Build and revise procedural 3D assets with your coding agent. Local MCP server, CLI and TypeScript engine with rendered review, editable source and GLB export.
Decisions on SPEC.md section 26
Historical decisions from the original scene workspace. References to its local review packets do not imply those files are included here. Current integrated source and staging instructions are in README.md.
Current coordination state: owner review revision 2, 2026-09-30 local date
The owner approved a representative level-1 floor at its measured south-west building location, an explicit entrance-to-floor transition and a bounded context cutaway. A continuous lobby/stair/lift route is outside this round. The registered context, corrected section viewing aperture, projected-size tree detail, Shift boost, separate touch driving controls and persistent Enter/Exit are implemented and technically qualified. The separate rear-repair checkpoint passes 173 full Foundry CPU tests plus a later focused vehicle test, both 8/8 Chromium contracts and desktop/touch/lifecycle checks. Its initial code is 1,525,112 B / 445,765 gzip under unchanged D-15; the 31 interior model files and interior code load only after Enter. REVISION2.md identifies the exact source, runtime and receipts.
The owner subsequently approved front-headlight repairs. New saved children, exact scene/site intake and affected qualification remain before the final combined candidate. The current rear-repair freeze explicitly records finalReviewPacket: false and ownerAccepted: false. The owner-approved humanoid repair/maintenance role and AMR/arm custody remain unchanged, as do the synthetic twin inputs and D-43 bounds. The initial candidate and accepted Farm r34 are preserved; the owner's new Farm hand/control changes proceed as bounded revision2 replacements. No commit, main movement, push, upload, deployment or publication is authorized.
First candidate coordination state: 2026-09-30, after the road marking correction (preserved history)
The full local-completion goal was reactivated after the documentation-only checkpoint below. FF3, revised campus planting and articulated freight are implemented and technically qualified as a frozen local scene candidate. The package has 163 passing CPU tests / 7,206,087 assertions, both full 8/8 Chromium contracts, local controls/no-graphics checks, final same-input KPI comparisons inside unchanged D-43 bounds, and replayed warm state/720 hashes. B-04 and D-15 are closed on the current bytes; the initial campus closure is 1,520,413 B / 444,177 gzip, with the interior deferred. packages/foundry-floor/PROGRESS.md records exact pack/build identities and limits. Final combined site intake and owner review remain.
The first-candidate test count above corrects an earlier transcription of 160 against its retained 163-test CPU receipt; this does not change that candidate's scope or acceptance.
The three revised plants, physical arm/AMR repairs and 24 standard-LOD child revisions preserve originals and are review candidates, not blanket owner acceptance. The installed per-part LOD workflow and final Blender/Unity consumer packet are qualified with explicit destination limitations; see ../engine-work/local-v09-review/consumer-final/README.md. The owner has explicitly chosen to keep humanoids on repair and maintenance, with AMRs and arms carrying the FOUPs. This settles the earlier D-26 role question for v0.9; a humanoid carrying route is not required. The shared plant/scale packaging proposal still awaits an answer. Farm r34 stays accepted; no new M5 was needed. No commit, main movement, push, upload, deployment or publication authority follows from this local qualification. The dated direction and prior checkpoint below remain historical records.
Earlier checkpoint: 2026-09-30, documentation before continuation
The owner requested documentation alignment and a proposed goal statement before further implementation. Preserve current work and its evidence; the latest per-workstream state belongs in packages/foundry-floor/PROGRESS.md and the coordinator's ../PLAN.md. The preceding local implementation interval did not confer acceptance. FF3 plus campus integration remain required for the first full owner review, and final builds, regenerated simulation evidence, both-backend qualification and site staging remain unfinished.
The implemented scene model uses AMRs and arms for actual container transfers and humanoids for real repair/PM visits. The current transfer seats fit the accepted arm/AMR geometry. Keeping humanoids in service roles is the coordinator's recommendation; the owner has been asked to choose that scope or require a separately qualified carrying route. The recommendation does not silently override historical D-26. D-43 remains synthetic regression evidence, and Farm r34 remains accepted without an unsupported M5.
Owner direction recorded at takeover: 2026-09-30
This dated update governs the remaining local review work. The decision table and past qualification below remain historical records; this update does not reopen accepted work or confer acceptance on unfinished work.
- First local review scope: finish FF3 and integrate the campus vegetation and freight vehicles before the owner's first full review of the site, scenes and assets. D-44's instruction to keep the site on FF2 was an interim staging boundary; it does not make FF2 sufficient for this review. The final local site must carry the qualified FF3/campus result, coordinated with the site owner.
- Real floor transport approved: add floor transport to the twin so AMRs and arms follow real simulated ownership and transfers, with humanoids in their repair/PM roles as clarified above. The earlier FF3 stop pending a model-change decision is resolved. Re-measure and revalidate after the change, including determinism, warm starts, conservation of FOUP ownership, hand-offs, sanity evidence and both-backend scene checks. Never animate a second copy of a FOUP that the OHT still owns.
- D-43 remains a regression decision: its accepted measured baseline and the superseded spec band stay in the record. A regression band around the current twin is not evidence of physical calibration. Floor transport will require a new measured baseline with the behavioral change explained; passing a re-based band alone does not validate the model against a real fab.
- Golden Gate delivery: retain the interactive Golden Gate Scene and separate downloadable bridge and Vehicles assets/packs. One does not replace the others.
- LOD review controls: per-part LOD overrides are owner-approved engine work. The scene and site consume the qualified engine behavior; they do not establish it by inventing separate LOD semantics.
