Files
UmaViewer/CLAUDE.md
T
ritenseki 546d8ef7ba Live: implement MirrorReflection — and it resolves the stage floor blow-out
Rebuilt Gallop.MirrorReflection from its original signature, so Unity deserializes the real
fields out of the bundle rather than anything being inferred. All 13 field names are copied
verbatim from the TypeTree, including the original's typo `_mirrorCliipPlaneOffset` — spelling
it correctly would silently receive no data.

The receiving end was looked up, not guessed. mtl_env_liveNNNNN_mirror000 uses
Cygames/MirrorAndShadow/ReceiveMirror, which declares exactly four properties — _MainTex,
_ReflectionTex, _Color, _ReflectionRate — and the component's _mirrorReflectionColor and
_mirrorReflectionRate line up with the last two name for name. Texture size, cull mask, clear
colour and clip-plane offset all come straight from the serialized values, which differ per
stage (256/12544 on live10149, 512/12288 on live10132), confirming real data is arriving.

Deliberately not implemented, because there is no ground truth for them:
  * _direction is an enum whose meaning is unrecoverable — IL2CPP metadata is encrypted.
    The mirror plane is derived from the mesh's own normals instead, which is geometry rather
    than a guess, and yields (0,1,0) on both stages. The field's value is logged as unused.
  * _mirrorDistortionTileOffset / _mirrorDistortionPower belong to the ReceiveDistortionMirror
    variant; these materials use the plain one.
  * _useMirrorTextureScale / _mirrorTextureScaleForBaseCamera — semantics unknown.

Rendering goes through RenderPipelineManager.beginCameraRendering and RenderSingleCamera.
Camera.Render() is not supported under SRP and would have failed silently.

This also settles the long-standing floor blow-out, and the diagnosis I had been running was
aimed at the wrong objects the entire time. ReceiveMirror's _ReflectionRate defaults to 1.0
while _ReflectionTex had never been assigned, and Unity substitutes a white texture for an
unbound sampler — a full-strength white reflection, i.e. a solid white floor. Binding the
reflection RT fixes it.

CLAUDE.md recorded the blown-out objects as plane_000 / stage_object_001 / specular_002 and I
took that as given, spending four rounds eliminating mechanisms on those three. The actual
culprit is mirror_a, which sits at the same position and scale under the same parent — visually
indistinguishable, entirely different name. The earlier "_EnvRate = 0 changes nothing" result
was consistent all along: it was toggling a different set of objects.

Both files gain the lesson: field values and shader property tables are hard data, but a
document's description of which object looks wrong is not, and a visual bug should start by
confirming which object the symptom belongs to.

Water surfaces are absent, not broken — none of the three downloaded stages has a ReceiveWater
material; each has exactly one ReceiveMirror.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-08-05 14:45:12 +08:00

18 KiB
Raw Blame History

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Project Overview

UmaViewer is a Unity application (2022.3.62f1) that loads and displays assets from Uma Musume: Pretty Derby. It reads encrypted asset bundles directly from the game installation without modifying game files.

Development

Open the project: Unity Hub → Add → C:\Users\riten\Desktop\UmaBuild Run: Open Assets/Scenes/Version2 scene → Play in Unity Editor Build: File → Build Settings → PC Standalone → Build

No CLI build commands — Unity Editor is required.

Game Data & Asset Pipeline

Game assets are encrypted. Key files:

  • meta (SQLite-MC ChaCha20 encrypted) — maps bundle names → hash filenames in dat/
  • dat/{2-char-prefix}/{hash} — encrypted asset bundles, each with a per-bundle XOR key from meta

Bundle decryption (UmaAssetBundleStream.cs): header (first 256 bytes) is plain; bytes ≥256 are XOR'd with FKey = ABKey[i] XOR bundleKey_bytes[j].

meta DB decryption: UmaDatabaseController.ReadMetaFromEncryptedDb() uses sqlite3mc_x64.dll with cipher=3 (ChaCha20). Final key = DBKey[i] XOR DBBaseKey[i % 13].

