Five architecture decisions behind the Workbench redesign
Five of the Workbench's design decisions are fonts that never call Google, a story whose strength is never stored, an editor that stores one mark, a glow drawn on a canvas of its own, and a stream that falls back to polling. Each is written down in an architecture decision record: a short note that states the context, the decision and the consequences, usually with the options turned down.
The repository keeps fourteen of these records. These are five of them.
Fonts stay on the computer
Each theme has its own fonts, and the research asked for them to be bundled locally, with “no request to Google Fonts while working on a sensitive case”. The record rejected Google Fonts links because every page view would tell Google which fonts the reporter's screen uses, and because offline work would lose them.
It adds six @fontsource packages, fonts packaged to install with the app, all 5.3.0 and under the SIL Open Font Licence (OFL) 1.1. It also rejects keeping the earlier substitute fonts, because they differ from the approved pictures. Its status is Proposed: I have not accepted it yet.
A story's strength is never stored
The story screen is a tree: a thesis, the sections it will say, the claims under each and their evidence. The record states that “nothing about strength is stored.” A branch's status is worked out when it is read, from its claims' statuses, so it can never disagree with them.
The same goes for the gates, the checks a story must pass: a sentence has a citation, innocent explanations were looked at, single-source statements are marked. “Ready to run” is refused until every gate passes, and cleared if the thesis changes.
On 19 October it gained version snapshots of the write-up. A snapshot is named by a person, or taken before a save at most once in ten minutes with the newest 30 kept, or taken just before a restore so the restore can be undone.
The editor stores one mark
The write-up editor is TipTap, a rich-text editor library, at 3.31.3, the version my own site already runs. It writes the format the server already stores, so the server did not change. The marks on a sentence (no source, rests on a disputed claim) are worked out in the browser and never stored.
The exception, amended on 19 October, is whether a paragraph came from an AI draft. That cannot be worked out afterwards, so a paragraph carries the day it was drafted, and whether a person kept it. Nothing inserts AI text into the story's write-up yet.
The glow is about a hundred lines of its own
In Rain Noir the AI run's lights get a bloom, a soft glow around bright light. I am one of the deciders on the record, and I built it the way the design research specified.
It uses a canvas of its own, about 100 lines, two shaders (small programs that run on the graphics card) and no new package, instead of my site's compositor, whose bloom is the lit edge of a glass pane. Verified steps glow dimmer with three tail sprites, the small images that trail a light. Problems glow at full strength with six.
The glow carries no information the SVG lights beneath it do not carry, so screen readers and calm motion lose nothing.
A stream that falls back to polling
The Drop, where files enter the app, used to ask the server for the whole drop every two seconds. It now follows a per-file stream that sends only what changed, and while connected it does not poll. After three failed reconnects in a row it polls every two seconds as before, and the “Reading” chip says “checking every 2 s”.
The record also lists what it did not change: the old Library panel and the batch page still poll.
Sources and notes (5)
- ADR-0008: theme fonts — the notes this was written from
- ADR-0010: story tree — the notes this was written from
- ADR-0011: TipTap editor — the notes this was written from
- ADR-0013: bloom on run lights — the notes this was written from
- ADR-0014: per-file stream — the notes this was written from
Journalism WorkbenchAll 12 parts →
