e588508436
11 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
| e588508436 |
LANE10: the whole shop — four rooms, stitched by its own portal graph
The hip-hop room has stock (7 racks, 47 bins, 2,865 records); the entry hall and back room are empty. Import all four anyway: the empty rooms are the connective tissue, and they are the only floor in the building big enough to get an office chair up to speed. A portal pair is one doorway seen from both sides, so the importer walks the graph and places each room by making its portal land exactly on its neighbour's. Every pair here sits on parallel walls with opposite facings, so each join is a pure translation — and when one isn't, it says so rather than quietly getting it wrong. It then recentres the union, emits the interior walls with a gap at each doorway, and puts the spawn in the entry hall, because that is the way in. hiphop-room <-> entry-hall <-> shop-floor <-> back-room 12.96 x 15.41 m - 36 racks - 325 bins - 302 real genre labels - 24,024 records Levels.record_store() now takes the walls and rooms from the import, hangs the street glazing on the entry hall's outer wall, and lays a light grid across whatever footprint the rooms came out as. probe_store.gd also checks the new facts: five doorways exist, and the spawn has nothing in it (excluding the player, who is supposed to be standing there — that false positive was the probe's bug, not the level's). All gates green: smoke clean, 7/7 sites 0.00 m/s, overlap CLEAN, all probes pass. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
|||
| 8e61640c38 |
LANE10: SITE 7 — the real shop, imported from its own database
The level the whole game started as, and the only one that isn't authored. robotmonster.party/store/ is assembled procedurally from Postgres (the wowplatter virtual_* schema, served by recordgod), and this game's sites are already data — so tools/import_store.py reads the live shop and emits assets/store/shop_floor.json: the actual 6.66 x 9.51 m floor, 35 racks across 21 archetypes, 278 bins carrying their real genre labels (HARD TRANCE, $12-$15 EURO/ITALO HOUSE), 21,159 records of stock. Levels.record_store() loads it; 'store' joins ORDER as the seventh site. The shop's schema IS the cascade. virtual_crate.rack_id already records which rack holds which bin, so bins arrive frozen and supported_by their rack and none of the structure is authored. To carry that, smashables gained two optional generic keys — 'id' and 'supported_by' — so any level can express support as data. Every existing level is unaffected. Two reconciliations, because a storefront is a renderer: - Mesh sizes are a lie: 21 archetypes share a few stand-in GLBs, so each piece carries 'fit' (the archetype's real dims from the shop's table) and _glb_piece scales the visual to it and builds the collider from it. Also fixed racks being flattened to the floor — pos_y is a piece's BOTTOM, which is what sit_on means, so the ceiling beams, aircon and wall shelves now stay up where they belong. - Nothing collides in a renderer: racks that sat inside each other for years get separated (weighted, so racks barely move and aisles keep their shape — 3 m of correction across 35), rack footprints inset 1.5 cm a side for the few genuinely wedged units, and the importer AUDITS ITS OWN OUTPUT with the same test probe_overlap runs. Bins are deliberately not separated from their rack: cargo inside a container is what the shop means, they spawn frozen, and the audit skips frozen bodies because a frozen body cannot be depenetrated across the room. dev/probe_store.gd: 22 racks hold bins, 255 labels survived, and smashing the rack carrying 22 bins releases exactly those 22 while the other 256 stay frozen. All gates green: smoke clean, 7/7 sites 0.00 m/s, overlap CLEAN, LANE9 probes pass. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
|||
| 81c9c4a349 |
LANE8: scales that swing, bags that don't open, and fruit you can throw
The three things the greengrocer was still missing, plus the physics bug that
finding them uncovered.
HANGING SCALES. Real PinJoint3D pendulums, not animations: a static dial and
rod, and a scale-pan rigid body pinned at its own origin so it can only rotate
about the pivot. They hang at chest height down the aisles — dial at eye level,
dish below — where a real shop scale hangs and where you keep walking into it.
The dish swings 0.40 m off a light knock and the pin holds the pivot to 1.1 mm.
The pan is steel so it never breaks; what you get is a heavy brass weight loose
in a room full of stacked fruit.
