The workflow ran everything in one job: check.sh, then build, then the e2e suite and the smokes.
The runner has two job slots and four cores, so the second slot lay idle for the whole run, and a
lint error was reported only after the full suite had passed in front of it - six minutes instead
of one.
Now `gate` and `suite` are two jobs without `needs:`. They start together, both have to pass.
The price is contention: while the gate runs, vitest and chromium share four cores, and the suite
measures text widths and frame counts, which a loaded machine reports differently. The gate
therefore caps vitest at two workers via GM_TEST_WORKERS, so two cores stay free for the browser
over the whole overlap. Locally the variable is unset and nothing changes.
Measured before the push: the cap costs vitest 4.4s (16.6s -> 21.0s), check.sh exits 0 with it,
build ok. Whether the split is worth it is a question for the runner - the expected net gain is
about 50s, and the acceptance is a measurement over at least two runs.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Fit measured the nodes alone: a group drawn around them was cut off, and a graph holding nothing
but a comment framed the empty origin. It now frames both sorts together.
The rubber band ignored boxes entirely. It catches them now, into the *same* selection as the
nodes — `nodeIds` and `annotationIds` side by side — so one drag moves all of it, one Delete
removes all of it, and one undo takes it back. A box counts only when the band holds all of it,
where a node counts on touching: a group wraps the nodes it holds and therefore overlaps every
band drawn over any of them, and since it drags those nodes along, catching it by accident would
be felt at once.
A node inside a selected group is moved once, not twice — `travellers()` collects everything one
gesture carries and lets each thing appear in it a single time.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
TextureFrame carries optional `border` (left/right/top/bottom) and
`fill` (stretch | repeat); absent keys mean "not sliced", so no format
bump and undo restores the plain frame instead of four zeroes.
The sliced sprite builds its own geometry instead of the shared
UNIT_QUAD: a 4x4 grid, and for `repeat` one quad per tile carrying the
frame cell's real UVs -- tiling in the geometry, not in the sampler, so
a frame out of a sheet cannot pull its neighbour in. Above 64 tiles per
axis it stretches, to bound the buffer.
A sprite narrower than left+right scales both borders down together
until the middle is exactly 0, so the corners meet at a seam and never
overlap.
The texture editor gets four draggable border lines plus the same four
numbers in the inspector. Hit order is resize handles, then the lines,
then the frame body -- handleAt reports 'move' anywhere inside a frame,
so lines placed after it were unreachable.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The graph could hold any number of nodes and said nothing about which of them
belonged together. Groups and comments are two separate things, as decided in
the issue: a group is a titled box that carries the nodes it encloses, a
comment is free text that carries nothing.
Both live as optional arrays on Graph, not as registry nodes. A comment that
were a real node would run through the interpreter, show up in the drop
palettes and be compiled into a shader — the wrong place for a piece of text.
As arrays on Graph they exist in all four graph kinds at once, and a document
without them means exactly "no groups", so no format bump.
Membership is computed when the box is grabbed, never stored: the center of a
node inside the box travels with it, and a node dragged out is simply out. A
box inside a box travels with the outer one, its own nodes included. A group
is grabbed by its title bar only — its body stays transparent to the pointer,
or a rubber band could never be drawn across the nodes it holds. Deleting a
box deletes the box; the nodes stay where they are, because a box that
silently took ten nodes with it would be the most expensive surprise in the
feature.
The color is a free picker, and the title color follows from it: relative
luminance per WCAG, threshold at 0.179 where the two contrast ratios meet —
not 0.5, which would pick white on mid grey where black reads better.
Resizing snaps the far edge to the grid, not the width, so the corner that is
not being dragged stays put. The minimum size rises in whole steps for that
reason: 80 is no multiple of 24, so a box squeezed to the stop would have sat
off-grid in exactly the mode meant to keep it on.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Set state and State branch derived their ports from the graph, so a machine
that came out of a variable or a list had no ports at all: nothing to set,
nowhere to branch. It could be passed around but not operated.
stateByName() resolves id first, then name; a name two states carry answers
with nothing, the same rule transitionTo already follows for two transitions
of one name — without something in the graph that decides between them, any
pick would be a guess.
Set state by name takes the machine and a text, and has a second exec output
not found, so a typo is not swallowed in silence. State branch gains a
derived else that fires whenever no state output matches — no state set, a
state of another machine, or a machine that arrives as a value and therefore
has no state outputs at all. A flow that enters State branch now always
leaves it somewhere.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The suite spent 88% of its step time asleep: 564 s of steps, 496 s of
them in sleep(n) - fixed spans estimated by hand and then copied from
step to step. That is why it was slow, and why it flickered.
The signal is afterFrames(count, timeout) in ui/render.ts, exposed as
__gm.afterFrames(). It wakes the loop itself, because an editor that
draws nothing would never redeem a promise of a frame, and it gives up
after a timeout instead of hanging the run. settle() in the harness
waits two drawn frames (~32 ms) instead of 150-300 ms and falls back to
sleep(150) when the hook is missing (an old build): silently not
waiting would make every later check a coin toss.
