img2threejs — Image to procedural Three.js

Rebuild the object visible in a reference image as a code-only procedural Three.js model, gated by a staged sculpting pipeline and an AI-vision self-correction loop. This is reconstruction-by-code, not photogrammetry, mesh extraction, or downloaded art packs.

Agent-agnostic: works under Claude Code, Codex, or OpenCode. Wherever this doc says "agent vision" or "agent browser tool", use whatever the host provides — native image reading, a browser MCP (playwright/chrome-devtools), the project preview, or a user-supplied screenshot.

Canonical shared checkout

Keep one checkout of this repository and let every host enter it through a symlink, so Claude and Codex execute the same code instead of drifting apart:

~/.claude/skills/img2threejs -> <your checkout>
~/.codex/skills/img2threejs  -> <your checkout>

When To Use

The user attaches/points to an object image and wants a procedural Three.js model, a reconstruction/animation/destruction plan, a sculpt spec, or code. Also for material studies, action-ready props, game objects, botanical/mechanical parts, and stylized reconstructions.

Core Promise

Sculpt from a photo, in order — never one-shot a mesh:

1. Run python3 forge/next.py <spec> first, or python3 forge/next.py --state .img2threejs/state.json. The state form reports the ordered local checklist, exact next command, evidence status, and bounded correction-loop status; it never replaces the spec/pass gates.
2. Use local state first. Initialize it once, then run python3 forge/next.py --state .img2threejs/state.json [<spec>] at every start/resume and before every correction iteration. Obey a hard stop; never continue from memory.
3. Validate the image is a suitable 3D target (grimoire/intake/validation_rubric.md).
4. Assess object class + complexity, then write a qualityContract before any code.
5. Spec it: component hierarchy, materials, lighting, pivots, sockets, action anchors.
6. Build pass-by-pass from blockout → structure → form → material → lighting → interaction → optimization.
7. Verify each pass with a screenshot compared against the reference; fail a pass if an identity-defining feature is wrong even when the global score looks fine.

State explicitly when output is approximate/stylized/low-poly. A single image cannot reveal hidden sides or guarantee exact geometry — say so instead of faking confidence.

Resumable local workflow

For a cross-agent or multi-session reconstruction, initialize the local state before intake:

python3 forge/state.py init --reference <image> --profile character --spec object-sculpt-spec.json
python3 forge/next.py --state .img2threejs/state.json
python3 forge/state.py mark image-analysis --evidence analysis.md

generic, character, and cs2 profiles insert their required intake gates in order. Every completed step needs evidence; every skipped step needs a reason. The state file is a resumability index, not visual evidence: renders, specs, review history, and deterministic gates remain the authoritative artifacts.

Transparency and Process Debugging (Critical — from Bowie Knife reconstruction)

The problem: When the user cannot tell what was done or where something went wrong, they cannot debug the process. Over-claiming (reporting success when features still don't match) destroys trust and makes iterative improvement impossible.

Rule: Be transparent + don't over-claim. State exactly what changed each pass, with evidence, and name what still doesn't match:

• After each pass, explicitly list what changed: "Updated guard shape to extend left edge from -0.56 to -0.48 for handle overlap"
• Provide evidence: reference the specific values, coordinates, or parameters that changed
• Name what still doesn't match: "Handle silhouette traced but still flat plane (no Z palm-swell), procedural crosshatch not reference's exact dot-grid knurl"
• Explain why a change was made: "Extended guard left edge because handle ends at X=-0.42 and guard ended at X=-0.20, causing visual gap"
• Never claim a feature is "done" when it's only "improved" — use precise language
• When a gate passes but visual inspection shows issues, explain the limitation: "2D gate passed (fidelity 0.83) but three-quarter render shows blade reads as toy (no grind wedge) — 2D gates are blind to 3D realism"

The user needs to be able to debug the process, not just the output. If something is wrong, they should be able to trace which decision led to the error and correct it. Opaque processes force restarts; transparent processes enable refinement.

Transparency and Process Debugging

Report what changed each pass with evidence (exact values/coordinates), name what still doesn't match, and never claim "done" when only "improved". A passing gate is not proof of 3D realism. Full rule + examples: grimoire/review/self_correction.md.

GLB-mediated v2 render-fidelity track (1.5 alpha)

When the user supplies a GLB as an intermediate reference, use the browser-rendered GLB as the structural and visual baseline, then author an independent procedural factory. The raw GLB is never pixel evidence and its topology/materials are never copied into the factory.

