The occ_service export worker has its own MEMFS — the editor's lazily-fetched
lib models were invisible there, so every export was a bare board (54
"Could not add 3D model" warnings on pic_programmer, 2 STEP products).
- models-bridge: collectBoardModelFiles(boardText) — scan refs, ensure via
the 0004 sparse source (IDB/R2, wrl->step fallback), read staged bytes
back, dedupe by real staged path.
- occ-service.ts: attach the collected models to every export request
(best-effort — prefetch failure still exports, misses reported by the
exporter); transfer the body buffers.
- occ-worker.js (shared app/harness): pass req.models through to occExport.
- occ_service_main.cpp: occExport(board, params, models) stages each entry
under PCBJAM_3D::MODELS_MEMFS_ROOT (path-sanitized) for the exporter's
staged-model probe (kicad 83645275ac), removed again after the export.
- tests: harness occ stub mirrors the prefetch against the page kicadLibs
provider + captures report/productCount; new occ-export-models.spec.ts
guards the delivery (green companion pins preconditions; guard asserts 0
missing lib models + component PRODUCTs). pic_programmer: 17/17 staged,
87 products @ 13.3 MB (was 2 @ 402 KB). models-bridge unit tests 13/13.
Known remainder (0007 step 4): project-local ${KIPRJMOD} refs still drop.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017UjpnviP3ZDxTM1Ap63Sqv
synced-source subscribes to its SyncStack: remote changes (self-save echoes
consumed via a selfPushed flag) debounce per kind into kicadLibsReload — a
new embind export (pcbjam_libs_reload.h, all three TUs) that drops the lib's
plugin cache (LIBRARY_MANAGER::ReloadLibraryEntry), reloads it, and mails
MAIL_RELOAD_LIB with the nickname so the symbol tree force-refreshes (the
plugin's modify hash is a pinned constant, so a plain sync would skip it).
After the reload, kicadLibsSymbolUsage (new eeschema embind: placed
SCH_SYMBOL count across unique screens) gates LIB_ITEM_UPDATED_EVENT, and
WasmTool shows an amber toast when a PLACED symbol changed — placed copies
keep the previous version until updated from the library.
syncedScopeLibsSource gives PROJECT sessions the synced source under
VITE_LIBS_SOURCE=synced (remote contract for lib listing/createLib, lazy
per-lib SyncStacks for item ops/presync) so realtime reaches open
schematics; previously project sessions silently fell back to the per-item
remote source. Unit tests cover reload debounce, self-echo skip, per-kind
routing, usage-gated event, and the no-Module no-op.
Bumps kicad (MAIL_RELOAD_LIB force-refresh payload).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QRWoXiM9uuo1enXGhAYku
emitViewportIfChanged deduped on scale+center only, so a canvas SIZE change
(boot layout settling after the bind-time seed, window resize) never re-pushed
{w,h} to JS. CommentLayer's worldToScreen maps through h/2, so every DOM pin
hit target (and its hover ring) sat vertically offset from its GAL dot by
exactly delta-h/2 css-px until the next pan/zoom finally passed the dedupe.
Size now participates in the dedupe and wxEVT_SIZE re-pushes post-layout
(CallAfter, after the GAL's own onSize). New regression e2e
comments-viewport-resize.spec.ts asserts the DOM pin re-aligns with GAL truth
(fresh kicadCollabGetViewport) across a window resize with no pan/zoom.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011x3h5AzkWDbHmCkeVYAuzU
New wasm/cli/sym_convert_stubs.cpp: first-definition-wins overrides
(--allow-multiple-definition) severing schematic load/save, the KIFONT
factory (drops newstroke + freetype/harfbuzz), and the four UI virtuals
whose bodies reference pruned typeinfo/data. New
wasm/bindings/sym_convert_embind.cpp: no-op kicadCollabOnSave so the
converter stops linking eeschema's embind object (and --bind) that rooted
the editor surface from .init_array. build-kicad-target.sh: per-app
EMBIND_LINK_FLAG + sym_convert now uses its own embind TU. Bumps kicad
(kiface prune, gated). Output byte-identical on the qa corpus.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H58sC87w12FotXEmxrJrQV
Follow-user: click a peer's roster avatar to mirror their viewport until
local input breaks it.
- collab_presence_core.h: CORE::fitViewport(cx, cy, halfW, halfH) — fit the
leader's world rect with CONTAIN semantics (zoom derived from the
follower's own canvas via the ToScreen ratio; GetScale is the zoom, not
px/IU). Exported as kicadCollabFitViewport from both editor TUs + the
merged dispatcher.
- presence-kicad.ts: publish the visible world rect (viewportRect) into
awareness, 100 ms trailing throttle; guarded for pre-0008 handles.
- follow-user.ts: createFollow — follows an awareness CLIENT (a tab, not a
user); applies leader rect changes via FitViewport, dedupes unchanged
republishes; break-on-interact compares local onViewport echoes against
the last applied rect (2% rel tolerance, echo-grace before the first fit
lands); unfollows on leader-left; pauses on eeschema sheet mismatch.
- PresenceRoster: avatars are follow toggles (ring on the followed peer);
WasmTool renders the "Following <name> — move to stop" banner.
- tests: 7 controller units (85/85 collab), fitViewport round-trip e2e in
both kicad presence specs (20/20), two-tab tests/web/follow.spec.ts
(converge → track → wheel-zoom breaks → subsequent moves ignored).
Chip depth-layer fix (user-reported): name chips washed out inside
low-alpha selection fills — chip rects shared the shapes overlay's single
depth, and same-depth fragments drawn LATER lose the depth test, so an
earlier-painted fill rejected the chip's pixels. Now three layers via the
fork's VIEW_OVERLAY::SetDepthOffset: text (0) < chips + pin dots (1) <
selection shapes (2). drawLabel/drawCursor/drawSelectionBox take the chip
overlay explicitly; comment-pin dots move to the chip layer too (the 0005
"drawn last so pins sit above" comment had the rule backwards). Verified
with a chip-inside-30%-fill pixel repro + the full presence suite.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013u9h8fkQktH7KRECaHJmUG
The tag-triggered release pipeline built the demo (and backed editor)
without --models-tag, so modelsSourceConfig() returned null and every
kind=model3d ensure logged "(unserved)" — 3D component models never
loaded on demo.pcbjam.com. The manual deploy-demo.yml already passed
it; mirror its MODELS_TAG env (10.0.3) in release.yml for both builds.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UQRBzfidkoTU5yc7HnF9x7
The eeschema/pcbnew binding TUs had ~1000 lines of copy-pasted collab code.