Every action goes through it now - click, doubleClick, wheel, drag,
dragAsset, press, typeText, select, mutate, undo, redo - and with that
about 550 sleep() calls behind those actions became redundant, helpers
that are nothing but a mutate in a coat included. What a frame cannot
answer got its own condition instead: waitUntil(expr), play()/stop()
waiting on __gm.state.mode rather than on 300 ms, saved() for the
autosave (the two hand-written polling loops sampled at 300 ms for
something that arrives after a second), and showTab() as a real tool -
47 places called evaluate("__gm.showTab(...)") raw, and 24 of them
waited for nothing at all and read the old panel.
564 s -> 243 s of step time, the waiting share 88% -> 76%. Three kinds
of wait were kept, and each one bit when it was taken away: loading
bytes (textures, fonts), game time that a repaint cannot stand in for
(time scale, shader and particle clocks, the live values a running
graph carries), and the inertia the scroll step measures on purpose.
Two things the new speed uncovered, both real: 25-animation read the
frame menu unfiltered and had frozen an incomplete list (the CI failure
on c990dbc), and 33-render-on-change measured idleness while the
autosave of the step before was still painting - it waits for
saveStatus to leave dirty/saving now, so it measures that nothing
happens any more instead of that it waited long enough.
#63: a shader Property node named "width" quietly took over the sprite
row of the same name. isSpriteProp lives in PropertySchema now and is
derived from FULL_SCHEMAS.sprite instead of being a second list that
drifts. isValidPropertyName rejects those ten names, and the compiler
reports it at the node as its own problem: "width" is a sprite
property, told apart from "not a word" and from a duplicate.
#64: Divide emitted (a / b) and handed NaN to the driver. It emits
divide(a, b) now, with the guard in the prologue next to straight(),
scalar and per channel: zero divided by zero is zero. That was already
the house rule in the behavior and particle graphs - only the shader
had never been told. Clamp comes along as a shader node for the same
reason: the behavior catalog has had it all along.
CI found the gap the hard way: with every kind of asset in the project
the new scene sits below the panel, and the right click for "Set as
start scene" landed somewhere under the list instead of on the row.
locate() only asked whether a text was visible, and a masked row still
is - it is merely clipped away.
locateAll and overlayTexts now walk the tree together (eachText) and
carry the clip of every ancestor, measured on the widget that owns the
mask: Pixi v8 hands a Graphics set as a mask to the mask effect and
reports empty bounds for it, so asking the mask itself declares every
masked text in the editor missing. Two hooks with two notions of
"visible" was the second half of the same bug.
The panel also scrolls the selected row into view now, the way the
hierarchy has always done - an asset selected elsewhere has to be
findable in the list, test or no test. And step 21 learned that scene
is a property kind.
Load scene sets a flag on the World like Stop and Restart do; the swap
happens after the tick, and the last wish of a tick wins. Exactly one
thing crosses over: the SharedState, so global variables and global
machines keep what they know while everything local is built anew.
Loading the scene that is already running rebuilds it, which is what
"reset the level" means when the score must survive.
A run starts either from the start scene (a new project field, set from
the asset panel and marked with an arrow in the list) or from whatever
scene is open. The toggle next to Play says which question is asked and
remembers the answer in localStorage: it is a way of working, not part
of the game.
Sounds no longer end because a scene ends. A voice belongs to the
behavior that started it and stops right after On destroy, so the scene
swap is no longer a special case at all. For the enemy that plays its
explosion and destroys itself in the same breath there is now a
detached input on Play sound.
The scene selector stays usable while a run plays, because a double
click in the asset panel already did the same thing unguarded. It moves
what the editor looks at, not what the run plays.
The height cap used to be a constant 420, and the check wrote that number down.
Since #43 the cap is the canvas minus a tenth at either end, which is 640 on an
800px screen - so CI failed on a picker that is behaving exactly as intended.
The check now reads the budget off the screen and reports rect, budget and
screen height when it trips, so the next change to the margin cannot strand it
again.
Typing "variable" into the node palette landed on `Run behavior`, because the
filter weighed label, type and description alike and then sorted by category.
`searchRank` in `src/ui/search.ts` now ranks a match: exact label or type, then
prefix, then label, then type, and only last a hit that lives in the
description alone. Category and alphabet order within a rank. The searchable
`Menu` uses the same function, so searching means the same thing everywhere.
The `+` menu grew to fifteen entries in declaration order. It is built in two
alphabetical groups now -- the types that go into a scene, then the ones that
stay in the library -- with a separator between them, the split `isPlaceableType`
already draws. `ASSET_TYPES` itself stays untouched: it is also the
serialisation and check order. A search flattens the groups and drops the lines.
Popups were capped at fixed heights (320 menu, 260 palette, 340 pickers) while
the canvas had room to spare. `Popup.roomFor()` measures its own chrome and
takes the window height minus a tenth at each end. Placement clamps
asymmetrically: the bottom margin is enforced, the top is only a budget, so a
menu anchored near the top stays at its button instead of being pushed off it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Sounds follow the road fonts already paved: the file goes to the server
content-addressed under its sha256, the project keeps a named asset that
points at it, and nothing is ever re-encoded. The format is read from the
file's own signature, not from what the browser claims - the same .m4a
arrives as audio/mp4, audio/x-m4a or as nothing at all.
