What is the most up-to-date open-source agentic game-dev stack in 2026 —
and how do we improve on PlayCanvas-for-3D + PixiJS-for-2D?

01Executive answer — the improved stack in one screen

One pick per layer. Where a layer legitimately has two answers (a zero-dependency default plus an optional dev-only add-on), both are named and the cost of the add-on is stated.

What changed versus the proposal being reviewed

02Improving on your proposal

Both halves of the original proposal are judged on fetched evidence, not on taste. One is replaced; one is kept and reinforced.

PlayCanvas-for-3D — replaced by Babylon.js 9.30.0

PixiJS-for-2D — kept, with the battery gaps named and priced

The dedicated 2D note answers this directly. Its verdict (§0): "Keep PixiJS 8.22.0 as the 2D renderer, and close every battery gap with a named MIT stack (planck/@dimforge/rapier2d, bitecs, @pixi/tilemap, @pixi/ui, @pixi/sound, pixi-filters) plus a real verification layer (pixelmatch + playwright + jsdom+@napi-rs/canvas)." 07-improve-2d.md §0

The honest summary of the whole review: the proposal's 2D half was directionally right but under-priced — it named a renderer and stopped. The 3D half was wrong, and wrong for a reason a feature matrix cannot show: a service contract, not a licence. 07-improve-2d.md 08-improve-3d.md §7

03Engine comparison table — live versions, dates, licences

Transcribed from 02-engines.md §2; every cell was fetched on 2026-10-09 (the note pins its live date from the Date: header of api.github.com/rate_limit). Licences were re-read from raw LICENSE files, not taken from GitHub's SPDX detection, which the note found wrong or ambiguous in five cases.

Adoption signals (last-month totals, fetched 2026-10-09)

three 78,674,917 · pixi.js 5,061,531 · phaser 1,754,552 · @babylonjs/core 1,623,073 · matter-js 1,343,879 · planck 722,224 · playcanvas 373,587 · @dimforge/rapier2d 88,746 · excalibur 39,522 · kaplay 33,800 · melonjs 7,071  // crates: bevy 7,878,259 total / 1,648,461 recent · fyrox 73,732 / 5,453 (tiny ecosystem)

The same packages on a last-week window (the basis 07-improve-2d.md used, so the numbers are not comparable to the month totals above): pixi.js 1,348,473 · phaser 479,112 · melonjs 1,652 · littlejs 3,299 · excalibur 11,558 · kaplay 9,312 · planck 204,123 · @dimforge/rapier2d 22,754 · matter-js 298,169 · bitecs 13,953 · pixi-filters 100,917 · @pixi/ui 21,143 · @pixi/sound 28,946 · @pixi/tilemap 5,468 · @napi-rs/canvas 28,876,913 · playwright 109,567,471. 07-improve-2d.md §2, §3

Install weight (npm dist.unpackedSize — a proxy for install weight and tree-shaking surface, not the runtime payload): phaser 107.3 MiB/3,593 files (112.5 MB on the 07 probe) · playcanvas 85.9 MiB/5,089 · pixi.js 72.7 MiB/3,745/10 deps (76.3 MB on the 07 probe) · @babylonjs/core 68.4 MiB/10,037 · three 19.5 MiB/1,263/0 deps · melonjs 9.0 MiB/650/0 deps (9.46 MB on the 07 probe). 02-engines.md §2 07-improve-2d.md §3

Ranked for an agent-driven build pipeline

Rubric from 02-engines.md §3: Text-only authoring [3] · Headless determinism [3] · Iteration latency [2] · 2D-action richness [2] · Performance [1] · Export breadth [1] · Maintenance evidence [2] · Docs for LLMs [1] → % of 75. Ties broken by text-only, then headless. The scores are the note author's judgement, each paired with the fetched evidence that moved it — the weights are stated so they can be re-weighted.