Factored into two header-only shared files (zero build-script changes — the
collab_presence_style.h precedent):
- collab_common.h (pcbjam_collab): toUtf8, runOnFiber (the CallAfter+COROUTINE
fiber idiom — was ~25 inline copies), the window.kicadCollab wire emitters
(onDelta/onItems/onCursor/onSelection/onViewport), frame-generic undo test
hooks.
- collab_presence_core.h (pcbjam_presence::CORE): PEER/PIN + all presence
state and machinery (start/canvas binds/lock query, setRemote/setPins/
setStyle, selection check + dedupe, overlay redraw loop, viewport push/pull,
releaseSelection, locks probe), written against the EDA_DRAW_FRAME +
SELECTION_TOOL base classes. Per-editor hooks: frame, selectionTool,
selectionEmitPayload, resolveItem, drawPeerShapes. One CORE instance per TU
(anonymous-namespace presenceCore()) so the merged image keeps per-editor
state separation.
- NEW per-editor resolveXsel(frame, peer): ONE cross-app resolver shared by
the ghost render AND kicadCollabTestGetCrossMapped — the mapping loop was
duplicated within each TU, letting the test probe drift from the pixels.
Deliberately NOT factored: the Yjs differ/apply halves (itemToJson/makeItem/
flushDiff/doApply*) — structurally parallel but the bodies encode per-editor
sync semantics and editor-specific asyncify devirtualization workarounds that
must stay visible. kicadOpenFile/kicadCollabOnSave keep the existing
KICAD_MERGED_EMBIND mechanism. TestClearSelection stays editor-typed
(ClearSelection is not on the SELECTION_TOOL base).
eeschema_embind 2203→1721 lines, pcbnew_embind 2518→1993. JS-facing names,
signatures and the kicad_editor_embind.cpp dispatcher are unchanged.
Verified: kicad_editor image builds clean; tests/kicad presence+locks 18/18
(incl. ghost-render pixel compares), collab+ysync-repros 31 passed/2 skipped.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013u9h8fkQktH7KRECaHJmUG
Chrome visibility is now a runtime toggle on any device instead of being
device-wired: mobile defaults to canvas-only, desktop to full UI, and the
floating top-right button (or Figma's Cmd/Ctrl+\ chord — free in KiCad,
only bare \ is bound) flips between them live.
- chrome-visibility.ts: module-global store (default isMobileMode(),
session-only) + pure hotkey matcher (rejects AltGr backslash + repeats)
- WasmTool: capture-phase hotkey (stopped before the wx layer), floating
toggle pill (matches the comment FAB design), useLayoutEffect apply with
sync first call + retry; overlays follow the toggle, capability-gated on
the kicad_editor bundle's kicadSetChrome export
- boot.ts mobile opt now installs touch gestures only
- kicadSetChrome: frame-keyed hide-time snapshot so restore re-shows ONLY
what hide took away (blanket Show(true) surfaced KiCad's default-hidden
Search/Properties/Net-Inspector panes); toolbars-only fallback otherwise
- tests: chrome-toggle.spec.ts (desktop hide/restore + geometric
restore-exactness ±3px), mobile toggle round-trip, 12 new unit tests,
test:web:mobile npm script
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
On a mobile device (or ?mobile=1) the editors run canvas-only with touch
gestures driving the view:
- touch-gestures.ts: pure recognizer (unit-tested) + DOM shim installed in
boot preRun — one-finger drag → synthetic middle-drag (pan), pinch →
synthetic wheel at the centroid (zoom-to-cursor), tap → left click.
preRun registration order is what lets stopImmediatePropagation suppress
the wx layer's single-finger→LEFT-drag touch mapping.
- kicadSetChrome(bool) embind: hides all AUI panes except DrawFrame + the
menubar/status bar via generic wx APIs (kicad fork untouched); boot polls
it after runtime init. Pairs with the wxwidgets IsShown layout fix.
- mobile-mode.ts: ?mobile=1/0 override or UA-CH/coarse-pointer autodetect;
shell hides its overlays and the inherent-to-mobile preflight warnings.
- e2e: mobile-chromium project (Pixel 7) + 4 specs (chrome-less, tap,
pinch, pan) with screenshot-invertibility assertions; also fixes
tool-switch.spec's stale pre-scope-refactor URLs (was broken on main).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Every KiCad editor's File→Quit used to destroy the wx top frame and
notify nobody, leaving the SPA stranded on the editor URL with a dead
canvas. It now behaves like the browser Back button: return to wherever
the user entered the editor from (project page, or a previous editor).
- WasmTool installs window.wxAppTopWindowClosed (the wx wasm port fires it
from the main frame's destructor — see the wxwidgets pointer bump). On a
real quit it history.back()s when there is same-origin in-app history,
else location.assign()s the project page (or "/" for lib editors).
- The navigation is deferred one task (setTimeout 0): it fires inside the
C++ destructor via EM_ASM under Asyncify, and a cross-document
location.assign() started there is aborted by the continuing teardown;
only same-document history.back() survives synchronously.
- Latches off on pagehide (the port also closes the frame on page unload
via UnloadCallback) and reloads on a bfcache pageshow so Forward never
restores a dead frame.
- e2e: quit-to-project.spec.ts covers both the history-back and the
deep-link fallback paths.
No kicad changes — the whole hook lives in the wxWidgets wasm port and the
web app, keeping the kicad fork clean for upstream syncs.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
build-editor.mjs already threads --plausible → VITE_PLAUSIBLE_DOMAIN; the
deploy-editor job now passes editor.pcbjam.com so the backed editor tracks
as its own Plausible site (separate from the demo's vars.PLAUSIBLE_DOMAIN).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014dasZuqo6FStgT3rkC85im
The cursor-emit tests asserted a fixed <25 emits for 30 awaited mouse
moves, assuming the sweep outpaces the 50 ms throttle. On a loaded
Ubicloud runner each Playwright->Firefox->WASM round-trip can straddle
a throttle window, so 28/30 moves emitted and the test failed (twice on
main, once from a docs-only commit).
