Open a project and it just renders + agents get long-term memory

The second public release, polished around "trust"

This day shipped v0.2.0 — the second public release after open-sourcing, with the homepage's "what's new" highlights updated. A version number means more than incrementing a digit; it packages the improvements accumulated over a stretch into a node that can be referenced, downloaded and trusted. And nearly everything done around this release this day points at one keyword: trust. Opening a project and rendering it faithfully makes results trustworthy; agents having stable identities and remembering you makes collaborators trustworthy; materials draggable straight to the AI without repeated description makes interaction trustworthy; Windows development working and bloom no longer dead makes the platform trustworthy in more environments. What a product just stepping into public view and being seriously tried by more and more people most needs to build is exactly this trust — "it does what it says, it remembers, it's dependable everywhere." This day firmed up that trust spot by spot.

Open a project, and it just renders that project

This day fixed a conceptual awkwardness, and what it really fixed was a kind of "unpredictability." Play used to first decide how to render by a guess that "prefer treating it as a certain special form" — a heuristic that, in some cases, made opening the same project give you not itself but something it "guessed" it into. The problem wasn't whether the guess was right but that it made results unpredictable: clicking something, you couldn't be fully sure what you'd see. This day changed it to the most direct, plainest behavior — whatever project you open is rendered faithfully as that project, with no hidden "preferred mode" interposed. Correspondingly, the editor can open a specific on-disk game to run with a single command, with that game's path supplied by the outer host instead of hardcoded into the editor's libraries (making the editor cleaner and more reusable as a pure tool). "What you open is what you get" sounds obvious, but a tool genuinely achieving "you open something and you definitely get it, no surprises" is exactly the cornerstone of predictability — and predictability is the precondition of trust.

Consistent agent identities + long-term memory that remembers you

The collaboration team became more "human" and more "lasting" this day. On one hand, all agents adopted a consistent identity presentation — named as "a Chinese role plus an English name," with their personas filled in, so what you face is no longer a string of function modules but a team of companions each with a name, a personality and a role. On the other, and more crucially: they gained long-term memory. Each agent in the workbench got a structured "long-term memory" tab and a browsable diary; the system automatically distills what's worth remembering at the end of each turn; and it maintains a "global user profile" that gradually learns your preferences, habits and what you're working on as you spend time together. Why does memory matter so much? Because an assistant without memory starts from a blank page each time — you must repeatedly explain the same background and preferences, and collaboration forever stays at "first meeting." With long-term memory, these agents remember you and past decisions and lessons, so collaboration grows more effortless and more attuned the longer it lasts. For a platform centered on "long-term collaboration with a team of AIs to make games," letting the AI "remember" is the key leap from "tool" toward "companion."

Drag into chat from anywhere + asset-pipeline improvements

Interaction got smoother again this day: you can now drag things straight into chat across surfaces. You can pull a batch of items from the asset browser or various panels into the chat box, where they line up as a queue of "reference pills" handed to the AI together. This removes a very common clumsiness — previously, to have the AI "use these models, this texture, that object in the scene," you had to describe each in words, laborious and error-prone; now you select and drag, and the AI knows exactly which you mean. The asset pipeline kept improving alongside: delete support, drag-to-chat, and regenerated internal identifiers for a batch of assets to resolve collisions when importing materials — collisions that had previously kept some resources from loading inexplicably. Upgrading "showing the AI" from "describing in words" to "dragging by hand" seems a small interaction but greatly lowers the friction of passing concrete materials between human and AI. Less friction, smoother collaboration.

Windows desktop dev + engine fixes

Desktop development extended to Windows this day: you can do desktop development on Windows (via Git Bash), with a Windows-specific linking issue fixed, so Windows users can develop and debug the desktop app locally, not just the web version. This continues the recent sustained investment in Windows compatibility — carrying the Windows users who hold most of the desktop market from "web version only" further to "desktop development too." The engine fixed two real problems: bloom (the effect that bleeds a soft glow from bright areas) had been fully dead due to three stacked bugs and is now properly repaired, bringing back the atmosphere; and the large vertex indices common in big models (a format that can address more vertices), previously rejected outright at parse, are now preserved correctly, so complex large meshes import without erroring. The startup path also gained an opt-in: when stale engine build output is detected, you can have it auto-rebuild (the default is still to prompt you to handle it, to avoid triggering a rebuild unexpectedly). These all help the platform run steadily across more environments, more materials and more complex scenes.

Engine: 2D GPU instancing + spot-light shadows

The engine advanced a step each on 2D performance and lighting this day. On performance, "chunk-level GPU instancing" landed for sprites and tilemaps: 2D games often have massive repeated elements — screens of tiles, sheets of bullets, many identical small tiles — drawn one by one before, slowing down as the count grows; now identical elements in the same chunk can be merged into a bucket and drawn by the GPU in one pass, sharply cutting draw calls and qualitatively changing large-scale 2D performance. On lighting, spot-light shadows were added: a spot light is a directional, cone-shaped source like a flashlight or a stage spotlight, and it can now cast shadows correctly, giving underlying support for effects like "a beam sweeping through the dark with shadows dragging along." The sprite rendering pipeline was also unified — folding a special path previously written just for sprites into the generic rendering path, one fewer fork to maintain separately. Some of these make 2D run faster, some make light and shadow richer, some make the engine simpler, together tamping the 2D line another layer firmer.

What this day means

The version number v0.2.0 marks the fruit of the first stretch of consolidation after open-sourcing. And the day's most moving line is really the agents' long-term memory — it gives "collaborating with AI" a dimension of time for the first time. A companion who remembers you and a tool that treats you as a stranger each time are two entirely different relationships; the former understands you more the more you use it, the latter stays in place forever. Paired with the predictability of "open something and it renders," the low friction of "just drag it over," the broad coverage of "Windows can develop too," and the engine's spot-by-spot fixes and speedups, this day pushed the whole platform another big step toward "worth trusting and living with long-term." A tool that genuinely wants to accompany users for a long time must both do what it says and be dependable everywhere, and also remember and understand you — and this day pushed on both at once.