.7 is headed.
Faster decisions
Before and during a turn, Whittl asks your model a few small questions: what a message is asking for, which function a change touches, which files matter. Each one waits two to three seconds on a cloud model. For OpenRouter users, a small decision model (Jev) can answer the same questions in under a second. It will be an opt-in setting, off by default, with today's way as the fallback.
Pick the platform first
A New Project dialog that starts with where the app will run: Windows, Linux, a phone, or plain Python. Whittl picks the right kind of project from there and says plainly what your machine can and can't build.
Groundwork
Each kind of project described in one place instead of spread across sixteen files, so new targets are easier to add and harder to get wrong.
Review a turn
A view of everything one turn changed: the files, the diff, and one button to undo the whole turn.
Outside edits, noticed live
Whittl already keeps edits made in other editors. A file watcher will show them as they happen.
A model benchmark
The same set of tasks run through each model, with the test gate as the judge: pass rate, rounds, cost and time. The numbers behind "works on every model".
- GTK4 and libadwaita apps, Linux first. GNOME-style desktop apps and Linux phone apps, with a way to run them on the phone straight from Whittl.
- Flet 1.0. Mobile projects stay on Flet 0.28.3 on purpose until Flet 1.0 is stable. After that, existing projects can stay where they are while new ones move up.
- Whittl Commons. Rules learned on one machine, shared with everyone who opts in, with a preview of exactly what would leave your machine before anything does.
- Signed updates. A signed update manifest, so the download server alone can't swap an installer.
- Whittl in other languages.
- Custom Agents and Consensus Mode, still on the list.
Most AI coding tools are a chat window over someone else's model, and their quality rises and falls with that model. Whittl puts a layer between you and the model: auto-fix rules learned from real bugs, skills checked against the real libraries, a test gate, and an agent loop that gives every model the same tools. The model writes the code, and Whittl makes it better at building Python apps, whichever model fits your budget.
It's also built to be owned. It runs locally, it's a one-time purchase, and your code stays as real files on your own disk.
- Ship and iterate. Real use beats brainstorming, which is why this page keeps changing.
- Check before it teaches. Every skill rule is tested on the real library before any model sees it.
- Every model gets the same agent. What a model can do decides what it gets, not who makes it.
- The model you picked does the work. Whittl never quietly switches models. Anything else, like the faster decision model, is yours to turn on.
- Never lose your work. Your edits survive failed turns, other editors and the Stop button.
- Five backends, every budget. OpenRouter (free models and 200+ others), DeepSeek (cheap), Gemini, Claude (premium) and Ollama (free, on your own computer).
- One-time purchase. No subscriptions, no cloud lock-in, your code on your disk.