Bound by measured sweep duration instead: one emit per 50 ms window,
+1 leading edge, +2 slack. Fast machines get a tighter bound (~9-13);
slow runners scale with actual elapsed time. A broken throttle still
fails.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Lg5jwWuhFH5dL8hEcBDuP2
New release.yml job `deploy-editor` (parallel to deploy-demo, same per-tag
WASM manifest): builds the standalone in REMOTE mode via the new
scripts/deploy/build-editor.mjs — projects/libs/auth from the closed API
(VITE_API_BASE_URL=https://api.pcbjam.com, VITE_LIBS_SOURCE=synced), Yjs
board rooms through the API host's path route (VITE_YJS_ENDPOINT=api origin,
doc source ydoc) — and ships it to the pcbjam-editor Pages project. The
demo deploy is unchanged. The closed stack itself (api./app.pcbjam.com)
deploys from pcbjam-private; when both sides change, that repo tags first.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014dasZuqo6FStgT3rkC85im
Ctrl+Z after a peer's edit no longer reverts (and re-broadcasts) the
peer's work, and the adopt undo-bomb is gone:
- doApply/doApplyItems (both editors) Push with SKIP_UNDO; the emit path
is unaffected (suppression keys off s_applyingRemote, not undo).
- With SKIP_UNDO no picker owns removed items — the bindings free them
after Push (explicit removals + upsert's remove-before-re-add; fields
excluded: CHT_REMOVE hides them, parent keeps ownership). Freeing stays
out of the fork commit classes so DRC's SKIP_UNDO callers can't
double-free.
- Test hooks kicadCollabTestUndo/UndoDepth, registered per-editor AND in
the kicad_editor dispatcher (merged image compiles out per-app
registrations).
- kicad pointer: eeschema UUID undo guard + SKIP_UNDO connectivity split
+ quiet stale-entry drop (ca8877324c).
- tests/kicad/collab-undo.spec.ts: 5 scenarios (no undo entry from remote
applies; selective undo; stranded replaced/deleted entries) — 5/5, plus
collab/ysync regression 30 pass.
- docs: ysync-review 20 fix record; 09 marked FIXED; overview indexed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019ejJEvS7ogef2o9gVTXjmp
Verdict: ~3-5 days. pcbnew undo already UUID-guards stale pointers; core
work = porting that guard to eeschema, splitting connectivity out of the
SCH SKIP_UNDO gate, plugging the SKIP_UNDO removed-item leak, silencing
the incomplete-undo modal. Cross-linked from 09 and the overview index.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019ejJEvS7ogef2o9gVTXjmp
While a peer has an item selected, local users can still select it for
inspection but move/drag/rotate/delete skip it with an infobar naming the
holder (native locked-item UX; enforced via the fork's PCBJAM_REMOTE_LOCK
query — kicad 81f9cd80fd, the epic's first fork-touching phase). Overlapping
holds (both grabbed inside the awareness propagation window) tie-break
deterministically: lowest (user.id, clientID) keeps the item, every losing
client auto-releases it.
- lock-tiebreak.ts: pure policy — beats(), remoteLocks() (union of ALL other
clients' selections incl. own user's other tabs, minus own-held-and-winning
uuids so the winner isn't blocked mid-release), contestedReleases()
- presence.ts: clients() (per-client view, no user dedupe) + self(); FIX for
a pre-existing flaky stack overflow — resolveCollision re-entered itself
synchronously via its own patch's awareness 'change' and could ping-pong on
stale same-user states (~1-in-3 unit runs); re-entrancy guard defers
re-resolution to the next genuine delivery
- presence-kicad.ts: locks ride the kicadCollabSetRemote snapshot
(`locks:[{uuid,name}]`); losing overlaps call kicadCollabReleaseSelection
- wasm bindings (both TUs + merged dispatch): g_locks map + fork query
install; kicadCollabReleaseSelection (cancelInteractive only when a tool
stack is live — bare ESC would clear the whole selection — then selective
RemoveItemFromSel + infobar + forced re-emit); kicadCollabTestGetLocked
- tests: lock-tiebreak unit suite; presence-locks e2e for both editors (real
move veto — pcbnew click+M hotkey since its default left-drag is
rubber-band select, eeschema real drag — each with an unlocked control);
two-tab tests/web/locks.spec.ts (lock propagation + deterministic tiebreak
release + unlock on clear, passing vs real partykit)
Spec: docs/features/collab-presence/0007 (closed repo).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Lg5jwWuhFH5dL8hEcBDuP2
Selecting a symbol in eeschema ghost-highlights the linked footprint(s) in
every pcbnew tab of the project, and vice versa — across users AND one
user's own two tabs. Native KIWAY cross-probe is inert in WASM (one frame
per page); this rides the presence layer instead.
- cross-app.ts: project-wide awareness-only room (presenceRoomId), publishes
full PresenceState at selection rate (cursor always null); peers() = other-
TOOL clients incl. own user's other tabs; window.__pcbjamCrossApp test handle
- presence-kicad.ts: parseSelectionEmit (bare array | {uuids,fpPaths}),
xselFromPeerState (pcbnew paths → symbol uuids; eeschema uuids verbatim),
cross peers appended to the kicadCollabSetRemote snapshot as
{id "<user>#x<client>", name "<user> · sch|pcb", xsel}
- C++ (zero fork changes): pcbnew emits {uuids, fpPaths} (FOOTPRINT::GetPath)
and ghost-renders xsel via path-tail suffix scan; eeschema resolves xsel via
ResolveItem gated to the CURRENT sheet (xsel arrives project-wide, unlike
room-scoped selections); ghostStyle = alphas × xselAlphaScale (0.55, tuner-
patchable); new exports kicadCollabGetSelectionFull / TestGetCrossMapped /
TestSelectComponent (skips power symbols — PWR_FLAG has no footprint) +
merged-image dispatch
- tests: presence suites extended (payload shape, ghost render pixel tests,
13/13) + new two-tab tests/web/cross-probe.spec.ts (passing vs real
partykit); eeschema pixel compares now target the #glcanvas-* GAL panel
(the whole-window #canvas compare flaked on the auto-dismissing version
infobar — also fixes the long-known presence-eeschema restore flake);
cross-app + presence-kicad vitest suites
Spec: docs/features/collab-presence/0006 (closed repo).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Lg5jwWuhFH5dL8hEcBDuP2
Library warm-up and the per-item fallback fat-load now run through a
shared asyncMap helper (counting semaphore, 8 in flight): the CDN
presync's hand-rolled 6-worker pool is replaced and the previously
fully-serial listItems+getItemBody fallback is parallelized (order
preserved).
