The project
Built to give a coding model eyes on the interface it was writing blind. What followed came from real use and from questioning models of every family — not from a roadmap, and not by accident either.
A language model was writing web interface code and could change it without ever seeing the result: it worked from what it was told, and the description is where the imprecisions got in. The operator’s intent was to give that model eyes — so it could look at the interface itself instead of taking someone’s word for what was wrong.
The first version did one thing: capture a page and hand back the image.
The shape of how it grew
Three stages, in that order — each one opened uses the previous one could not reach:
Seeing came first — a screenshot, then numbered interactive elements, then the accessibility tree. Each addition because the previous one was not enough to act on, and each one widened what a scenario could cover.
Acting followed — filling, clicking, navigating, and then the awkward cases that only appear in real use: elements the browser refuses to click, forms that report success and never submit, pages that never finish loading.
Trusting came last and is the largest part. The encrypted directory, the declared identity, the operations log, the reading lock. Part of it was anticipated; part of it was forced by a model mishandling what it had been given.
The development log
The development log runs in reverse chronological order and the field notes carry more than a hundred numbered frictions. Read together they show how the tool actually grew.
Sessions 1 to 17 are recorded outside the public repository, in the project’s own working notes; from session 18 onward the log is public and continuous.
Who builds it
Three participants, with distinct roles, and the repository has said so from the start.
Ronan Davalan — architect and arbiter: direction, security requirements, validation. Every architectural decision goes through him.
Claude Code (Anthropic) — systems engineer: the source code, the implementation, the technical documentation.
Gemini (Google) — synthesizer and strategic advisor: independent architectural analysis, cross-validation, resolution of logical conflicts.
Around them, models of other families are consulted — for review, for role-play, for maintenance passes — and local perception models do the visual work inside the tool itself.
Three things that did not change
It runs on your machine. Captures, analyses, credentials — the default has never been anything else.
It says what it is. No version of Diwall has shipped with disguise as the default.
It executes, it does not decide. Every action comes from a file you wrote. The tool has never been given the authority to add a step.
Where it is now
Six commands, an encrypted directory, scenarios validated against a schema before a browser starts, four languages of documentation, and a manual page in each. Debian, Ubuntu, Fedora, openSUSE and Arch Linux by package, or by Git clone on any of them.
And a list of limits that grows deliberately, because a limit that is written down is one that can be pushed back.