Add the Axiom Relay roof group (all 10 models) and a shared axiom-roof-kit - #5
reposlayer wants to merge 6 commits into
Conversation
All ten briefs under docs/assets/reusable/architecture/roof-pieces had no model source. This adds all ten, on one new support item. axiom-roof-kit fixes the group's convention: every module is authored about the deck's top surface at y = 0, with structure hanging below into negative Y. A roof fitting is positioned by where it sits, not by how thick the deck under it happens to be, so a pad dropped onto a thicker deck later does not need re-authoring. The kit carries the shared deck, curb, outlet and parapet builders, and a wear tier with grime above the rest of the kit's and rub below it, because nothing rubs a roof and nobody cleans one. The recurring detail across the group is that nothing sits flat on the deck. Fittings stand on curbs with the membrane turned up against them; the railing and the antenna mast are ballasted rather than bolted, because drilling a roof puts penetrations through the one layer keeping water out. Modules: flat-roof, roof-parapet, rooftop-railing, roof-access-box, roof-machinery-pad, rooftop-hvac-foundation, roof-skylight, roof-antenna-mount, vented-roof-panel, sloped-sci-fi-roof. typecheck, build, and the 49 existing tests pass; the registry builds 179 items from 174 models. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
i just merged #6 which is in conflict with this and others |
# Conflicts: # registries/scifi-kit/src/build.ts
|
Thanks for the heads-up — already handled, and our timing just missed: you commented at 22:10:42Z and the fixes went up at 22:12–22:13Z, so you'd have been looking at the pre-fix state. All four (#2, #3, #4, #5) have The conflict was a single line in Resolution is to keep both rather than pick one — the list is additive, and each kit is needed by its own models. Verified on all four after merging: Worth knowing for the merge order: the remaining three will re-conflict on that same line the moment any one of them lands, for exactly the same reason. Nothing avoids it while they're cut independently — but it's a one-line resolution each time and I'll re-run it as soon as you merge, or tell me the order you want and I'll stack them so they go in without you touching a conflict. 🤖 Posted by Claude Code |
Every batch appends its kit to one hardcoded array in this file, so two batches in flight always collide on that single line even though their kits have nothing to do with each other. That is what put four of these PRs into conflict the moment the interiors batch landed, and it would do it again for every batch after them. A prototype directory with no model.ts is a shared kit rather than a model, which the directory walk here already has to determine in order to collect the models. Registering those directly removes the hand-maintained list, so a batch that adds a kit no longer touches this file at all and there is nothing left to conflict on. Both lists are sorted, so the output stays deterministic. The registry builds the same items as before; only their order within the lib tier changes, since discovery is alphabetical where the list was chronological.
|
Follow-up: rather than stacking these so you merge them in a fixed order, I removed the thing causing the collisions.
} catch (error) {
if ((error as NodeJS.ErrnoException).code !== 'ENOENT') throw error
supportIds.push(entry.name) // no model.ts -> shared kit
}A batch that adds a kit no longer touches this file at all, so there is nothing left to conflict on — now or for any future batch. The same commit is on all four branches, textually identical, so git resolves it as one change rather than four competing ones. Verified by simulating your merge order — merged all four into Merge them in any order, no conflicts. Both lists are sorted so output stays deterministic; the only behavioural difference is that the lib tier is now alphabetical where the array was chronological. 🤖 Posted by Claude Code |
Absence of model.ts marks a shared kit, but it is equally true of a directory that is simply empty - one left behind by an abandoned model. Registering that published a support item with zero files, which the schema accepts and an install would silently resolve to nothing. Discovery now requires at least one TypeScript source, which is what actually distinguishes a kit from a leftover. Identical to the change on the other batch branches, so they still merge cleanly in any order.
PreviewsEvery model in this PR, rendered so you don't have to check out the branch and run the renderer.
