Colour is a preference, so it is a setting: hex in, hex out, one spelling,
refused if it is not one. A module handed "red" in one place and "#f00" in
another guesses wrong somewhere.
Image and mask are NOT settings and deliberately have no type. They are
data -- tensors the app uploads and the graph wires -- and ComfyUI already
has socket types for both. Putting a drawn mask in a JSON settings payload
would be inventing a second way to move something that already moves.
Runs now have an id. Previews, progress, results and log lines all have to
be attributable to one generation, and ComfyUI gives a node no way to ask
which run it is in. So FunPack marks its own, every log record carries it,
and the history can be filtered to a single run. Readable rather than a
uuid, because it is going to be read: "run 7" is findable in a wall of
output and a hex string is not.
The colour element was caught by the registry-driven rule test before it was
finished -- an element with no catalogue demo is neither shown nor attacked,
so coverage keeps up with the registry instead of lagging behind it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
[Log] Info: FunPack Diffusion Model Loader | flux.safetensors loaded, attention sage
[Log] Alert: MiniMaxH3ImageToVideo node | no "first_image" input is connected
[Log] Warning: FunPack LoRA Loader | style.safetensors is empty and will not load
[Log] Critical: torch | cannot allocate 14 GB (3 GB free); the run will not start
The difference between the middle two is the one that matters. Alert is a
choice you may not have made on purpose. Warning is something that will NOT
do what its name says -- the level v4 needed and did not have, which is how
560 handlers came to swallow a failure and carry on, so a feature could be
switched on, do nothing, and never say so.
Info is not decoration either: "loaded on GPU using sage attention" is how
an attention setting that did nothing gets noticed months before the
pictures look wrong enough to investigate.
Every call site moved -- there is no note() left to be vague through -- and
the loaders now say what they actually loaded rather than only what they
were asked for. Records are kept in a bounded ring as well as printed, so a
long rental session cannot turn the log into a memory leak, and served at
/funpack/api/log so a person can read it somewhere other than a terminal
full of ComfyUI's own output. There is a page for it.
Also dropped the preflight this was traded for: min and max are already
declared and already enforced at both doors, and a second protection layer
around a check that exists is ceremony.
Two mistakes worth keeping: a log line reached into the model's shape to
describe it and could have failed the load it was only narrating; and both
new pages passed `children` to an element that takes `rows`, which the kit
accepted and silently dropped. It throws on an unknown element and not on an
unknown prop.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
First real generation through FunPack's nodes: SD1.5 on this Mac, real
weights, a real picture of an apple. Everything before this was static.
What it proved, on a real model rather than a fixture: the empty latent
takes its shape from SD1.5's own format with no family branch anywhere
(1, 4, 64, 64, an IMAGE latent from a pack built for video); the traits read
off the live model are predict_eps and spatial_latent; and ALG is correctly
absent, saying "needs temporal_latent", rather than running on a model with
no time axis.
What it found: the absolute-core-import bug again, this time INSIDE A
FUNCTION BODY. It imported fine, registered fine, passed 258 tests, and
raised at execute time -- during the generation. A suite that imports every
module cannot see an import that only happens when a node runs, so the check
for it is now static and walks the AST: no module may name `core` anywhere,
at any nesting, except through the shim.
The dev preview loaded one module in nine. It runs system python, which has
neither ComfyUI nor torch, so every module that imports either was absent
and the app looked almost empty while reporting nothing. It now re-execs
under ComfyUI's interpreter. Looking at the app is meant to be how the app
gets tested; it has to see what ComfyUI sees.
ALG asked for a mount point no region offered, so its panel vanished in
silence. The shell now offers one per stage, and a modules page lists what
loaded, what asked for a mount nothing offers, and what failed to import --
absence is correct behaviour and indistinguishable from "never installed",
so somewhere has to say which.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
ComfyUI's V3 entrypoint replaces NODE_CLASS_MAPPINGS, which must not exist
at all: nodes.py checks it first and reaches comfy_entrypoint only in an
elif, so an empty dict registers nothing and reports no error.
A module declares NODES for graph structure and SETTINGS for preferences,
and the two never mix -- a saved workflow stores class_type and input
names, so letting a new setting add a socket would rot existing graphs.
Traits now come from two places: facts core can read off any model, and
providers announced by a model's own module. Core names no model, so
supporting a new one is a folder rather than an edit here.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A cycle rejected everything still waiting, not just the modules on it. One that
merely says 'after B' is not itself circular, and once B is dropped its relation
has nothing to wait for -- the same as naming a module that was never installed.
Now only the ids that can reach themselves are dropped, and a dropped id stops
being a dependency so the damage stops rather than spreading downstream.
A relation could also overturn the stage order: 'load after conditioning' was
honoured, putting a whole stage out of sequence to satisfy one module. Stage is
the coarse order; an impossible relation is dropped and logged, and the module
itself is kept.
panels.js now keeps what a module's setup() returns, which is its unsubscribe --
discarding it leaves a live listener behind on every re-mount.
One declaration does three jobs: renders the panel, carries the defaults, names
the values that come back. A module writes no DOM and reads no DOM.
Validation refuses rather than repairs -- unknown type, a renderer its type does
not have, a label that is not text, a default outside its own bounds, a when
naming a key the module never declared. A module that is wrong does not get a
best-effort panel; it does not load, and says why once.
Ordering is by declared relation and stage, never priority numbers, and a cycle
drops only the modules in it. Compatibility is by trait, never model name.
The shell hands a module { composer, values, on, shell } and nothing else: no
node, no document, no kit internals. It cannot style anything because it is
never given anything that could.
catalogue/hide.html puts one good module and five broken ones through the real
shell and shows the result: one panel, five absent, no placeholders. The lint
fixtures make the scanner prove it catches each rule, so a clean scan means
something.