Per-scenario picks (verbatim from the source note)

• Web 2D → 1) Phaser 4.2.1 · 2) PixiJS 8.22.0 · 3) melonJS 20.8.0 · runner-up LittleJS 1.27.0. "Choose Phaser when the agent should write as little engine code as possible; choose PixiJS when the agent should own the architecture and you want the smallest, most deterministic install." refined by 07-improve-2d.md: for this repo PixiJS wins outright, because the swap cost is total and the headless story is the one thing Phaser cannot match.
• Web 3D → 1) Three.js 0.186.1 · 2) Babylon.js 9.30.0 · 3) PlayCanvas 2.23.1. superseded for a game building loop by 08-improve-3d.md, which re-orders this to Babylon → Three → PlayCanvas on headless evidence.
• Desktop 2D → Godot 4.7.2 · Bevy 0.20.0 · MonoGame 3.8.6 · raylib 6.0 · Defold 1.13.2 (only if the non-OSI Defold License is acceptable).
• Desktop 3D → Godot 4.7.2 · Bevy 0.20.0 · Stride · O3DE; avoid Flax for OSI reasons.
• Highest-ceiling open-source 3D → Godot 4.7.2; Bevy if you care about CPU-side scale. "There is no true AAA ceiling inside OSI-libre licensing."

04The 3D verdict in detail

This is the one layer where the proposal is wrong, and it is wrong for a reason that never appears in a feature matrix. The note's verdict (§0): "PlayCanvas is not the 3D layer for an agent-driven game-building pipeline… Choose Babylon.js 9.30.0 as the 3D engine (Apache-2.0, batteries included, NullEngine + Havok + glTF all verified working in Node with no GPU, official MCP authoring servers, ships its own TypeScript types) with Three.js 0.186.1 as the renderer-first alternative." 08-improve-3d.md

The PlayCanvas Editor licensing / ToS finding

Agent tooling exists but is editor- or session-bound: @playcanvas/editor-mcp-server@0.7.1 (2026-09-09) needs "an open Editor session"; third-party playcanvas-mcp@1.0.3 (2026-10-01) drives the REST API and needs PLAYCANVAS_ACCESS_TOKEN, i.e. a real account. MIT-editor optics are the trap here: nothing is a licence trap for the code; the trap is a service contract. 08-improve-3d.md §1, §7

Empirical headless results (all three engines booted in Node, no GPU)

The note author installed all three engines in a scratch directory (%TEMP%\gg3d, Node v24.21.0) and ran them. Raw run outputs: research/tavily/3d-harness-results.json.

Headless ranking: 1) Babylon (scene graph + animation + physics + glTF asset validation, one documented xhr2 polyfill, no DOM) → 2) Three (no logic harness; real offscreen WebGPU rendering is now achievable but only with hand-written self/canvas/device shims; the WebGL path is dead in Node) → 3) PlayCanvas (good logic harness, but asset loading needs jsdom and physics needs a wasm you supply, and its own docs disclaim browser-free rendering). For pixels, all three converge on the same answer: a headless browser pass (Playwright/Chrome) — identical work for all three. 08-improve-3d.md §2

Babylon.js vs Three.js vs PlayCanvas on the rest

Start-from-here package set (Babylon)

// tools/scene.mjs - prove the loop before any graphics work (from 08-improve-3d.md §6)
// npm i @babylonjs/core@9.30.0 @babylonjs/loaders@9.30.0 @babylonjs/havok@1.3.14 xhr2
import * as B from '@babylonjs/core';
import '@babylonjs/loaders';
import HavokPhysics from '@babylonjs/havok';
import xhr2 from 'xhr2';
import fs from 'node:fs';
import assert from 'node:assert';
globalThis.XMLHttpRequest = xhr2.XMLHttpRequest;      // documented Node requirement

export async function boot({ glb } = {}) {
  const engine = new B.NullEngine();                   // no GPU, no DOM
  const scene = new B.Scene(engine);
  const ground = B.MeshBuilder.CreateGround('ground', { width: 20, height: 20 }, scene);
  const box = B.MeshBuilder.CreateBox('box', { size: 1 }, scene);
  box.position.y = 4;

