Prompt Markup strips it before the encoder; Apply Prompt Markup puts weights
and timed windows on attention and blends on the conditioning. Where the words
sit is the H3 module's (text_tokenizer, prompt_rows, video_time_rows). Wired
into both H3 presets.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Both replace what a block allocates, which crashes aimdo's compiler (seen on a
rental with v4's shadow negative). Attached to the model, so any sampler.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
H3's latent upscaler loads through the H3 module (ComfyUI's loader has no
branch for it); an upscale resizes H3's anchor pin so pass 2 stays pinned.
FunPack Sampler takes an optional wired schedule, run exactly as written.
New preset: H3 Reference to Video + second pass.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The H3 module offers anchor_pin, so ALG blurs the pin's latent without
learning H3's payload layout.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
core.patching.strip now also removes our object patches, so a feature switched
off does not keep its patch on a model that goes round again.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
v4 read the prompt length from transformer_options['c_crossattn'], which
ComfyUI never sets, so every block fell back to stock. It also stood down on
any i2v scene (per-token mod rows). Both fixed; tested on a real tiny H3.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Block repeat's video-only mode now really keeps audio single-pass: H3 blocks
write in place, so v4 restored audio from an already-repeated stream.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
[[project_v5_ui_requirements]]: "any action in the UI can be assigned to a
slot... full wheel or half wheel... half wheel positioned left or right...
both draggable... default: middle mouse button, anywhere."
app/legacy/actions.js: ported the composer tree's tiny action registry
(app/shell/actions.js, an ES module, unreferenced from boot) as a classic
global window.FunPackActions -- offerAction({id,label,icon,run}), offered(),
run(id). Whatever announces itself is what the wheel (or anything later --
a command palette, a keymap) can show.
app/legacy/wheel.js: middle-mouse (anywhere) opens a radial menu of
whatever's currently on it. Full wheel spaces items evenly around 360°;
half wheel spaces them over a 180° arc on the chosen side, with drag
locked to the vertical axis only (full wheel drags both axes) per the
spec's "half wheel moves up/down, full wheel up/down and left/right" line.
Dragging repositions the CURRENTLY OPEN wheel only, not persisted -- read
"draggable" as "don't get stuck somewhere awkward this time" rather than a
remembered dock position, since the spec doesn't say which and this is the
simpler reading. Customize dialog (reuses pinned_buttons.js's existing
.pin-modal/.pin-slot CSS) toggles shape, side, and which offered actions
are on the wheel. pinned_buttons.js is NOT removed -- the spec says this
supersedes it, but retiring a working control felt like a separate
decision from adding a new one; both coexist for now, confirmed to not
conflict (separate storage keys, no shared DOM, never both open at once).
app/legacy/wheel_actions_seed.js: seeds 9 starter actions (Generate,
Generate selected, Render, Auto Montage, Undo, Redo, Models & Pipeline,
Engine settings, Sampler quick-access) so the wheel isn't empty on first
use.
Two rounds of extensive_testing, plus a bug I found myself in live
testing before either round: the half wheel's fixed radius, sized for a
full wheel's 360° spread, packed the same item count into half the
angular room and produced heavily overlapping items -- fixed by growing
the radius only as far as needed to keep adjacent items from touching,
confirmed visually before/after.
Round 1: no functional break found in outside-click-vs-item-click
ordering, listener leaks on repeated opens, or coexistence with
pinned_buttons.js.
Round 2: found the real defect -- enabledIds()'s first-call snapshot
("default to everything offered so far") could run in the gap between
wheel.js executing and wheel_actions_seed.js finishing (classic <script>
tags block sequentially, and this window widens on a slow/cold-cache
load), permanently persisting an empty or partial action list with no
error, no indication -- the user would just experience "the wheel doesn't
work" and have no way to know why. Fixed by deferring the snapshot to the
window `load` event (guaranteed after every script has run); before that,
enabledIds() shows whatever's offered so far live without writing it
down as the permanent default. Also cleaned up a dead if/else in the
half-wheel angle math flagged as a "looks like a bug that got half-fixed"
smell.
Live-verified after both fixes: real physical drag (manual mousedown/
mousemove/mouseup sequences, not just JS calls) confirms half-wheel drag
is genuinely locked to the y-axis (x stays exactly put) while full-wheel
drag moves both, and neither drag opens Customize -- only a press that
never crosses the drag threshold does. npm test: 691 pass, 1 pre-existing
unrelated flake (shell/tests/projects.test.js:285).
No test file exists yet for actions.js/wheel.js/wheel_actions_seed.js --
a real, named test gap, not fixed here. Real hardware middle-click and
cross-browser autoscroll-blocking are also unverified (browser automation
has no native middle-click) -- declared, not chased further without
physical hardware.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
[[project_v5_ui_requirements]]: "steps, scheduler and sampler are reached
for constantly, and in v4 they are several clicks deep... a thin hovering
window above the timeline, which keeps them visible without spending
timeline height." Nothing in v5 built this yet -- the only existing path
was Settings > Engine > Sampling, a modal several clicks deep.
app/legacy/sampler_quickbar.js: a toggle button in the timeline header
(next to Assets/Preview/Properties/Composer) opens a fixed-position panel
above the timeline showing only category="sampling" module settings.
Hidden until PipelineState confirms at least one installed module has a
sampling setting (the "absent, not greyed" convention this codebase
already uses for a module/role that doesn't apply).
The quickbar needed to read/write the SAME module-values tree
engine_settings.js already edited. That tree gets serialized into one
opaque JSON blob and written into a pipeline slot by the server
(core/graph.py's place()) with no server-side merge -- a second, naive
editor sending a partial tree would silently erase every other module's
settings. To avoid two independent caches of the same server data
diverging (the exact bug class pipeline_state.js's own header comment
already describes for `slots`, from an earlier fix), extracted the
shared pieces instead of duplicating them:
- PipelineState gained currentValues() (always-fresh: defaults +
valuesAlreadyPlaced()) and setModuleValue(moduleId, name, value)
(patches one field, sends the whole tree). Single source of truth,
used by both editors.