Before any factory edit:

1. Run forge/stage1_intake/probe_glb.py and inspect semanticDecomposition. A merged one-node/one-mesh/one-primitive/one-material asset is insufficient for reliable semantic labels; connected-component/curvature/normal/UV segmentation is hypothesis evidence only. Request a multipart GLB or capture a browser semantic-ID pass before claiming exact regions.
2. Author and validate one shared render-profile.v2 with forge/stage4_review/validate_render_profile.py. Both GLB and procedural routes must use the same output color space, linear working space, tone mapping/exposure, PMREM environment, viewport/DPR, camera, background and lighting settings.
3. Capture six passes for every admitted view: beauty, alpha-silhouette, semantic-id, depth, normal, and roughness-material-id. Use forge/stage4_review/compare_region_passes.py for deterministic global/per-region evidence; missing semantic-ID data blocks per-region confidence instead of falling back to whole-image color or silhouette scores.
4. Use region-specific continuous geometry/material strategies. Do not replace a face/head volume, cloth shell, kasa, staff, or tail with a generic collection of floating primitives when the region's silhouette or attachment requires a continuous surface.
5. Run one correction group per loop in this order: camera → silhouette → face → clothing → accessory → materials → lighting. Recapture the full pass set after each group and record the changed group, hashes and score. Never combine groups when diagnosing improvement.

The machine-readable contract lives in docs/specs/render-profile.v2.schema.json and docs/specs/render-profile.v2.example.json. The executable manifest bridge accepts --render-profile and exposes record-pass; its glb-mediated-v2 validation is fail-closed.

Required Inputs

• one image path / screenshot / URL / attached image (if missing or unreadable, ask)
• intended use: prop, game object, hero render, playable/destructible object, animation rig (default: real-time browser prop with interactive performance)
• for a CS2 request, an authoritative classification record (family/subtype and evidence refs) or an explicit request for the user/vision provider to supply one; heuristic detection alone is not enough to select a geometry adapter

Mandatory Local State Gate

Conversation context is disposable; .img2threejs/state.json is the local checklist authority. Initialize it once per reconstruction:

python3 forge/state.py init --state .img2threejs/state.json --reference <img> --profile <generic|cs2|character>

At every fresh start, resume, or correction loop, run python3 forge/next.py --state .img2threejs/state.json [object-sculpt-spec.json] before touching code. It prints the current step, pass, incomplete mandatory steps, exact next command, and loop/max. Exit code 3 or status=stopped is a hard stop: report the reason and request input. Never bypass it by reconstructing progress from chat history.

After evidence exists, record it with python3 forge/state.py mark <step-id> --state .img2threejs/state.json --evidence <path>. Mark a non-applicable step skipped only with --reason; silent omission is forbidden. Loop counts are derived from reviewHistory actions refine-spec/refine-code, not agent memory. Defaults are 3 corrections per pass and 6 total.

Profiles add mandatory gates rather than changing the core order: cs2 requires classification, manifest, and a machine-readable CS2 review before AI review; character requires the character contracts and landmark evidence. Every profile records suitability, projection applicability, and material-evidence applicability; conditional steps require evidence or an explicit skip reason.

The Loop (scripts do enforcement; agent vision does judgment)

Run scripts from the skill root (forge/...). Pure Python 3.10+ stdlib, no pip installs. Full flags: grimoire/scripts.md. Never let a script score visuals — that is the agent's job.