  const havok = await HavokPhysics({                                            // pass the wasm,
    wasmBinary: fs.readFileSync('node_modules/@babylonjs/havok/lib/esm/HavokPhysics.wasm') // not fetch()
  });
  scene.enablePhysics(new B.Vector3(0, -9.81, 0), new B.HavokPlugin(true, havok));
  new B.PhysicsAggregate(box, B.PhysicsShapeType.BOX, { mass: 1 }, scene);
  new B.PhysicsAggregate(ground, B.PhysicsShapeType.BOX, { mass: 0 }, scene);

  if (glb) await B.ImportMeshAsync(glb, scene);                                 // asset smoke test
  return { engine, scene, box };
}

export async function step(scene, frames = 120, dt = 1 / 60) {
  const zero = scene.getEngine().getDeltaTime();
  for (let i = 0; i < frames; i++) scene.render();                              // deterministic ticks
  return zero;
}

if (import.meta.url === `file://${process.argv[1].replace(/\\/g, '/')}`) {
  const { scene, box } = await boot();
  const y0 = box.position.y;
  await step(scene);
  assert.ok(box.position.y < y0, 'gravity did not apply');                     // logic assertion
  console.log('OK headless: y', y0.toFixed(2), '->', box.position.y.toFixed(2));
}

Then, in order: (1) this harness + npm test; (2) a Playwright spec that boots the same scene in Chromium and screenshots it (the pixel gate); (3) one glTF from the Khronos sample set as the asset smoke test; (4) only after that, gameplay. Keep assets out of git or use LFS — the zero-external-asset discipline of the 2D project does not survive 3D. 08-improve-3d.md §6

05The 2D verdict in detail, with the add-on stack

1. The decisive new finding: PixiJS v8 renders pixels in pure Node

The common assumption — "PixiJS v8 is WebGL2-only, so headless pixels need a browser or a compiled gl" — is false. PixiJS v8 ships a real CanvasRenderer (type: RendererType.CANVAS), its autoDetectRenderer order is webgpu | webgl | canvas and it will drop to Canvas2D; GlContextSystem tries getContext("webgl2") then falls back to getContext("webgl") (v8 supports WebGL1). Driving the canvas path with jsdom + @napi-rs/canvas produces engine-drawn pixels in plain Node with zero native build. 07-improve-2d.md §1.1 (row I), §1.2

2. What was actually executed (verbatim results in research/tavily/2d-harness-results.json)

3. Per-candidate headless verdict

4. The add-on stack: package + version + licence + what it closes

PixiJS is a renderer: everything below is code the agent must otherwise write and verify. Three packages are load-bearing — planck (or rapier2d), bitecs, and the verification pair around pixelmatch. 07-improve-2d.md §2, §6

What the agent still owns after this stack (and why that is acceptable): the game simulation (steering AI, wave budget, upgrade rarity/clash rules, projectile pooling, XP curve) — that is the actual depth of a roguelite and no engine hands it over; Tiled/TMX parsing (@pixi/tilemap renders, it does not parse); determinism discipline; and netcode (out of scope, single-player). 07-improve-2d.md §6

5. The alternatives, priced as decisions

• Phaser 4.2.1 — the "write less engine code" pick: arcade physics, tilemaps, tweens, particles, scenes, input, DOM UI and a loader out of the box; MIT; no editor. Costs: it is a framework, so adopting it is a rewrite to its Scene/GameObject model; v4.2.1's LLM corpus is still overwhelmingly Phaser 3, so an agent emits v3 API that no longer exists; the Node HEADLESS type gives logic only (verified), so pixels still need a browser; the package is 112.5 MB; cadence is the slowest of the active web engines (last push 2026-08-21).
• melonJS 20.8.0 — the only serious candidate that is both batteries-included and verified to render headlessly in pure Node: physics (SAT or matter/planck adapters), Tiled + glTF, audio, UI, cameras, tweens, particles, WebGPU→WebGL2→Canvas2D auto-fallback, ~250 KB minzip, 0 deps, 22 stable releases/6 mo, agent skills (npx skills add) and an llms.txt. Costs: 1,652 downloads/week → effectively no LLM training corpus; docs and examples assume the Tiled + browser workflow; the Node-Text interop bug.
• LittleJS 1.27.0 — 4 K-line, MIT, Canvas2D/WebGL, a real headlessMode (confirmed in source). Fine for a tiny prototype; too thin for an arena roguelite with 400+ entities, and no corpus.
• Excalibur 0.32.0 / Kaplay 3001.0.19 — pleasant APIs, BSD/MIT, but stable channels frozen since 2025-12-23 / 2025-06-15 while dozens of prereleases ship. Cadence risk for a build pipeline.
• Godot 4.7.2 web / Bevy 0.20.0 WASM / Defold 1.13.2 HTML5 — desktop-class engines that each impose a build/export step a no-build ESM repo does not have. Godot's web export is WebGL2-only and cannot export C#; Defold's licence is not OSI-approved; Bevy is the best of the three for CI (first-party headless_renderer.rs gives a real image diff) but pays a Rust compile tax per agent cycle.