- Campus assets: the two Sol author runs completed all 13 outputs (8 vegetation, 5 freight); they are not still running. Author-side acceptance wording is not blanket owner acceptance. The owner's conifer and ornamental critique requires redesign before vegetation acceptance. Shared plant assets with scale presets remains the coordinator's proposal awaiting the owner's answer, not a required architecture. See the dated updates in the campus briefs.
- Importer evidence: the completed 13 outputs passed headless Blender 5.2.0 LTS and Unity 6000.2.3f1 / glTFast 6.20.0 baseline import checks. Plain imports show LOD0, not automatic LOD switching: Blender excludes lower tiers in Orphan Nodes; Unity creates no LODGroup. This does not qualify the unfinished engine candidate, GPU appearance, animation, visibility or a player build. Evidence:
../engine-work/dogfood-v09/codex-takeover-imports-20260930/summary.json. - Farm remains accepted at r34. Its completed hub evidence and adopted hitch gate remain in force. D-34's conditional M5 has no demonstrated trigger from this audit; no new Farm optimization round is authorized merely because other scenes remain unfinished.
- Boundary: finish and verify locally for owner review. This update authorizes no commit, main update, push, deployment or package publication. The proposed final-local-completion goal is a separate owner decision.
Recorded by the coordinator (Claude Opus 5.5) on 2026-09-29, while the owner was asleep. The owner asked the coordinator to finish the night's work and review it in the morning. Each decision below is reversible, and the morning summary lists them.
This file is read-only for the builder. Where it and SPEC.md differ, this file wins. Record every effect on your work in PROGRESS.md.
| Id | Decision | Why |
|---|---|---|
| D-01 | Adopted. Include the optional SceneExtras fields. The required contract is unchanged. |
They are additive, and every one is optional. |
| D-02 | Adopted. Unknown or mid-range touch devices start on economy and learn upward. The S24+ class starts on high. |
This is safe when nothing is known about a device. |
| D-03 | Adopted. Accept the WebGL2 start-up cost for now. Mobile WebGL2 uses minimal (D-05). |
It cannot be measured tonight, and a cheaper tier bounds the worst case. |
| D-04 | Adopted. TypeScript 7.0.2, with the typescript-6 alias check (check:site-compat). |
The coordinator verified that the website pins typescript 6.0.3 (site/package.json). The website's REPORT says this is because @astrojs/check 0.9.10 declares TypeScript ^5 || ^6 peers. Sources the website imports must therefore typecheck under 6.0.3 as well. |
| D-05 | Adopted. Define minimal in the kit and in the Farm tier tables. Use it only for the D-03 cases. |
|
| D-06 | Adopted. Apply the tangent-safe derivative to the woodland, with the B-06 A/B captures. | |
| D-07 | Decided by the owner. Unchanged from SPEC.md. | |
| D-08 | Adopted. Exclude @react-three/drei from every scene package. |
|
| D-09 | Adopted. Expose no review views publicly. Offer overview camera buttons. The destination menu and the tour stay unexposed unless the owner asks. | |
| D-10 | Keep "Rowan" in the door status text, as the pilot has it, in one named constant. | This is parity with the owner's shipped pilot text. Changing it is a one-line edit if the owner prefers "the farmer". |
| D-11 | The coordinator updates the website copy when the Farm scene is wired into the site branch, and only if B-09 (mobile play) passes. The builder does not touch the website. | |
| D-12 | Owner action, not tonight. Uploads to the asset host and its CORS headers wait for the owner's explicit approval. Everything tonight is staged and served locally. | Uploads are outward-facing, and the owner has said nothing is deployed tonight. |
| D-13 | Adopted. Keep @kiln-scenes/*: private, never published. |
|
| D-14 | Adopted. Each package decides its tier knob names in its own M0. | |
| D-15 | Adopted. Set the bundle budget after the first measured build at M1, at the measurement plus 10 percent. Amended 2026-09-29, 12:45 (owner, question tool): the ceiling is per scene. Each scene freezes its own ceiling from its first passing build plus 10 percent; the kit demo keeps 1,557,399 B / 452,687 B gzip, and the Farm's is frozen at its M2 public chunk, 1,574,581 B / 460,130 B gzip, so ceil(x 1.1) = 1,732,040 B / 506,143 B gzip. | Bytes are static, so the load on this PC does not affect them. |
| D-16 | Adopted. Public builds honour no URL flags. | |
| D-17 | Adopted. MIT, Copyright (c) 2026 Matthew Kissinger. | The coordinator verified that this matches the owner's existing MIT notices, for example farm-pilot/scene/field-grass-source/LICENSE. |
| D-18 | See below. | |
| D-19 | Adopted. No credits UI inside the Farm scene. | |
| D-20 | Owner direction (2026-09-29). Parity with the pilot is the baseline, not the goal. M2 closes on the parity evidence already gathered; no threshold tightening or chasing of residual differences. The presented look stays standard glTF PBR under conventional tone mapping (D-18), the way three.js, Godot and Unity show the assets; stylized rendering such as cel shading is not part of the presentation (at most an off-by-default toggle costing under a day). M3 adds a look and performance exploration: engine-typical post-processing as runtime toggles (anti-aliasing, ambient occlusion, bloom on emissive materials; vignette and grading off by default; AO off on the tablet tier) and performance implementations that do not change the look, each measured per tier for absolute numbers on the hub (the laptop) and the tablet, this PC giving indicative numbers only while agents run on it (owner, 11:35), presented as one contact sheet for the owner. D-18 stands until the owner chooses. | The site shows developers what their import will look like, so the default look must be portable; post-processing that every engine offers is representative, stylization is not. |