- app/legacy/settings_field.js: the field-widget builder (checkbox/
slider/enum/color/text, the `when`-clause visibility gate) extracted
out of engine_settings.js verbatim, parameterized on a values tree and
an onChange callback instead of a private cache -- one place renders a
setting, one place saves one, for every editor of this tree from now on.
- engine_settings.js refactored to use both instead of its own private
values/seeded/setValue/ensureSeeded/commit/controlFor.
Two rounds of extensive_testing. Round 1 found a real, live-reproduced
defect: setModuleValue's "always fresh, no cache" design had its own
race -- two calls fired before either's round-trip resolved could
silently lose the earlier edit, because the second call's "read the
whole tree" snapshot didn't yet include the first call's in-flight
write. Fixed with a `pendingValues` overlay in pipeline_state.js,
written synchronously before every read, reconciled (per-key, not
wholesale) once valuesAlreadyPlaced() actually confirms it server-side.
Round 2 re-verified the original race closes -- checked against
valuesAlreadyPlaced() specifically (decoded straight from the server's
own response, not the optimistic overlay) for both a two-different-
modules race and a three-calls-same-field race -- and found one more
thing, currently inert: setModuleValue's returned promise doesn't mean
"this specific edit is confirmed" for every concurrent caller, only for
whichever one owns save()'s queue at that moment. Documented in place
(both real callers only re-render from shared state, which is already
correct regardless of promise timing) rather than reworking save()'s
shared queue for a guarantee nothing currently needs.
Live-verified: button appears only when a sampling setting exists;
opens positioned above the timeline; editing in the quickbar changes
only the target field (confirmed via full before/after tree diff);
same change visible in the Engine Settings modal and vice versa; real
outside-click closes it; 5 open/close cycles leave no DOM leak.
npm test: 691 pass, 1 pre-existing unrelated flake
(shell/tests/projects.test.js:285). No test file exists yet for
pipeline_state.js/engine_settings.js/settings_field.js/
sampler_quickbar.js -- a real, named test gap, not fixed here.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
v4 always presented an overlay lane whether or not anything used it.
The user's own v5 UI requirements explicitly reject this: "the
timeline starts with no overlay tracks. One appears when the user
adds it explicitly." store.js's normalizeOverlayLanes() force-created
a default "Overlay 1" lane the moment overlay_lanes was empty -- on
every project load AND every timeline render -- so a fresh, empty
project always showed one anyway.
Fixed at the two real sources:
- normalizeOverlayLanes() now only auto-creates a lane to house
ORPHANED tracks (old-format projects, or a lane removed out from
under its tracks) -- never just because nothing asked for one yet.
- defaultOverlayLaneId() (reached only from addImageOverlay/
addTextOverlay when the user is explicitly adding their first
overlay) now creates the first lane itself at that exact moment --
the "user adds it explicitly" trigger the spec asks for.
Traced every overlay_lanes/overlay_tracks consumer across store.js,
timeline.js, inspector.js, view_bus.js, tour.js: none assume >=1 lane
always exists (removeOverlayLane's own guard already handles "never
below 1 once lanes exist"; drag-drop onto a lane row is naturally
unreachable before one exists, matching intended behavior).
Live-verified: a fresh project (window.Store.newProject) renders zero
overlay rows; addTextOverlay("Hello") creates exactly one lane, wires
the new track to it, and the row appears with a live text-overlay
preview on the canvas. JS suite: 691 pass, 1 pre-existing unrelated
flake (shell/tests/projects.test.js:285).
One round of extensive_testing (adversarial review) came back clean.
Declared, not fixed: a project saved under the old behavior keeps its
already-persisted empty "Overlay 1" lane -- a forward-only fix, no
migration was asked for and the lane is now an ordinary, removable one.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Round 9 emptied demoProject()'s scene_renders to stop a real unmocked
network fetch against the live backend. Round 10's adversarial review
caught the sibling: tour-s1 still had rating: "good" in the demo data,
but the rating badge (timeline.js's sceneRatingRaw -> genUnitHasRender)
is gated on hasRender(), which now reads false for every scene since
scene_renders is empty -- so the badge silently never renders on the
scene the tour spotlights, self-contradicting the stored rating.
Cleared the rating to "" to match the no-render state honestly. Traced
every step/FAQ reference to tour-s1 and ratings: the FAQ answer is
generic ("after a render, rate a generative scene"), no step claims
tour-s1 specifically shows a badge, so this is a pure consistency fix
with no user-visible regression.
Live-verified: fresh tour boot, tour-s1.rating reads "" via
window.Store.get(). JS suite: 691 pass, 1 pre-existing unrelated flake
(shell/tests/projects.test.js:285).
Round 10 of the tour extensive_testing chain otherwise came back clean:
no dangling onCancel overlay (PipelineSetup's cancel-button variant is
unreachable in tour mode -- onboarding.js explicitly skips opening
during the tour), no crash from the empty scene_renders object across
every reader (store/timeline/inspector/player/montage_dialog/view_bus/
history), and the patchApi()/patchStore() mock surfaces still match
their consumers' expected shapes on a fresh full read.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Round 9's module-definition sweep came back clean (the four standalone
modules found in rounds 5/7/8 are all still correctly covered, and no
fifth exists) -- but it caught two real bugs my own round-8 live
verification missed by not waiting long enough:
- Both restartComfy() implementations call FunPackRestart.waitForReload()
after API.restart() resolves. That function polls API.health() starting
3.5s later and calls location.reload() once it reports ok -- and
API.health is ALREADY mocked to ok:true for the Settings About panel's
own unrelated reasons, so the very first poll after a restart click
reloads the real page mid-tour, silently resetting all tour progress.
Nothing actually restarted, so there's nothing to wait for -- fixed by
short-circuiting waitForReload() in tour mode instead of polling toward
a reload.
- demoProject() pointed a fake scene_renders entry at a real-shaped
ComfyUI /view URL for a file that doesn't exist. Since bootstrap()
notifies immediately on tour start, player.js's render() (a plain Store
subscriber, never blocked or mocked) picks it up and issues a real GET,
then a real HEAD retry, against the live backend -- zero clicks, always
404s, but still a genuine unmocked reach into the real install. Fixed at
the source instead of patching player.js's internals: an always-
unrendered demo scene is an honest trade-off the tour's own FAQ text
already accounts for.