THE BAG YOU CANNOT OPEN. Tearing one off the roll is the easy half — that's the
setup. Then you have to open it, and the bag has two ends, only one of which
opens, and nothing tells you which. Rubbing alternates arrow keys, because
mashing one key is not rubbing. Either the meter climbs or it doesn't, and the
only way to learn which end you're holding is to have already lost several
seconds to the other one. SPACE turns it over. That is the whole solution and
the game never says so.
CARRY AND THROW. E picks up anything under 6 kg that isn't a record; LMB throws
it at 11 m/s, six times any brittle threshold in the game. Thrown fruit bursts
on landing through the same brittle_speed path a collapse uses — no separate
thrown-object code at all.
Then the part that took the longest. Round produce got sphere colliders (a box
on an apple is why nothing ever rolled) and every display in the shop instantly
fell over. Three separate things were wrong:
- the stack radii were TYPED, not measured. The table said a cabbage was
84 mm; the model is 213 mm, so that pyramid was built with every row driven
a third of the way into the row below it. Main._glb_radius() reads the asset.
- the lattice was box geometry. For spheres the rise per row is
sqrt(4r^2 - step^2/2) — the height at which a fruit touches all four
beneath it. Anything else spawns every row above the first in mid-air.
- there was no tray. A pyramid of spheres on a bare flat table cannot stand;
nothing holds the bottom row in, so the weight above wedges it outward and
the display walks itself apart in a second. Main._stack_tray() frames each
pile in four low timber walls sized to its base row, which is what every
greengrocer on earth already does.
All 310 bodies asleep within 3 s.
While chasing that, the spawn guard fired on a cardboard box in the OFFICE. The
guard is a net, not a test: it only catches a pair Jolt happens to resolve
violently on the frames it's watching, and this one had been interpenetrating
in four levels for weeks. dev/probe_overlap.gd now finds them by measurement,
comparing every pair of dynamic colliders across all six sites. It found 15,
including a row of filing cabinets 0.70 m apart that are 0.80 m wide, and a
stapler inside a monitor. The box asset is 1.35 m across, so boxes are now
stacked into piles rather than dotted about, and the Backrooms scatter uses
rejection sampling with a 1.6 m minimum. All six sites read CLEAN.
Also: the rage veins were drawing every branch from its own start point
regardless of how far the trunk had grown, so at low rage you got disconnected
fragments floating mid-screen that read as biro scribble rather than blood. A
branch now can't appear before the trunk carrying it, and trunks are thick and
dark where capillaries are fine and pale.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|||
| c96c2f1a67 |
LANE8: the greengrocer — soft stock, stacks that collapse, and litres
The only site whose contents are mostly SOFT, and that one material change rewrites the
level. `produce` is a new Smashable material that throws no shards at all — Splat.gd
bursts it instead: coloured pulp, a puddle that STAYS, a squelch, and litres counted,
because "how much juice did you make" is a far better score for a greengrocer than "how
many objects did you destroy".
Two knock-ons make the level play itself, both falling out of existing systems rather
than needing new ones:
* produce has a low brittle_speed, so anything heavy landing on it squashes it — which
means a pyramid crushes its OWN bottom layer as it collapses, with no special code.
Knock one off the top and the pile does the rest. Shove test: 257 items -> 45, and
42 litres of juice out of the collapse.
* puddles are slippery, cutting ACCELERATION rather than top speed so it reads as
sliding rather than as being slowed down. The more mess you make, the less the floor
cooperates.
Stacks are built bottom-up by Main._build_stack and nothing is glued or frozen — they
stand because they are stacked, so pulling one out of the bottom row does what you'd
hope. 258 produce items spawn and settle to 0.00 m/s with no spawn crush. The glass
bottles behind the juice bar are the deliberate exception: one shelf in the room that
still rewards a proper swing, and the contrast between litres and shards is the joke.
Round produce is authored CENTRED, not floor-centred like every other prop in this repo,
because a rolling body whose origin sits on the floor plane wobbles like a loaded die.
_glb_piece grew a `centred` flag for it.