| D-21 | Owner direction (2026-09-29, 13:41). Scenes are built in R3F but trend engine-agnostic: the truth of a scene is data (the asset GLBs plus a layout of placements, instance parameters, paths, lights and cameras in metres) that Unity, Godot, Blender or another engine could consume; R3F components are one renderer of that data and hold no scene truth of their own. Canonical assets are plain glTF 2.0 PBR that load without a decoder; compressed or engine-specific forms are build derivatives. Each scene package can export its static composition as one scene.glb (kit composeSceneGLB) for import elsewhere and documents behaviours (traffic, controllers) as data plus a short spec. |
LOD: the MSFT_lod files degrade to LOD0 in every importer without the extension; per-tier derivative GLBs serve engines with their own LOD systems. Terrain: meshopt tiles are a web derivative (GG-002). |
| D-22 | Owner direction (2026-09-29, 15:45). Every scene gets controls made for touch rather than on-screen buttons: a virtual joystick for movement (when driving, push forward to accelerate, pull back to brake and, once stopped, to reverse; left and right steer), pinch to zoom in and out (orbit distance, chase distance), one-finger drag to look or orbit, and at most one context action (Interact, Leave) as a tap target. The Golden Gate drive is third person: a chase camera behind a car the visitor controls, and on touch the joystick is the only driving input. The owner left the exact scheme open ("or something like that"); this row was the coordinator's reading; the owner confirmed it at 16:15 ("agree with all, go"). | Golden Gate: the Accelerate and Brake pedals and the Turn-around button give way to the joystick model; a turn-around at a deck end is a held pull-back at rest (or one tap target if that reads badly). Farm: the M3 mobile work (B-09, B-10, B-13) qualifies the joystick, drag-look and pinch-zoom on the tablet; buttons stay only for Interact and Leave. Kit: VirtualJoystick gains a bidirectional throttle mode and pinch handling lives in the kit controls so every scene shares them. |
| D-23 | Owner decision (2026-09-29, 22:05). X-04 and prewarm: SPEC 6.4 changes so a scene compiles every pipeline its first complete frame needs before it declares ready (behind the loading screen), and X-04 is the time to the first stable frame (the first frame after ready without a compile stall), measured the same way for the pilot and the rewrite, against the 1.20 limit. Farm M4 implements and re-measures; the hub re-run is the coordinator's. | Cause on record: 234 WebGPU pipelines compiled after the first frame, onReady fired after the stall (evidence/perf/x04-attempt/README.txt). |
| D-24 | Owner decision (2026-09-29, 22:05). Foundry Floor model sanity: the pilot line is loaded to about 85 percent utilisation of its busiest families by raising the release rate so the CONWIP cap binds; tool counts, MTBF, MTTR and PM stay as specified; the sanity band (cycle time 14 to 23 d, WIP 57 to 91, lots out 110 to 130, moves 45 to 75 per hour) stays. FF2 implements, re-measures days 30 to 60 with seed 1 and regenerates the warm snapshot. | The twin at the spec release rate ran at x-factor 1.2 (packages/foundry-floor/evidence/sim/retunes.json). D-18 is unchanged (MSAA 4x, ACES 0.95 stay the default look). |
| D-25 | Owner decision (2026-09-29, 22:10). Tablet tier and mobile look: the Tab S9 FE GPU class (Mali-G68, WebGPU adapter arm / valhall) is classified to the minimal tier on WebGPU as on WebGL2, so the 33.3 ms p95 target (SPEC 20.2) holds for it; the WebGL2 minimal-tier play positions captured on the tablet (evidence/tablet-2026-09-29/play-webgl2/device/) are the recorded mobile look under D-20, with no grass or shadow requirement on minimal. X-08 and X-09 stay deferred (no phone). |
On economy the tablet gave p95 49.8 ms with 2 tasks over 100 ms; on minimal 16.6 / 33.3 ms (evidence/tablet-2026-09-29/summary-play.json, soak/). Farm M4 implements the classification change. |
| D-26 | Humanoid work robots are a mover class in the Foundry Floor twin: hand-carry FOUP moves where the rail does not reach (litho zone, metrology, stocker to under-track storage, a configurable share) and service visits to tools in the down or PM state; the asset is generic (no brand, livery or likeness). FF3 wires it; FF2 keeps the mover wiring data-driven and the technician stub behind one interface. | Owner direction 2026-09-29 22:25 ("Both now"; role: material handling and service visits); the pack honesty rule forbids naming or copying any real product. |
| D-27 | Owner decision (2026-09-30, 01:04). Evidence captures stay in the local branch commits as delivered (Golden Gate round 2 added 432 MB; scenes-root evidence PNGs are git-ignored). Before any push the coordinator filters the capture files out of history or moves them to LFS, and the owner approves that rewrite then. Nothing is pushed before that. | Keeps the record complete while the branch is local; the cost is paid once, at push time, with the owner's approval. |
| D-28 | Owner decision (2026-09-30, 01:04). The hitch rule proposed in M4 (per 60 s run at every tier: 0 frames over 100 ms, 0 over 50 ms, longest frame at most 50 ms; tablet row at most 2 over 50 ms and 0 over 100 ms) becomes a gate once the hub X-05 runs of 2026-09-30 confirm it holds on absolute numbers. Until then it stays report-only in compare. Adopted 2026-09-30, 05:40 (coordinator): the idle hub's X-05 runs held the rule on 160 of 160 runs (32 cells of 5; longest frame 33.4 ms, high/WebGPU/living-orbit; evidence/perf/hub-2026-09-30). The rule lives in packages/farm/scripts/hitch-rule.ts with its unit test; perf compare now reports hitchRule per cell, an overall pass with per-gate verdicts (schema kiln.farm-hub-comparison/3) and exits 1 on any miss. The tablet row stays a device measurement. |