Verified live: clicked both restartComfy implementations, waited the full
5s past the 3.5s poll delay this time (my round-8 verification hadn't
waited long enough to catch this) -- confirmed no reload, still on
?mode=tour; confirmed zero /view?filename= requests; full tour sequence
still completes cleanly.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Round 8's module-definition sweep found a fourth standalone module (after
GenerateBridge, PipelineState): window.FunPackGit (git_actions.js) and an
independent window.FunPackMaintenance.restartComfy (menubar.js) both call
window.MovieEditorAPI.restart() directly -- a REAL server restart, killing
any generation in progress -- one click from the same .settings-win
sidebar rounds 5-7 already proved reachable (Custom Nodes' inline
"Restart ComfyUI" button, and the built-in "Updates & ComfyUI" section,
which also fires a real GET /api/git/status the instant it mounts, zero
clicks needed -- same shape as pipeline_state.js's ensureLoaded()). That
same section's "Switch branch"/"Update FunPack" reach real git
checkout/pull. Worse than any prior finding: not a stray read, not even
round 7's directory-delete, but restarting the live production server and
mutating the FunPack install's own git state.
While fixing this, found two more real, unmocked, irreversible operations
in the same file (menubar.js's FunPackMaintenance, not caught by round 8's
own report): deleteRefinementKey and clearGlobalTaste, both described by
their own confirm dialogs as "This cannot be undone." Currently dormant --
v5's own api.js has these as stubs ("not built in v5 yet") -- same shape
as saveClipToMediaBin/renameMedia, closed pre-emptively rather than
waiting for a future backend to wire them up for real.
Fixed by mocking all ten functions in patchApi(): restart, gitStatus (also
cleanly short-circuits git_actions.js's own _ensureStatus() gate via its
existing alert path), gitUpdate/gitCheckout/gitRollback, and the five
refinement-key/taste functions.
Verified live in a genuinely fresh tab: called both restartComfy
implementations directly (auto-accepting their confirm dialogs) plus
refresh/update/switchBranch/deleteRefinementKey/clearGlobalTaste -- zero
/api/git/* requests in the network log, zero errors, page stayed on
?mode=tour rather than reloading, full tour sequence still completes
cleanly.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Round 7 finished round 6's admittedly-incomplete api.js-vs-patchApi() sweep
and found two more real, live gaps -- both on the same reachable surface
(the models-modal step's sidebar) rounds 5-6 already proved live:
- pipeline_state.js (window.PipelineState) is a standalone module, like
GenerateBridge -- talks to MovieEditorAPI directly, never through
window.Store, invisible to every Store-shaped sweep. ensureLoaded() fires
real, unmocked API.modules()+API.pipeline() the INSTANT the models-modal
step's panel mounts, zero clicks needed. Any ordinary widget edit on that
same screen then calls API.editPipeline(), a real POST.
- custom_nodes.js registers its own Settings section in the same sidebar.
Its panel calls API.customNodes() on mount, and its buttons reach a real
git clone (customNodeInstall), git pull (customNodeUpdate), and an
irreversible directory delete the panel's own confirm dialog describes
as "cannot be undone" (customNodeRemove) -- worse than any prior finding,
since it can destroy real files on disk rather than just make a stray
network call.
Also closed temp_browser.js's listTemp/downloadTempFile (lower severity,
read-only, but still a real read of the live server's actual temp files).
Fixed by mocking all eight functions in patchApi(), matching each one's
real response shape (pipeline_state.js needs slots/incomplete/refused/
queueable; editPipeline's mock echoes back whatever slots were sent so the
demo panel behaves sanely rather than erroring).
Verified live in a genuinely fresh tab (a stale tab's request log falsely
showed these firing before the fix landed -- reproduced clean in a new tab
to be certain): opened the Models modal directly (the zero-click trigger)
and the Custom Nodes panel, zero /api/modules, /api/pipeline, or
custom-node requests in the network log, zero console errors, full tour
sequence still completes cleanly.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Round 6's sweep of raw API.* calls that bypass Store entirely (a layer
none of the prior five rounds checked) found the most easily-triggered
gap yet: models.js's "Export settings..." button calls API.settingsCard
directly, a real unmocked POST /api/settings-card -- and it's on the
FIRST screen the tour's own "models-modal" step shows, no sidebar
navigation needed (unlike round 5's Combo Sweep panel, which needed a
click into Settings first). Fixed by mocking API.settingsCard with a real
Blob (the function resolves a Blob and feeds it straight to
URL.createObjectURL, so a JSON mock like every other entry uses would
break the button instead of blocking it) -- a 1x1 transparent PNG is
enough for the card to render.
Also closed the secondary finding: applyGlobalPromptQuiet has TWO
internal closure-scoped callers (setEditorSetting's anchorEnabled toggle,
applyPromptTemplate) besides its exported Store property -- the same
"block the export, miss the internal call" shape as commit()'s original
vulnerability, so a patchStore() block-list entry alone couldn't have
closed it. It reaches a real, unmocked API.parsePrompt call. Fixed with
the same root-cause pattern as commit()'s guard: a window.__FUNPACK_TOUR__
check directly inside the function, which also covers
scheduleGlobalPromptApply transitively since that just debounces into a
call to this same function.
Verified live: API.settingsCard resolves a real 68-byte PNG Blob (not an
error); toggling anchorEnabled and calling scheduleGlobalPromptApply
directly, waited past both debounce windows, zero prompt/expand or
settings-card requests in the network log; full tour sequence still
completes with zero thrown errors.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Round 5's exhaustive sweep of store.js's Store export found the serious
one: runComboSweep ("Run sweep" in the Settings window's "H3 Combo Sweep"
panel) reaches a real ComfyUI job queue completely outside every mock and
guard built so far. It calls window.GenerateBridge -> app/shell/session.js's
real queue-and-poll pipeline directly -- a separate module tree from
window.MovieEditorAPI that patchApi() never touches, and with zero
__FUNPACK_TOUR__ awareness anywhere in that chain. The "models-modal" tour
step spotlights the entire .settings-win element (sidebar included), so
the Combo Sweep panel is genuinely openable and runnable mid-tour -- not
theoretical.