Individual previews (1024×1024)
Same deterministic rig for every model in the wave — that is what a contact sheet is for, since the defect that matters at pack scale is whether the props read as one catalogue rather than as individually plausible objects. Images live on the 🤖 Posted by Claude Code |
|
which clanker has produced them? they look very low quality to me |
|
claude opus 5 on ultracode lol |
Review called the batch low quality and was right. The cause is countable: across the ten models and the kit there are 97 uses of the dark tier (graphite, ink, steel, deck) against 5 of the light one, and none of accent or cobalt. Ten models built from one value read as ten grey slabs, which is what the contact sheet shows. Copings and curb counter-flashings move to the light tier, which is also what they are - capping metal is genuinely lighter than the upstand it sits on. That gives every parapet, curb and skylight in the group a horizontal highlight to read against, and it is a one-place change in the kit rather than ten model edits. Wear rub was set near zero on the theory that nothing rubs a roof. True of a membrane, false of every edge a boot, a ladder or a service trolley crosses, and suppressing it removed the edge definition that separates one plate from the next. Rub now lands on the parts that are actually walked on, and steel gets a profile at all, which it did not have. typecheck clean, bun test 49/0, registry unchanged in count.
|
Fair, and you were right about this batch specifically. I went and counted rather than argued:
So the whole group was built out of one end of the palette and came out as ten grey slabs. The doors and windows batches are the other way round, which is why they read and this one didn't. Two fixes, both in the kit rather than in ten models:
Before / after:
Pushed. Still weak, and I'd rather say so than have you find it: 🤖 Posted by Claude Code |
The two models flagged as still weak after the palette pass. Both needed geometry, not colour. roof-antenna-mount was a 70 mm tube on a slab and read as a flagpole. A mast on a roof is a lever resisted by ballast, and it is a lattice for a real reason: a triangulated frame carries the same moment at a fraction of the wind area. It is now a three-leg lattice on hinged base plates, with a ring and a diagonal at every bay. The equipment is modelled rather than implied - headframe, three sector panels, an offset dish, a lightning finial above everything it protects, and the cable bundle running the full height into a gland on its own curb. sloped-sci-fi-roof ran a 0.9 m rise across a 4 m bay in six 60 mm steps, so each riser was shorter than the panel was thick and the whole thing read as a flat card on a slab. The rise is now 1.7 m in five steps, which makes each riser a real vertical face - and that face is the entire silhouette. Risers and the ridge closure sit in the light tier so the slope reads as a stair of highlights rather than one grey wedge. typecheck clean, bun test 49/0.
|
Both done, pushed.
🤖 Posted by Claude Code |
What this is
All ten briefs under
docs/assets/reusable/architecture/roof-pieces/had no model source. This adds all ten, so that family is now complete.The kit
assets/prototypes/axiom-roof-kit/fixes one convention and shares four builders.The convention: every module is authored about the deck's top surface at y = 0, with structure hanging below into negative Y. A roof fitting is positioned by where it sits, not by how thick the deck under it happens to be — so a machinery pad dropped onto a thicker deck later needs no re-authoring.
The builders are
deck,curb,outletandparapet, plus a wear tier with grime above the rest of the kit's and rub below it, because nothing rubs a roof and nobody cleans one.The recurring idea
Nothing on an Axiom roof sits flat on the deck. Every penetration stands on a curb with the membrane turned up against it and counter-flashing over it; the railing and the antenna mast are ballasted rather than bolted, because drilling a roof puts a hundred penetrations through the one layer keeping water out of the building. Fittings drawn sitting directly on a deck are the clearest sign a roof was modelled rather than built, and avoiding that drove most of these ten.
Modules
flat-roofroof-parapetrooftop-railingroof-access-boxroof-machinery-padrooftop-hvac-foundationroof-skylightroof-antenna-mountvented-roof-panelsloped-sci-fi-roofVerification
bun run vibe:model preview.bun run typecheck,bun run build, andbun test(49 pass, 0 fail) all pass.bun run build:registriesadds 10 models and 1 support item to the registry — 190 items from 184 models against the currentmain. Stated as a delta rather than a fixed pair, because the absolute count moves every time one of these batches lands.Note for review
This is the fourth architecture batch cut independently from
main(doors, windows, walls, roof). All four touch the onebuildSupportItemlist inregistries/scifi-kit/src/build.ts, so whichever land after the first will want a trivial merge on that list.🤖 Generated with Claude Code