6. Migration cost from this repo, with real counts

Measured over src/** on 2026-10-09: 07-improve-2d.md §5

Pure-renderer-bound code that must be re-authored in any other framework's drawing model — core/textures.js 181 + core/textures/shapes.js 410 + core/particles.js 660 + core/postfx.js 331 + core/camera.js 311 + ui/rendering.js 492 + ui/hud.js 512 + ui/dom-ui.js 437 + ui/dom-panels.js 266 + ui/overlay.css 434 + dev/fxlab.js 593 = ~4,627 LOC (38 %) — and that is a lower bound, because every entity's view/layer wiring inside src/game/** is written against Pixi Container/Sprite/Graphics. A swap would also invalidate tools/check.mjs's import-map and "no PixiJS-v7 API" checks, rewrite docs/SPEC.md §(b)(e)(f), and replace the index.html import map. 07-improve-2d.md §5

7. Determinism fit, and what this stack does not solve

Why determinism survives: PixiJS is only the renderer; the simulation stays the repo's own 60 Hz fixed-step loop over world.rng. Adding planck (explicit world.step(dt), no hidden clock) or bitecs (entity ids and iteration order are explicit arrays, not hash-map order) strengthens reproducibility rather than endangering it — provided RNG draws stay inside the sim and out of render. Costs: a physics engine adds a float path (planck is deterministic per build for the same step sequence; rapier2d's WASM floats are not guaranteed bit-identical across platforms, so pin it and keep gameplay RNG out of it); and the Node-canvas pixel path validates layout/geometry/HUD placement, not the bloom look. The recommended stack adds no build step and every package is either 0-dep or prebuilt, so the repo's "no bundler, no transpiler, import map" identity survives — with a vendor/-style copy per package, exactly as it already does for pixi.js. 07-improve-2d.md §7, §8

06Verification harness — the layered recommendation

Keep the three existing tools as the default npm test path and add layers around them, adding at most two dev-only dependencies. Layer order matters: each layer is cheap and fast, and a failure in an earlier layer invalidates the later ones. Layers 1–3 take seconds; 5–7 take minutes and belong on a schedule, not on every edit. 05-qa-harness.md §3

npm test
├─ 1  node --test tools/unit/            → pure logic: math, RNG, pool, waves, upgrades   [0 deps]
├─ 2  node tools/check.mjs               → static/policy/DOM/import contract              [0 deps]
├─ 3  node tools/smoke.mjs               → boot, schema, stateHash + rngDraws determinism [0 deps]
├─ 4  node tools/input.mjs               → real keyboard/mouse paths (NEW, dev-only dep)
├─ 5  node tools/shots.mjs               → content assertion + stability gate + diff     [+pixelmatch]
├─ 6  node tools/perf.mjs                → wall-clock frame pacing, heap, long tasks     [0 deps]
└─ 7  node tools/soak.mjs                → peak caps, pool return, heap growth           [0 deps]

engine-side pixel path (no browser, from 07-improve-2d.md)
└─    node tools/pixels.mjs              → PixiJS CanvasRenderer + jsdom + @napi-rs/canvas  [prebuilt]