The load-screen sync line stops flashing: progress reports on lib
COMPLETION only (no "current lib" — with 8 in flight there isn't one),
the label is a fixed prefix ("Syncing symbol libraries — ") and the
counter is space-padded to the total's width in font-mono +
whitespace-pre, so 2→3 digit growth doesn't shift the text. Same
padding for the fat-load overlay's "x / y libraries" line.
Verified in demo mode against the live R2 CDN: 7-8 overlapping
manifest/bundle requests, 222 cold symbol libs in a few seconds.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JgThWXtdvrYLK47EDFoGdq
The full root-cause chain of the empty/mangled badges:
- GAL text justify is painter-residue → PRESENCE_TEXT_OVERLAY pins TOP-LEFT.
- VIEW_OVERLAY::ViewDraw hard-sets EVERY overlay to GetMinDepth(), so the
shapes and text overlays always collided at one depth, where later-drawn
fragments lose (and bitmap glyphs are textured quads whose transparent
cells also write depth — punch-through produced cell-shaped holes instead).
Fix rides the new fork VIEW_OVERLAY::SetDepthOffset: shapes at min+1,
labels at min — 'rect first, text on top' now holds regardless of paint
order. Verified: chips contain crisp names (dark-on-light, white-on-dark)
for selection tags and cursor labels; presence suites 10/10.
- kicad pointer bump (fork 24c5854d5b).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CvqUd4QsJSGHN28aunJRTq
Two root causes of the wandering/invisible badge text:
- A plain VIEW_OVERLAY draws with whatever text justify the LAST painter left
in the GAL (CENTER is only the reset default) — anchoring was
nondeterministic. PRESENCE_OVERLAY (VIEW_OVERLAY subclass, replaces
MakeOverlay) pins TOP-LEFT justify before executing its commands; the label
math is written against that.
- The whole overlay draws at ONE depth and same-depth fragments drawn later
LOSE the depth test — a chip rect over its text erased the text. Draw the
text FIRST, the chip rect AFTER: the rect is rejected exactly on the glyph
pixels, punching the text through.
Verified live: chips contain their names on light (dark text) and dark
(white text) user colors, for both selection tags and cursor labels.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CvqUd4QsJSGHN28aunJRTq
- presence.ts: colors claimed by ARRIVAL ORDER — each client takes the lowest
free palette slot (no birthday-problem hash collisions; the best palette
colors go first; N ≤ palette-size editors always distinct). Coordination-
free: claims ride awareness; simultaneous-join collisions converge (lower
clientID keeps, the other re-claims), same-user tabs adopt one color,
claims are sticky across eeschema sheet rebinds (skeletons reuse them).
colorOf(userId) resolves live colors; comment pins/popovers follow it
(offline authors fall back to the hash). +2 unit tests (verified live:
alice=slot0, bob=slot1, roster/pins consistent).
- CommentLayer: single comment icon expanding into a horizontal bar — new
comment · list · show/hide all (eye empties the GAL pin set + DOM targets;
re-shown on new-comment/panel-jump). Pins are DRAGGABLE: live LWW anchor
writes while dragging (peers + the GAL dot follow), nearest-item re-snap on
drop, click-vs-drag by 4px threshold (verified live: drag synced to the
peer tab exactly). Shared setThreadAnchor + controller moveThread/
setPinsVisible/colorFor.
- collab_presence_style.h: chip text color by background luminance (dark on
light chips, white on dark — BitmapText draws with the stroke color).
- comments/presence-roster e2e updated for the bar + passing; collab units
56/56; shared 122/122; shared pointer bump.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CvqUd4QsJSGHN28aunJRTq
- collab_presence_style.h: GAL BitmapText CENTERS on its position (confirmed
in GAL::ResetTextAttributes — the mispositioned nameplates); labels/chips
now hand GAL the block center. New selection shape 5 'exact outline':
pcbnew hugs real geometry (footprint bounding hull, TransformShapeToPolygon
for the rest, padding inflates the polygon); eeschema falls back to rect.
- kicadCollabTestDemoSet (both TUs + merged): labeled demo groups — smallest
+ largest footprint and the two busiest nets' segments (symbols + wire
bundles on sch) — so the style preview covers the real range of shapes.
- PresenceTuner: Colors section rebuilt as explicit modes (per-user / fixed /
palette) with preset palettes (default, pastel, vivid, okabe-ito), buffered
hex editing + Apply (the old always-filtering textarea ate keystrokes), an
'overlay only' hint; demo injection consumes the varied demo set; 'exact
outline (pcb)' in the shape list.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CvqUd4QsJSGHN28aunJRTq
Parametrizes every visual knob of the presence overlay so we can pick the
shipped look live, then wire the winners into the defaults:
- collab_presence_style.h: shared STYLE struct + drawing (now used by BOTH
editor TUs — no more duplicated overlay code): selection shape (rect /
corner brackets / underline / rounded rect / filled-only), border width +
alpha, infill alpha, padding, corner radius; name tag show/size/chip-
background/inside-outside/top-bottom/start-end-center/offset; cursor shape
(cross / pointer / circle+dot), size/width/alpha + label knobs; fixed-color
and palette-by-name-hash overrides (try palettes without changing what
senders publish); pin radius/ring/alphas. Defaults == shipped look.
- kicadCollabSetStyle(json) live-patch export + kicadCollabTestListItems(n)
(real KIIDs for synthetic previews); merged dispatch; pins now carry the
author name so palette overrides recolor them consistently.
- PresenceTuner.tsx: floating dev panel (env-gated, tree-shaken otherwise) —
grouped sliders/selects, demo peers+pins injection for SOLO tuning,
localStorage persistence across reloads, Copy JSON export, reset.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CvqUd4QsJSGHN28aunJRTq
Hybrid pins: the wasm draws the dot (kicadCollabSetPins rides the presence
VIEW_OVERLAY, author color + white ring, drawn above selections; zero
kicad-fork changes), the DOM owns interaction —
- comments.ts: controller gluing the MIT kdoc_comments helpers to the editor:
anchor resolution per tool IU (pins track item moves via kdoc_items
observation), throttled pin snapshots, anchorAt nearest-item snap, jumpTo
via new kicadCollabSetViewport; rebinds per sheet like presence.
- CommentLayer.tsx: comment mode (click catcher + composer), pin hit targets
over the GAL dots, thread popover (reply/edit/delete own, resolve/reopen,
delete thread), panel with resolved filter + jump-to (popover centers when
the pin is off-screen). Resolved pins drop figma-style.