This is a fourth caller of the same generation funnel generate/
generateMontage/generateSelected already block, just reached through
Settings instead of the timeline -- the kind of sibling path this whole
review chain keeps finding one at a time. Also closed renameMedia, a
second dormant landmine of the exact shape saveClipToMediaBin was
(currently inert -- API.renameMedia is a stub -- but already reachable
from the media bin's rename control).
Verified live: both return the block toast directly (confirmed the exact
Generate toast fires for runComboSweep, not a queued job), zero
prompt-queue network requests, full tour sequence still completes with
zero thrown errors.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Round 4's review confirmed the commit() guard genuinely closes every
project-save path (keyboard shortcuts included), but caught the fix's own
comment overreaching: "every mutating path in this module funnels through
here" is false. store.js's shortcut/category/transition CRUD functions
(saveShortcut, deleteShortcut, addCategory, importShortcuts,
clearShortcuts, saveTransition, deleteTransition, importTransitions,
clearTransitions) each call their own real, unmocked API.* endpoint
directly and never touch commit() at all -- a second, separate funnel the
guard doesn't cover.
Currently dormant (no Settings > Shortcuts panel is wired up in v5 yet,
confirmed by grep -- these functions have no caller anywhere outside
store.js/api.js today), same shape as the saveClipToMediaBin landmine
rounds 2-3 already closed pre-emptively rather than waiting for it to go
live. These write GLOBAL state (the on-disk shortcut/transition library),
not sandboxed per-project state like everything commit() guards, so if
left open they'd be worse than a stray project PUT once a Shortcuts UI
ships.
Fixed with the same block-list pattern, and corrected both comments
(store.js's commit() guard, tour.js's blocklist) to state their actual
scope instead of the overclaim.
Verified live: all nine now return the block toast instead of reaching
Store's real implementation; full tour sequence still completes with zero
thrown errors.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Round 3's review found the enumeration approach (patchStore()'s blocked
list) was still incomplete, and worse, live rather than dormant: keyboard
shortcuts (Delete/Backspace/S) bypass the tour's click-blocking backdrop
entirely, since CSS pointer-events does nothing to a window-level keydown
listener. timeline.js's shortcut handler calls removeSelectedScenes(),
removeAudioTrack(), removeSelectedOverlay(), and splitScene() -- none of
them in the blocked list -- and every one ends in scheduleSave(), whose
setTimeout calls the closure-scoped commit() directly, reaching a real,
unmocked PUT /funpack/api/projects/<id>. tour-s1 is pre-selected from the
moment the tour boots, so pressing Delete at any point immediately
triggers this -- no click needed, no specific step required.
The blocked-list pattern (three rounds of fixes: parsed.scenes crash,
syncFromPreview, saveClipToMediaBin) was closing this one caller at a
time and kept finding another. Fixed at the root instead: every mutating
path in store.js funnels through the single closure-scoped commit()
function before it can reach the network, so that's the one place a
tour-mode guard covers every current AND future caller by construction,
without needing tour.js to know the name of anything that might call it.
Verified live: dispatched a real Delete keydown during the tour with
tour-s1 selected (the default boot state) -- the scene still gets removed
locally (harmless, it's fake demo data), but zero requests to
/api/projects appear in the network log across the full debounce window
after the fix, versus a real PUT before it. Confirmed the guard is a
no-op outside tour mode (window.__FUNPACK_TOUR__ is false there) and the
full 16-step tour sequence still completes with zero thrown errors.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Round 2's review of the syncFromPreview fix came back clean but flagged an
adjacent, currently-dormant gap while tracing the same pattern: the "Save
video" button (Store.saveClipToMediaBin -> API.importClipToMediaBin) is
neither in patchStore()'s block list nor patchApi()'s mock list, and it
already renders during the tour (tour-s2 satisfies
clipSaveableToMediaBin()). It's harmless today only because
importClipToMediaBin is itself a stub ("no render/stitch stage yet") --
the moment that stage lands, this becomes a real, unblocked PUT during the
tour with nothing prompting anyone to notice. Closing it now with the same
established pattern rather than leaving it for a future regression.
Verified live: Store.saveClipToMediaBin now returns the block toast
directly, and a full tour run completes with zero thrown errors.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adversarial review of ee2c0d5 (the parsed.scenes crash fix) found the crash
fix's own side effect: stopping the crash meant "Sync scenes from preview"
became clickable for the first time during the tour. That button calls
syncFromPreview() -> scheduleSave() -> a 5s-debounced closure-scoped
commit() -> a real PUT /funpack/api/projects/<id>. patchStore() already
blocks "commit", but only as the exported Store.commit property --
scheduleSave()'s internal setTimeout calls the closure-scoped commit
function directly, bypassing that patch entirely. This directly
contradicts the tour's own welcome copy ("nothing here touches your real
files"). Fixed by blocking syncFromPreview itself, the same way every
other mutating entry point (generate/save/delete/import/...) is already
blocked -- consistent with the existing pattern rather than trying to
patch scheduleSave's internals.
Also fixed a cosmetic mismatch the review caught: the demo preview's
middle scene had prose text while its actual timeline clip (tour-s2, a
video clip) has none -- now both agree.
Verified live: clicked "Sync scenes from preview" mid-tour, confirmed the
block toast fires and no PUT request is made (watched network across the
full 5s+ debounce window), then re-ran the full tour sequence to
completion with zero thrown errors.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Verified the ported Tour feature live end-to-end (entry via ?mode=tour and
via Help > Welcome tour..., all 15 spotlight steps, FAQ landing, exit) and
found the one real bug: tour.js's canned preview.parsed was `{scenes: 3}`
-- a count, not an array. inspector.js's renderSplit() does
`(parsed.scenes || []).forEach(...)`, which throws the moment the Scene
preview panel renders during the tour. Confirmed identical in v4's
original tour.js -- a pre-existing bug carried over verbatim by the
byte-for-byte port, not something the port introduced.