The fix needs no baseline, no decoder and no new dependency: add a ?shotcheck=1 mode that draws the live canvas into a small 2D context, quantizes to a 16×16 grid, and writes {nonBackgroundRatio, distinctBuckets, gridHash} into #autotest-report. --dump-dom then yields a machine-readable assertion that the canvas is not blank, not one flat colour and not the wrong palette — exactly the bug class that passes today. Note that the 2D note's Node path uses the same idea natively: its probes returned {"nonBg":59644,"distinctColors":3,"ok":true} in the browser and {"nonBg":9600,"distinctColors":2} in Node, i.e. a content assertion with no baseline at all. 05-qa-harness.md §3 Layer 4 07-improve-2d.md §1.1

Anti-flake rules for canvas screenshots

1. Pin the rasterizer, not just "headless". --use-gl=angle --use-angle=swiftshader + --disable-gpu + --enable-unsafe-swiftshader (Chromium's documented SwANGLE recipe). The repo already passes the last two — it is on the right side of this. SwiftShader's automatic fallback is deprecated; the opt-in flag is required. The 2D harness confirmed the full recipe works here (row D).
2. Never compare across renderers. Baselines must carry the renderer + Chrome version in the name.
3. Multi-pass capture. Capture twice back-to-back and require zero changed pixels (includeAA: false) before accepting either capture — the manual equivalent of Playwright's "wait until two consecutive screenshots yield the same result".
4. Double-rAF before capture so the compositor has promoted the WebGL back buffer into the surface a screenshot reads; this replaces setTimeout(200) with a wait exactly as long as needed. In headless Chromium the timer-query/fence path is degenerate: SwiftShader's synchronous pipeline makes clientWaitSync report signalled immediately.
5. Lock the frame state. Fixed window size, deviceScaleFactor: 1, scale: "css", --force-color-profile=srgb, fixed seed. Do not copy --disable-lcd-text/--font-render-hinting=none from the text-site case study — they are font-antialiasing flags and are irrelevant to a canvas.
6. Mask or clip. Mask the HUD for arena shots and clip to regions, so one HUD change does not invalidate every baseline.
7. Assert content before diffing (see the ?shotcheck=1 mode above, and the Node pixel probe).
8. Thresholds are a sensitivity trade, not a fix. A threshold loose enough to absorb SwiftShader/ANGLE variance is also loose enough to hide a missing glow layer. Prefer deterministic rasterization + a tight threshold: start at threshold: 0.1 and a low-single-digit-percent maxDiffPixelRatio.
9. Quarantine by renderer, not by retry. If a shot flakes, capture it on the renderer CI uses — retries hide the mutation.
10. Do not add a WebKit project for the visual layer. There is an open Windows-WebKit bug (#42885, opened 2026-09-23) where page.screenshot() renders the page background where a WebGL canvas should be after setViewportSize() resizes the drawing buffer — Windows fails 5/5, macOS passes 5/5, with the GL context healthy and gl.readPixels() returning the correct colour. Chromium is the only browser where canvas capture is defensible today.

What is honestly not automatable headless

Machine-readable artifacts — the self-correction affordance

• Every tool gets --json, emitting one file under .tmp/report/ with a shared envelope: {layer, ok, seed, exitCode, findings:[{code, severity, file, line, msg, reproduce}], metrics:{}, artifacts:{}}.
• Every finding carries a reproduce command and the seed (node tools/shots.mjs --only=play-20s --seed=1234). That is the difference between an agent that can self-correct and one that can only report.
• Findings keep the existing path:line: message shape so they stay greppable and diffable.
• Keep exit codes stable and documented (0 pass, 1 findings, 2 harness/setup failure) so a caller can tell "the game is wrong" from "the harness is broken" — today smoke.mjs conflates a missing Chrome with a failed assertion.
• Extend the determinism fingerprint with stateHash (string), rngDraws (number), entityIdMax, frameSteps, computed over a sim world serialized sorted by entity id with floats quantized. Both parsers read the schema from docs/SPEC.md, so extending the doc extends both checks automatically. Write the seed + input log + final hash to .tmp/replay/<seed>.json.