REAL BUG FOUND, and it was costing hits in every level: _strike took the NEAREST collider
of any kind, so a static surface could eat a swing. Leaning over a display table, the
table edge is a few centimetres nearer than the fruit piled on it, so every swing hit the
table and nothing broke — 0/306 for the whole first take. It now prefers a Smashable and
only falls back to loose bodies if there isn't one. The offices were quietly losing hits
on desks and shelves the same way.
tools/gen_produce.py: apple, orange, tomato, watermelon, cabbage, banana hand, glass
juice bottle, display crate. The melon's stripes were radial boxes first time and it came
out looking like a sea mine; they're thin lenses through the middle now.
dev/probe_grocer.gd verifies stacks settle, that produce yields litres and NO debris,
that a collapse crushes, and that none of it leaks to the next site.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|||
| a806272a7f |
LANE7: there are two of them, and the room changes one thing every loop
THE OTHER ONE (Entity.Kind.MIRROR)
After three loops a second one turns up, and it shares the first one's silhouette on
purpose — same height, same suit, same dark — so at range you cannot tell which you are
looking at. Its rules are the hunter's rules INVERTED:
hunter the other one
while watched frozen IT COMES
while unwatched it closes inert
noise hunts it deaf
correct play keep it in view LOOK AWAY
Every instinct the first one trains into you is the way the second one kills you, and
that is the entire design. The tells are close-range only and all wrong in the same
direction: head thrown back so it stares at the ceiling instead of you, arms raised, and
feet that never reach the carpet (Backrooms floats it 14 cm).
It also makes no static and doesn't breathe, so the only warning the tape gives for it
is the edge closing in — even while you are looking straight at it.
ONE THING PER LOOP (Backrooms._disturb)
Exactly one change per wrap. Never two. That constraint IS the effect: change nothing
and a loop is just a teleport, change several and it reads as a new room, which is the
opposite of what this place is for. One change means you are never sure whether you
noticed something or imagined it — and since the room is otherwise identical, when you
ARE sure it is worse.
Escalating: early loops shift a prop half a metre, turn one, or kill a bank of lights.
Later ones hang a totem that wasn't there, permanently reveal an object nobody dusted,
or put the hunter exactly where you were standing a moment ago.
The shift is half a metre, not one — a metre is enough to shove a neighbouring prop and
read as TWO changes, and it's too obvious anyway. Half is "…was that there?"
dev/probe_two.gd verifies the inversion with numbers (unwatched 0.00 m, watched 4.39 m,
noise heat 0.00) and that the disturbance fires per loop. Note the probe has to wait out
WRAP_COOLDOWN between laps or only the first wrap ever fires.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|||
| 006c324f1f |
LANE7: LEVEL 0 gets an entity, a tape, and a loop
THE ENTITY (Entity.gd) borrows from three places and each borrowing does a job:
SLENDERMAN it does not move while you look at it, and has closed the distance every
time you look back. Looking is ALSO what builds the static — so the safe
option (keep it in view) kills you slowly and the fast option (run) lets
it catch up. That's the trap, and it's why Slender worked.
XENOMORPH it hunts SOUND, and the thing you came here to do is smash furniture.
Every break and every puff of powder drops a marker it walks toward.
Playing well is what feeds it. A loud enough noise overrides the freeze,
because staring is a valid answer to a stalker and not a valid answer to
something that just heard a sledgehammer hit a filing cabinet.
BLAIR WITCH the level wraps: walk far enough and you come out of the opposite wall
into the same room, because every room here IS the same room. A wrap is
also when it repositions, so "I've been here before" and "it's already
here" land in the same second.
It can't be fought. Getting caught doesn't kill you — it costs a lungful, a spike of
rage, and the knowledge that it can do that whenever it likes.
Deliberately unresolved design: 2.5 m, faceless, arms past the knees, cranium swept
backwards. Slenderman from the front, something else in profile. A silhouette your brain
finishes beats a monster it recognises. Blair Witch stick totems hang in the maze.
THE TAPE (Dread.gd) — one shader: grain, scanlines, chroma split, a wandering tracking
tear, an edge that breathes when it's near but unseen, and Slender interference driven
straight off the stare. The static IS the read-out; there's no meter for it.
TWO BUGS WORTH THE COMMIT MESSAGE
1. `seen_now` was a pure view-cone test with NO line of sight, so it built static
through walls — and in a maze that's most of the time.