First slice: modules/system/appearance. Two settings, an optional ui.js, and
clicking Light actually turns the app light.
A rung compacts when a claim is released, so an overlay's z can move without it
doing anything. Six of seven call sites read z once at open time and kept a
number that was later handed to the next overlay: two of them at one level, DOM
order deciding which wins and which eats the clicks. Reproduced with two menus
and a third opened after closing the first -- both ended up at 501.
claimFor(rung, node) binds the node, so there is nothing left to forget, and the
raw claim() stays for the one case with no node. The test walks the registry
rather than naming overlays, because the bug was one site remembering and six
forgetting.
Also drops a no-op ternary that picked the modal rung either way.
Three bugs in wheel.half, all found by looking at it:
The hub was centred on the edge, so the edge sliced it in half -- half a cancel
target is not a target. It is pushed inward by its own radius now, and the
wedges start beyond it. The ring is an inset shadow rather than a border,
because a border pushes every absolutely positioned child inward by its own
width, which put the hub a pixel past the screen edge.
Highlighting fired anywhere the pointer went, including well outside the wheel,
promising a pick that release would not deliver. The hit test is bounded by the
same rIn/rOut the wedges are drawn from, so the lit shape is the region.
An edge panel did not follow its edge: it kept the pixel position it opened at
and drifted inward when the window grew. It re-derives its anchor on resize,
and on visualViewport resize for pinch-zoom and mobile chrome.
One GEOMETRY object per variant now drives the wedges, the labels, the hub and
the hit test together, so they cannot disagree.
wheel.half pins the centre to a screen edge and fans the items inward, for a
control that lives at the edge and has nowhere to open a full wheel. The arc
geometry is now shared with the full wheel, so the two cannot disagree about
which item a direction means. The hub is a glyph sized to the dead zone: the
cancel target you see is the one you hit.
list.rows takes a real control per row. A bare number printed beside the reorder
arrows reads as a stepper for that number, so moving a LoRA looked like editing
its weight.
setBusy re-derived disabled from the constructor's value, so finishing a busy
spell re-enabled a button that had been disabled since; busy and disabled are
tracked separately now.
An emptied number field committed as 0 and snapped to min, discarding the last
good value -- a browser normalises invalid text to empty, and Number('') is
finite so it sailed past the isFinite guard.
A context menu anchors to a POINT, and that stand-in reached the dismiss stack,
which threw on the first click anywhere.
55 element groups across text, buttons, choice, input, slider, layout, status,
overlays, gallery and the wheel picker. Each lands in a registry-driven
catalogue that shows every variant in both themes, so an element that fails to
register is visibly absent rather than quietly missing.
rule_content_is_data iterates the REGISTRY, not a list: a new element is
attacked with markup the moment it calls define(), and one without a demo fails
the suite. That is announcement-not-enumeration applied to the tests.
Fixes found while building: checkbox silently ignored its label (inverted
logic left it with aria-label=toggle), the overlay root set z-index 0 and so
became a stacking context that trapped every overlay it was built to free, and
raising a window left its peers' z stale because z is derived from position --
claim() now notifies a rung when a peer moves.
The wheel highlights the actual sector rather than a box around the label: the
lit shape is the region that picks the item, which is what makes it aimable.
raise() could send a claim to the BACK. Each claim owned a fixed offset, so on a
nearly-full rung there was no free slot above the top and the fallback took the
lowest free one instead: a floating window clicked to front went behind its
peers. Only reachable once a rung fills, which is a long session, not a test.
A rung is now an ordered list and z is derived from position, so raise() is a
move to the end and release() compacts.
drag.js shipped with no tests. 12 now, including the ones that matter: a second
pointer cannot hijack a drag, capture is released on cancel, and dispose stops a
drag already in progress.
register/zlayer/anchor/motion/density are pure cores over plain data, so most of
the suite runs with no DOM at all; jsdom covers only what is genuinely DOM.
100 tests, zero elements.
These five are why v4 had 11 hand-built modals and 7 hand-built popovers: there
was nothing shared to reuse. register makes an element exist only because a file
defined it; portal gives every overlay one stacking parent; zlayer owns the only
z-index in the app; dismiss keeps one stack so Escape pops the top and not the
modal underneath; anchor measures, flips and clamps once.
Three bugs the tests found: unmount() called removeChild on null when no portal
existed yet, focusables() checked aria-hidden on the element instead of its
ancestors so a hidden subtree stayed tab-reachable, and the density ratios v4
hand-tuned are now derived from the gap they were meant to allow for.
Self-hosted, so the editor is correct offline and behind firewalls that block
Google's CDN -- v4 linked the CDN and silently rendered in a fallback there.
133 KB: Figtree and Archivo are variable, one file per subset; latin and
latin-ext only.
The OFL texts ship beside the fonts because redistribution requires it, and a
test fails if a bundled family arrives without one.
A missing face degrades silently into the fallback stack, so the faces declared
in fonts.css are checked against what is actually on disk, both directions.
Cleaned scales from v4's measured usage: 11 control heights to 4, 19 radii to 6,
9px named as a real type step, one reduced-motion guard because every duration
is a token now.
Derived tokens are declared on [data-theme] as well as :root -- a custom
property's var()s are substituted where it is declared, so on :root alone a
scoped light subtree inherits the dark accent.
The catalogue reads the tokens out of the stylesheet instead of listing them, so
it cannot drift; anything it has no renderer for shows up as Ungrouped.
Static serving routes through a pure resolve(), so the extension allowlist and
the traversal guard are tested without aiohttp. modules/ serves .js only: a
module cannot ship a stylesheet even if one is written.