05-qa-harness.md §3

07MCP / tool-interop picks for this repo

The picks for this repo

Windows note (from the official servers README): wrap npx entries as "command": "cmd", "args": ["/c", "npx", …]; uvx entries need no wrapper.

Adjacent standards

• AGENTS.md — 25k★, the standard repo states it is "used by over 60k open-source projects"; repo pushed 2026-09-10. The cheapest single thing to add so any harness reads docs/SPEC.md/docs/INTERFACES.md.
• Agent Skills (SKILL.md) — anthropics/skills ~180k★, pushed 2026-10-05; the highest-momentum of the three; now reaching engines (Unity ships unity skill install). Caveat: the repo has no root licence — the README says "many skills… are open source (Apache 2.0)" while the docx/pdf/pptx/xlsx skills are explicitly source-available, so check per skill before vendoring.
• A2A — 26k★, moved to AAIF Growth Stage on 2026-08-27 (so MCP, A2A, goose and AGENTS.md are now siblings under one foundation), a2a-sdk 1.2.2 (2026-10-05), 150+ organisations. Completely orthogonal to an agent building a single game: it links agents to each other, not to engines.

08Asset pipeline — keep procedural, and the traps to avoid

The verdict is unambiguous: keep the procedural, zero-asset approach as the default. Adopt generated assets only as an optional, build-time-baked, permissively-licensed polish layer — never as a runtime or build dependency. 04-asset-pipeline.md §Recommendation

The permissive tier (the only one to use, offline)

Licence traps, ordered by how likely an agent is to walk into them

1. Qwen-Image 2.1 — the biggest trap. The most-downloaded current image model (7,472,445 downloads on the ComfyUI repack, 2026-09-15) and its weights are under the "Qwen RESEARCH LICENSE AGREEMENT" (Release Date: September 20, 2026): "Non-Commercial shall mean for research or evaluation purposes only… You shall not use the Materials for any commercial purpose without obtaining a separate commercial license." An agent that picks "the top model" ships an asset it has no right to. Qwen-Image v1 is Apache-2.0; 2.1 is not.
2. FLUX dev lines. FLUX.2-dev / FLUX.1-dev / FLUX.2-klein-9B have HF cardData.license_name literally flux-non-commercial-license. Only FLUX.1-schnell, FLUX.2-klein-4B, FLUX.2-klein-base-4B and FLUX.2-small-decoder are Apache-2.0 in the family.
3. MusicGen and MMAudio weights are CC-BY-NC-4.0 while the AudioCraft/MMAudio code is MIT. The whole facebook/musicgen-* family (1,911,195 downloads on medium alone) is non-commercial for the weights.
4. The Stability AI Community Licence is revenue-gated, not free. Free only for "individuals or organizations that generate under $1M of annual revenue"; above that an Enterprise Licence is required. It is also revocable on violation. Covers SD3.5, SDXL Turbo, Stable Audio 3.0 and Stable Fast 3D.
5. Hunyuan3D is territory-limited. The Tencent Hunyuan 3D 2.1 Community Licence states in its second line: "THIS LICENSE AGREEMENT DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA." Covers Hunyuan3D-2, 2.1, 2mini, Omni and HunyuanWorld.
6. Non-OSI tool licences. Aseprite is not open source — its EULA.txt is an End-User Licence Agreement from Igara Studio S.A. ("licensed, not sold"), and GitHub reports no SPDX licence. Use LibreSprite (GPL-2.0) instead. pngquant is GPL-3.0-or-later (dual-licensed commercially): fine as a build tool, don't link it into a proprietary binary. ComfyUI is GPL-3.0 — Copyleft applies to ComfyUI, not to the images it produces, so never ship ComfyUI itself as a component.
7. Gated downloads are an automation blocker, not just a legal one. FLUX.2-dev, all stabilityai/* weight repos and Lightricks/LTX-Video return HTTP 401 without an authenticated HF account and often require accepting a licence plus name/email/country. An unattended agent pipeline will fail on these unless the token and gate acceptance are handled explicitly.
8. "Open weights" ≠ "open licence". Every family above is code-permissive (flux Apache-2.0, generative-models MIT, audiocraft MIT, LTX-Video repo Apache-2.0) while the weights are restricted. That is why the source note carries two licence columns.

