The scenes were never destroyed. Pressing Play with only Bootstrap open means
GameBootstrap loads Systems and the level at runtime, and Unity discards
runtime-loaded scenes when Play ends, restoring whatever the Editor had open —
Bootstrap alone. The Hierarchy comes back nearly empty. Every scene file was
byte-identical to HEAD throughout, still text.
Tools > Nightclub Arcadia > Open Game Scenes opens all three with the level
active, which is the authoring setup. With them already open, GameBootstrap's
LoadIfNeeded skips them and they survive Play. GameSceneWorkflow also captures
the scene setup before Play and restores it afterwards as a safety net, skipping
the restore when Unity already got it right and ignoring scenes deleted meanwhile
— an exception in that callback would leave the Editor in a worse state than the
problem it fixes. Toggleable under the same menu.
Also fixes `make lock`, which had a self-matching bug since Phase 1: `pgrep -f`
on the full binary path matched the very shell running the check, because the
pattern appears in its own command line. It reported the Editor as running when
nothing was, and would have blocked every make test/build once anything else
matched. Now `pgrep -x Unity`, which matches the process name and cannot match
the shell.
EditMode 42/42, PlayMode 6/6, YarnCheck 9 files / 35 nodes.
Co-Authored-By: Claude Opus 5 <[email protected]>
Fixes the "no Yarn Project has been configured" error coming back after running
the setup menus.
Last commit fixed the wrong layer. Retargeting YarnDemoSetup at a sandbox
stopped it destroying the level, but left the deeper problem: all of these
builders assume the single-scene layout, and after the split there is no target
path that makes them correct.
Two failures, both observed rather than reasoned about. Running the Yarn setup
left a generated scene open whose DialogueRunner had no project and auto-started
— that is the error the user kept seeing, coming from Dev/YarnSandbox.unity, not
from Systems. And retargeting the other three at Systems made PlayerControlSetup
pull navigation in with it; the baked NavMeshData is embedded rather than an
asset, so Systems.unity silently turned binary, 61KB text to 109KB binary. That
is trap 1 recurring on a scene that had just been fixed.
So the three scene builders now return early through LegacySceneSetup.Blocked,
with a message saying what they would have done and what to do instead. They are
not deleted: UILayerSetup alone is 818 lines that still record how the UI layer
is wired, and that record matters until prefabs replace it in Phase 5.
Set Up Skill System keeps its useful half. It regenerates the SkillDefinition
assets from JSON, which touches no scene; only the WireWorkingLevelScene call is
dropped, because those objects already exist in Systems.unity. Set Up Candidate
System was always asset-only and is untouched.
While diagnosing, YarnDemoSetup's own bug was found and fixed even though the
menu is now blocked: it set yarnProject through a SerializedObject on a prefab
instance without recording the override, so the assignment was dropped on save.
It now records the modification and refuses to save a scene whose runner has no
project.
Verified by running all five menus headlessly: the two asset generators run and
produce no diff, the three builders refuse, and no scene or asset changes.
Systems.unity restored to text, the level scene intact at 5 roots.
EditMode 42/42, PlayMode 6/6, YarnCheck 9 files / 35 nodes.
Co-Authored-By: Claude Opus 5 <[email protected]>
Fixes "Can't start dialogue: no Yarn Project has been configured."
Cause, and it was mine. Phase 3.7 said "retarget YarnDemoSetup.cs or retire it";
I retargeted it. But that menu does not modify a scene — it builds one from
scratch and calls SaveScene over its target. Pointing it at
SC101_ConferenceHall.unity armed it at the real level, and running it replaced
the scene with a bare Yarn demo: ten of fourteen roots gone, including the
player, the lighting and the navmesh. The leftover Dialogue System prefab
instance it left behind had no yarnProject override, auto-started, and produced
the error.
The scene was committed, so nothing was lost; it is restored from HEAD at 14
roots with its sun reference intact.
The menu now targets Assets/Scenes/Dev/YarnSandbox.unity and refuses to run if
its target is not under Scenes/Dev, via ScenePaths.IsDisposable. It is renamed
to "Set Up Yarn Sandbox Scene" so the menu says what it does. The other three
setup menus open and modify their target rather than regenerating it, so they
were never affected and are unchanged.
The test suite could not have caught this, because it asserted only that a
DialogueRunner existed. TheDialogueRunner_HasItsYarnProject now asserts there is
exactly one runner across the loaded scenes and that its project is set and
compiled — which reproduces the failure exactly when the scene is broken.
Also here: four UI prefabs pick up m_EditorClassIdentifier changes from
Assembly-CSharp to NightclubArcadia.Game, written by Unity after the assembly
split in the previous commit. Expected and correct.
EditMode 42/42, PlayMode 6/6, YarnCheck 9 files / 35 nodes.
Co-Authored-By: Claude Opus 5 <[email protected]>
No project code is in Assembly-CSharp any more. Five assemblies: Core (no
dependencies), Locomotion.Math, Skills, Game (everything else under Scripts),
and StarterAssets.
Game is one assembly rather than one per area, which is what the plan assumed.
Measuring the dependency graph first found three cycles, all through UI —
CharacterPanelView reaches into Dialogue, PlayerControlLock into Player, and
UILayerBootstrap into Cinematics — plus two edges a using-scan cannot see,
because NPCStandIn is in the global namespace. Assemblies cannot be circular, so
splitting further means relocating those three files. That is a code-movement
task, not an asmdef task, and nothing needs it yet.
StarterAssets had to get an assembly of its own. It had none, so it lived in
Assembly-CSharp, and an asmdef assembly cannot reference the predefined
assemblies — four runtime files use it, and all four would have stopped
compiling.