- WasmTool: controller lifecycle beside presence; live viewport feed;
window.__pcbjamComments test handle (threads persist in the room ydoc).
- e2e tests/web/comments.spec.ts: two-tab create → reply → resolve → panel
filter → delete, passing vs real partykit; presence suites + collab units
stay green; shared pointer bump (0004 model).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CvqUd4QsJSGHN28aunJRTq
Each remote selection rectangle now carries the selector's name above its
top-left corner (9px glyph, peer color) and the outline width goes 1.5→2.5px.
Verified live two-tab + presence-pcbnew.spec.ts 5/5.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CvqUd4QsJSGHN28aunJRTq
Zero kicad-fork changes — all in the embind layer:
- pcbnew_embind.cpp: presence section — canvas wx Bind() triggers (motion/
leave/up/key/wheel) + COLLAB_LISTENER piggyback → post-settle selection
check (dedupe) → onSelection; throttled cursor emit via VIEW::ToWorld;
viewport push/pull (px-per-IU via the GAL matrix — GetScale() is the zoom
and sized the first cut's overlay nm-small); kicadCollabSetRemote renders
peers' cursors (cross + name) and selection bbox outlines into one
per-user-colored VIEW_OVERLAY (CallAfter+COROUTINE), never touching local
selection; PresenceStart/GetSelection/TestSelectFirst/TestClearSelection.
- kicad_editor_embind.cpp: merged-image dispatch (pcb-only until 0003).
- presence-kicad.ts: bindKicadPresence — routes emits into awareness (0001)
and pushes trailing-throttled peer snapshots into the wasm; wired from
WasmTool.startPresence (pcbnew-gated). +5 unit tests.
- tests/kicad/presence-pcbnew.spec.ts: 5 e2e — programmatic + real box-select
emit, throttled cursor, remote render with no-leak + pixel restore,
viewport unit band. Existing pcbnew-collab/items-bridge suites stay green.
Verified live: two tabs over partykit — peer cursor cross + label + selection
outline visible on the other tab's canvas.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CvqUd4QsJSGHN28aunJRTq
- provider.ts: expose awareness on YjsProvider (partykit/hocuspocus surface
their own; BroadcastChannel gets a sibling-channel relay via awareness-bc.ts)
- presence.ts: createPresence — publish local PresenceState, subscribable
deduped peers() roster, setCursor/setSelection for P2, pagehide fast removal
- PresenceRoster.tsx + WasmTool wiring: facepile chip next to SourceChip;
eeschema rebinds per active sheet (sheet-manager ActiveSheet now carries the
room's provider); config presenceUser() = slug + colorForUser
- new dep y-protocols; unit tests (presence.test.ts) + e2e
tests/web/presence-roster.spec.ts (two tabs, verified vs real partykit)
- bump pcbjam-shared: presence-wire schema (MIT)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CvqUd4QsJSGHN28aunJRTq
The merged kicad_editor bundle made the chooser's footprint side reachable
from a schematic session; this wires up the data:
- boot.ts/constants.ts: every kicad_editor frame seeds BOTH sym-lib-table and
fp-lib-table (+ placeholder files; a created user lib joins both lists) —
the eeschema frame used to write fp-lib-table empty, leaving the selector
dead. TOOL_LIB_KIND remains only the presync/primary-kind lever.
- publish-libs.ts + kicad-pretty.ts: publish fp-index.json per tag —
[name, uniquePadCount] per footprint (countUniquePads mirrors KiCad's
GetUniquePadCount(DO_NOT_INCLUDE_NPTH)); index-only top-up mode for
already-published immutable tags.
- source.ts/cdn-source.ts: new bridge op "index" (source-global, dispatched
before the lib-id parse) + LibsSource.getFpIndex; the CDN source fetches
<tag>/fp-index.json once (404 ⇒ null ⇒ C++ default-only fallback).
- dev-demo.mjs: --libs-local serves a local publish-libs layout same-origin
at /libs-cdn (mirrors --models-local).
- tests/web/eeschema-fp-selector.spec.ts: e2e — chooser opens in --frame=sch,
selector fills from ONE index crossing, clicking a row per-item-gets the
body and the cross-face GAL preview renders; adaptive for index-less
sources (asserts crash-free default-only selector).
Submodule bumps: kicad (index-backed filterFootprints + preview AsyncLoad
fix + modal-pump crash guard), wxwidgets (modal pump logs e.stack).
Doc: pcbjam-private docs/features/libs/0014-eeschema-footprint-selector.md
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SCWVaCRM9T847PdPwYYajX
The closed API now membership-gates /parties/sync-room/* per request via
the session cookie and no longer puts a bearer token in sync-stack layer
descriptors. Pass a credentials:include fetch into SyncStack so live-layer
HTTP ops (manifest/bodies/put) carry the cookie; the realtime WebSocket
already gets it automatically (same-site handshake).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014dasZuqo6FStgT3rkC85im
The example backend's reflected-origin + allow-credentials pairing is safe
solely because this server holds no ambient credentials by design (default
origin is the explicit :3048; '*' is an operator opt-in). Make that a stated
invariant so adding any auth here forces revisiting the wildcard mode.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016TxciQ5VkNYUZesgMSEMUr
Backends with real auth (session cookies) need credentials:'include' on the
editor's cross-origin fetches — ts-rest clients, project file bytes/upload,
drift keepalive, lib item GET/PUT, sync-stack resolve. Cookie-less setups are
unaffected (the thin identity headers still ride along and same-site Lax
cookies simply don't exist). CDN/static-gallery fetches stay credential-less
(wildcard CORS rejects credentialed requests). The example backend's CORS now
sends allow-credentials — origin reflection was already exact, so this only
lets browsers accept those responses.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016TxciQ5VkNYUZesgMSEMUr
waitForRenderStable now runs two phases (still entirely in-page, still just the
canvas hash — no CDP screenshots):
1. Fast per-rAF convergence — hash every animation frame until stableFrames are
identical (~48ms). Catches high-frequency motion, cheap.
2. Wide confirmation — then re-hash confirmFrames times, each `interval` ms apart
(default 2 × 250ms ≈ 500ms). A slow async repaint (e.g. a file list arriving
after an asyncify readdir) that a few 16ms frames would sail past now forces a
reset back to phase 1.
This closes the hole where three ~16ms samples could declare "settled" ~200ms
before a delayed paint lands. Genuinely-animating states still never converge and
are captured at the timeout, as before.