AAppendix — headless pixel matrix (verified 2026-10-09)

The three 2D engines that 07-improve-2d.md §9 left UNVERIFIED (Excalibur 0.32.0, Kaplay, LittleJS 1.27.0) were run headlessly on 2026-10-09, so every 2D candidate in the table is now resolved: Excalibur and LittleJS render real pixels in both a GPU-free browser and plain Node, and Kaplay resolves the other way — browser pixels yes, Node impossible without a compiled WebGL. 11-headless-pixel-matrix.md

One correction for anyone installing the last row: the npm package is littlejsengine — registry.npmjs.org/littlejs returns {"error":"Not found"}. 11-headless-pixel-matrix.md §2.1

10The chosen stack, actually built — and what building it proved

Everything below was installed, executed and measured on this machine on 2026-10-09, not selected from a comparison table. The build changed two of the conclusions this report previously carried, confirmed one, found that one layer cannot deploy where it was assumed to go, and discovered a hard platform limit that only surfaced when the deploy was attempted. The numbers are what came back from the runs.

Five things the build changed or exposed

1. The repo's headless-Chrome recipe cannot detect a render loop at all, and one test of ours would have passed vacuously because of it. Appendix A §3 proves that the existing --dump-dom --virtual-time-budget flag set runs zero requestAnimationFrame callbacks (six flag variants tested), while still writing a non-zero-byte PNG — each engine draws one synchronous first frame. The 3D slice's first pixel test was written against that recipe, and the raw-GL control experiment showed why it was worthless: the readback returned the clear colour only (0 non-background pixels, mean luma 17.155 = exact background) because Babylon skips meshes whose shader is still compiling. It was rebuilt to drive Chrome over CDP with real wall-clock time and to assert that pixels came from a frame after the burst. Any agent-authored visual test must use a real pixel readback over a ticking clock; a PNG size check is not evidence. 11-headless-pixel-matrix.md §3 · VERIFY.md

2. A hosted image model cannot replace procedural sprites, and that is a property of the model, not a prompting failure. Muse obeyed hex colours exactly (a requested #808080 came back 128,128,128) and framing, but across 12 generations it never once produced a flat sprite — 20,098 to 285,038 distinct colours — and every file was colour type 2, i.e. no alpha channel is possible. The requested seamless floor did not tile (border deviation 149.1 vs 2.8 interior), and the endpoint exposes no seed, so runs are not reproducible. Keep sprites procedural; use hosted generation for textures, backdrops and UI only. 12-muse-assets.md

3. Generative 3D output is not deployable without a downscale step, and the platform limit is what proves it. The Pages deploy failed before uploading anything: Pages only supports files up to 25 MiB in size — assets/model.glb is 33.7 MiB. Note the trap: decimating the mesh is the intuitive fix and it is the wrong one — remesh cut geometry by 24× (1,136,438 → 28,788 tris) and the file fell only 26%, because it re-encoded the textures to PNG at 4K and made them 130% larger. 96% of the remaining bytes were textures. Downscaling the three maps to 1024² JPEG is what took it to 2.01 MiB. Measure the byte split before optimising. 15-meshy-lod.md · 16-asset-optimization.md

4. A low correlation was a broken measurement, not a broken asset — and the negative control is what proved it. The first analysis measured base-colour agreement with the original at r = 0.102 and honestly flagged possible corruption. Comparing whole images in atlas space is meaningless when the two maps are not aligned there. The tell was the control: deliberately mismatched atlases also collapse to r = 0.0946. Re-measured properly at surface points by mapping vertices back through their UVs, the shipped base colour agrees with the original at r = 0.9266, and with the LOD's own agreement at 0.9339. Every fidelity claim in this report should carry a negative control for this reason. 16-asset-optimization.md

