platform/THEMES-NOTES.md

Themes decoupling — working notes

Branch themes-decoupling, worktree platform-themes. Nothing here is on master.

Read this before the diff. It is the state of the work, the decisions taken and why, and the things deliberately left alone.

---

Where it got to

The visual layer is out of the app and into src/workspaces/theme. The application no longer contains any colour: it imports one stylesheet and mounts one provider.

Proven by the two things that matter:

• Duck renders identically to before. Panels, wallpaper, grid fills, glass, logo — checked against screenshots taken before anything was touched.
• Plain renders as white with black ink, with no component edits. Panels turn white, the teal avatar turns black, grid fills disappear. That is the decoupling working.

Switch themes from the avatar menu, top right.

---

The shape

src/workspaces/theme/
  css/base.css              structure — Tailwind, the @theme contract, element defaults. NO values.
  css/themes/duck-light.css  the app's original appearance, moved unchanged
  css/themes/duck-dark.css   ditto, the old .dark block
  css/themes/plain.css       white/black — an instrument, not a design
  css/chrome.css            scrollbars and overlays, now in theme variables
  css/index.css             the single import an app makes
  src/themes.ts             the registry: id, name, family, applyDarkClass. Nothing about appearance.
  src/ThemeProvider.tsx     sets data-theme, toggles .dark, persists the choice

A theme is a CSS file scoped to [data-theme='<id>'] plus one line in the registry. Nothing else.

---

Decisions, and why

Light and dark are separate themes, not a mode. The owner's call, and it removes a concept from the app. Duck Light and Duck Dark are siblings. Written out in full rather than as a diff — a theme that only makes sense in the presence of another cannot be authored by a third party or sold to a customer.

.dark survives as a compatibility flag. 396 dark: utilities are written into 33 components and they resolve against that class. A theme sets applyDarkClass and they keep working. Shim with a known end date; removing those 396 retires it.

duck-* utility names are kept, pointed at neutral variables. ~1,470 occurrences across 108 files. Renaming them and extracting in the same change would make a visual difference impossible to attribute. They now resolve through --brand-primary, --ink, and friends, so a non-Duck theme can define them sensibly even while the spelling is wrong.

duck-dark is --ink. It was never a brand colour: #14532d in light, near-white in dark. 946 components use it to mean "text". That misnaming is why it is the most-used token in the app.

Surfaces are complete CSS values, not HSL triplets. --panel-bg can be rgba(), a gradient or an image. A panel background that could only be a flat colour would rule out the glass the app has now.

The old provider file is a re-export. Six call sites keep working untouched, so the extraction is one variable.

---

Blockers and open questions — for the morning

1. The reference repo (vimazing) has not been read yet. The switching mechanism (data-theme on <html>) and the packaging format are deliberately the dumbest thing that works, because that repo will dictate them. Everything above is independent of that choice.

2. Leaks the Plain theme exposed. These are components painting themselves rather than asking the theme. The screenshots in .snapshots/plain/ are the punch list:

• The wallpaper. Still renders under Plain, so it is set by a component, not --page-bg. Probably the dashboard layout. Needs to become a theme-owned surface.
• Panel radius. --panel-radius: 0 had no effect — panels hardcode rounded-*. Shape is themeable in principle and is not yet.
• The logo. A fixed image. Fine for Duck, wrong for white-labelling: a customer's build needs its own.
• ~985 raw Tailwind palette usages (text-emerald-400, bg-black/10) and ~119 hex literals. The worst concentration by far is officerdev/src/apps/Soulseek/ (~285). MetricCard.tsx is the highest-leverage single fix — a shared component hardcoding its palette.

3. Server settings still store appearance.colorMode, not a theme id. useThemeSync used to stamp the server's colour mode over the local theme on every load — picking Plain and reloading silently reverted it. Now the server value is only honoured when the browser has never chosen. The proper fix is storing a theme id, which needs a settings migration.

4. A pre-existing bug, unrelated but adjacent. The server's default is appearance: { theme: 'light' } while the client reads appearance.colorMode — so a fresh account's server default has never been read. Left alone.

5. rebuildFrontend() reports "Bundle failed" and discards the bundler's actual errors. Cost twenty minutes tonight. Same shape as the agent bugs from earlier — a failure that refuses to say why.

6. Theme ASSETS have no packaging story yet — deferred deliberately. A theme publishes an asset as a quoted path (--logo-src: '/officer-logo.svg') and useThemeAsset wraps it, because Bun's CSS bundler resolves url() at build time while these files are served from public/ at runtime. That works for themes we ship. It does not answer where a THIRD PARTY's wallpaper and logo live. The owner's intent is a convention under public/ — something like public/themes/<name>/ — so an author drops files in a known structure and references them by path. Not designed yet; the token shape above already fits it.

7. Landing and marketplace are untouched. They have their own copies of the palette; duck-teal has three different definitions across the three apps. Folding them into this layer is a natural follow-up and was not in tonight's scope.

---

On verification, and what the screenshots can and cannot tell you

The byte comparison is dominated by wallpaper noise. The dashboard background rotates per page load, so any screen showing it differs from itself between runs — a control capture of the same build against itself had 4 of 16 differ, and which 4 changes each time. Between 5 and 6 screens come out byte-identical to the pre-refactor baseline on any given run; the rest are the wallpaper.

So the authoritative instrument is scripts/theme-vardiff.ts, which compares the COMPUTED value of every theme variable between the old and new builds in the same colour mode. It reports all 53 identical. That is a much stronger statement than a screenshot: it says the token layer resolves to exactly the values the old stylesheet produced.

It also found the one real regression, which no amount of looking would have: six --duck-* variables came out EMPTY in the new build. The utilities were remapped, but prose.css reads var(--duck-teal) DIRECTLY in eleven places for chat markdown — links, blockquotes, inline code, checkboxes — and landing/styles.css does the same. Fixed by keeping the legacy names as aliases in every theme, which also protects any plugin or inline style doing the same thing. They retire with the utility rename.

Tools built

• scripts/theme-snapshot.ts — captures 8 routes × 2 modes, or a pinned theme with --theme. Compares two directories byte-for-byte. Its first version silently photographed light twice; it now warns when a light/dark pair is identical.
• scripts/theme-vardiff.ts — computed theme variables, old build vs new, side by side. The instrument that actually proved the extraction faithful.
• scripts/theme-preview-server.ts — serves this worktree's build against the RUNNING platform's API, so the new frontend can be seen with real data without starting a second platform beside the live one.

Byte comparison is too strict for before/after across two builds: this worktree has no plugins, so the dock differs on every page, and live data moves between captures. Use it to find which screens changed, then look at them.