Asset bundle paths (key formats):

  • Live cutt timelines: cutt/cutt_son{music_id}/cutt_son{music_id}
  • Live effect prefabs: 3d/effect/live/pfb_{effectName}
  • Stage prefabs: 3d/env/live/live{bgId}/pfb_env_live{bgId}_controller000

Python tools (tools/) — rebuilt 2026-08-04 after the old UmaCrack\ toolchain was deleted. Run with ~/.venvs/umatools/bin/python3 (needs UnityPy + apsw-sqlite3mc, not plain apsw; see tools/requirements.txt).

Script Use
uma_common.py Library: key constants, decrypt_bundle_bytes(), MetaDb (path→hash, UnityPy load)
dump_meta.py plain → out/meta_plain.db · find <LIKE> · info <path> (hash, dat path, deps, [NOT DOWNLOADED]) · songs
dump_bundle.py list / tree / json / save / gameobjects, --deps
dump_cutt_typetree.py Authoritative track schema. --fields <song> · --tree <field> <song> · --sample <field> <song> · --groups
scan_live_tracks.py All 59 songs → out/scan_keys.csv, scan_groups.csv, scan_summary.txt, scan.json (~21 s)

Notes: meta decrypts via sqlite3mc PRAGMA cipher='chacha20' + PRAGMA hexkey='<64 hex>' (the key=x'...' and raw: forms both fail); VACUUM INTO re-encrypts the target, so dump_meta.py plain copies schema+rows explicitly.

Cutt bundles ship full TypeTrees, so field names and keyframe counts from these tools are ground truth. All 70 worksheets across all 59 songs share one 128-field schema — no version drift. Only son<id>_camera and type_01..type_10 carry the trees; the cutt_son<id> prefab and data bundles are stripped.

Getting a track's real field name without the dump tools — Unity matches bundle TypeTree fields by name, case-sensitively. A field whose name doesn't match is never populated (it does not become null — Unity initialises every serialised List<T> to an empty list, so null-vs-empty tells you nothing). To test a candidate name, declare it on LiveTimelineWorkSheet and check whether it receives keys:

LiveTimelineWorksheetDiag.Probe(sheet, "candidateFieldName");

Unsure about casing? Declare both spellings side by side — only the correct one fills. This is how WashLightList (capital W) was confirmed; note upstream katboi01 uses washLightList, which never populates.

Live Timeline Architecture

The Live playback system is in Assets/Scripts/umamusume/Gallop/Live/:

Director.cs              ← singleton, wires everything together
Cutt/
  LiveTimelineControl.cs ← dispatches timeline events each frame (~1900 lines)
  LiveTimelineWorkSheet.cs ← all track data fields (ScriptableObject)
  LiveTimelineDataList/   ← one .cs per track type
  UpdateInfo/             ← structs passed to event handlers
StageController.cs        ← manages stage GameObjects, StageObjectMap

Track implementation pattern (repeat for each new track):

  1. Add data class in LiveTimelineDataList/ (extend LiveTimelineKeyWithInterpolate for key, add group wrapper implementing ILiveTimelineGroupDataWithName)
  2. Add [SerializeField] public List<...> trackList; to LiveTimelineWorkSheet
  3. Add public event Action<TData, TKey> OnUpdateXxx; in LiveTimelineControl
  4. Call AlterUpdate_SimpleListControl(workSheet.trackList, d => d.keys, OnUpdateXxx, _currentFrame) in AlterLateUpdate()
  5. Subscribe _liveTimelineControl.OnUpdateXxx += handler in Director.InitializeTimeline()

Important: LiveTimelineKeyWithInterpolate already declares frame, attribute, interpolateType, curve, easingType — do NOT redeclare these in subclasses.

Stage shader colour channels (enumerated at runtime via Shader.GetPropertyType, confirmed on stage live10149 / song 1177):

shader colour brightness other
Gallop/3D/Live/Stage/DefaultNoAmbient _MulColor0 _ColorPower
Gallop/3D/Live/Stage/DefaultEnvMapNoAmbient _MulColor0 _ColorPower _AddColor, _EnvMap
Gallop/3D/Live/Stage/DefaultTransparentNoAmbient _MulColor0 _ColorPower _AmbientColor
Gallop/3D/Live/Stage/LightBlinkBlend _BlinkLightColor _ColorPower _SrcBlend/_DstBlend
Gallop/3D/Live/Stage/LightBlend / LightAdd1 _MulColor0, _MulColor1 _ColorPower, _ColorPowerMultiply
Gallop/3D/Live/Stage/Laser _MulColor0, _MulColor1 _ColorPower _RandomValue
Gallop/3D/Bg/BgMirrorBall _MulColor0 (none)
Gallop/3D/Live/Stage/StageMirrorBallShine (none)
Gallop/3D/Live/Stage/StageTransmittedLightMask (none) _TransmittedLightMaskScale