Cost (local, 4 workers): e2e 62s->103s, kicad 155s->212s — pure waiting, no extra
CPU. Both suites green (e2e 340, kicad 92). Tunable via confirmFrames/interval.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BVX1pHMvRPYHdp6ZfEawrk
The determinism conversion silently dropped/renamed screenshots vs the committed
baselines, so CI flagged ~35 baselines as "removed". Restore them:
- pl_editor-*: re-add the "pl_editor-" prefix the conversion dropped (5).
- Buffer screenshots that lost their path: (eeschema draw-wires/loaded-css,
pcbnew draw-lines/loaded-css, pcbnew-move, dark-mode, eeschema-crosshair,
aui-resize) re-add path:'test-results/<name>' so they SAVE and still feed the
in-test pixel-diff analysis (17).
- Dropped animating/loading captures re-added via stableShot / raw screenshot:
timer-03/05/06, dialogs-timer-running/started, wxgrid-02/03, 01-loading (7).
Delete 5 obsolete wizard baselines (eeschema/gerbview/pl_editor/symbol_editor/
generic -wizard-00-initial): every app is now seeded to bypass the first-run
setup wizard (that was the point of removing the wizard for/if loop), so those
states are no longer produced — pl_editor.spec asserts "no wizard appears"
instead.
Net: every committed baseline has a producer again (0 "removed"); determinism
guard green; affected specs pass (e2e 40, kicad 12) and emit the restored PNGs.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BVX1pHMvRPYHdp6ZfEawrk
Make the Playwright e2e + kicad suites deterministic so screenshot flake stops
tracing to timing races.
- Blind page.waitForTimeout -> condition waits (expect.poll, web-first
assertions, waitUntil) + readiness helpers (waitForWxApp, waitForCanvasApp).
Remaining sleeps are documented interaction dwells (annotated).
- Defensive "if element exists" branches -> loud asserts; label-fallback chains
-> normalized clickMenuItemByText. First-run wizard for/if loops removed by
seeding calculator/gerbview/pcbnew HTMLs.
- Screenshots: new stableShot(page, name) settles the render in-page (canvas
hash over rAF) then writes a raw PNG to test-results/ for the existing offline
gate (tools/screenshots vs baseline-screenshots). Replaces toHaveScreenshot,
which did inline compare + its own baselines and had decoupled the specs from
the real gate. scale:'css' pinned.
- retries: 0 in both configs.
- Guard: tests/tools/lint-determinism.ts (npm run lint:determinism) bans blind
sleeps / toHaveScreenshot / inline retries / swallowed catches in specs;
documented exceptions carry a marker. Rules in tests/TESTING.md.
Assertions, coverage, and renders unchanged (semantic-equivalence reviewed;
captures pixel-identical modulo inherent timer/timestamp/3d-raytrace variance).
Both suites green at retries:0 (e2e 340, kicad 92); ~35-61% faster.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BVX1pHMvRPYHdp6ZfEawrk
Miss 08A: the binding stores wire-carried (lib_symbols …) definitions in
kdoc_libsymbols and prefixes them on apply wires — a joiner that never saw a
symbol adopts it WITH its definition (new e2e ysync-libsymbols.spec.ts).
Miss 08B: registerSaveHook gains onSavedText; WasmTool routes saved-file text
to syncLayoutToY (per sheet room via the manager's syncLayoutFromSave, or the
single-room doc) so title block / paper / setup edits converge instead of
drifting. Opt 12 (TS half): zod off the observer hot path (yToItemUnchecked),
children index built once per conversion. Opt 13: seed()'s adopt diffs the
editor snapshot against the doc view and applies only the doc-authoritative
difference — clean rebinds apply nothing, the adopt undo entry shrinks to the
real changed set. Opt 14 deliberately deferred (doc 18). All TS-side; no wasm
rebuild (the C++ blob/findLib sides already carried definitions).
Verified: shared 107, standalone 79 (+2 known pre-existing wasm-assets),
ysync e2e 21/21 chromium, collab regression 21/3-skip firefox.
Bumps: web/pcbjam-shared (lib_symbols channel + syncLayoutToY + opts).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JgThWXtdvrYLK47EDFoGdq
All seven ysync-review bugs fixed and verified; every repro's expected-fail
marker removed (they now run as regression tests). Also fixes two bugs found
while verifying (doc 17 F5/F6): file-seeded Y bodies are re-upserted in the
editor's serialization (doc-16 F4 was an artifact of bug 01), and the
0008-era "asyncify-fragile envelope parse" was really wrapInBoardEnvelope
emitting display layer names — canonical LSET::Name() fixes track/via/zone
v2 applies; makeFromBlob now logs parse errors instead of swallowing them.
Verified: shared 98, standalone collab 37, ysync e2e 20/20 (chromium),
collab regression set 21 passed / 3 pre-existing skips (firefox).
Bumps: kicad (board_commit child-removal listener notification),
web/pcbjam-shared (slot prune + arbitrated seed).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JgThWXtdvrYLK47EDFoGdq
The 2026-07-02 sync review (docs/features/ysync-review, ysync-review branch)
found 7 bugs and that the two-tab e2e only exercised the DEAD legacy scalar
wire. This lands plan doc 15 in full; results + empirical findings in doc 16.
- tests/collab/browser-entry-v2.ts (+build.mjs): the PRODUCTION v2 stack
bundled for e2e (connectKicadDoc + attachKicadCollab, kdoc_* keys), with
in-page renderActiveDoc/singleSeedRender/driftReport helpers and yjs forced
to ONE copy (the two web pnpm workspaces otherwise bundle two
instanceof-incompatible instances).
- tests/kicad/ysync-two-tab.spec.ts: pl_editor green baseline (A↔B edits,
ITEM-level drift silence) + divergent-uuid adopt; bug-01 pcb/ee fresh-room
repros (Chromium-only: two kicad_editor tabs exceed Firefox's per-process
wasm budget); bug-06 concurrent-seed race; bug-03 Y-half.
- tests/kicad/ysync-repros-{pcbnew,eeschema}.spec.ts: bugs 02/03/05 + the
bug-04 matrix (anchor-centred fp rotation, pad resize, endpoint drag,
symbol rotation, Value-field edit), each with green landed-preconditions;
the "local move emits" controls double as headless-emit probes — GREEN on
both tools, so every emit-dependent repro is a live test.fail.
- wasm/bindings: 7 local-edit test hooks via real commits
(CallAfter+COROUTINE) — TestRemoveItem/TestRotateItem (both tools,
dispatched in the merged image), TestSetPadSize/TestMoveEndpoint (pcbnew),
TestSetFieldText (eeschema).