Fixed by giving the demo preview a real 3-scene array matching the demo
project's own scene texts. Re-verified live: zero thrown errors across
the full tour sequence after the fix.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
extensive_testing round 5 did a systematic audit rather than one more
spot check, and found the one remaining gap: row VALUES are wrapped to
card width before drawing (_wrap()), bounding each draw call's cost, but
"head" block titles (built from slot id/group/node, each individually
_cap()-allowed up to 400 chars per round 4) were drawn as one unwrapped
string. PIL's text-render cost scales with string length regardless of
visible width, so 2000 slots each with a 1200-char (id+group+node)
concatenated title -- every documented cap individually honored, nothing
near any threshold -- cost 20.5s of pure font-render CPU.
Fixed by truncating head-block text to the same rendered width row values
already get, before it's ever handed to the font renderer. Round 5's own
systematic pass (tracing every slots-derived value to its draw/embed
destination, checking the wired-link path, non-string id/group/node,
by_id construction cost, host reachability, and the render kwarg) found
nothing else -- everything else that reaches the raster or the JSON embed
is capped.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Found while re-reading collect() after round 4's fix, same pattern as every
prior round: an unaudited site where caller data reaches the card uncapped.
- collect()'s per-input loop caps `value` via _short() but never capped
`name` -- the input dict's KEY is caller data too (the whole `inputs`
dict comes from the client), and flows verbatim into the drawn row label
and the embedded JSON.
- _rows_from_list's per-entry join (`f"{k}={_short(v)}"`) caps `v` but not
`k` -- a huge key still gets fully materialized into the string before
the joined result's own _short() call ever runs, so it's real per-row
cost even though the final printed string comes out short.
Both fixed with the same _cap() helper already used for slot id/group/node.
Added regression tests for both.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
extensive_testing round 4 found every cap added in rounds 1-3 bounded ROW
content only. A slot's id/group/node class, and project_name, were never
run through any length cap -- they reach render_png as "head"/"title"
block kinds, which the round-2 line-truncation loop skips entirely (it
only counted kind == "row"). Worse, collect()'s row-budget check never
fires for a slot with zero rows (empty inputs), so slot COUNT itself was
unbounded whenever each slot carries no rows.
Reproduced: 3000 slots with empty inputs and 4000-char id/group strings --
total_rows stayed at 0 the entire time, so none of the three existing caps
ever engaged -- cost 193.6s render time and produced a 12.1MB embedded
tEXt JSON, with zero truncation declared anywhere on the card.
Fixed by: capping slot id/group/node/project_name/wired-from-label length
via a new _cap() helper (same truncation as _short(), without its
display-value formatting so "no id"/"no group" fallbacks still work);
capping slot COUNT directly in collect() rather than relying on row count;
and making render_png's line-truncation loop count "head" blocks toward
the same budget as "row" blocks, as defense in depth against any other
block kind reaching render with no accompanying row. Added a test
reproducing the exact empty-input/huge-id shape, asserting on render_png's
wall-clock time and both the PNG size and the embedded JSON size.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
extensive_testing round 3 found round 2's fix was itself incomplete: it
bounded row count and per-key value length, but not the JOINED row string
_rows_from_list builds from a dict with many keys. A funpack_list entry
like {"k1": "x"*400, ..., "k150": "x"*400} produced one row whose value had
no length cap of its own -- row count and per-key length both stayed under
every existing budget while the report JSON embedded in the PNG's tEXt
chunk (which scales 1:1 with it) reached 123MB, reproducing round 2's exact
number through a different attack shape, with zero truncation declared
anywhere on the resulting card.
Fixed by running the joined row string back through _short() (closing the
length gap) and capping how many keys get processed per row at 50 before
that join even happens (closing a related CPU-amplification gap: an absurd
key count still cost one f-string+join per key before any length cap could
apply). Added a test reproducing the exact many-key-row shape and asserting
on render_png's actual wall-clock time and output size, not just collect()'s
data.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
extensive_testing round 2 found the round-1 render-cost fix was broken:
- _MAX_ROWS and _MAX_VALUE_CHARS each bounded one factor but never their
product -- 2000 rows of 400-char no-space values honored both caps
individually and still cost ~14s/~10MB per render.
- collect()'s card-wide row budget was checked once per SLOT, so one slot
with many funpack_list inputs (each independently capped at _MAX_ROWS)
could blow the total budget by 50x with no declared truncation -- 67.8s,
~123MB, and a PNG that trips Pillow's own decompression-bomb guard on
reopen. Worse than the original round-1 finding, not fixed by it.
- none of round 1's three new tests called render_png -- the function whose
cost was the actual finding -- so both regressions passed all 21 tests.
Fixed by bounding the actual cost driver directly: render_png now caps
total rendered lines at 1200 regardless of how collect() produced them
(defense in depth), and collect()'s row-budget check now runs per input,
not per slot, so one slot's inputs can no longer accumulate past the card
budget before the next slot's check fires. Added tests that call render_png
against both worst cases and assert on wall-clock time, not just collect()'s
output data.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Round 1 of extensive_testing on 6d49625 found two real defects:
- Unbounded server-side render cost/memory from a client-supplied
pipeline. render_png's image height grows with wrapped-line count,
so one crafted long input value amplified a small request into a
77,400px-tall, ~250MB render taking several seconds of single-
threaded PIL draw calls -- reproduced directly. This route is
unauthenticated, same as every sibling pipeline route, and reachable
over a rental's plain-http IP -- a real DoS surface on the box
actually doing GPU sampling. Fixed with two bounds: _short() now
truncates any one value past 400 chars (a card is a compact summary,
nobody reads a 200k-char row on it), and collect() caps the whole
card at 2000 rows, both declared on the card itself via a "…N more
omitted" note rather than silently dropped. Verified: the 200k-char
reproduction case drops from 3.2s/250MB to 0.09s/22KB; a 200,000-slot
pipeline renders bounded at ~1.7s/5MB with the truncation stated.
