MeshBay — read-only public mirror https://git.meshbay.org/meshbay.git/atom?h=0.15 2026-09-06T18:46:39Z 2026-09-06T18:46:39Z Christophe Besson cbesson@gmail.com 2026-09-06T18:46:39Z urn:sha1:37225762109ad8d11e073f8661359860d678f84a Two defects from one screenshot, and a test for the class behind the first. **The modal was transparent.** `.ftp-panel` asked for `var(--bg-panel, var(--bg))` and this stylesheet defines neither — the palette is `--bg-base`, `--bg-surface`, `--bg-raised`. An unknown custom property makes the whole declaration invalid while the rule around it still applies, so the panel simply had no background and the page showed straight through it. Three more of mine were the same: `--bg-hover`, `--bg-input`, `--danger`. **The field had no layout.** It was built on `.settings-row`, which is `display:flex; justify-content:space-between` — so a label, a hint and a value inside one end up spread across a single line in source order, which is how it read as three unrelated fragments per app. It is its own block now, and the chosen folders are a table borrowing `.shared-dirs-tbl`: these are lists, and a wrapped run of chips gives nothing to scan and nowhere to put a per-row remove. The operator is looking at two lists of directories on one page and they should read alike. `test_css_variables.py` is the general form. CSS fails silently and generously here, and nothing checked. It found four more that predate this branch: the unread-count badge (`--danger`) had white text on nothing, a transfer link had no colour, a notification card had no rounding and no unread marker. Fixed, and `--warn` and `--accent-bg` are promoted from literal fallbacks to real palette entries at exactly their current light values. Two of its own regexes were wrong before they were right — a scoped definition written inline, and one preceded by a comment, were both reported as undefined. A third check comparing the two palettes fired on `--border-focus`, a focus ring the themes share deliberately; a heuristic that has to be explained away on its first run is worse than no test, so it is gone rather than exempted. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011pvMdvLBG92jyhvD5pD6us 2026-09-06T17:03:22Z Christophe Besson cbesson@gmail.com 2026-09-06T17:03:22Z urn:sha1:ab44526a291fa673aa2850d105f6412a70a5341f Each app's settings were inlined in `group-settings.js` — TMDB, MusicBrainz, and one folder picker per app, each with its own draft state and save handler saying the same thing about a different key. They are one file per app now, reached through the `apps.js` registry, and the page that renders them names no application at all: adding one is a registry entry and a settings file. The line between the two is what makes that true. What every app has — folders — the page does generically, through one `saveDirectories` bound to the app. What one app alone has, its pane does itself with the transport it is handed. An app that only needs directories touches neither `group-settings.js` nor `group-page.js`, which is `test_app_settings_plugin.py`'s subject. `settings-ui.js` exists because a pane importing the page that renders it is a cycle, and ES modules answer that with a temporal-dead-zone ReferenceError at first render — a component that silently does not appear, the fault already recorded in CLAUDE.md about hook ordering. The flat depth-indented `<select>` of every folder in the library becomes a modal tree. It asks the node for nothing: the tree is derived from paths the client already holds, so it shows exactly what the group's index contains and adds no folder-browsing protocol. For Chat's attachment folder — the one directory that is written to rather than read — read-only roots are greyed out, so the node's refusal arrives before the operator picks rather than when somebody sends a file. Videos and Music take a list of folders. A library on two drives could not be described before; the only recourse was pointing the app at a parent containing both, which pulls in everything else under it. The scalar shapes survive on the wire alone, for a node speaking MNP 1.0, and the client reads them as a one-element list. Two things the tests caught that I would not have: `test_asset_versioning` — six new modules were missing from `_ASSETS`. Reached through the registry rather than imported by name, they are exactly the files nothing else would notice changing, and a stale one is served from cache with no version bump. And `node --check foo.js` does **not** reliably report a module syntax error: it accepted `${/* ... */''}` — htm template syntax pasted into a plain object literal — and reported success. A `.mjs` copy forces the module parser and reports it. The suite had no syntax check at all, which is how that reached a file; `test_spa_syntax.py` does it for every module now, and pins that the loose path is not what it uses. Suite: 12 failures, all pre-existing. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011pvMdvLBG92jyhvD5pD6us