The payoff is immediate: BootstrapTests no longer needs reflection to reach
NPCStandIn and CinemachineFollowsPlayer, which is exactly why Phase 3 wanted
this.
Assets/Editor/PlayerBuild.cs is the headless build entry point. It reads the
enabled scenes, asserts Bootstrap is scene 0 because the player opens scene 0 on
launch, honours the -buildOutput the CLI forwards from -o, and calls
EditorApplication.Exit(1) on anything short of Succeeded — without which Unity
exits 0 on a build that produced nothing and CI goes green on it. Verified end to
end: a 172MB .app in about two minutes.
Two renames, both different from what the plan described. DialougueSkillComparison.cs
contains a class called SkillFunctions — the filename never matched the type, so
it is now SkillFunctions.cs. The two .yarn files with spaces in their names lost
the spaces rather than becoming kebab-case, which would have made them
inconsistent with Bartender.yarn and SC101.yarn; spaces were the actual problem.
Both renames preserved their .meta GUIDs, and the yarnproject graph keys and the
two character sheets were updated.
EditMode 42/42, PlayMode 5/5, YarnCheck 9 files / 35 nodes, make build green.
Co-Authored-By: Claude Opus 5 <[email protected]>
The game now starts from Bootstrap.unity, which additively loads Systems (the
dialogue runner, skills, UI layer, camera rig and the player) and then a level
(geometry, light, navmesh, reveal cameras). The level becomes the active scene so
new objects and lighting land there.
The player moved into Systems. It had been parented under Room_101 — under level
geometry — and it is persistent content, not level content. Moving it also
removes most of the cross-scene breakage on its own.
Unity nulls any serialized reference that crosses a scene boundary. Measured
rather than guessed: the pre-split scene was pulled from git and its null
references diffed against the split result. Exactly seven were lost — both NPCs'
dialogueRunner, dialogueUI and player, plus the reveal camera's tracking target.
A first naive audit reported 210, which turned out to be pre-existing Unity
defaults like Image.m_Material; the baseline diff is what separated the two.
Each of the seven now resolves at runtime. SceneServices.Resolve fills in a null
Inspector reference by searching the loaded scenes, keeping an assigned value if
there is one; CinemachineFollowsPlayer binds the reveal camera once the player
exists. Both follow the pattern already used by CameraFramingVolume and by
NPCStandIn's player lookup, rather than introducing a new one.
Assets/Tests/PlayMode is new, and it immediately paid for itself. BootstrapTests
loads Bootstrap and asserts the whole game comes up; on its first run it caught a
real bug the split had introduced. The player's NavMeshAgent lives in Systems
while the NavMesh is baked into the level, so during load the agent exists
off-mesh and ResetPath logs an error. ClickToMoveController now guards on
agent.isOnNavMesh rather than a bare null check. No EditMode test could have seen
that, because none of it has run yet.
Two of those checks reach their types by name through reflection: NPCStandIn and
the Cinematics namespace live in Assembly-CSharp, and an asmdef test assembly
cannot reference the predefined assemblies. Per-area asmdefs remove the need.
Build settings list all three scenes with Bootstrap at index 0, which
LoadSceneAsync by name requires.
EditMode 42/42, PlayMode 5/5, YarnCheck 9 files / 35 nodes.
Co-Authored-By: Claude Opus 5 <[email protected]>
The invocations in CLAUDE.md were correct but long, and retyping them is how
flags get dropped. `make check` is the gate a change should pass: Yarn compiles,
voice sheets are in sync, EditMode is green.
`make test` and `make build` depend on a `lock` target that fails if the Editor
has the project open, which removes the most common confusing failure. It
filters out AssetImportWorker children, which are not a second Editor.
tools/ci/report_tests.py parses the NUnit results and exits non-zero on failure,
because Unity documents no common exit-code definition across the components
under test — the XML is the authority, not $?.
`make merge-driver` configures UnityYAMLMerge for the .gitattributes rules added
earlier. Note the binary lives in the Editor bundle under Contents/Helpers, not
Contents/Tools as most guides say; there is no Tools directory in Unity 6 on
macOS.
Also ignores the local .test-results/ and build/ output.
Co-Authored-By: Claude Opus 5 <[email protected]>
CLAUDE.md gains the writing/ layout, the voice-sheet pipeline and its direction
of truth, the sync commands, and a pointer that Ground Truth must never be
copied into this repo. The writer-owned/code-owned table now covers writing/.
The restructure plan records Phase 2 as executed, and closes its first open
question: the writing handbook does exist, in the Obsidian vault, so STYLE.md
became a digest with pointers rather than the reconstruction the plan assumed.
Also noted there: the vault's opening-scene spec records SC-101's tone check as
BLOCKED on a one-page Voice Sheet that was never written. STYLE.md now holds
that sheet with four fields explicitly unset.
Co-Authored-By: Claude Opus 5 <[email protected]>
CLAUDE.md documents repository layout, the Unity command line, the dialogue
content pipeline, validation, conventions, and known traps. AGENTS.md points at
it for non-Claude agents.
docs/restructure-plan.md covers both halves of the restructure: moving character
and voice prose out of JSON string fields into a writer-facing writing/ tree, and
migrating off the single-scene DialogueTest scaffold into Bootstrap/Systems/Level
scenes. Phase 0 is recorded as executed; everything from Phase 1 on is proposal.
Two findings in it were corrected after first being got wrong: the Unity CLI is
installed (just not on PATH for non-interactive shells) and does have build and
test commands, and the binary scene is not a stale save but a standing Unity
behaviour where Force Text converts only when the setting changes.
Co-Authored-By: Claude Opus 5 <[email protected]>
lennartandClaude Opus 5