A gate needs absolute numbers from the idle hub, not dirty-PC numbers; the tablet allowance reflects its measured 2 frames over 50 ms in a 210 s orbit. |
| D-29 | Owner decision (2026-09-30, 01:04). A governor locked above the ladder's floor by a forced tier never changes the tier but still publishes its advice (the tier it would step to), so the UI can show a hint; SPEC 8.3 says so. | Honours the user's forced choice while making its cost visible; a forced-high tablet at 18 FPS is the user's call, not a silent one. |
| D-30 | Owner decision (2026-09-30, 01:04). No same-session pilot run on the tablet for the X-07 comparison; X-07 passes at minimal (D-25) and the comparison is deferred. | The pilot comparison adds nothing to the gate and would cost another quiet, cool tablet session. |
| D-31 | Owner decision (2026-09-30, 02:04). The Kiln v0.9 pass runs tonight as the full pass: the foundation on kiln-oss main (F0 to F9, 173 uncommitted paths) is committed on a new branch codex/v0.9 (never on main), the review-lighting rig merged, all logged issues triaged and the engine ones fixed, docs, skills and tools aligned, runtime bundles rebuilt, every gate run, a fresh clone proved and the npm tarball produced. Nothing is published, pushed or deployed. |
K-1 placed the foundation in this pass and the site's 0.9 docs already describe it. |
| D-32 | Owner decision (2026-09-30, 02:04). Opus 5.5 at max effort builds the engine pass; the coordinator reviews. | The delegation note assigns hard builds to Opus while Codex is out. |
| D-33 | Owner decision (2026-09-30, 02:04); indexing superseded by D-77 (2026-10-26). If FF2 has not landed when the site is rebuilt, the Foundry Floor scene page carries an in-progress noindex preview of what FF2 has delivered, labelled as such for the owner's local review, and is replaced when FF2 is accepted. | The owner reviews every scene locally tomorrow. |
| D-34 | Owner decision (2026-09-30, 02:04). Fix rounds may start tonight without asking: a bridge-fix review 4 (Kiln author, cap 10 dollars) if the round-3 read report says the pylons and anchorages do not read; a bounded Farm M5 if the hub numbers fail X-01, X-04, X-05 or the hitch rule; FF3 for the humanoid mover class and the FF2 review findings; site fix rounds until the local preview is clean. | Keeps the night moving without waking the owner. |
| D-35 | Owner decision (2026-09-30, 02:04). The vehicles pack and the Foundry Floor pack carry the Farm pack's licence, CC0-1.0 with the same scope qualifier (authored asset content only). | Consistency; nothing else on the site is relabelled. |
| D-36 | Owner decision (2026-09-30, 02:04). Five long subagents may run tonight (Golden Gate round 3, FF2, the hub runner, site round 2, the engine pass); the adversarial site reviews run as streams land. No timing is taken on this PC. | The owner accepted the load for the night. |
| D-37 | Owner decision (2026-09-30, 02:04). Blind agent and dogfood runs run tonight after the issue triage and the docs and skills alignment: Agy with Gemini 3.8 Flash (subscription) and OpenCode with Meta Muse Spark 1.3 Contributor and DeepSeek V4.1, several runs (setup from scratch and asset authoring), findings logged in KILN-ISSUES.md and the pass report. | Owner's words: "several blind agent and dogfood runs after you triage issues and update docs and skills to make sure everything is aligned". |
| D-38 | Owner decision (2026-09-30, 07:31). Bridge-fix review 5 runs under a raised cap, $25 and 60 minutes (D-34 gave review 4 $10 and 45 minutes), on the same author session, to remove the depth fighting in the review-4 lift construction and finish both tiers; the coordinator set the amount. | Owner's words: "If you need to deploy another author you can just increase the spend cap". Review 4 stopped at its cap with the web tier saved and the full tier unsaved. |
| D-39 | Owner decision (2026-09-30, 09:25). The site's Golden Gate page shows g5 (bridge-fix review 5 accepted and restaged, D-38 and the coordinator's review-5 record); the hub does not time the g5 kit before the deploy. g4 stays timed only by its g3 ancestor's hub session; a g5 timing run is a later, owner-scheduled job. | Owner's words: "Site shows g5, no hub timing first". The g5 change is the anchorage housings and shaft lift lines only (web and full tiers), measured clean by the coordinator; the public chunk is byte-identical to g4. |
| D-40 | Owner decision (2026-09-30, 09:25). The transit bus review-2 revision is approved: the "Awaiting owner review" flag comes off the vehicles pack page and its licence line reads like the other five vehicles. | Owner's words: "Approve it". Sent to the site round-3 builder; if round 3 cannot absorb it, the coordinator applies it in the review round. |
| D-41 | Owner decision (2026-09-30, 09:38). The hitch rule keeps the limits adopted under D-28: per 60 s run at every tier 0 frames over 100 ms, 0 over 50 ms, longest frame at most 50 ms; the tablet row at most 2 over 50 ms and 0 over 100 ms. It stays a gate. | Owner's words: "Keep as adopted". The limits held on 160 of 160 idle-hub runs with the longest frame at 33.4 ms, so the gate binds only on a real regression. |
| D-42 | Owner decision (2026-09-30, 09:38). The subfab pump and abatement kit is placed by enlarging the final section module's kit zone in the scene to take the accepted kit as built (3.6 x 6.0 x 2.4 m); the accepted GLB (r_d0ff94d7e74c487b9b822f61a1ccd171) is not revised. FF3 carries the change. | Owner's words: "Enlarge the zone". No author spend; the zone is scene layout, not asset geometry. |
| D-43 | Owner decision (2026-09-30, 09:38). The model-sanity band is re-based on the twin at the adopted 4,500 wafer starts a month: the band becomes a regression band around the measured values (cycle time 8.32 d, WIP 50.1, 180 lots out, 61.8 moves an hour over days 30 to 60, seed 1), the three test.failing criteria become passing tests that guard drift, and the spec's band (14 to 23 d, 57 to 91, 110 to 130, 45 to 75) is recorded as superseded with the report's finding that the release rate alone cannot meet it. FF3 carries the change; the band width is set from the release sweep in evidence/sim/retunes.json. |