- web/standalone ysync-repros.test.ts: bug-01 units (C++-faithful fake gating
emit on ensureBridge) + bug-07a/b (stale DOWN hook, real sheet-manager gap).
- web/pcbjam-shared bump: bug-03/06 unit repros.
Convention: every repro asserts the CORRECT behavior and is expected-fail
(test.fail/it.fails) naming its bug doc; a fix flips it to "unexpected pass",
forcing marker removal — the repro becomes the regression test. Every
expected failure verified (JSON reporter) to fail at its documented assert.
Suite state: 39 passed / 0 failed / 0 flaky / 5 skipped (2 firefox guards,
2 pre-existing legacy two-tab skips, 1 pre-existing roundtrip fixme).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DPfrVhfYgPPgtawjSssZfn
New doc 16: the full repro-suite map (per-bug unit/e2e paths with verified
failure sites), the phase-C probe outcome, and four findings only the RUNNING
system revealed:
- F1: bug 03's sending half emits NOTHING — a child-only delete commit never
triggers a flush at all (worse than the doc's predicted bare removed-wire);
the GetWidth-assert tracer evidence and the fix implication.
- F2: Firefox cannot host two kicad_editor tabs in one context (per-process
wasm budget) — bug-01 two-tab repros are Chromium-only.
- F3: headless emit WORKS on both pcbnew and eeschema — the legacy two-tab
skip rationale and items-bridge localEdit omissions are stale.
- F4: drift-detect is strictly ITEM-silent on the green path (no writer-
formatting false positives).
Cross-updates: 00 index + verdict note; 01/04/05/06/07 Verification sections
gain their repro paths; 02 upgraded to runtime-CONFIRMED; 03 gains the F1
empirical correction; 11 (no v2 e2e coverage) CLOSED with a point-by-point
status update — only the legacy retirement remains.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DPfrVhfYgPPgtawjSssZfn
Now that demo.pcbjam.com is live, the landing page leads with it instead
of "coming soon":
- Hero: "Open the live demo" is the primary CTA; the coming-soon panel is
replaced by a DemoLaunchCard (screenshot + play overlay, new tab — the
editor needs COOP/COEP so it can never be iframed here); waitlist demoted
to a ghost button under a "Want early access to multiplayer?" lead-in.
- Removed the Gerber demo and positioning-map sections, then restored the
positioning map per review — recolored for light theme and without the
multiplayer badge/legend ("Everyone picks one. We do both.").
- Light theme: flipped the :root ink tokens, darkened accent/signal for
contrast on white; fixed two dark-assumption bugs it exposed (header CTA
text losing to .nav>a specificity, DemoVideo play scrim derived from --bg).
- Copy: positioned around what's live (real KiCad in a tab) with
multiplayer explicitly "coming — join the waitlist"; cut untrue claims
(GitHub import, self-host, offline PWA -> roadmap); cut the jam metaphor
outside the product name; TrustBar fake-stat grid -> plain claims strip;
footer tagline + fallback meta description aligned.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01REwWZGkAz1AZUJw8yqxkCN
Trim the waitlist microcopy on the standalone demo and the marketing site's
CTA/hero forms to just the value prop, removing the "no spam, unsubscribe
anytime" reassurance line.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01REwWZGkAz1AZUJw8yqxkCN
The three 3D-viewer shots changed because the specs now screenshot the
COMPLETED render (pixel-gated waitForThreeDRender) instead of a fixed-sleep
mid-render frame. The rest is the accumulated unpromoted backlog: eeschema/
symbol-editor drift, and first-ever baselines for the models-spec renders
(tail re-enabled in 561d050), frame-runtime, occ-export/probe and
xface-preferences from earlier merges.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The 3d-webgl merge (kicad eb13ff3bdc: the viewer now defaults to the real
OpenGL renderer via wasm/gl1, and occ-split moves STEP parsing into the
occ_service worker) made the raytracer-era orchestration on this branch moot —
main's chromium-ci phase is green at 15-way parallelism (28666407570 /
28698861536). Drop what no longer earns its complexity, keep the diagnostics,
fix main's live flake, and make the deadlock spec test what it was written for.
- REVERT the chromium-ci-3d serial project, the two-phase test:kicad:ci, the
SwiftShader GPU-process flags, and the resize-drag/models skips: config and
package.json are byte-for-byte back to main's shape. The raytracer contention
they guarded is no longer on the CI path.
- FIX main's live flake: run 28698861536 is green only via retry
(3d-viewer.spec:26 flaky) and 28666407570's deadlock red sampled an ALL-ZERO
pixel signature — the viewer's first frame lags the canvas's creation on
software WebGL under parallel load, and sampling too early reads an all-black
backbuffer. New waitForThreeDRender() gates render assertions on actual
pixels (1s-interval full-frame CPU reads) instead of fixed sleeps, used by
3d-viewer.spec:26 and the models render tail.
- KEEP the storm-proofed samplers (one full-frame getImageData on a
willReadFrequently canvas replacing 256 per-pixel GPU round-trips per sample
— the "GPU stall due to ReadPixels" trigger) and the logThreeDDiag
instrumentation: engine-independent, and they de-risk every remaining
software-GL pixel read.
- models spec: bridge assertions stay front-loaded (the protocol regression
signal is independent of the render); the occ_service parse verdict is now
POLLED — it lands async relative to the bridge ensures, so asserting it
immediately raced the worker; the render tail runs again everywhere. (The
pre-webgl raytracer+models renderer-death documented in a17f3be does not
affect the OpenGL default path — the raytracer-toggle+models combination
remains untested product surface, tracked outside this branch.)
- deadlock spec: the deadlock it guards is raytracer-specific and the viewer
now defaults to OpenGL — on the GL engine it either passes vacuously (fast
renders make every liveness assertion trivial, 28698861536) or fails on the
black first frame (28666407570). It now flips the engine via the "Use
raytracing" toolbar toggle (loud assert if the toggle moved) and
cross-checks engagement by requiring the canvas pixels to CHANGE after the
flip with no input in between (the raytraced frame is lit differently; a GL
re-render reproduces identical pixels; heap growth is unusable — mimalloc
satisfies the raytracer from freed arena pages). That guard immediately
caught a REAL defect: on the webgl-era wasm build the toggle is INERT (the
click lands and "Reload time" updates, but the canvas never changes —
suspects: DoRePaint's silent catch(runtime_error) freezing the canvas after
a raytracer Redraw throw, or ToggleRaytracing writing m_boardAdapter.m_Cfg
while RenderEngineChanged() reads GetAppSettings<…>(), possibly different
instances in the merged bundle). The spec is therefore test.skip-annotated
as a KNOWN ISSUE with the full engine-force machinery in place — unskipping
it self-validates the product fix. The CI-skip also stays (raytracer
liveness needs real-GPU pacing; the Worker-boot deadlock mechanism is
covered on CI by the standalone wx harnesses).
