abgen

Standalone Decentraland asset-bundle converter + ab-cdn-compatible JIT server. The converter is a clean-room Rust reimplementation of the legacy asset-bundle-converter: GLB/GLTF content from any catalyst content server in, Unity AssetBundles layout-compatible with the production ab-cdn.decentraland.org CDN out (byte-level parity posture in PARITY.md). See Provenance.

Quick start

Prebuilt binaries ship two ways: release archives on the tagged GitHub releases, and npm - npx @dcl/abgen runs the server directly (linux x64/arm64, windows x64/arm64, macOS x64/arm64; the platform binary installs as an optionalDependency of @dcl/abgen, and require('@dcl/abgen').binPath() resolves it for embedding tools - see npm/abgen). From source:

git clone <this-repo> abgen && cd abgen
cargo build --release            # no postgres, no openssl, no protobuf
cargo build --release --examples # bundle dump tools (objdump/texdump/matdump/texcmp/texpng)
scripts/bootstrap-runtime.sh     # integrity-check the vendored runtime data

There is also a wasm build: crate/abgen-wasm/ compiles the converter to wasm32 — see its README.

Binaries

target/release/ after cargo build --release:

Bin	What
abgen	ab-cdn-compatible HTTP server: serves a corpus dir + JIT-converts misses
abgen-build	single-file local converter CLI (glb -> bundle, --expect-hash verify)
abgen-corpus	batch corpus builder: manifest / --entity-ids (add --fetch-missing to pull content from a catalyst) / --world <name>[,...] (resolve + fetch + convert a world; upstream via --worlds-url or ABGEN_WORLDS_URL) / --live-mode / --collection-urn / --from-reference; full-corpus runbook in docs/BATCH.md
abgen-verify	parity differ (ours vs reference bundles, ppm-bits, --tolerant)
abgen-lod	LOD lane CLI: bundle, compare, placements, assemble, atlas, simplify, generate
abgen-lambda	AWS Lambda handler (deployment event in, bundles + manifests to S3; from the lambda/ workspace member) - see lambda/README.md

Plus bundle-inspection examples (texdump, matdump, objdump, texcmp, texpng) under target/release/examples/.

Server

ABGEN_CATALYST_URL=https://peer.decentraland.org/content ./target/release/abgen
curl -s localhost:5147/health | jq .   # expect template_ok:true, templates_missing:[]

Served routes - the full asset-bundle delivery surface an explorer needs:

• GET /manifest/{entity}_{platform}.json - bundle manifest, JIT-converting on a miss
• GET /{version}/{cid}/{file} - versioned bundle blobs (JIT on miss)
• GET /LOD/{level}/{file} - classic per-scene LODs
• GET /lods-unity/manifests/* - ISS LOD descriptors
• POST /entities/active, POST /entities/versions - the asset-bundle registry index routes the explorer's registry client calls to resolve bundle versions
• GET /health, /readyz, /livez, /metrics, /ping

An explorer can point both its asset-bundle CDN base and its registry base at this one host (what the optimized-assets client flag targets). The unsigned registry surface from the in-tree dcl-contents crate is always mounted (POST /profiles, POST /profiles/metadata, GET /entities/status/{id}, GET /worlds/{world_name}/manifest), sourced from a connected content DB or proxied from ABGEN_CATALYST_URL. The signed/write registry routes (/denylist*, /queues/*, /registry, /flush-cache) do mount, but only when CONTENT_PG_CONNECTION_STRING is set in URL form - so a converter deployment that leaves it unset serves none of them. Setting it exposes admin surface: gate it with API_ADMIN_TOKEN and DENYLIST_MODERATORS, and do not put it on a public listener unauthenticated. Route-by-route parity, header semantics, and the DB-vs-proxy source selection: docs/ROUTES.md.

Tests

ABGEN_ROOT="$PWD" cargo test --workspace --lib -- --test-threads=1

--test-threads=1 is required: the lib tests share process-wide ABGEN_ROOT state.

wasm

crate/abgen-wasm/ compiles the converter lib (default features off) to wasm32-unknown-unknown behind a hand-rolled C ABI; the JS runtime in crate/abgen-wasm/js/ (worker pool + WebGPU bridge) is what a host page embeds. It is a plain cargo package with its own committed Cargo.lock, excluded from the workspace: nothing in CI or the release matrix ever needs a wasm toolchain (the pinned one lives in crate/abgen-wasm/toolchain/flake.nix). crate/abgen-wasm/README.md documents the build, the headless driver, the native-vs-wasm byte-parity gate and its decoder contract. Layout caveat: this repo places template/ at the repo root, while the wasm lane's helper scripts and the wasm32-only include_bytes! template paths in crate/src/builder/templates.rs assume the source layout where it is a sibling of the crate — an in-repo wasm32 build needs those relative paths bumped one level (../../template/ -> ../../../template/). No workspace target compiles for wasm32, so the divergence never touches CI.