Brightness is a separate channel — do not fold it into the colour. Write the key's colour to the colour property and its power to _ColorPower; let the shader combine them. Only fall back to colour * power where _ColorPower is absent (BgMirrorBall).

Dump any shader's real property table with:

# tools/, via ~/.venvs/umatools/bin/python3
env = MetaDb().load("shader")          # 499 shaders
pf = obj.read_typetree()["m_ParsedForm"]   # m_Name, m_PropInfo.m_Props

_MulColor0 is on every stage shader and is BgColor1's channel; _AmbientColor belongs to BgColor2. They are separate properties on the same shader, so the two tracks must not write the same one. Same naming family as WashLight's MulColor0 and UVScrollLight's mulColor1.

⚠️ A material's m_SavedProperties in the bundle lists properties it no longer has — it retains entries from whatever shader it was authored against. Material.HasProperty() asks the current shader. Always enumerate the shader, never trust the bundle's saved-property table.

Track group names are not always GameObject names. Four separate silent failures were traced to this in one session, so check the mapping before assuming a track is unimplemented:

  • a whitelist filter in the dispatcher dropped 92% of BgColor1's keys (validBgColorNames)
  • StageObjectMap dedupes by name, so a crowd of 66 same-named objects only ever got its first member
  • the Object track's neonsign is a _stageObjectUnits unit name, not a GameObject — no object of that name exists
  • ObjectUpdateInfo.OffsetType was populated by the dispatcher but never read by StageController, so Add-relative positions were written as absolute

All four failed silently, with no error. Treat a track that "works" but looks wrong as a routing problem first.

StageObjectMap: Stage child GameObjects are indexed by name. Objects whose names match IsTimelineControlledLight() (blinklight, spotlight3d, _wash_, laser) start SetActive(false) — their handlers call SetActive(true) when the track fires. All other objects (neonsigns, glow meshes, UV scroll objects, etc.) keep their prefab default state.

Ground-truth sources: what is recoverable and what is not

Mapped 2026-08-05. Check this list before declaring something "unknown" — several things previously written off as un-reversible are in fact dumpable, and one whole class of problem was solved by realising the game's own scripts can be re-bound.

Available

来源 给出什么 怎么取
cutt bundle TypeTree 128 个轨道字段名 + 全部关键帧值,59 首全带 tools/dump_cutt_typetree.py
舞台 prefab TypeTree GameObject 层级、组件构成,以及缺失脚本 MonoBehaviour 的字段名和值(离线可读) MetaDb().load('3d/env/live/liveNNNNN/pfb_..._controller000')
shader bundle 499 个 shader 的真实属性表(名/类型/默认值/range) 枚举 m_ParsedForm.m_PropInfo.m_Props
material bundle 材质的 shader 绑定与存档属性(⚠ 存档表含过期条目,见下) sourceresources/3d/env/live/liveNNNNN/materials/mtl_*
AnimationClip 曲线的 path / 目标属性 / 关键帧时间与值 m_RotationCurves 等
MonoScript 类名 + namespace + 程序集名 见下,这是最重要的一条

⭐ 用原签名重建脚本,可直接拿到真实序列化字段

舞台 prefab 引用了 11 个脚本类,UmaViewer 起初只实现了 StageController,其余 892 个 MonoBehaviour 实例脚本为空。但这些脚本是可以「接管」的 —— Unity 按 类名 + namespace + 程序集名 解析 MonoScript,而 Assets/Scripts/umamusume.asmdef 的程序集名正是 bundle 里写的 umamusume。所以只要按签名建类,字段就会被真正反序列化进来。

已验证:新建 Gallop.Live.BillboardController 后,[StageScripts] 的「脚本丢失」 从 892 降到 728(正好 −164 个实例),_rotationType 等字段直接可读。

