Skip to content

A fast 2D and 3D game engine — TypeScript SDK, C++/WebAssembly core, visual editor. Ship one project to web, desktop, WeChat MiniGames, playable ads, and native Android / iOS.

docs: v0.40.0 release notes

esenginec34884b2ed02
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 28a12eb08..4abb9a479 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -14,8 +14,64 @@ published separately; it ships inside the editor.
## [Unreleased]
+## [0.40.0] - 2026-08-02
+
+The agent went from working to being pleasant to work with. It answers from the keyboard
+now, shows you the values a batch is about to write rather than how many there are, keeps
+what you were typing when you close the drawer, and asks before it drops a conversation.
+It renders tables, which it turns out models reach for constantly when comparing a few
+entities.
+
+Underneath, the UI layout stopped re-deriving every node's box every frame: Yoga has
+tracked what changed all along, and resetting its nodes each pass was throwing that away.
+Editing one field now costs a quarter of what it did.
+
+The last part is the one that makes the rest keep: performance has a snapshot in the repo
+and a gate in CI, the way the API surface has had for a while. Building it turned up that
+eleven of the fifteen benchmark files had not run since the ECS moved directories — and
+that `vitest bench` skips a file it cannot resolve and still exits 0, so nothing had said
+so.
+
### Added
+- **Performance has a snapshot now, guarded the way the API surface is.** This repo gated
+ correctness hard — thousands of tests, pixel regressions, an API-surface snapshot, cycle
+ and layer checks — and performance not at all: fifteen benchmark files, no CI job running
+ any of them, no baseline anywhere. The UI layout re-derived every node's box every frame
+ for months inside exactly that gap. `tools/perf-guard.mjs --check` asserts two things
+ chosen to survive running on a different machine: **ratios** between benchmarks in the
+ same run (the CPU cancels out, and each ratio states an architectural invariant — "editing
+ a field is far cheaper than adding a node" *is* the incremental layout working), and
+ **coverage**, how many cases produced a number at all. Tolerance is 30% deliberately: it
+ cannot see a 5% slowdown, it is there for the 2×-and-up kind. Accept a change with
+ `--update`, the same as the API snapshot.
+
+- **The agent can be answered from the keyboard.** A passage that stops a run takes focus
+ when it arrives — that is where the keyboard belongs while a run waits on it — and Enter
+ is its primary action, Shift+Enter the for-this-run one, both labelled on the buttons
+ rather than left to be found. The batch preview also shows the *values* it wants to
+ write, as the same `before → after` the change set uses, where it used to say "3 fields"
+ — which is exactly the part a preview exists to show. Strike-all takes a batch apart in
+ one gesture, and an emptied one disables Apply instead of offering "Apply 0", which the
+ kernel would have treated as a decline anyway.
+
+- **The agent's answers render markdown tables.** Asked to compare a few entities or
+ field values a model reaches for one, and the fallback was a paragraph of raw pipes
+ wrapped in a 384px column — the least readable thing the renderer could produce. Column
+ alignment is honoured, an escaped `\|` stays content, a ragged row is fitted to the
+ header rather than shifting its cells, and cells go through the same inline parse as
+ prose, so a code span naming an entity is still a way into the scene. The table scrolls
+ inside its own box: the transcript never scrolls sideways as a whole. As everywhere
+ else in this parser, a half-arrived table is not an error — a header stays a paragraph
+ until its divider lands, so nothing flickers through being a narrower table first.
+
+- **A run that was cut off says so, and can be carried on.** The turn loop is bounded so a
+ model that keeps calling tools cannot spin forever, and reaching that bound was reported
+ as an ordinary end of turn — on screen, a run cut off mid-task looked exactly like one
+ that had finished, and the sentence that would have said otherwise is the one it never
+ got to write. Both endings a run could have gone further from, that and Stop, now offer
+ to carry on from where it stopped.
+
- **UI zoom, on Ctrl/Cmd +, − and 0.** On a 2K display the editor's numbers were too
small to read, and the one setting that could fix it sat in a dialog with no menu row,
no shortcut, and nothing on screen to say it existed. It is now a View-menu group and
@@ -26,34 +82,6 @@ published separately; it ships inside the editor.
to capture a zoomed shell), because Chromium persists per-origin zoom across restarts
and one left behind by an ordinary session would silently rebase every later capture.
-### Fixed
-
-- **UI zoom no longer blurs the viewport or misplaces what you click.** It was applied as
- a CSS `zoom` on the body, which leaves `devicePixelRatio` at 1: every canvas — the
- viewport, the profiler graphs, the waveforms, the node graphs — kept its backing store
- at the unzoomed size and was upscaled to fit, and `getBoundingClientRect` drifted a
- whole zoom factor away from `clientWidth`, so viewport picking and gizmo dragging
- missed by more the further right and down you aimed (at 150% the right third of the
- viewport could not be hit at all). Chromium applies the zoom now, and raises the dpr
- with it: measured at 80/100/125/150/200%, canvases stay pixel-exact and a dragged gizmo
- lands under the cursor.
