makepad/apps/arcade/README.md
Admin 0ac21bdc9d A real cast: 36 rigged characters across 3 shared rigs — and Kenney IS rigged
THE FINDING THAT CHANGES THE PREMISE: "1 rigged model in 4,442" was measured
with the BROKEN GLB probe (the gLUF magic bug). With the probe fixed the
library holds 36 rigged models across three rigs:
  41 joints — 9 KayKit heroes + undead, up to 95 clips
   7 joints — 22 KENNEY civilians (male/female a-f, orc, human, archer, shop
              employees, skaters, soldiers), 32 clips
   6 joints — 5 Kenney platformer characters, 25 clips
So a village can be populated with 22 visually distinct civilians TODAY, with
no third-party pack at all. Every conclusion drawn from that probe before it
was fixed needs re-checking, not just this one.

KayKit: 9 characters fetched (Adventurers + Skeletons), pinned by commit +
sha256, 37 MB, gitignored. CC0 verified by READING LICENSE.txt at each pinned
commit, recorded in the script header and CREDITS.toml.

THE SHARED RIG HOLDS ACROSS PACKS, proven rather than assumed: hashing the
joint-name list of all nine files yields the SAME digest — 41 joints, same
names, same order — despite two separate repositories. Skeleton clips are a
strict superset (95 = the adventurers' 76 + 19 undead extras: awaken,
resurrect, spawn, taunt). So a clip authored for the knight plays on the
skeleton warrior and one animation path drives the cast. A test pins this,
including that both packs are present, so a version bump cannot silently break
it.

The texture trap that cost the Kenney fetch three attempts does NOT apply:
KayKit GLBs EMBED their atlas (image/png in a bufferView), verified by parsing
all nine.

tests/rigged.rs parses all 36 rigged models through makepad_game_render::skin
— the loader the app actually runs — and asserts the index's joint and clip
counts match it. Deliberate: the index's own probe was wrong for the entire
library once and survived because the fixture encoded the same error.

find_cast groups by JOINT COUNT rather than pack, because the valuable fact is
cross-pack interchangeability. Cast states are the INTERSECTION, not the union
— advertising a state one member cannot perform is worse than a shorter list.
Added the state words the skeletons needed (spawn/resurrect/taunt/use):
Skeletons_Awaken_Floor previously matched nothing, so "an undead that rises
from the ground" was unfindable.

One bug found in its own work: casts_to_json emitted a doubled closing brace —
malformed JSON that still looked fine in a log. Fixed with a structural test.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-08-03 11:43:42 +02:00

5.3 KiB
Raw Permalink Blame History

Makepad Arcade

Networked AI game sandbox (plan: repo-root game.md). Run with:

cargo run -p makepad-arcade

Stock assets

Arcade ships with a searchable library of CC0 models and sounds. The binaries are not vendored in git — fetch them once:

./apps/arcade/download_assets.sh            # core packs, ~75 MB
./apps/arcade/download_assets.sh --packs=all   # everything, ~185 MB
./apps/arcade/download_assets.sh --list        # show packs, download nothing

The default fetches a core set (~1900 models) so a fresh clone isn't forced to pull the lot; --packs=all gets the full Kenney 3D catalogue (4669 models across 47 packs), and --packs=nature-kit,car-kit picks specific ones. Add the 556 sounds and the nine rigged KayKit characters and the full library is ~5000 searchable assets — of which 36 are rigged and animated.

Everything is gitignored and pinned — starter kits to exact GitHub commits, the rest to content-hashed kenney.nl URLs — and every file is sha256-verified, so a moved or tampered upstream fails loudly instead of silently changing the library. Downloads are sequential with a delay: this is someone else's bandwidth.

Mirroring

resources/MIRROR.toml lists every pack with its canonical URL, sha256, size and model count. Because these assets are CC0, anyone may re-host them, and that file is what makes a mirror reproducible and verifiable. Point at one with:

ARCADE_ASSET_MIRROR=https://your.host/assets ./apps/arcade/download_assets.sh
./apps/arcade/download_assets.sh --mirror=https://your.host/assets

A mirror is expected to serve <base>/<slug>.zip. The sha256 from MIRROR.toml is verified identically whichever host served the bytes — a mirror is never trusted more than upstream; the hash is the authority.

Everything degrades gracefully without them: the demo runs with primitive shapes, and tests that need real assets skip with a hint.

Audio is Ogg Vorbis. Every Kenney audio pack ships .ogg only — no WAV variant exists upstream — and this tree has no vorbis decoder. The sounds are therefore indexed and searchable by the AI but not yet playable; adding a decoder (or running download_assets.sh --transcode, which converts to WAV when ffmpeg is installed) is what closes that gap.

Finding assets

The library is queried by description, never by filename, via makepad-game-assets. The agent gets a find_model tool and a one-paragraph summary — it searches, it never receives the catalogue (5000 entries would not fit in a prompt, and the summary stays ~480 characters however large the library grows). Ids look like kenney/racing/vehicle-truck-yellow and are stable across re-downloads, because generated game code writes them.

Findability is built in three layers so it scales past 4000 models: per-pack theme curation (~55 rows, giving every model its setting), filename token parsing (free, and Kenney's names are systematic), and a hand-curated query-time synonym table (~240 rows) that applies to the whole catalogue at once. Item-level curation is spent only on the few hundred most-requested things. At the full 4,999-entry catalogue: index build ~120 ms, a search ~0.2 ms, ~2.1 MB of heap (release).

Query-side stemming means an inflected request still reaches a base-form alias ("smashing" → the smash alias), and when two entries tie on score the kind the query implies wins — an unqualified noun like "spaceship" is an object request, while "metal clang" or "win music" wants something audible.

Credits

Every asset here is CC0 (public domain). Attribution isn't required by the licence — we credit anyway, because these libraries exist because someone chose to give them away.

  • Kenneyhttps://kenney.nl/assets — 47 model packs totalling ~4,670 models (nature, city, castle, space, food, furniture, vehicles, dungeons, characters and more), the five starter kits, and all seven sound packs (impact, interface, sci-fi, music jingles, UI, RPG and digital audio). Thank you for the extraordinary breadth of free, consistent, genuinely usable game assets — this library is most of what Arcade can build with.
  • KayKit / Kay Lousberghttps://kaylousberg.itch.io/ — nine rigged and animated characters: five adventurers (Knight, Barbarian, Mage, Rogue, Rogue_Hooded) and four skeletons (Warrior, Mage, Rogue, Minion). All nine share one 41-joint skeleton, so an animation authored for any of them plays on every one — the adventurers ship 76 clips and the skeletons those same 76 plus 19 undead extras. Thank you for giving away rigs and animation sets, which are the expensive part.

Rigged characters

find_cast lists the animated casts. A cast is a set of characters sharing one skeleton, so any state works on any member and a part can be recast without touching animation code:

rig members clips notable states
41 joints (KayKit) 9 7695 block, dodge, hurt, dance, sleep, carry
7 joints (Kenney mini) 22 2532 walk, run, jump, attack, sit, wave
6 joints (Kenney platformer) 5 25 walk, run, jump, attack

Pick members from ONE cast for a scene, and use different members so a crowd isn't clones — the same rule that applies to props.

Machine-readable attribution lives in resources/CREDITS.toml so a published game package can carry credit with it.