Owner's words: "Re-band to the twin". Throughput equals the release rate, and the twin's variability does not produce the band's x-factor short of saturation. |
| D-44 | Owner decision (2026-09-30, 09:38). FF3 (the floor robots' moves and hand-offs, the humanoid mover class, with D-42 and D-43) starts today, once FF-C1 lands and passes the coordinator's review, on the fifth long-subagent slot of D-36, in parallel with the engine contracts. The site keeps the FF2 page until the owner switches it. | Owner's words: "After FF-C1 lands and passes review". The brief is scenes/TASK-FF3.md. |
| D-45 | Owner decision (2026-09-30, recorded 12:09). The campus gets a vegetation set (three broad-crown trees, two conifers, an ornamental tree, a shrub and a hedge segment, each with three declared LOD tiers) and freight vehicles (a day-cab tractor, dry-van, flatbed and tanker trailers, a delivery van), authored now in parallel by Codex gpt-6.1-sol at high effort in kiln-init workspaces on the 0.9.0 tarball (showcase/authors/sol-ff-vegetation, sol-ff-freight; briefs pack2-research/terafab/briefs/campus-vegetation.md and campus-freight-vehicles.md; sessions launched 12:06; events and final messages under engine-work/dogfood-v09/codex-runs/). After they land and pass the coordinator's review, a campus round (FF-C2) plants them from the campus data (avenue verges on the split road, windbreak belts behind the parking loops, parking aisle islands, the roundabout island with a ring of twelve conifers, a central grove of five broad trees and low hedges inside the kerb) and adds the truck classes to the traffic with a trailer follower. Owner's words: "dogfood should go to sol 6.1 on high - update codex cli to find this model and author now in parallel. everything is happening today." E-7 in engine-work/v09/DECISIONS.md records the dogfood split behind it. |
|
| D-46 | Owner decision (2026-10-01, about 22:00). Draw optimization runs in its own worktree off main (C:/Users/Mattm/X/kiln-draw-optimization, branch draw-optimization). Scene work goes ahead now. Model-facing engine changes (tool output, authoring and QA skills) wait for v1-readiness to merge, then rebase. |
v1-readiness changes no file under scenes/, so scene work cannot collide with it. Plan of record: docs/plans/2026-10-01-draw-optimization-cycle.md. |
| D-47 | Owner decision (2026-10-01, about 22:00). The cycle covers Phases 0-1 of the draw-call research roadmap, then stops and reports. It covers all scenes, and its findings feed a report that upgrades or aligns the tooling, projects and skills for authoring scenes. | Owner's note on scope. |
| D-48 | Owner decision (2026-10-01, about 22:05). The success measure is frame time first. Draw count is also reported as a headline. Timing must come from clean data, because this PC is often under load. | Owner's words: draw count is "a metric that a lot of devs look at". |
| D-49 | Owner decision (2026-10-01, about 22:05). Small shadow changes are acceptable: tiny parts may stop casting, and stand-in outlines may differ slightly. Everything else stays at parity with the D-20 baseline, and every difference beyond parity is shown as images. | |
| D-50 | Owner decision (2026-10-01, about 22:10). Count metrics (draws, triangles, pipelines, CPU draw-call counts) may run anytime on this PC. Timing comes only from quiet hosts, with interleaved before/after (A/B/A/B) runs. | Counts are deterministic under load; frame times are not. |
| D-51 | Owner decision (2026-10-01, about 22:10). The coordinator may schedule tablet timing runs, using the device kit only while the tablet is quiet and cool, and removes the port mappings after each run. | |
| D-52 | Owner decision (2026-10-01, about 22:10). The tablet rule stays p95 ≤ 33.3 ms. Reports also give the share of frames within one and within two display refreshes, so that any restatement can be decided with data. | |
| D-53 | Owner decision (2026-10-01, about 22:10). X-02 (draws and triangles within 2% of the sealed pilot at high) is re-baselined on the optimized build. Pixel parity against the pilot remains. | The optimized build lowers draws on purpose, so a ±2% band around the pilot would fail by design. The re-baseline itself is recorded as a later decision. |
| D-54 | Owner decision (2026-10-01, about 22:15). Moving objects get a cached static sun shadow plus a small live map for movers, and both maps are sampled. | |
| D-55 | Owner decision (2026-10-01, about 22:15). The MSAA store-op suspicion must be confirmed by an adversarial review before anything is done. If it is confirmed, the scenes get an elegant fix, and a private artifact covers the issue, how it was found and what a fix would be. No three.js issue or PR is opened. | Owner's words: "i don't want to slop spam AI generated PRs or issues for three.js". |
| D-56 | Owner decision (2026-10-01, about 22:15). The shadow work applies to all three scenes: Farm, Foundry Floor and Golden Gate. | |
| D-57 | Owner decision (2026-10-01, about 22:15). The tablet stays on the minimal tier (D-25). The cycle brings economy before/after numbers; any tier change is the owner's call. | |
| D-58 | Owner decision (2026-10-01, about 22:18). Commits stay local on draw-optimization, made in logical steps. No push: push and PR are decided by the owner afterwards. |
|
| D-59 | Owner decision (2026-10-01, about 22:18). No changes to site/ and no deployment. Scene builds and staging happen only inside the worktree. |
|
| D-60 | Owner decision (2026-10-01, about 22:18). The end report is a private artifact page (per-scene before/after, quiet-host timing, parity images, open decisions, feedback map) plus a dated docs/reviews/ record on the branch. |