- 180s viewer-open waits kept as pure CI headroom (never slow a passing run).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The headed Mesa-llvmpipe switch (ce44636) was built on a misdiagnosis. With
--use-gl=angle --use-angle=gl under Xvfb, Chromium gets NO WebGL AT ALL on the
GPU-less CI VM: runs 28652367347 + 28664038296 both log
`glcanvas count before opening 3D viewer: 0` (even the pcbnew board's own GAL
canvas fails to create) and every failure is the FIRST viewer-open wait timing
out. "Timeouts instead of crashes" meant no GL context existed to crash, not
that llvmpipe was stable-but-slow. The Mesa/GLX path works for the Firefox
project, but not for Chromium on this image.
Headless SwiftShader is the environment where GL provably works — the
serialized run 28649537489 opened the viewer and rendered/sampled 56 colours
(viewer:26/130/209 all passed single-tab). Its remaining failures split into
two families, each addressed on its own evidence:
1) SwiftShader GPU-process kills (deadlock spec's lost canvas/black frame):
- chromium-ci-3d back to headless + --enable-unsafe-swiftshader, plus
--disable-gpu-watchdog and --disable-gpu-process-crash-limit: heavy churn
stalls the software-GL GPU process until the watchdog shoots it
(CONTEXT_LOST -> removed GL canvas -> "Target crashed"); with the watchdog
off a slow op completes, and without the crash limit a lost GPU process
keeps restarting for later tests. Same args locally (no-ops on a healthy
real GPU).
- Storm-proof every canvas sampler (3d-viewer, models, deadlock specs): one
full-frame getImageData on a willReadFrequently (CPU-backed) 2D canvas +
JS grid sampling, replacing 256 per-pixel getImageData GPU round-trips per
sample — the "GPU stall due to ReadPixels" trigger, worst in the deadlock
spec's 1.5s-interval settle polling. Sampled pixels and sig order are
unchanged.
- edge-resize spec: keep the open + 5-resize-handles assertions on CI, then
test.skip the drag — each resize step re-raytraces synchronously on the
wasm main thread and the 12-step drag blocked mouse.move past the 240s
budget even single-tab. Real-GPU-paced; still runs locally (deadlock
spec's CI-skip from d9900fe unchanged, same rationale).
2) KNOWN PRODUCT BUG uncovered in the models spec (NOT SwiftShader, NOT CI):
raytracing a scene WITH component models kills the Chromium renderer
process outright ~6s after scene build — silent process death with a FLAT
wasm heap (531-637 MB watched at 2s intervals; not OOM), no console error,
no wasm abort, no crash report. Deterministic on a real GPU (5/5),
fixture-independent (700 KB USB-C and 61 KB USON-8 STEP crash identically);
the same death hits CI SwiftShader at ~66s. Board-only raytraces complete
and stay alive, so the bug is specific to the model path. The spec's one
historical green ended at raytrace-age ~6s — inside the death window by
luck; every CI run of it ever died. Fix here: front-load the bridge
assertions (ensure requests fire during scene BUILD, before the raytrace),
so protocol regressions still fail the suite everywhere, and test.skip the
render tail unconditionally with the bug documented at the skip site. The
raytrace->blit pipeline stays gated on CI by 3d-viewer.spec's board-only
render. Bug tracked for a dedicated fix session.
Correct the now-wrong llvmpipe comments (config, open-wait rationale, deadlock
skip reason). Expected CI matrix: viewer open/render/stacking/titlebar +
models-bridge on SwiftShader serial; edge-resize drag + camera-move deadlock
on real GPUs locally; the Worker-boot deadlock mechanism stays covered on CI
by the standalone wx harnesses.
Locally validated on a real GPU (--project=chromium): 5 passed + 1 skipped
(the documented models render tail).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Follow-up to the headed-llvmpipe switch. That run (28652367347) stopped CRASHING —
every failure became a plain timeout, confirming llvmpipe gives a stable WebGL context
(no more CONTEXT_LOST / "Target crashed") but software raytracing is just slow:
- The first board raytrace takes ~60s on llvmpipe, right at openThreeDViewer's 60s
wait → raise it to 180s (shared helper + the models spec's inline copy). Real GPU
returns in ~2s, so it is only CI headroom.
- The deadlock spec asserts the wasm main thread stays responsive within ~15s DURING a
raytrace — a premise only true on a fast (real) GPU. Under slow software rendering a
legitimately-slow raytrace is indistinguishable from a deadlock, so skip it on CI
(test.skip on process.env.CI). The Worker-boot deadlock MECHANISM stays covered on CI
by the standalone coroutine-pthread-ondemand / raytrace-threads wx harnesses; the spec
still runs locally on a real GPU.
- Gate the headed + llvmpipe launch flags on CI so chromium-ci-3d behaves like a normal
bundled-Chromium project locally.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Serializing the 3D specs (prior commit) proved the crashes are NOT concurrency: with
--workers=1 (one tab at a time) the same three heavy-interaction specs still failed
(run 28649537489). The [DIAG] instrumentation + trace console pinned the real cause —
under HEADLESS SwiftShader on the GPU-less CI VM, the CPU raytracer's WebGL blit plus
the tests' drawImage(glCanvas) pixel reads trigger "GPU stall due to ReadPixels", and
Chromium's GPU watchdog then resets the context (CONTEXT_LOST_WEBGL) or crashes the
renderer ("Target crashed") — even single-tab.
Run chromium-ci-3d HEADED under the CI's Xvfb with Mesa llvmpipe (--use-gl=angle
--use-angle=gl) — the same software-WebGL stack the Firefox CI project already relies
on — and disable the GPU watchdog / crash-limit so a slow ReadPixels completes instead
of being killed. --enable-unsafe-swiftshader kept only as a last-ditch fallback. The
serial --workers=1 pass + phase-1 worker cap stay (cheap, and they removed concurrency
as a variable).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