-- **The viewport canvas is sized from its own layout box**, not `floor(clientWidth × dpr)`
- — which lost up to a pixel at fractional dpr, the case UI zoom makes ordinary.
-- **A message typed while the agent was working is no longer lost.** The composer said
- "this will be the next message"; the host refuses a send mid-turn, and nobody held it —
- so the box cleared, the send was rejected, and what you typed was gone, replaced by a
- red banner. It is queued now, shown in the transcript while it waits, and sent when the
- turn ends. Stopping the turn drops the queue, because stopping means stop.
-- **The agent's typing caret sits after the last character** instead of on its own line
- below the paragraph — it was rendered after the whole markdown block rather than inside
- it, which read as a stray blinking box rather than as text being written.
-- **A tool row's result column says what came back.** It showed the first ~20 characters
- of the raw result, which for a JSON answer is `[{"id` — and the disclosure under it was
- the *same* string, flattened and cut at 160 characters, so expanding a row taught you
- nothing. The row now carries a one-cell summary (a list is its length, an object leads
- with its first named string) and the disclosure keeps the full result with its line
- structure intact.
-
### Changed
- **The agent transcript animates the way the prototype does.** Runs, reasoning, tool
@@ -83,30 +111,6 @@ published separately; it ships inside the editor.
- **Reasoning shows how long it took**, a stopped or refused run explains itself in the
transcript rather than only in its header badge, `@` offers project assets alongside
entities, and re-ask is on the answer as well as on the run header.
-- **Performance has a snapshot now, guarded the way the API surface is.** This repo gated
- correctness hard — thousands of tests, pixel regressions, an API-surface snapshot, cycle
- and layer checks — and performance not at all: fifteen benchmark files, no CI job running
- any of them, no baseline anywhere. The UI layout re-derived every node's box every frame
- for months inside exactly that gap. `tools/perf-guard.mjs --check` asserts two things
- chosen to survive running on a different machine: **ratios** between benchmarks in the
- same run (the CPU cancels out, and each ratio states an architectural invariant — "editing
- a field is far cheaper than adding a node" *is* the incremental layout working), and
- **coverage**, how many cases produced a number at all. Tolerance is 30% deliberately: it
- cannot see a 5% slowdown, it is there for the 2×-and-up kind. Accept a change with
- `--update`, the same as the API snapshot.
-- **The benchmarks had not run since the ECS moved.** Eleven of the fifteen files still
- imported `../src/world` / `../src/query` / `../src/component` — the paths from before
- `src/ecs/` existed. `vitest bench` skips a file it cannot resolve and still exits 0, so
- the suite reported success while measuring nothing: 27 cases produced a number, 147 did
- not. It is **230** now, and the guard above fails if that number ever falls again. Fixing
- the rest turned up the API drift the silence had been hiding: `physics` instantiated
- through `locateFile`, which under the test environment fetches against localhost;
- `wasm-vs-world` re-declared Transform/Sprite/Camera locally and hand-wrote payloads that
- had all drifted (Camera's four viewport scalars are one `Vec4`; `Sprite` had gained five
- fields, so every embind case died on a missing one), and still measured `UIRect`, which
- `UINode` replaced; `spine` called `extractMeshBatches`, since replaced by the
- allocation-free `forEachMeshBatch(id, visitor)`, and submitted through a function renamed
- when that path was generalised to skeletal.
- **The UI layout keeps Yoga's work, not just its allocations.** Yoga already tracks what
changed — every style setter is `if (old != new) { set; markDirtyAndPropagate(); }`, and
a solve skips the subtrees still clean — but the layout pass reset every node and rebuilt
@@ -165,32 +169,49 @@ published separately; it ships inside the editor.
names it is not a smaller edit but a failed batch. Since authoring is one undo step, the
gate is neutral rather than a warning, and "Allow for this run" turns it off for the
rest of a long build.
-- **The agent can be answered from the keyboard.** A passage that stops a run takes focus
- when it arrives — that is where the keyboard belongs while a run waits on it — and Enter
- is its primary action, Shift+Enter the for-this-run one, both labelled on the buttons
- rather than left to be found. The batch preview also shows the *values* it wants to
- write, as the same `before → after` the change set uses, where it used to say "3 fields"
- — which is exactly the part a preview exists to show. Strike-all takes a batch apart in
- one gesture, and an emptied one disables Apply instead of offering "Apply 0", which the
- kernel would have treated as a decline anyway.