签名清单(Assets/Scripts/ 下任意位置均可,只要在 umamusume 程序集内):

namespace 类 live10149 实例数
Gallop.Live AnimationObjectController 282
Gallop.Live BillboardController ✅已建 164
Gallop.Live UnityLensFlareController 205
Gallop.Live WashLightController 27
Gallop.Live LightProjection 4
Gallop.RenderPipeline CustomLensFlare 200
Gallop.RenderPipeline MirrorBallProjector 4
Gallop.RenderPipeline CustomProjector 4
Gallop MirrorReflection 1
Gallop.Live.ShaderParam ShaderParamController 1

字段名必须和 TypeTree 逐字一致(大小写敏感),dump 一下即可。 这条同样适用于任何其它 bundle 里的 MonoBehaviour。

Blocked

IL2CPP 元数据拿不到。 游戏是 IL2CPP 构建,但 global-metadata.dat 既不在 umamusume_Data/il2cpp_data/(那里只有 Resources/),也没嵌在 GameAssembly.dll 里 —— 魔数 0xFAB11BAF 零命中,BillboardController / _rotationType 等标识符字符串 也全部搜不到。这是该游戏已知的反篡改,元数据运行时才还原。

因此以下永远拿不到标准答案,只能靠交叉验证或承认未知:

  • 所有枚举的语义 —— _rotationType(0/2)、LightBlendMode、CmnColorType0/1、 BlinkLight 的 pattern、ColorType 等。字段值读得到,含义读不到。
  • 脚本的行为本身 —— 重建签名只能拿到数据,拿不到原方法体。 例:MirrorBallLoopRotationSpeed 的值可读,但原版怎么用它不可读。
  • 速度/系数一类的常量 —— 如果原版在代码里对动画做了调速,那个倍率无从得知。
  • 未随包发布的 shader(PostFilm 的定义性 shader 就在 app 二进制里)。

⚠ 现象描述不是 ground truth

字段值、shader 属性表是硬数据;「哪个物件看起来不对」不是 —— 文档里的现象描述往往 没经过验证。地板发白排查栽在这上面:文档记着 plane_000 / stage_object_001 / specular_002 发白,整轮排查都围着这三个转,最后查明白的其实是叠在同一位置的 mirror_a(同在 pfb_env_live10149_main000 下、局部变换全为单位,肉眼分不出)。 它用 Cygames/MirrorAndShadow/ReceiveMirror,_ReflectionRate 默认 1.0 而 _ReflectionTex 从没赋值 —— Unity 对未绑定采样器代入白贴图,满强度白反射。

期间那个「把 _EnvRate 归零无变化」的实验因此也是误导:切的是另一批物件。

排查视觉问题的第一步应该是确认现象归属于哪个物件(让代码把可疑物件隔离/高亮), 而不是从文档里的名字出发。

间接手段(已被证明有效)

拿不到直接答案时,相关性交叉验证多次奏效,比猜测可靠得多:

  • BlinkLight 调色板槽位规则:靠「物件名 lightNNN_/alphaNNN_ 前缀的种数」与 「调色板里去重颜色数」在每一组都精确相等推出,7 组无一例外。
  • BgColor1 的通道 _MulColor0:靠枚举 shader 属性表 + 排除法确认。
  • 「某属性存在」≠「目标 shader 上有该属性」—— _Color 空写连续栽了三次 (BlinkLight / Spotlight3d / UVScrollLight),扫一遍 shader 属性表即可批量排查。

Working rule: no ground truth, no implementation

Field names, keyframe values and shader property tables are all recoverable from the bundles (tools/), so verify before you write code — and when something genuinely cannot be recovered, leave it unimplemented with a TODO rather than guessing a value.

Guessing has cost real time in this project: the PostFilm vignette geometry was guessed twice (full-screen wash, then a huge border) before it turned out the defining shader isn't shipped in any bundle. PostFilmRendererFeature._enableRendering is now false for exactly this reason.

Note the distinction that has caught me out: "the property exists" is not "this field maps to that property". _ColorPower is present on the stage shaders, but there is no evidence that BgColor1's power field drives it — so that write was reverted while BlinkLight's was kept (there the shader LightBlinkBlend is named for the track and pairs _BlinkLightColor with _ColorPower, which is a much tighter fit).