Features

One compile-time feature flag: server (on by default) gates the abcdn HTTP server + registry stack (axum/tokio/sqlx and the dcl-contents crate); the abgen bin has required-features = ["server"]. Library consumers - lambda/, crate/abgen-native/, crate/abgen-wasm/ - build with default-features = false and get the converter without the server stack. Under the default features every capability below builds into every native binary; wasm32 builds target-gate the server, content-DB, and GPU code off automatically. Each capability is activated at runtime, not at build time.

Capability	Built	Enable at runtime
tokio/axum HTTP JIT server (abcdn module + abgen bin)	always (native targets)	run the abgen bin
catalyst content-DB index (sqlx/postgres, via the in-tree dcl-contents crate) for real timestamps + deployer on /entities/* and the unsigned registry routes (/profiles*, /entities/status/{id}, /worlds/{name}/manifest)	always	set a content-DB connection (CONTENT_PG_CONNECTION_STRING or POSTGRES_CONTENT_*); without one the built-in content-client fallback serves the index routes with timestamp: 0, empty deployer, and the registry routes proxy the upstream catalyst
CUDA GPU BC7/BC5 encode path (libcuda dlopen'd at runtime, no toolkit needed to build; PTX kernels vendored in crate/src/gpu/kernel.ptx, regen from crate/kernel-ptx/)	always	opt in per run with --gpu / ABGEN_GPU=1, backend pick via ABGEN_GPU_BACKEND=auto|cuda|wgpu|off
portable wgpu compute backend (Vulkan/Metal/DX12), same dispatch + qualification gate	always	ABGEN_GPU_BACKEND=wgpu (or auto, which tries CUDA first)

GPU backends self-qualify per device at enable time: the selected backend (auto tries CUDA, then wgpu) must reproduce the CPU BC7 encoder bit-for-bit on a probe matrix (sizes x srgb x perceptual x profile, full mip chains) or it is disqualified and --gpu/ABGEN_GPU=1 exits 2 instead of silently degrading - a qualified GPU run produces byte-identical output to the CPU path. ABGEN_GPU_QUALIFY=0 skips the probe.

The slim, async-free converter lib + CLIs are the wasm32-unknown-unknown target build (the server, content-DB, and GPU stacks are cfg(not(target_arch = "wasm32"))).

Toolchain

• rustc/cargo (edition 2021; the vendored draco_decoder is edition 2024)
• cc + a C++ toolchain (vendored crunch, libjpeg9c)
• cmake + make (draco_decoder builds upstream draco C++; missing cmake is the #1 build failure)

On NixOS: nix develop, or nix shell nixpkgs#cargo nixpkgs#rustc nixpkgs#gcc nixpkgs#gnumake nixpkgs#cmake nixpkgs#pkg-config. No openssl (ureq uses rustls), no protobuf. Optional: libturbojpeg (dlopen'd, or set TURBOJPEG_LIB=/path); without it JPEG decode falls back to the vendored libjpeg9c path - valid output but not byte-parity with production; startup logs turbojpeg_available.

Platform matrix

Platform	Build
Linux x86_64	native (nix develop or bare toolchain)
macOS (Apple Silicon)	nix develop (clang stdenv)
Windows	WSL2, checkout on the WSL filesystem

Runtime data

Vendored in-repo and sha256-pinned; a fresh clone is zero-config. scripts/bootstrap-runtime.sh verifies the sha256 of both sets and, when a payload is missing, prints the git-history commands to restore it - it never fetches or re-downloads:

• template/*.windows.bundle - 4 typetree-donor bundles read by builder.rs::load_template()
• crate/shader/scene_ignore_{windows,mac} - the compiled per-platform DCL/Scene shader bundles (the canonical ab-cdn URLs 404; the vendored copies are the source of truth, provenance in PROVENANCE.md). The server self-primes them on first request - no bucket pre-seeding step; the lit_ignore_*/texarray/linux shader names 404 by design (absent upstream too)

Environment variables

Server (crate/src/abcdn/config.rs)