2. Fixing that exposed the real one: the generated maze was so dense there was no
sightline longer than a few metres ANYWHERE, which kills the entity outright, since
a stalker you can never see can never be stared at. Thinned the generator — skip a
third of the lattice lines, much bigger gaps — which also matches the reference
photographs far better. They're mostly open floor with occasional slabs, not
corridors.
dev/probe_horror.gd verifies all of it headlessly, including hunting for a clear
sightline first (otherwise the Slender test silently tests nothing).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|||
| 4bcd7b49d9 |
LANE7: four workplaces, built from data
Office.gd is gone. Floorplan.gd consumes a spec Dictionary — bounds, palette, walls
with doors and glazing styles, window runs with mullions and blinds, ceiling style,
light grid and style, slabs, static props, desk clusters, chairs, level-specific
smashables and spawn — and Levels.gd holds four of them.
Making this data rather than four subclasses means a site is authorable in minutes,
levels can be diffed, and a real generator can later emit the same structure (a BSP
split of the footprint -> rooms -> doors on shared walls -> fittings by room type).
That was the actual answer to "can we use this git to generate plans": the repo John
found is OpenSCAD SVG->STL for 3D printing, with no generation, no plan parsing, no
room polygons and GPL-3.0, so it can't help — but Office.gd was already most of a
parametric plan builder, and lifting its numbers out is the real path.
The four sites are deliberately not reskins; they differ in the three things a player
actually reads — palette, light, and what the walls are made of:
SCRANTON magnolia, grey carpet, drop ceiling, fluorescent troffers, daylight
down one glazed wall. The baseline.
PAWNEE civic beige and blue-grey, low partitions everywhere, a public
counter, pinboards. Municipal and over-partitioned.
THE INCUBATOR timber floor, white walls, 3 m ceiling, PENDANT lights, glass wall
onto a pool. Nobody has an office; they work at a dining table.
SUB-LEVEL 4 concrete, NO windows at all, exposed services, bare strip lights, a
wall of server racks. The light is green and everything is junk.
L cycles sites in-game; _load_level tears down the shell, rebuilds, re-registers task
stations and re-arms the shift.
tools/gen_level_props.py adds ten more procedural props. The filing cabinet is the
important one: it exports as a CARCASS plus a separate DRAWER, each with its own
floor-centre origin, so the Gauntlet can pull a drawer out as its own rigid body and
spill the files. Also file folder, desk phone, guillotine, shredder, server rack,
sofa, wastebin, stapler.
dev/probe_levels.gd builds every level and reports residual motion after a full
second. All four read 0.00 m/s.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|||
| d17e2314aa |
LANE6: LEVEL 01 is an office — Dunder Mifflin floor plan, and a spawn guard
The game had no room. Props sat on a grey disc in a black void, which quietly broke the premise: a game about wrecking your workplace needs a workplace. Office.gd — LEVEL 01 - Open-plan office laid out after The Office (US) floor plan: bullpen of facing desk pairs behind cubicle partitions, reception, glass-walled manager's office, conference room, break room with vending machines, copier alcove, warehouse roller door. - Drop ceiling on a T-bar grid with fluorescent troffers that ARE the light sources (the old scene lit an interior with one outdoor directional lamp, which is exactly why it read as "props on a plane"), window wall with blinds, magnolia and carpet. - Office owns the shell + static fittings; Main._populate() owns everything smashable and asks Office where things go. Walls/desks/counters are boxes because they ARE boxes; chairs, vending machines, microwave and plant come from a new Blender generator (tools/gen_office_props.py) because a box reads as wrong for those. The level no longer falls over on its own - Records were stacked 3 cm apart vertically while being 30 cm TALL, so Jolt resolved 27 cm of interpenetration explosively on frame one and shoved the furniture over before the player touched anything. They now stand side by side. - Per-material mass (MASSES): everything was 1 kg, so a thrown record could tip a filing cabinet. - Spawn guard: placing ~40 props by formula guarantees an occasional overlap, and depenetration is violent (a chair left the building at 500 m/s). Dynamic bodies are speed-limited for 0.75 s, and each offender is named once with the position it was PLACED at, so the cause stays visible instead of being papered over. It then found the real bug: bullpen rows 3.6 m apart left the two rows' chairs meeting back-to-back with 3 cm to spare. Rows are now 4.8 m apart. - dev/DemoDriver.gd act 0 touches nothing for 5 s and prints total body speed. Now reads 0.00 m/s with zero spawn warnings. Two real melee bugs found while testing the level - The hit test was a SPHERE parked at `reach`, so anything CLOSER than the weapon's reach fell in front of it and was missed — you could stand against the printer with a sledgehammer and swing straight through it. Now a capsule swept from the camera. - Swings started at eye height (1.6 m), so a carton on the floor was ~1.5 m away even standing over it and short-reach weapons could never touch anything on the ground. Swings now originate at hand height. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
|||
| 61757d24bc |
LANE6: rigged FPS viewmodel, weapon loadout + material matrix, 4 HUD styles
Hands/POV - Cut a real first-person arms rig out of the GODVERSE modular character kit (tools/gen_fps_arms.py): ch01 hands + per-side sleeves on the full 65-bone mixamorig skeleton, so all 20 finger bones per hand are poseable at runtime. Textures shrunk to 1k; 4.2 MB. - ViewModel.gd instances that rig ONCE PER HAND and places each instance so its own hand bone lands on the grip — no IK. Grip orientation is measured off the rig at load (pinky->index knuckle = bore axis, elbow->hand = forearm dir), so it survives re-tuning grip_rest_rot instead of needing new euler angles. - Motion layers: look-sway with spring-back, walk bob scaled by speed and weapon heft, idle breathe, landing dip, weapon lower/raise on swap. - Sleeve material overridden (donor asset is a fantasy leather bracer); hand material forced non-metallic (its spec/gloss maps rendered skin as bronze). Weapons - Weapon.gd replaces MeleeAttack: 6 weapons, 4 swing archetypes, and the weapon-vs-material matrix from the founding chat. - Smashable moves from binary hits_to_break to hp/toughness, giving three outcomes: break, dent, or futile (dead clank, no score, HUD nudge). The box cutter genuinely shreds cardboard and genuinely cannot hurt a filing cabinet. - The hit lands at Weapon.contact THROUGH the swing, not on the click — that delay is most of why the sledge feels different from the cutter. - Slots on 1-6 / wheel / Q; game modes moved to M, HUD cycle to H. HUD - Hud.gd: Arcade, Minimal, Work Order (a corporate destruction docket that fills in line items) and Dev, over one shared data feed. Scoring + combo multipliers. Dev harness (not shipped) - macOS screen-recording perms aren't available to the CLI, so the game records itself: dev/demo.tscn + DemoDriver.gd drive a scripted tour for --write-movie, and dev/probe_*.gd print rig/scale/placement numbers. Fix: tools/gen_viewmodel.py box() scaled by size/2 on top of primitive_cube_add's already-unit side length, halving every box — which detached the bat's blade. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
|||
| 51210215e2 |
LANE4 (Godot): real fractured-chunk destruction + printer boss
Consume Lane 2's pre-fractured props in the Godot game. Smashable.shatter() now prefers a <name>.fractured.glb sibling: it instances the fractured template and turns each chunk_00…chunk_NN mesh into its own debris RigidBody3D that inherits the intact body's velocity + the smash impulse (box collider per chunk, debris group so the HUD count + Juice keep working), falling back to primitive shards when no fractured sibling exists. _glb_piece auto-wires the sibling, so every GLB prop upgrades for free. Smashable gains hits_to_break + is_boss: the office PRINTER is the level boss — it sits on a desk, soaks 5 hits (rocking with each), then bursts in a bigger juice cloud into its 16 GLB chunks. Six more Lane 2 props (crt-monitor, filing-cabinet, water-cooler, cardboard-box, turntable, office-desk) are scattered as smashable furniture; the store crate/rack also pick up real 8-chunk shatter from their mirrored fractured GLBs. Assets pulled into game/assets/store/ (intact + .fractured.glb): office-printer, office-desk, crt-monitor, filing-cabinet, water-cooler, cardboard-box, turntable + record/crate/rack fractured siblings (generated via meshgod's Voronoi fracture CLI). Verified headless: smoke test clean (0 script errors), printer boss soaks 4 hits then spawns 16 chunk bodies, a crate spawns 8 — both real GLB chunks, not primitive shards. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> |
|||
| 5852322606 |
Restructure into monorepo: game/ is the Godot lane
Repo root moves up from game/ to the project folder. Lanes live as top-level dirs (game/ now, web/ planned); LANES/ will hold parallel-session specs. 3D-STORE asset vault stays out of git. History preserved as renames. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> |