- Reopening "Export settings…" via the toolbar button (not the modal's
own ✕/Close) leaked the previous render's blob URL -- close()'s
revoke only runs on ITS OWN overlay, and the top-of-function DOM
cleanup only removed old .sc-overlay nodes, never revoked their
blobs. Carried over unfixed from v4's identical pattern. Fixed by
stashing the blob URL on its own overlay element so the cleanup line
can find and revoke it. Verified live: two clicks without ever
hitting Close produces two distinct blob URLs and exactly one
revokeObjectURL call, one overlay left in the DOM.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
MEDIUM-tier v4 feature ("Which model was that?" answered three days
later): a button in Models & Pipeline renders the currently loaded
pipeline as a PNG -- every slot's node class and typed-in values, wired
inputs marked as wired rather than printed as a stale value, and the
host facts (torch/CUDA/GPU/attention) that decide whether the setup
reproduces elsewhere. Downloadable, copyable to clipboard, and the
report is embedded in the PNG's own tEXt chunk for later machine
reading.
core/settings_card.py is a hard cut of v4's version, not a verbatim
port: v4 spent most of its length (_OWNED_BY, _EDITOR_DEFAULTS,
_pass_rows, _sampling_sections) describing FunPackStudio and the
LTX-AV Scene Chain Sampler's own hardcoded knob tables -- neither node
exists in v5's module-driven pipeline, so there is nothing left to
describe there. What's genuinely generic (walking a slot's real
inputs, detecting a wired one via core.graph.is_link, the PNG
renderer) is what's kept, adapted to v5's slot shape ({id, node,
group, inputs}, a wire arriving as [source_id, index] rather than a
separate input_sources/wires lookup) and v5's existing core/sysinfo.py
for host facts (whose field names already matched v4's needs).
New POST /api/settings-card route: stateless like /api/pipeline, the
CLIENT sends its own live pipeline (window.PipelineState.slots()) since
that's the one copy that includes any unsaved edit; the server supplies
only what it alone knows (host facts) and the render.
Frontend: app/legacy/api.js's settingsCard stub (present since the
Phase A port, explicitly marked "not built in v5 yet") replaced with a
real POST+blob call; app/legacy/models.js's Export-settings modal
ported from v4 near-verbatim, reading PS.slots()/project name instead
of a project id the old per-project models config needed. Verified
live: real pipeline, real host facts (this Mac's actual Apple MPS/
Python/ComfyUI versions), wired inputs correctly marked.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Round 2 of extensive_testing on 548d99a found no exploitable defect --
has_block's "contains, not is" semantics are shared with minimax_h3's
own equally consequential is_h3() gate, and nothing in this codebase
nests diffusion_model in a wrapper that could trigger a false positive
today. Recorded as a one-line limit rather than left implicit, per the
round's only (non-blocking) suggestion.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Round 1 of extensive_testing on 79ca8cf found is_h3_model() matched on
diffusion_model's bare class name ("MiniMaxH3Model"), inherited
unchanged from v4. This codebase's own core/traits.py.has_block()
documents exactly why that's unsafe for a name this generic: a
same-named class from an unrelated custom node or a future FunPack
model would be silently treated as H3 and sparsified at 0.90 with
H3-specific audio-prefix pinning -- the identical LTX-collision risk
this file's own docstring already warns about, just not guarded
against strongly enough. modules/models/minimax_h3 already qualifies
its own identical check (MODEL_CLASS = the dotted path); this port
carried over the private, weaker duplicate instead of reusing that
pattern.
Switched to core/traits.py's has_block() with the same qualified dotted
name, kept as a local string (not an import of modules/models/minimax_h3
itself, which pulls in comfy_extras at import time -- this file
deliberately stays importable without ComfyUI on the path). Added a
test exercising the exact collision the bare match would have missed:
a class named MiniMaxH3Model from an unrelated module.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Loader-level attention option: mean-pool Q and a smoothed K into
blocks, score them with one small matmul, attend only the top
1-sparsity fraction of key blocks per query block, via a vendored
Triton kernel. Runs alongside the chosen attention backend rather than
replacing it (one override slot; the backend becomes SLA's dense
fall-through for the text refiner, masked calls, and trailing dense
steps), and refuses cleanly -- naming why -- on anything that is not
MiniMax H3 or a machine without CUDA+Triton.
modules/loaders/sla_block_map.py and sla_kernel.py are near-verbatim
ports of v4's vendored files (no FunPack-specific imports, so nothing
to adapt); sla_attention.py is the loader-facing glue (H3 gate,
override composition, per-step wrapper), same content as v4's with a
"not a module" note added to match gguf_support.py's convention.
Wired into FunPackDiffusionModelLoader as six new optional widgets,
following v4's loaders.py wiring but expressed in v5's io.Schema style.
The Triton kernel itself is unexercisable on this Mac (no CUDA) --
matches how real generation is verified in this project (rented GPU,
not local). Ported the CPU-testable half of v4's test suite (the
override contract, the H3 gate, the per-step wrapper); the three
kernel-correctness tests carry the same CUDA+Triton skip guard v4 used.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Round 3 of extensive_testing found ae319cb's _noticeIsStaleRenders flag
had the exact footgun it was built to avoid: exportSelected() and
saveSelectedToMediaBin() also write state.notice (through the generic
set() helper, spelled `notice:` in an object literal -- invisible to a
`state.notice =` grep) and never cleared the flag, so a pending
stale-render flag could survive into an unrelated export/save
confirmation and have _recordSegment wrongly clear THAT instead.
Every future writer of state.notice would have needed to remember to
maintain a parallel boolean by hand -- two already didn't. Replaced the
flag with a regex test against the stale-render sweep's own known
message template: _recordSegment now clears state.notice only when its
current text actually IS that message, derived from data instead of
tracked by convention. No writer needs to know about any other writer.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Round 2 of extensive_testing on 2e738a8 found the fix only closed the
race at queue time. state.notice is shared with a third writer:
_recordSegment unconditionally blanked it on every successful render
(state.notice = "" whenever recordedSceneIds.length), which exists to
retire ITS OWN "N saved renders no longer on disk" message once a
regen fixes it -- but fires on every successful run regardless, so it
silently ate the reference-unwired notice the instant the very run it
described finished. A person who wasn't watching the screen during
generation would find the notice gone by the time they looked back.