Var	Default	Meaning
ABGEN_HTTP_HOST / ABGEN_HTTP_PORT	127.0.0.1 / 5147	bind address (legacy aliases HTTP_SERVER_HOST / HTTP_SERVER_PORT still honored)
ABGEN_OUT_ROOT	./data/ab-generator/out	corpus root served + JIT write-back target (may start empty)
ABGEN_CATALYST_URL	http://127.0.0.1:5141/content	content server (standalone: https://peer.decentraland.org/content)
ABGEN_CONTENT_DISK	unset	optional local content-store root (disk-first client)
ABGEN_CACHE_DIR	./abgen-serve-cache	JIT conversion cache
ABGEN_VERSION	v49	advertised AB version
ABGEN_MANIFEST_CONTENT_SERVER_URL	https://peer.decentraland.org/content	contentServerUrl stamped into manifests
ABGEN_ROOT	repo root	root containing template/
ABGEN_METRICS_BEARER_TOKEN	unset	if set, /metrics requires this bearer token
ABGEN_JIT_FAIL_TTL_S	60	failure negative-cache TTL for entity JIT builds; concurrent identical requests single-flight per {entity}:{platform}
ABGEN_JIT_CONTENT_DIGEST	0	freshness for content servers whose declared hashes are not content-addressed (the @dcl/sdk-commands preview server keys them by file path). Defaults to ON when ABGEN_CATALYST_URL points at a loopback host (a dev preview server), OFF otherwise; explicit values always win. When on, manifest requests re-download the entity's convertible content, sha256 it, and prune stale conversions when bytes changed under an unchanged hash. Changed non-glb files (textures/.bin) also invalidate every glb bundle of the entity. Debounced per entity via ABGEN_REVALIDATE_DEBOUNCE_S (default 2)
ABGEN_UPSTREAM_AB_CDN	unset	read-through to a production ab-cdn (e.g. https://ab-cdn.decentraland.org): manifest/bundle/LOD/ISS requests that 404 through every local and JIT lane are streamed from upstream without persisting anything locally. Lets a client point its whole optimized-assets base URL here (wearables/emotes keep working) while only local scene entities are built locally. Paths containing b64- preview ids are never proxied
RUST_LOG	abgen=info,tower_http=info	tracing filter
CONTENT_PG_CONNECTION_STRING (or POSTGRES_* parts)	unset	content DB; URL form also mounts the signed registry routes

S3 space cache (read-through + write-back)

Enabled when ABGEN_S3_BUCKET is non-empty (or ABGEN_USE_SPACE=1):

• credentials, first match wins: ABGEN_S3_ACCESS_KEY/ABGEN_S3_SECRET_KEY, AWS_ACCESS_KEY_ID/AWS_SECRET_ACCESS_KEY, or the ECS container role (AWS_CONTAINER_CREDENTIALS_RELATIVE_URI or _FULL_URI)
• session token: ABGEN_S3_SESSION_TOKEN or AWS_SESSION_TOKEN
• endpoint ABGEN_S3_ENDPOINT; region ABGEN_S3_REGION|AWS_REGION; ABGEN_S3_PATH_STYLE for path-style addressing
• ABGEN_S3_READ_ONLY=1 - read-through only; server write-back is refused
• ABGEN_FALLBACK_VERSION (default v41) - extra version prefix tried on space-cache lookups
• ABGEN_WORLDS_CONTENT_URL - worlds-content-server fallback for entity/content fetches that miss the primary source (default public worlds server; 0/off/empty disables)

Uploads carry each key family's production writer's object metadata, derived from the key by space::object_headers: bundles application/wasm + public,max-age=31536000,immutable (cdn-uploader's comma-joined spelling), scene sources (.js/.json/.crdt) the direct-upload spelling public, max-age=31536000, immutable, manifests (manifest/…) application/json + private, max-age=0, no-cache, ISS descriptors (lods-unity/manifests/…) public, max-age=31536000 (lod-generator-unity's spelling), .br keys public,no-transform,max-age=31536000,immutable + Content-Encoding: br (as cdn-uploader sets).

Asset-reuse mode (upstream converter parity)

ON by default, matching the ab-cdn deployment's asset-reuse naming from v49 onward: scene glb/gltf bundles use the upstream converter's canonical naming and shared bucket layout (applies to the JIT server and abgen-corpus). Set ABGEN_DEPS_DIGEST=0 to fall back to legacy {hash}_{platform} names and entity-scoped space keys — needed only for parity runs against pre-v49 reference trees (e.g. --live-mode sampling of v15–v41 vintages, whose manifests list non-digest names).