Playing is two nodes. "Play sound" takes the asset and may raise or lower
its volume; every call is its own voice, so two jumps in a row are two
jumps heard. Its loop input is a choice of three - from the asset, on,
off - because a checkbox has to spend one of its two positions on "not
set" and would lose a case: an asset that loops could then never be
played once.
Below the graph core stays browser-free: NodeContext hands out an asset
id, BehaviorSystem resolves it against the project and folds the asset's
own volume and loop into the call, and only viewport/SoundLibrary knows
about Web Audio. A run without a sink plays nothing instead of throwing.
The CI browser is muted, so the ear is no witness: the debug hook counts
started and stopped voices, and step 34 reads that counter.
Four issues that all reach for the same line: how the current state is
read and written. Splitting them into four commits would mean four
states of the tree that do not compile, so they arrive together.
#55 start state — the machine asset carries an optional `start` holding
a State node's id. Right-click a state to set or remove it, a ring in
the node's title says which one it is. A dead id (the state was deleted)
reads as "no start", so nothing has to clean up after a deletion.
#53 entered/exited — `On state entered` and `On state exited`, one exec
output per state. Every change now goes through a single `setState`
that reports it; BehaviorSystem delivers it after the exec flow that
caused it, `exited` of the old before `entered` of the new, round after
round until the queue is empty, with a cap that stops a ping-pong graph
with an endless-loop error. A local machine tells the instance that
switched, a global one tells every instance holding a matching event
node. The start state fires `entered` too: per instance for a local
machine, once per run for a global one.
#57 reading — `Current state` with `name` for showing and `id` for
comparing, `Is in state` with one boolean per state. Neither has a
`run`, so `read` learned `def` and `lookup`, and the catalog test knows
this third kind of node.
#54 seeing it — during a run the state machine tab marks every state
some instance is holding faintly, and the one held by the selected
scene object's instance strongly.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The renderer is taken off the ticker; the ticker itself keeps running, so
the clock, a live world and the panels still get their step. Drawing now
happens inside a 200 ms trailing window that every pointer, wheel and key
event, every Widget.invalidate/resize, every Stack.relayout, every
EditorState.emit and both library onChange hooks reopen. The window rather
than a single frame is what makes the ~70 direct redraw() calls that never
pass through invalidate() safe, along with repaints an event only starts:
a microtask, a texture upload, a resolved font.
Editor.movesByItself() keeps asking for every frame while Play or an editor
run is on, a time shader is on screen, or the particle or animation preview
is showing.
The FPS counter now counts drawn frames over the time the renderer was
awake. Idle seconds are not slow seconds, so it answers "how smooth while
it moves" and holds its last answer instead of falling to 0 at rest.
Measured on the built stand, 20 s per variant, CPU of all chromium
processes from /proc: kept awake 75.5 % of a core at 60 fps, render on
change 8.8 % with 0 frames drawn, awake again 83.9 % as an order probe.
No colour check can see a missing repaint: pixelAt extracts from the stage
on demand and is just as green on a screen frozen a minute ago. Step 33
watches the new __gm.renderStats() counter instead.
`Sqrt` already exists in the behavior graph (`data/Sqrt.ts`) and in particles
(`ParticleSqrt`); the shader was the only graph without it. `Pow` with exponent
0.5 does the same job, so this is convenience, not a missing capability -- a
radial falloff no longer needs a node whose purpose has to be read off its
parameter.
Like `Pow` it cannot go through `unary()`: GLSL leaves `sqrt` undefined below 0,
so the value is held there with `max(value, 0.0)`. That form takes a float
second argument for a `vec3` too, so unlike `pow` the exponent broadcast has no
counterpart here.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`Variables` pushed `Font` under the menu's mask edge: `locate` still finds the entry, a click there
lands on nothing, `__gm.fontPicker()` is null and reading `.captions` throws. Only step 31 carried
the scroll workaround, so step 30 fell. The helper moves to `lib/scaffold.mjs` as `pickAssetType()`
and both steps use it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`nodeUnder()` recomputed the node's face as `0 <= x <= w`, while the pointer events follow the
`hitArea`, which reaches `PORT_HIT_RADIUS` past the edge. An input port sits at exactly x = 0, so the
global point rounded back landed a hair outside: the drop counted as one on empty canvas, the node
palette opened and swallowed every later click – four checks of step 6 at once. `NodeView` now
answers from that same `hitArea`, and the step also checks that no palette is open.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The default of the Edge param and the address mode the shared sheet physically
carries were the same constant; changing the value would have turned every
sentinel with it and given each shader sprite a full copy of its texture.
SHEET_EDGE ('mirror', what TextureLibrary uploads) is now separate from
DEFAULT_EDGE ('clamp'). A graph whose Sample reads nothing but its own pixel
draws with the plain sheet whatever the Edge says, so the common shader costs
no second copy.
Without a shader the quad's coordinates never leave 0..1, and mirror-repeat
reads like clamp within a texel - so clamp is what a sprite shows anyway.
adminandClaude Opus 5