The first carve, and deliberately the riskiest: app.js IS the plugin loader (the
R0 host rails), so it goes first while the module graph is still one edge deep.
static/js/plugin-loader.js (829 lines) — bodies VERBATIM. app.js 12,217 → 11,439.
Core's first `static/js/` module, exactly as constitution II anticipates.
CLOSURE (measured with acorn, not regex — brace-matching stripped source drifted):
the block at app.js:11246-12031 is contiguous and self-contained. It needs only
TWO things from the rest of app.js, and exports only TWO:
exports: loadPlugins (the window contract), bootstrapPluginsAndUi (boot)
inbound: window.showScreen — already the public host contract (constitution II),
so it is called through `window`, not re-coupled as an import
_populateVizPicker — injected via configurePluginLoader()
WHY A SEAM, NOT AN IMPORT. plugin-loader must not import app.js: app.js imports
it, so that would close a cycle. I checked whether _populateVizPicker could just
move into the module instead (which would delete the seam entirely) — it drags 9
further symbols (_canRun3D, _autoMatchViz, _showPromotionNag, …), i.e. a whole
viz cluster. That is its own carve, so the seam stays.
THE SEAM'S DEFAULT IS LOUD, ON PURPOSE. A no-op stub is the classic silent
failure for this pattern (see the editor's setHostHooks trap, hit twice): drop the
wiring call and the loader keeps working while the viz picker quietly stops
refreshing — no test, no boot check says a word. The default now console.errors,
so the smoke harness catches it. VERIFIED BY BITE TEST: removing
configurePluginLoader() from app.js surfaces
"[plugin-loader] host seam not configured" at boot. The seam IS exercised on the
plugin-startup path, so an unwired hook cannot pass silently.
no-cycle is now LIVE on core's own graph for the first time. eslint.config.js
gains `static/app.js` + `static/js/**` to the module block — app.js now `import`s,
so parsing it as a script would be a syntax error. VERIFIED BY BITE TEST: making
plugin-loader import app.js back fails with "Dependency cycle detected".
HARNESSES (the R3a note said budget one conversion per carve — it was five):
retargeted capability_inspector_nav, plugin_hydration_wipe,
plugin_loader_script_type, plugin_style_injection, legacy_shim_hits (SPLIT — one
test needs the loader, one still needs app.js) + test_plugin_runtime_idempotence.
legacy_shim_hits was missed by a symbol-name grep because it greps for a code
STRING; only the failing run found it. test_capability_events' NEGATIVE asserts
now span app.js + the loader — carving code out of app.js would otherwise make
them vacuous instead of failing.
VERIFIED: A/B against origin/main in two browsers — mounted plugin screens, 14
loaded plugin scripts, the 3 module plugins injected as <script type="module">,
37 capability participants, 14 shims, window.loadPlugins: IDENTICAL, zero
console/page errors on both. /static/js/plugin-loader.js serves 200; R0 rails
intact (src/main.js 200, conditional GET 304, script_type passthrough).
pytest 2396, node 1032/1032, ESLint 0, Codex 0.
Codex preflight caught a REAL [P1] first pass: static/js/plugin-loader.js was
untracked, so a checkout would have served an app.js importing a nonexistent
module — a failed static import kills the whole module and every window handler
with it. Now tracked.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(plugins): don't treat transient absence from /api/plugins as uninstall
The backend clears its plugin registry at the start of load_plugins()
and repopulates it incrementally while HTTP stays up, so every backend
restart (desktop: Audio Quality soundfont switch, LAN toggle, update
restart) serves a window of partial — even empty — /api/plugins
responses. loadPlugins() treated absence from the current response as
an uninstall, with three destructive consequences for still-loaded
plugins:
1. Their settings-panel and screen DOM were wiped while their
_loadedPluginScripts entry survived, so the NEXT refetch failed the
DOM-existence check and re-evaluated the plugin's screen.js
mid-session. For the desktop audio_engine plugin that re-ran init()
against the surviving native audio chain and exactly duplicated
every VST/NAM/IR stage (the alpha testers' "chain duplicates after
leaving the Audio menu" / blown-out gain reports).
2. _reconcilePluginStyles dropped their stylesheet, leaving them
visible but unstyled until they reappeared.
3. The stale-contribution sweep unmounted their UI contributions and
unregistered their capability participant with no re-registration
path (plugin scripts don't re-run thanks to the loadedScripts
guard).
Absence is now a non-signal everywhere in loadPlugins: the DOM wipe and
style reconcile are scoped to plugins the response actually names, and
the absence sweep is removed. Present plugins still fully re-sync via
_registerLegacyPluginUiContributions each round; failed plugins are
present in the response and still cleaned up; nav is rebuilt from the
response so genuinely uninstalled plugins drop out of it, and their
(un-unloadable) already-evaluated scripts keep their DOM until reload.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test: update idempotence contract to the absence-is-not-uninstall invariant
The removed-plugin sweep contract pinned the old behavior this branch
deletes; pin the new invariant instead (no absence sweep + respondedIds
scoping on the DOM/style reconcilers).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