• glb/gltf bundles are named {hash}_{depsdigest}_{platform} — the digest is a 128-bit hash of the glb's resolved (file, hash) dependency pairs (.bin + textures), so a glb whose dependency set changes lands at a new name/key; textures stay {hash}_{platform}
• space keys move from entity-scoped {version}/{entity}/{bundle} to the shared {version}/assets/{bundle} layout (entity-scoped keys remain a read fallback)
• before building, the space is HEAD-probed at the canonical key; hits are listed in the entity manifest without rebuilding, so bundles are converted once across entities
• a glb whose deps can't be resolved is skipped (upstream skipped-assets semantics) unless ABGEN_MAGENTA_MISSING is on, in which case unresolvable deps are dropped from the digest and the build substitutes placeholder textures

Display-only files — autogenerated-thumbnail.png anywhere in the content map, and whatever metadata.display.navmapThumbnail points at — are dropped from the conversion input, matching the prod converter's -skippedHashes flag: no bundle, no manifest entry, and no contribution to the manifest exitCode (the client reads those files as raw bytes from the content server, so a bundle for them is a dead artifact). Applies to the JIT server and the lambda, independent of asset reuse.

Registry index eager build

On POST /entities/active|versions, entities missing a converted bundle are queued for conversion; the request waits up to the deadline, builds finish in the background. Knobs: ABGEN_INDEX_EAGER_BUILD (default on; 0/false/no disables), ABGEN_INDEX_BUILD_PLATFORMS (default windows,mac), ABGEN_INDEX_BUILD_CONCURRENCY (default: CPU count), ABGEN_INDEX_BUILD_DEADLINE_MS (default 20000), ABGEN_INDEX_BUILD_MAX_QUEUE (default 0 = unlimited; when exceeded, new builds are skipped).

LOD JIT lane

ABGEN_LOD_JIT=1 enables JIT LOD builds on GET /LOD/{0|1}/... misses; needs gltfpack (ABGEN_GLTFPACK or $PATH), fails closed without it. Knobs: ABGEN_LOD_MANIFEST_BUILDER (unset limits JIT to scenes with a published ISS descriptor), ABGEN_LOD_CACHE_DIR (default: ABGEN_CACHE_DIR), ABGEN_LOD_JIT_TIMEOUT_S / ABGEN_LOD_JIT_FAIL_TTL_S (defaults 600 / 3600), ABGEN_LOD_BUILD_CONCURRENCY (default 1). Builds stage in a per-build workdir; only gate-passed output is promoted into the serving root, so rejected bundles are never servable.

Conversion / parity knobs

• ABGEN_SHADER_BUNDLE - path to scene_ignore_windows (default crate/shader/scene_ignore_windows; siblings like scene_ignore_mac resolve from the same dir)
• ABGEN_CONTENT_ROOT - local sharded content store root (default ./content)
• compat: ABGEN_V38_COMPAT, ABGEN_V38_TIMESTAMP, ABGEN_COLLECTION_MODE, ABGEN_REAL_TEXTURES, ABGEN_MAGENTA_MISSING, ABGEN_JPEG_TURBO_BOX, ABGEN_JPEG_GLB_9C (live.rs::Proxy::new sets some process-wide; not thread-safe before the first build)
• debug/dev: ABGEN_BC7_CACHE, ABGEN_BC7_SCALAR, ABGEN_BC7_NO512, ABGEN_BC7_CAPTURE, ABGEN_LZ4_DUMP, ABGEN_TEST_CRN_OURS/_REF/_DUMP

Deployment

The container image is built straight from the pinned Nix flake - nix build .#dockerImage (dockerTools.buildLayeredImage, tini init, no base OS) - producing an image that runs abgen as an unprivileged user on :5147 with template/ and shader/ baked in and ABGEN_ROOT/ABGEN_OUT_ROOT preset. The image job in .github/workflows/release.yml builds that image and pushes it to ghcr.io/decentraland/abgen:<tag> (and :latest) on every v* tag. The workspace also ships an AWS Lambda consumer in lambda/ (the abgen-lambda bin); nix build .#lambdaImage produces its container image, which the lambda-image-push job pushes to ECR on the same tags - see lambda/README.md. The org services-pipeline that publishes the same service to quay (service-name abgen) is configured externally in the private decentraland/definitions repo, not in this tree. Cutting a release - tagging, what the pipeline does, and what to check afterwards: DEVELOPMENT.md.

Parity posture and clean-room provenance are documented separately: PARITY.md and PROVENANCE.md.

Provenance

This repository is a generated standalone export: the server crate maps to crate/ with a small overlay; the content component + unsigned registry surface ship as the shared crate/dcl-contents crate, copied verbatim from the internal tree. Mechanical differences from the internal source: signed/denylist/queue registry routes deleted; the server bin is abgen and the local build CLI is abgen-build.

License

Apache-2.0 - full text in LICENSE. Contributions are accepted under the same terms. Vendored third-party licenses and shader/template bundle provenance: PROVENANCE.md. Note: the lib sets #[global_allocator] mimalloc; any downstream embedding the lib inherits it.