|
| D-61 | Owner decision (2026-10-01, about 22:18). Feedback is applied now to skills/kiln-compose-scene/ and its runtime-scenes reference. Changes to authoring and QA skills, tool output and project conventions are written as proposals in the report. |
v1-readiness does not touch the compose-scene skill. |
| D-62 | Owner decision (2026-10-01, 22:35). The cycle analyses and implements across all scenes, not only the Farm. Foundry Floor and Golden Gate (robots, cars) give more useful performance data and a better feedback loop for judging whether the approach is right. | Owner's words: test the other scenes to get "more useful information and feedback into our feedback loop". |
| D-63 | Owner decision (2026-10-01, 22:45). Beyond Phases 0-1, the cycle also implements the scene-side merges that each scene's diagnosis proves look-neutral, in scene-kit and scene code only, with no engine changes. Examples are the Golden Gate bridge parts merged by material, and a Foundry-style anchor merge for the Farm's moving heroes with each moving group kept separate. Stand-ins cover every extra pass: the shadow pass and Golden Gate's planar reflection. | |
| D-64 | Re-baseline under D-53 (2026-10-02, recorded by the coordinator). X-02 compares each scene against its optimized draw-after build, keeping the rules: draws and triangles within ±2%, pipelines at most +5%.• Farm: packages/farm/fixtures/x02-baseline.json. 16 views at high, 1280×720, one fresh page per view, settled shadow, WebGPU and WebGL2. Farm and kit sources at 0174e7d. Pixel parity against the sealed pilot remains.• Golden Gate: the perf-local.ts baseline, recorded in packages/golden-gate/PERFORMANCE.md. It has no pilot.• Foundry Floor: packages/foundry-floor/fixtures/count-baseline.json. Its first count baseline, checked by scripts/check-foundry-counts.ts. |
The optimized builds cut draws by design, so ±2% against the old counts fails everywhere: the Farm's old build is flagged on 16/16 views, and Golden Gate's old baseline passes 0/32 high views. Cross-build parity against draw-base: Farm 86/86, Golden Gate 488/489 (one thin-line case held for owner review), Foundry Floor 160/160. Corrected 2026-10-02: Golden Gate's 488/489 is too high. S5's parity run covered golden side-full only at high on WebGPU; the view also fails at economy and balanced (193 px) and on High WebGL2 (200 px). |
| D-65 | Owner decision (2026-10-02, 12:53; sweep D16). Sweep IDs (D1 to D34, D7b, D11b) are those of the 2 October decision sweep as docs/plans/2026-10-02-draw-optimization-landing.md lists them, not D- numbers of this file. The tablet bar. p95 at most 33.3 ms (D-52) and the D-41 hitch limits stay the gate. Adopting a change on the tablet also needs, in 3 of 3 interleaved pairs, the share of frames within one refresh not to fall by more than 1 point and busy × clock not to rise by more than 3%. A miss needs the owner's explicit acceptance. The bar applies to this tablet, the only X-07 device measured, and a higher GPU clock at the same frame rate counts as more GPU work. The timing runner's Mali clock and busy × clock columns: adopted; implementation follows in a later PR after the owner's trip. |
On a 60 Hz tablet, frame time stays at 16.7 ms until the GPU runs out of headroom, so p95 moves only when it is too late. The Mali governor changes clock with load, so busy × clock, not busy alone, is the leading signal of GPU work. |
| D-66 | Owner decision (2026-10-02, 12:53; sweep D17). Re-baselines. A change that an owner decision adopts re-records its scene's count baseline in the same commit, without a decision of its own (D-53). The commit records provenance (build, commit, views, backends) and adds a one-line DECISIONS note naming the decision. | The owner already decides the change; the baseline records its consequence. D-64's bands still catch drift between decisions. |
| D-67 | Owner decision (2026-10-02, 12:53; sweep D18). Thin lines in B-06. The thin-line budget becomes part of B-06 for every scene, defined here: at most 100 pixels that differ by more than 32 levels beyond the repeat capture, per 1280×720 view. The accepted Golden Gate residuals are ratified: golden arrival 74 px, day span 52 px, low shore reflections 21-35 px. The check runs today as --over32-budget 100 in the three draw-parity tools; moving it into compareParityImages for all three: adopted; implementation follows in a later PR after the owner's trip. |
Tile-mean B-06 passed the Golden Gate kerb stipple and the tower tie; the thin-line count caught both. B-06's text is in kiln-commons, which is not versioned and is not edited, so the rule is defined in this repository. |
| D-68 | Owner decision (2026-10-02, 12:53 and 13:08; sweep D20). Farm pilot parity. Before/after parity gates every change. At each Farm release a paired audit runs on WebGPU and on WebGL2: the new build and the last accepted build are captured at the same pilot phase, and a view fails only where the new build fails and the accepted build passes. This amends "pixel parity against the pilot remains" in D-53 and D-64. The sweep's run is this cycle's pilot evidence: against the sealed r34 pilot, draw-after and draw-base got the same verdict in 17 of 17 same-phase pairs. The tool work (the busy-port fix, a paired mode promoted from the sweep's prototype script, and the re-baseline of B-07's texture golden: 39 at every view since the prewarm pass, 41 with the shadow cache) follows after PR 1, before the next Farm change. | Standalone pilot runs failed a different view for each build, for reasons in the tool (a loose phase capture, animations it cannot align). A pair cancels both and still audits drift from the pilot at every release. |