Added _noticeIsStaleRenders, tracking which case currently owns
state.notice, so _recordSegment only clears its own message.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Round-1 adversarial review of 1abf136 found two real conditions:
- _buildQueueInputs set state.notice for an unwired reference directly,
before GB.generate() ran check() -- whose own "warn" note writes the
same field via GB.on("warn"). Whichever fired last won, so a person
with any Engine Settings configured on a pipeline with no settings
sink (the default pipeline is exactly that) would see "module
settings didn't apply" instead of ever learning their reference was
dropped. Fixed by applying the reference notice AFTER GB.generate()
resolves, so it always wins that race instead of losing it silently.
- The Resolution-source picker's collapsed view can't tell "nothing
picked" from "picked something since deleted from the bin" -- both
render as MediaPicker's noneLabel placeholder. The References list
right next to it already shows "(missing)" for this case; the
source picker didn't. Added the same declaration as an insp-hint.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Fills in the placeholder left in _buildQueueInputs: v5's real per-scene
fields (core/projects.py's Scene.source_image/.references) already
round-trip through api.js's getProject/saveProject today -- it's a raw
whole-object PUT/GET, and the server's own Scene.from_dict()/asdict()
extract and return them regardless of what else v4's project shape carries.
There was just no UI to set them and nothing reading them at generate time.
Added:
- store.js: setSourceImage/setReferences, thin patchScene wrappers (same
pattern as the existing setSourceTrim).
- inspector.js: a "Reference media" section on the scene panel, right after
v4's own i2v anchor/Source field, using the same window.MediaPicker
thumbnail-grid component already used throughout this file for image
picks -- a single Resolution-source picker plus an add/remove list for
References.
- generate_bridge.js: exposes reference_wiring.js's wireReferences()
(untouched, already tested, no DOM dependency).
- store.js's _buildQueueInputs: sets assets.source_image from the run's
first/anchor scene, and calls wireReferences() for its references --
same "first scene stands for the whole run" convention its prompt-text
branch already uses for a multi-scene (whole-timeline) run. An unwired
reference (no matching slot in the current pipeline) surfaces as a real
notice in the timeline header, not a silent drop.
Live-verified against the dev server end to end: picked a real reference
image from the Media bin, watched it persist server-side
(GET /funpack/api/projects/:id showed it in the saved file), then clicked
Generate and confirmed the pipeline check ran with it -- the current
default pipeline has no reference slot, and the app correctly said so
("1 reference could not be wired into this pipeline and will not be
used.") rather than silently dropping it. 691/692 JS tests and 548/548
Python tests pass (the one JS failure is the pre-existing, unrelated
projects.test.js import-race flake).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Round 5 of adversarial review found a real ordering race: generate_bridge.js's
module-level wire() call starts reattach() (a /queue or /history round trip)
the moment the module evaluates -- before store.js's classic script has even
run, let alone reached _wireGenerateBridge() (called deep inside init(), after
seven sequential awaited setup calls). On a reload during or just after a
generation, "adopt" routinely fired and was gone before store.js ever called
GB.on("adopt", ...): _genRunSceneIds stayed [], so the adopted run's eventual
result was never recorded onto its scene (_recordSegment's own length guard
silently skipped it), and _genRunActive/_genAdoptedRun never got set, leaving
that run invisible to the overlap guard this whole review thread has been
hardening.
Fixed by giving "adopt" -- and only "adopt", since it carries the sole copy
of which scene(s) a reattached run belongs to, unlike the transient "say"/
"warn"/"hold"/"release" status text, which is fine to miss -- replay
semantics: the bridge remembers the last adopt payload and hands it to any
listener that registers afterward, immediately and synchronously, the same
way it would have if the listener had been there in time. Verified with a
standalone reproduction of the exact closure logic (late subscribers,
including a second one, both receive the replayed payload correctly) and
against the real dev server (boots clean, no regressions). 691/692 JS tests
and 548/548 Python tests pass (the one JS failure is the pre-existing,
unrelated projects.test.js import-race flake).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Round 4 of adversarial review found that the previous commit's overlap
guard, keyed off state.gen.state, produced false refusals: Export, Save to
Media bin and Render Final all set state.gen to "running" for their own
entirely local work, so clicking H3 Combo Sweep's "Run sweep" button while
any of those was in progress got refused with a message borrowed from that
unrelated job ("Stopped at ...: Exporting 3 clips…") -- not silent, but
actively misleading about the real cause.
Switching the check to GenerateBridge's own run.state.phase (considered as
the obvious fix) would have reopened the very race round 3 closed: phase
stays "idle" for the whole check() network round trip, before run.start()
ever fires, so a second call could still slip through during that window.
Fixed with `_genRunActive`, a flag dedicated to exactly one question --
"does GenerateBridge have a run of this page's making in flight" -- set
synchronously the instant a call is accepted (before any await, so two
calls fired back to back still serialize) and cleared either by the
terminal-phase subscription (covers a run that actually started, click or
adopted) or by _generateRun's own `finally` (covers a run refused before
ever leaving "idle", which that subscription never sees).
Live-verified: a state.gen "running" left over from an unrelated job no
longer blocks a real generate() call, while two genuinely concurrent
generate() calls still correctly serialize (one proceeds, one is refused
with the accurate reason). 691/692 JS tests and 548/548 Python tests pass
(the one JS failure is the pre-existing, unrelated projects.test.js
import-race flake).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Round 3 of adversarial review found that GenerateBridge now holds exactly
ONE run's shared state (_genRunPrefix/_genRunSceneIds/_genAdoptedRun,
window.GenerateBridge's own module-level ranFor/ranForList), but only the
main Generate/Selected/Render buttons are gated against starting a second
one while the first is still queuing/running (actionbar.js's busy(st)). H3
Combo Sweep's "Run sweep" button has no such gate and calls _generateRun
directly -- if clicked while another generation (including one adopted on
reload) is already in flight, it silently overwrote that run's scene
attribution and status message, got refused deeper inside (session.js's own
starting/phase guard), and left the REAL run's eventual result recorded
onto the sweep's scene instead of its own when it finished -- persistent
data corruption, not just a UI glitch -- and could leave _genAdoptedRun
stuck false, reopening round 1's "adopted run's clock never stops" bug.
