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.

chore: release v0.41.0

yhhb7c2a4a937b5
diff --git a/CHANGELOG.md b/CHANGELOG.md
index e6dae965f..fdc358cae 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -14,8 +14,44 @@ published separately; it ships inside the editor.
## [Unreleased]
+## [0.41.0] - 2026-08-03
+
+The editor's own agent is the headline, and this is the release where a user can
+find out it exists: a conversation now lives with the project — its transcript and the
+model's memory of it — so quitting no longer throws away what the agent was told, earlier
+conversations can be picked back up, and a long one says out loud which of its runs have
+been folded out of the model's reach. You can hand it a picture, and re-ask a run in
+different words. All of it is documented, in both languages, for the first time.
+
+Underneath, two things that only show up after you ship: a game's own `.json` data is an
+asset now rather than a path you fetch and discover missing in the build, and a save
+written by an earlier build is carried into the durable directory this release moves saves
+to. Android's floor drops to 7.0.
+
### Added
+- **The agent's conversation outlives the window, and there is a guide to all of it.**
+ Quitting the editor threw away everything the agent had been told, and nothing led back
+ to yesterday's work. A conversation is kept with the project now, under
+ `<root>/.esengine/agent/` beside the autosaves and for the same reason — what was asked
+ and what it did are about THIS project. **Earlier conversations** lists them and picking
+ one up restores both halves: the transcript you read and the memory the model answers
+ from. You can drop or paste a **picture** into the composer (in a 2D editor "make it look
+ like this" is most of what anyone wants to say, and until now only the agent could show
+ pictures), and **re-ask** a run in different words — it opens the question prefilled
+ rather than firing the identical words again, since wanting to put it better is the usual
+ reason to run a turn twice.
+
+ A long conversation also stops forgetting in silence. Past its context window the oldest
+ runs fold away — what you asked survives word for word, the tool traffic does not — and
+ the transcript now says so on the line where it happened, because runs still on screen
+ that the model can no longer answer about read as a model ignoring you rather than one
+ that was made to forget.
+
+ None of this was documented anywhere: the only agent the guides knew about was MCP. There
+ is now a **Built-in Agent** guide in both languages, screenshot-led, including the table
+ of what it may do unasked — the tiers are drawn where Undo stops working.
+
- **A game's own data is an asset.** A `.json` that is not a Spine skeleton or a
DragonBones pair — a level table, a tuning file, dialogue — was not an asset at all:
no `.meta`, no uuid, nothing in the registry, so the only way to read one was to fetch
@@ -33,6 +69,43 @@ published separately; it ships inside the editor.
nothing that was already a skeleton becomes data; and the project's own configuration
(`package.json`, `tsconfig.json` and their kind) never becomes an asset.
+
+- **A UI node can say that it arrived on screen, or left it.** Everything a panel does when it
+ appears — play the entrance, start the timer, refresh the count — had no event to hang on,
+ because a node does not go away by itself: an ANCESTOR's `display` changed, and every node
+ under it stopped being drawn without being told. The only way to notice was to walk the parent
+ chain of every animated node, every frame, re-deriving what the layout pass had already
+ computed. `shown` and `hidden` are those two moments on the same channel as `click`, so the
+ answer is an EventBinding row rather than a system, and the whole subtree is told rather than
+ just the node whose field changed.
+
+ Visibility is read as the engine's own resolved bit, the same one rendering and hit-testing
+ use, so it cannot drift from what you can see. That bit is reachable one entity at a time,
+ which makes the scan the entire cost of the feature — so it is gated twice, and on a steady
+ frame it is nothing: no listener, no scan (`EntityEventQueue.hasListenersFor` makes that O(1)),
+ and no scan on a frame where no UINode was written and nothing moved in the hierarchy — the
+ same pair of signals the layout solve and the physics reconcile already gate on. Measured in
+ a real UI of 112 nodes: 0 reads per idle frame, one full pass on the frame a panel is toggled.