5. Tool ergonomics are where an agent pipeline actually loses time. Three verified examples: the genmedia CLI silently drops the --flag=value equals form, so an identical prompt returned 2240×1120 instead of 2048×1152 (and silently cost a PNG-as-WebP); @babylonjs/core has no UMD build, so a no-build-step repo needs a deliberate one-shot vendor step rather than a CDN; and the cf CLI's OAuth token expires hourly, which is why an early wrangler call failed with 9109/10000 while cf auth whoami succeeded minutes earlier — and /user/tokens/verify rejects OAuth bearer tokens outright, so it is a red herring when debugging. 12-muse-assets.md · this build

What this build does not prove

No human or model has visually inspected any generated asset — the image-understanding tool was unavailable for the whole session (HTTP 429, weekly usage limit), so every art-quality statement here is numeric. The 3D slice ran on SwiftShader only (no hardware GPU), so its absolute frame rates are not a user-facing benchmark. Determinism is proven on this machine and between the page and Node (0 m difference over 240 steps), but not across OS or architecture. Colyseus on Cloudflare Containers is INFERRED from first-party documentation and an example, never executed. Nothing here was tested under load, and no long-run leak or stability test was performed. 12- · 13- · 14- · 15- · 16-*.md, VERIFY.md

09Sources & evidence — every claim linkable, with honesty about gaps

Each claim above is traceable to one of sixteen research notes. The notes are not published alongside this page, so they are shown as filenames in text; the external URLs below are the real links (all fetched/observed 2026-10-09).

External evidence index (the real links)

registry.npmjs.org/pixi.js · registry.npmjs.org/planck · registry.npmjs.org/@dimforge/rapier2d · registry.npmjs.org/bitecs · registry.npmjs.org/@napi-rs/canvas · registry.npmjs.org/@pixi/ui · registry.npmjs.org/pixi-filters · registry.npmjs.org/melonjs · registry.npmjs.org/@babylonjs/core · registry.npmjs.org/playcanvas · registry.npmjs.org/three · api.npmjs.org/downloads (playcanvas) · api.npmjs.org/downloads (@babylonjs/core) · api.npmjs.org/downloads (pixi.js, month)
playcanvas.com/terms · PlayCanvas: engine standalone · developer.playcanvas.com/llms.txt · playcanvas/editor README · playcanvas null-graphics-device source
doc.babylonjs.com — server-side / NullEngine · Babylon WebGPU · Babylon MCP servers · @babylonjs/core 9.30.0 NullEngine d.ts
three 0.186.1 WebGLRenderer source · gl 8.1.6 README (WebGL 1.0.3 only) · threejs.org/docs/llms.txt
Godot releases.atom · Godot command-line tutorial · Godot web export (WebGL2 only) · crates.io bevy · Defold bob manual · Defold AI-agents position
registry.npmjs.org/pixelmatch · registry.npmjs.org/vitest · Spector.js · fast-check model-based testing · registry.npmjs.org/@modelcontextprotocol/sdk · pypi.org/pypi/mcp · official MCP registry · registry about (preview)
chrome-devtools-mcp · playwright-mcp README (a11y-tree, not pixels) · Playwright release notes · Playwright snapshots · Playwright issue #42885 (Windows WebKit blank canvas)
Chromium SwiftShader doc · Chrome headless CLI (virtual time) · Web3D Survey: timer query support · Spector.js README
Z-Image-Turbo · FLUX.2-klein-4B · Qwen-Image 2.1 licence (non-commercial) · Hunyuan3D-2.1 licence (EU/UK/KR excluded) · Stability AI Community Licence ($1M gate) · ACE-Step 1.5 · TRELLIS.2-4B · UniRig · ComfyUI (GPL-3.0) · Aseprite EULA
AAIF formation (2025-12-09) · Unity CLI replaces its MCP server · Unreal MCP plugin (Experimental) · Blender Lab MCP server · AGENTS.md · A2A · anthropics/skills

Confidence and gaps — nothing UNVERIFIED is presented as fact