Fixed at the root: _generateRun itself now checks state.gen.state before
touching any of that shared state, and refuses (loudly, via console.warn)
rather than clobbering it. Fixes every current and future caller in one
place, not just block_sweep.js -- consistent with every other caller
already being a sequential, awaited loop that this guard doesn't affect.
Live-verified: two overlapping Store.generate() calls now produce exactly
one real attempt and one logged refusal, with the first call's status
never overwritten. 691/692 JS tests and 548/548 Python tests pass (the one
JS failure is the pre-existing, unrelated projects.test.js import-race
flake).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Round 2 of adversarial review on the previous commit found that
_genClockStop() firing unconditionally on every run's terminal phase (added
to fix an adopted run's clock never stopping) broke the OTHER case it didn't
account for: generateMontage()/generateSelected() call _generateRun several
times under one press for a chain-sampler pipeline (_runs() splits the
timeline into several runs whenever PipelineCaps.usesChainSampler() is
true, which is the default). genWallStart is documented to span the WHOLE
batch, started once by _clockedMontage/_clockedSelected's own wrapper --
stopping it after run 1 of N left genElapsed() returning null, and nothing
in the batch loop ever restarted it, so the visible timer next to Generate
disappeared for the rest of an in-progress batch.
Fixed with a `_genAdoptedRun` flag: only a run actually ADOPTED on reload
(nothing else will ever stop its clock) clears it from the persistent
subscription; a normally-clicked run's terminal phase leaves the clock for
its own caller's `finally` to stop, same as before this whole feature
existed. 691/692 JS tests and 548/548 Python tests pass (the one JS
failure is the pre-existing, unrelated projects.test.js import-race
flake). Live-verified against the dev server.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adversarial review of the previous commit found four real defects, all
confirmed:
1. A run covering more than one scene (the whole-timeline H3 shape, or any
single-scene click whose gen-unit has sibling subclips) only ever sent
the FIRST scene id as extra_data.funpack_scene_id. On reload mid-run,
reattach recovered just that one id, so _recordSegment only ever wrote
the finished media onto scene 1 -- every other scene the job actually
rendered silently got nothing recorded, no error shown.
Fixed by adding funpack_scene_ids (plural) alongside the existing
singular field, threaded through client.js's queuedFor and session.js's
reattach() (both additive -- a new optional third onAdopt argument and
a new optional response field, ignored by every existing caller such as
app/boot.js) down to generate_bridge.js/store.js, which now record onto
every adopted scene.
2 & 3. A run adopted on reload never primed pollStart/genWallStart, so
_elapsed() and the transport button's own clock read time since the
Unix epoch instead of real elapsed time, and never stopped afterward --
wedging genElapsed() for the rest of the session (its own guard is "if
already started, do nothing"). Fixed: the "adopt" handler now primes
both clocks (flagged approximate, same as the deleted
resumeRunningGeneration used to), and every terminal run phase now
stops them.
4. check()'s `notes` field (e.g. "N settings didn't reach the pipeline")
was fired through transport.warn() but nothing ever subscribed to
GenerateBridge's "warn" event -- silently dropped, reproducing the
exact bug class the deleted code's own comment said must not recur.
Fixed: wired into state.notice, same as the old code did.
New tests added for both the plural-scene-id path (client.test.js,
session.test.js) and the fallback-to-one-scene case. 691/692 JS tests and
548/548 Python tests pass (the one JS failure is the pre-existing,
unrelated projects.test.js import-race flake). Live-verified against the
dev server: a multi-scene GenerateBridge.generate() call still correctly
surfaces a real pipeline refusal with no console errors.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds generate_bridge.js, a small ES-module shim over the untouched
app/shell/run.js + session.js + pipeline.js + client.js, and rewires
store.js's generate()/generateMontage()/generateSelected()/interrupt()/
init() reattach to go through it instead of the now-permanently-stubbed
API.generate/status/active. PipelineState (pipeline_state.js) stays the
one client-side copy of the live pipeline; the bridge only adds the
run/queue/socket layer on top of it.
Single scene (or the whole active timeline, for a non-chain pipeline like
H3) per real queue -- v5 has no chain-sampler/montage stage. REINS captures
and the anchor-guide i2v bypass are said loudly (console.warn) rather than
silently dropped, since neither has a v5 equivalent yet. Reference-image
wiring is left as a placeholder: legacy's saveProject/getProject adapter
doesn't round-trip source_image/references yet, so there is nothing to
wire in from there today.
Live-verified against the real dev server: a real Generate click built the
scene's prompt, POSTed it through /funpack/api/pipeline's check, and
surfaced the genuine refusal reason (an unset model loader) in the existing
status UI, with the button correctly re-enabled after. 688/689 JS tests
and 548/548 Python tests pass (the one JS failure is the pre-existing,
unrelated projects.test.js import-race flake).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Review found the save() queue's pendingBody merge was shallow: two queued
edits that both touch the top-level "inputs" key (two different slots, or
even two fields on the same slot from models.js's per-call-fresh objects)
had the second's inputs object silently REPLACE the first's rather than
combine -- the earlier edit vanished with no error, the field just
reverted on the next render. engine_settings.js's own calls happened to be
immune by accident (one long-lived mutated `values` object, so the last
queued snapshot already contained every earlier edit) -- that immunity
was never a property of save() itself, and models.js's fresh-object-per-
call shape hit it immediately. Fixed with a one-level-deeper merge for
"inputs"/"values" specifically. Verified live via two real race
reproductions run through the browser console: two edits on different
slots queued with no await between them, and two edits on the SAME slot --
both survive now, confirmed against the actual saved graph state.
Also closed a load race: ensureLoaded()'s only guard was `slots !== null`,
which doesn't cover the window while a first load is still in flight -- two
panels mounting close together could both start a fetch. Now every caller
awaits the same in-flight promise, so exactly one fetch ever goes out per
session.
Documented (not fixed, deliberately) that Models & Pipeline's node-level
pin target is currently unreachable: pinTarget() only ever returns a
group-level pin today, since there's no "currently open node" concept in
this flat slot list yet -- the openNode/ctx.sub("node:...") plumbing stays
for whatever old pin data might still reference it, but nothing produces
a fresh one. Real gap against v4's per-node pinning, not a regression.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