| D-69 | Owner decision (2026-10-02, 12:54; sweep D15). Golden Gate reversed depth. Golden Gate adopts reversed depth through a scene-kit renderer option, off by default, that Golden Gate opts into, with three adaptations under unit tests: the sky is drawn first without a depth test; the reflector's oblique projection is rebuilt for reversed depth; the polygon-offset sign is flipped. The 17 thin-line flags it raises against draw-after are accepted as corrections under D-67. It removes the tower depth tie, so sweep D7 and D7b are moot (D-75). Adopted; implementation follows in a later PR after the owner's trip. | The skyline flicker the owner saw on the hub and the tablet is depth precision: a standard buffer with a 0.5 m near plane and a 160 km far plane steps about 3 m at 5 km and 48 m at 20 km. The prototype took the tower tie from 193 px over 32 levels to 0 px over 8. |
| D-70 | Owner decision (2026-10-02, 12:54 and 12:59; sweep D5 and D6). Golden Gate High reflection. The High reflection drops the traffic, including the driven car, and the fog banks, as WATER-SPEC says (D5). The fog banks get a hidden flag that their update respects, in its own commit (D6). Both: adopted; implementation follows in a later PR after the owner's trip, with the count re-baseline under D-66. |
WATER-SPEC leaves traffic and fog banks out of the reflection. The cycle kept them there (F2), because removing them is a visible change for the owner to decide. |
| D-71 | Owner decision (2026-10-02, 13:08). Vegetation cards in the reflection. Vegetation impostor cards stay in Golden Gate's High reflection. WATER-SPEC's "cards" is read as the fog cards it names at its line 36. No code change. | kiln-commons is not edited, so the reading is recorded here. |
| D-72 | Owner decision (2026-10-02, 12:57; sweep D8). Farm economy small casters. Economy exempts the trailer from the small-caster rule: keepSmallCasters: ['farmer', 'trailer']. Adopted; implementation follows in a later PR after the owner's trip, with image pairs, and the counts if they move. |
At economy the 2-texel threshold stopped the trailer drawbar, hitch ring and hitch pin from casting (146 px over 8 levels). |
| D-73 | Owner decision (2026-10-02, 12:58; sweep D11 and D11b). Farm woodland at minimal. Minimal keeps its 64 woodland cells; woodland packing is not adopted. The gate's reason is the tablet's GPU cost: packing raised busy × clock by 4.5% to 27.2%, higher in 12 of 12 pairs. That replaces SPEC section 15's D-03 reason. No code change. | Packing saved up to 48 draws, but D-65 counts GPU work on the tablet, and packing added it in every pair. |
| D-74 | Owner decision (2026-10-02, 12:58; sweep D14). Tablet tier. Minimal stays the tablet's tier in every scene; D-25 and D-57 stand. Golden Gate at economy may be revisited after its flyover stall is fixed (D-75). No code change. | Golden Gate at economy held 97.7-98.2% of frames within one refresh against 100% at minimal, about 2 points, outside D-65's margin. |
| D-75 | Owner decisions (2026-10-02, 12:54 to 13:06). Answers that change nothing: sweep D7 and D7b (moot: reversed depth removes the tower tie, so there is no tower split); D9 (the live map stays at full tier size); D10 (the Farm hero merge stays variant A); D12 (no hero instancing across placements); D13 (the Foundry Floor campus structures stay as they are); D28 (the tractor's sealed kinematics stay). Backlog: D22 (the three pre-existing stalls, after PR 1; the Golden Gate flyover fix comes before any Golden Gate tier change); D25 (an R8 shadow colour target); D26 (scene-kit correctness soon after PR 1: anchor names, and morph-target and positionNode motion detection); D27 (Golden Gate loads data/layout.json at run time after PR 1; its other deferred items stay in the backlog); D33 (the next performance cycle targets GPU work: per-tree culled packing, then vegetation LOD). |
The owner took the recommended option on every question. |
| D-76 | Owner decisions (2026-10-02, 13:00 to 13:06; sweep D1, D4, D19, D21, D24, D29, D30, D31, D32). Landing. Two PRs (D1, D4). PR 1 carries every draw-optimization commit except e72b610, the records corrections (D30) and the compose-scene corrections (D24). PR 2, the engine work (e72b610, the optimize: 'full' wiring, D2, D3, D23 and D34), follows 0.10.0, which merged as 9375d04; docs/reviews/2026-10-02-draw-optimization-engine-handoff.md describes it. The Farm ships at the next site build after the site's own Farm build check; Golden Gate and Foundry Floor are resealed later in one site change with an approved upload (D29, D-12). The leaked browser profiles were removed on 2 October (D32).D19 (the portable runner includes scenes/scripts/tests, Golden Gate's count baseline JSON, glb-world's staged path, a release checklist), D21 (the 14 hub cells at balanced and economy and the Golden Gate hero orbit, on a quiet hub), D31 (one compose-scene dogfood session each in Agy and OpenCode) and D32 (launchChrome removes its profile on close, with a test): adopted; implementation follows in a later PR after the owner's trip. |
The owner split the landing on 2 October: "avoid tedium and ceremony; get this over the line". |
| D-77 | Owner decision (2026-10-26, Matthew). The scene and pack pages linked from the kilnstudio.tools home page that were still unindexed are indexed and listed in the sitemap: /scenes/golden-gate/, /scenes/farm/, /scenes/foundry-floor/ and /packs/foundry-floor/. His sign-off: "i sign off on all those and the holodeck scenes". This changes site decision S-2 (Foundry Floor scene noindex, nofollow, pack noindex, follow) and the indexing part of D-33; the Foundry Floor pages keep their In production label and no-affiliation line. The Foundry Floor asset pages under /gallery/foundry-floor/ and the archive item pages stay noindex. The Holodeck pack, asset and world pages were already indexed. |
foundryFloorErrors in site/scripts/static-validation-core.mjs now requires both Foundry Floor pages to be indexed and in the sitemap; the scene pages drop their noindex/nofollow props. |
D-18: the look of the scenes
Rule: a scene uses the tone mapping its assets were qualified under.