-- **The agent's answers render markdown tables.** Asked to compare a few entities or
- field values a model reaches for one, and the fallback was a paragraph of raw pipes
- wrapped in a 384px column — the least readable thing the renderer could produce. Column
- alignment is honoured, an escaped `\|` stays content, a ragged row is fitted to the
- header rather than shifting its cells, and cells go through the same inline parse as
- prose, so a code span naming an entity is still a way into the scene. The table scrolls
- inside its own box: the transcript never scrolls sideways as a whole. As everywhere
- else in this parser, a half-arrived table is not an error — a header stays a paragraph
- until its divider lands, so nothing flickers through being a narrower table first.
-- **A run that was cut off says so, and can be carried on.** The turn loop is bounded so a
- model that keeps calling tools cannot spin forever, and reaching that bound was reported
- as an ordinary end of turn — on screen, a run cut off mid-task looked exactly like one
- that had finished, and the sentence that would have said otherwise is the one it never
- got to write. Both endings a run could have gone further from, that and Stop, now offer
- to carry on from where it stopped.
-
### Fixed
+- **The benchmarks had not run since the ECS moved.** Eleven of the fifteen files still
+ imported `../src/world` / `../src/query` / `../src/component` — the paths from before
+ `src/ecs/` existed. `vitest bench` skips a file it cannot resolve and still exits 0, so
+ the suite reported success while measuring nothing: 27 cases produced a number, 147 did
+ not. It is **230** now, and the guard above fails if that number ever falls again. Fixing
+ the rest turned up the API drift the silence had been hiding: `physics` instantiated
+ through `locateFile`, which under the test environment fetches against localhost;
+ `wasm-vs-world` re-declared Transform/Sprite/Camera locally and hand-wrote payloads that
+ had all drifted (Camera's four viewport scalars are one `Vec4`; `Sprite` had gained five
+ fields, so every embind case died on a missing one), and still measured `UIRect`, which
+ `UINode` replaced; `spine` called `extractMeshBatches`, since replaced by the
+ allocation-free `forEachMeshBatch(id, visitor)`, and submitted through a function renamed
+ when that path was generalised to skeletal.
+
+- **UI zoom no longer blurs the viewport or misplaces what you click.** It was applied as
+ a CSS `zoom` on the body, which leaves `devicePixelRatio` at 1: every canvas — the
+ viewport, the profiler graphs, the waveforms, the node graphs — kept its backing store
+ at the unzoomed size and was upscaled to fit, and `getBoundingClientRect` drifted a
+ whole zoom factor away from `clientWidth`, so viewport picking and gizmo dragging
+ missed by more the further right and down you aimed (at 150% the right third of the
+ viewport could not be hit at all). Chromium applies the zoom now, and raises the dpr
+ with it: measured at 80/100/125/150/200%, canvases stay pixel-exact and a dragged gizmo
+ lands under the cursor.
+- **The viewport canvas is sized from its own layout box**, not `floor(clientWidth × dpr)`
+ — which lost up to a pixel at fractional dpr, the case UI zoom makes ordinary.
+- **A message typed while the agent was working is no longer lost.** The composer said
+ "this will be the next message"; the host refuses a send mid-turn, and nobody held it —
+ so the box cleared, the send was rejected, and what you typed was gone, replaced by a
+ red banner. It is queued now, shown in the transcript while it waits, and sent when the
+ turn ends. Stopping the turn drops the queue, because stopping means stop.
+- **The agent's typing caret sits after the last character** instead of on its own line
+ below the paragraph — it was rendered after the whole markdown block rather than inside
+ it, which read as a stray blinking box rather than as text being written.
+- **A tool row's result column says what came back.** It showed the first ~20 characters
+ of the raw result, which for a JSON answer is `[{"id` — and the disclosure under it was
+ the *same* string, flattened and cut at 160 characters, so expanding a row taught you
+ nothing. The row now carries a one-cell summary (a list is its length, an object leads
+ with its first named string) and the disclosure keeps the full result with its line
+ structure intact.
+
+
- **A reloaded window rejoins the agent conversation instead of coming back to an empty
drawer.** main owns the session, so it survives a reload; the transcript did not, and
nothing could ask for what it had missed. Worse and silent: re-asking a run passed the
@@ -2622,6 +2643,7 @@ not kept before this file was introduced — see the Git history at
commit on 2026-01-25.
[Unreleased]: https://github.com/esengine/estella/compare/v0.39.0...HEAD
+[0.40.0]: https://github.com/esengine/estella/compare/v0.39.0...v0.40.0
[0.39.0]: https://github.com/esengine/estella/compare/v0.38.0...v0.39.0
[0.38.0]: https://github.com/esengine/estella/compare/v0.37.0...v0.38.0
[0.37.0]: https://github.com/esengine/estella/compare/v0.36.0...v0.37.0