+
+### Changed
+
+- **Android's minimum is API 24 (Android 7.0), lowered from 29.** The floor was 29
+ because the font path called `AFontMatcher_create` and nothing below it could
+ answer; on 24 through 28 the engine now picks the font out of `/system/fonts`
+ itself, asking each candidate whether it has a glyph for the character rather
+ than trusting a family name. Read the Vulkan requirement alongside this: it is
+ unchanged and still `required="true"`, and it filters far more devices than the
+ API level does, so what 24 adds is Android 7/8-era hardware that has a Vulkan
+ driver — not every phone on those releases. The compatibility matrix now starts
+ at 24; emulators below API 28 may have no Vulkan, so those rows can report "no
+ data" rather than a verdict.
+
+ Which is also the honest limit on this: the `/system/fonts` path has not run on a
+ real 7-or-8-era device. The emulators that could have covered it are the ones with
+ no Vulkan driver, and the phones in the device lab are all newer. What is verified
+ is that the floor builds, installs and boots; if you ship to API 24–28, test text
+ rendering on real hardware before your players do.
+
### Fixed
- **An export's output is no longer mistaken for the project's content.** `dist/` and
@@ -77,39 +150,6 @@ published separately; it ships inside the editor.
screen for the second before the first byte report, right where the "did my click do
anything" doubt lives.
-### Added
-
-- **A UI node can say that it arrived on screen, or left it.** Everything a panel does when it
- appears — play the entrance, start the timer, refresh the count — had no event to hang on,
- because a node does not go away by itself: an ANCESTOR's `display` changed, and every node
- under it stopped being drawn without being told. The only way to notice was to walk the parent
- chain of every animated node, every frame, re-deriving what the layout pass had already
- computed. `shown` and `hidden` are those two moments on the same channel as `click`, so the
- answer is an EventBinding row rather than a system, and the whole subtree is told rather than
- just the node whose field changed.
-
- Visibility is read as the engine's own resolved bit, the same one rendering and hit-testing
- use, so it cannot drift from what you can see. That bit is reachable one entity at a time,
- which makes the scan the entire cost of the feature — so it is gated twice, and on a steady
- frame it is nothing: no listener, no scan (`EntityEventQueue.hasListenersFor` makes that O(1)),
- and no scan on a frame where no UINode was written and nothing moved in the hierarchy — the
- same pair of signals the layout solve and the physics reconcile already gate on. Measured in
- a real UI of 112 nodes: 0 reads per idle frame, one full pass on the frame a panel is toggled.
-
-### Changed
-
-- **Android's minimum is API 24 (Android 7.0), lowered from 29.** The floor was 29
- because the font path called `AFontMatcher_create` and nothing below it could
- answer; on 24 through 28 the engine now picks the font out of `/system/fonts`
- itself, asking each candidate whether it has a glyph for the character rather
- than trusting a family name. Read the Vulkan requirement alongside this: it is
- unchanged and still `required="true"`, and it filters far more devices than the
- API level does, so what 24 adds is Android 7/8-era hardware that has a Vulkan
- driver — not every phone on those releases. The compatibility matrix now starts
- at 24; emulators below API 28 may have no Vulkan, so those rows can report "no
- data" rather than a verdict.
-
-### Fixed
- **The mock-wasm test harness left every engine-calling plugin inert.** `bootMockApp` connected
the world but not the App, and `engineApi(app)` answers with the App's module — so plugins that
@@ -128,6 +168,13 @@ published separately; it ships inside the editor.
reclaimed instead of growing forever. Writes go through a temp file and a
rename, so a kill mid-write cannot truncate a save.
+ A save written by an earlier build is in the old directory, and moving where
+ storage reads would have lost it on exactly the version that exists to keep it.
+ So the first launch after the update reads the cache copy when the durable store
+ has nothing, and writes it through — once, and only when there is nothing durable
+ to prefer, so a stale copy can never overwrite newer progress. Nothing to do on
+ your side if you have already shipped a native build.
+
## [0.40.0] - 2026-08-02
The agent went from working to being pleasant to work with. It answers from the keyboard
diff --git a/desktop/package.json b/desktop/package.json
index 6be601992..433426aac 100644
--- a/desktop/package.json
+++ b/desktop/package.json
@@ -1,6 +1,6 @@
{
"name": "@estella/editor",
- "version": "0.40.0",
+ "version": "0.41.0",
"private": true,
"license": "Apache-2.0",
"description": "Estella Editor — visual scene editor for the Estella 2D WASM engine",