- Farm keeps ACES Filmic (exposure .95, INV A.5), for parity and after it.
- The r33 and r34 Farm assets were reviewed and sealed in the pilot's ACES look.
- The owner decided tonight to leave the Farm alone.
- Switching its tone mapping could expose colour choices made for that look.
- Golden Gate and Foundry Floor use three's
NeutralToneMapping(Khronos PBR Neutral).- Their assets are being authored tonight under Kiln's calibrated review rig,
review-neutral-v1. That rig also uses a PBR Neutral tone map, and the owner chose "accurate lighting everywhere" for it. AgXToneMappingis the documented fallback, used only if B-00h shows Neutral failing on either backend.
- Their assets are being authored tonight under Kiln's calibrated review rig,
- What the kit must do:
- support
ACESFilmicToneMapping,NeutralToneMappingandAgXToneMappingasRendererLook.toneMapping; - have B-00h test Neutral and AgX on both the WebGPU and WebGL2 backends of 0.186.0, and record the result in PROGRESS.md.
- support
- Owner evidence, without changing the default. At M3, capture three Farm views twice, once with ACES and once with Neutral at matched exposure: the hero view and two others that show saturated materials. Store them under
evidence/look-ab/. The owner decides in the morning. The default does not change.
Coordinator notes for the builder
- The Golden Gate builder is a separate session. It starts only after PROGRESS.md carries
KIT-M1: READY <date>(section 21.5). - After that line exists, answer entries in
KIT-REQUESTS.mdin PROGRESS.md, following section 2, rule 4.
Conflicts resolved after the builder's first stop (2026-09-29, 03:50)
| Id | Conflict | Resolution |
|---|---|---|
| C-01 | Section 15 sets the Farm high-tier pixel-ratio floor at 1.0, but the section 8.3 ladder example reaches .75. |
Keep 1.0. The governor never steps below a tier's floor, so the example does not apply to Farm high. |
| C-02 | Under the 8 percent benefit rule, a constant 30 ms trace reverts, but U-16 expects it to reach the floor. | Keep the benefit rule. U-16 reaches the floor with an improving trace, and a separate case asserts the revert under constant load. |
For the rest of the night the builder resolves further conflicts itself, provided every MUST, owner decision and boundary stays intact, and logs each one in PROGRESS.md. See site-build/FEEDBACK-scenes-review-1.md.
STOP-03 resolved (2026-09-29, 05:45)
| Id | Conflict | Resolution |
|---|---|---|
| C-03 | React DOM 19.3.0 installs one shared, non-passive selectionchange listener on the document at the first createRoot. root.unmount() never removes it. Section 5.2(1) forbids document listeners, and sections 5.3 and B-05 require teardown. |
See below. |
| C-04 | R2-13: a predeclared initialisation protocol for B-05 and X-03 (warm-up, then a forced garbage collection, then ten measured cycles at under 5 percent). | Endorsed as predeclared. The builder also runs and reports a third block of ten cycles, for the trend only. If the third block grows by at least as many bytes as the measured block, that is a finding to investigate before M4, not a gate result. |
C-03, in detail: a narrow exception for React DOM's listener.
- What is exempt. Section 5.2(1) governs the listeners the kit and the scenes add. React DOM's listener is page infrastructure that React owns. The website already loads React DOM for its own components, so every page that hosts a scene carries this listener anyway. It is exempt from 5.2(1) and from the teardown requirements of 5.3 and B-05, subject to three conditions:
- After full teardown from a cold mount, the only document or window listener left is that one non-passive
selectionchangelistener, identified by its type and by React's marker on the document. Nothing else remains. - Later mount and unmount cycles never add a second one: the count stays at one through every B-05 cycle.
- REPORT.md lists it as a known limitation of the pinned React DOM, citing the source locations in
evidence/m1/stop-rule-03.md. The website handoff notes it too.
- After full teardown from a cold mount, the only document or window listener left is that one non-passive
- No change of mounting architecture.
Standing rule for pinned third-party infrastructure. When a pinned dependency (React DOM, R3F or three) forces a deviation from the contract, the builder may apply the same pattern and continue, provided that:
- there is no supported fix;
- the behaviour the page can observe (the signatures, callbacks, error codes and events of section 5) is unchanged.
The pattern is to identify the deviation exactly, bound it, assert that it does not grow, and document it. List every such exception in the final message. Any change the page can observe still stops work.