Open items and why each is blocked are tracked in LIVE_TRACKS.md → 「已知未解问题」.

Currently Implemented Tracks

Track Data class Notes
Camera (pos/lookat/fov/roll/switcher) existing fully working
CharaMotionSequence existing
Facial / LipSync / FormationOffset existing
GlobalLight (48) existing sets rim light shader props on characters
BgColor1 existing two branches. Groups named CharaCenter/CharaLeft/CharaRight/CharaColor (LiveTimelineControl.CharaBgColorNames) → character toon props (_CharaColor, _ToonDarkColor, …). Every other group name → stage renderers, resolved by walking the whole stage hierarchy for same-named Transforms (Director.ResolveStageTargets), writing _MulColor0 via MaterialPropertyBlock.
BgColor2 existing sets _AmbientColor on stage renderers via Lerp(color1, color2, value)
Transform / Object existing handled by StageController.cs
Effect (60) LiveTimelineEffectData loads prefab from 3d/effect/live/pfb_{name}
GlobalFog (49) LiveTimelineGlobalFogData sets RenderSettings.fog*
Spotlight3d (68) LiveTimelineSpotlight3dData looks up keyData.assetName in StageObjectMap
UVScrollLight (46) LiveTimelineUVScrollLightData sets texture UV on stage materials
VolumeLight (37) LiveTimelineVolumeLightData data deserialized, no visual (component absent)
LightShafts (50) LiveTimelineLightShaftsData data deserialized, no visual (component absent)
Particle (41) LiveTimelineParticleData sets ParticleSystem.emission.rateOverTime
ParticleGroup (42) LiveTimelineParticleGroupData sets FlickerLightRate
ChromaticAberration (73) LiveTimelineChromaticAberrationData URP ChromaticAberration.intensity ← power;per-channel offset 无法映射
HdrBloom (38) LiveTimelineHdrBloomData dead code — 0/58 songs carry data (hdrBloomKeys empty in every scan and in the song-1177 dump). Safe to delete.
ColorCorrection (61) LiveTimelineColorCorrectionData URP ColorAdjustments.saturation + ColorCurves RGB;depth/blend curve 无 URP 对应
BlinkLight LiveTimelineBlinkLightData SetActive + child Light color/intensity (56/58 songs). Biggest lighting payload by far — 全语料 19420 keys / 57 songs, and over half are named *_wash_* (wash_truss_a 147, wash_truss_b 147, wash_ground_a 51, …). color1Array / LightBlendMode / isReverseHueArray still unused.
WashLight LiveTimelineWashLightData SetActive only (5/58 songs). Low value: real "wash light" output comes from the BlinkLight *_wash_* groups above, not this track — song 1177 has just 1 group / 11 keys here.
Laser LiveTimelineLaserData SetActive + position/rotation/scale — blink/raycast not implemented (6/58 songs)

Track Coverage (from 58-song scan)

The pre-2026-08-04 scan data in LIVE_TRACKS.md was wrong and has been re-measured with tools/. Two systematic errors: a batch of tracks marked "WorkSheet 无 X 字段" do have the field (MobControlKeys, CyalumeControlKeys, nodeScaleList, sweatLocatorList, monitorCameraPos/LookAtKeys, facialNoiseKeys, charaMotionNoiseKeys, …), and a batch marked "全 0 keyframe 空占位" carry thousands of keys (postFilmKeys/postFilm2Keys/postFilm3Keys ≈20 200 keys over 59/59 songs, radialBlurKeys, tiltShiftKeys, charaFootLightKeys, …). Rows tagged ✅实测 are trustworthy; untagged rows are still old data. Re-verify with tools/ before acting on any untagged row.

Coverage numbers measure what exists in the bundles, not what the C# consumes — the two diverge. When a track looks under-powered, check the dispatcher in AlterUpdate_* for filters before assuming it's unimplemented (BgColor1 was dropping ~92% of its keys that way).

Measuring what a song actually carries: LiveTimelineWorksheetDiag.Dump() is called from Director.InitializeTimeline() and logs, per worksheet, every track field with its group names and key counts. Confirmed on song 1177: worksheetList.Count == 1, SheetType == MainLive, so reading worksheetList[0] loses nothing.