1. Analyze the image first (agent vision, before any script): work the layered observation protocol in grimoire/intake/image_analysis.md — identify/classify, decompose macro→meso→micro, map part relationships, name materials in PBR terms, list identity-defining features, and flag what the single view hides. Observation before inference; controlled 3D vocabulary; 3D object-space not 2D image-space. This is generic for any subject and feeds every field below. Then probe local images: forge/stage1_intake/probe_image.py <image> (metadata only, not a visual check). 1a. Local Spec Search — after image analysis and before writing or refining a spec, local evidence is a pipeline stage, not an optional memory lookup, whenever the request needs domain-specific anatomy, PBR, wear, geometry, runtime, or physics specifications. The pre-spec command automatically runs BM25, chooses cs2 for CS2 targets and core_3d otherwise, and writes a localSpecSearch evidence bundle into the assessment: python3 forge/stage2_spec/new_pre_spec_assessment.py "Name" --image <img> --out assessment.json. Add observed terms with repeatable --spec-query "<term>"; use --collection <collection> only when the automatic collection choice is insufficient. new_sculpt_spec.py --assessment carries that bundle into the final spec, including snippets, source_refs, and evidence_refs. For extra focused retrieval, the direct CLI remains available: python3 forge/stage1_intake/search_specs.py "<query>" --collection <collection> --limit 3 --snippet-chars 250 --json. For CS2, include the anatomical and the colloquial name, for example --spec-query "safety ring finger ring" or search_specs.py "roughness matte" --collection cs2. Expand queries with object names, component names, material/finish terms, behavior terms, and known aliases; retry focused alternatives when the first result is incomplete. Build the spec from returned evidence and do not invent domain specs when local evidence exists. Search caches are local/generated only; preserve JSONL records and source provenance rather than replacing them with cache output. 1aa. Optional fidelity evidence adapters — use them only when they improve an observed weak point; the stdlib core remains authoritative. Route thin/complex masks to local SAM2, character face/pose evidence to MediaPipe, and weak front/back cues to Depth Anything V2: python3 forge/stage1_intake/run_vision_adapter.py <segment|landmarks|depth> .... Every adapter emits provenance. Confirm SAM2 selected the intended component; treat monocular depth as relative only; review landmarks before copying them into anatomy. For browser work, prefer Chrome DevTools MCP for live console/network/performance diagnosis, use threejs-devtools MCP read-only to inspect scene/material/renderer state, and reserve Playwright MCP for cross-browser or host fallback. MCP-only scene mutations never count as implementation: write the proven change back to the spec or TypeScript, rebuild, and recapture. Use Context7 only with the target project's installed Three.js version; local types, typecheck and runtime smoke tests override live docs. Full routing and commands: docs/integrations/reference_fidelity_tooling.md. 1b. CS2 intake manifest — for a CS2 request, create and validate cs2-intake.json before pre-spec authoring. Run admission and probing for every source view, record the heuristic signal as non-authoritative evidence, attach the classification record, resolve the supported family, and choose route independently from exactnessTier. Missing classification, insufficient coverage, or a contradictory high-confidence class is request-input; unsupported families do not continue into spec generation.
2. Pre-Spec Assessment Gate — classify + score complexity + write the quality contract: forge/stage2_spec/new_pre_spec_assessment.py "Name" --image <img> --complexity <simple|moderate|complex|ultra-complex> --out assessment.json. Rules: grimoire/intake/quality_contract.md. Set objectClass.primaryDomain (object | character | hybrid) and fill the seeded detailInventory (its targetMinDetails scales with complexity). Supported CS2 knife skins and Glock-18 assets: always pass --cs2, which defaults the complexity tier to ultra-complex (targetMinDetails 16) — the finish/wear/hardware is the item, so CS2 is held to the top fidelity bar; targetMinDetails never drops below the 9 floor even if downgraded by hand. Author procedural GEOMETRY (blade/guard/grip profiles) but make the FINISH a de-lit reference-crop PROJECTION, not a procedural finish material — projecting the photo's own pixels is what reaches reference fidelity for patterned skins (Doppler/Gamma/Marble/Fade), and is what the v1.3 baseline demos do; a procedural finish for a patterned skin reads visibly wrong against the reference. Take the projection path in step 2c (it generalizes from characters to any reference-matched surface). Procedural finish is the fallback ONLY when live view-dependent response matters more than matching this one reference. Finish routes + rulebook: grimoire/build/cs2_finishes.md; optional exact-texture acquisition: grimoire/intake/cs2_texture_acquisition.md. 1a. Local Spec Search — after image analysis, before writing or refining a spec, pull local domain evidence (anatomy/PBR/wear/geometry/runtime/physics) rather than inventing it: python3 forge/stage2_spec/new_pre_spec_assessment.py "Name" --image <img> --out assessment.json (auto-runs BM25, auto-picks cs2/core_3d collection, writes a localSpecSearch bundle that new_sculpt_spec.py --assessment carries into the spec). Full query-expansion recipe (bilingual terms, focused search_specs.py retrieval, cache rules): grimoire/intake/local_spec_search.md. MUST read it before retrying an incomplete or domain-specific query. 1b. CS2 intake manifest — for a CS2 request, create and validate cs2-intake.json before pre-spec authoring (admission, heuristic signal, classification, family/route resolution). MUST read grimoire/intake/cs2_intake_contract.md completely before creating the manifest or running pre-spec assessment.
3. Pre-Spec Assessment Gate — classify + score complexity + write the quality contract: forge/stage2_spec/new_pre_spec_assessment.py "Name" --image <img> --complexity <simple|moderate|complex|ultra-complex> --out assessment.json. Rules: grimoire/intake/quality_contract.md. Set objectClass.primaryDomain (object | character | hybrid) and fill the seeded detailInventory (its targetMinDetails scales with complexity). Supported CS2 knife skins: always pass --cs2, which defaults the complexity tier to ultra-complex (targetMinDetails 16, floor 9) — the finish/wear/hardware is the item, so CS2 is held to the top fidelity bar. Author procedural GEOMETRY but route the FINISH through the projection path in step 2c — a procedural finish for a patterned skin (Doppler/Gamma/Marble/Fade) reads visibly wrong against the reference. Finish routes + rulebook: grimoire/build/cs2_finishes.md; optional exact-texture acquisition: grimoire/intake/cs2_texture_acquisition.md. 2b. Detail inventory (do not skip for detailed subjects) — scan zones and enumerate every identity-defining small detail (gloss, bevel, fasteners, linework, contours, stains): forge/stage1_intake/build_detail_inventory.py <image> --mode grid-3x3 --out-dir <dir> --out di.json. Each detail MUST map to a component.localFeatures or material.localOverrides entry — never prose only. Taxonomy + 3D-term recipes: grimoire/intake/detail_inventory.md. 2c. Projection-first fidelity (characters AND reference-matched surfaces — supported CS2 skins, decals, painted patterns) — when the goal is matching a specific reference's surface, put the photo's own pixels on the mesh instead of approximating them procedurally. This is the single biggest fidelity lever; a procedural material for a patterned surface is the #1 reconstruction failure. Recipe (grimoire/character/likeness_maximization.md — its two levers, align-mesh+camera and project-the-photo, generalize past characters): solve the camera (stage1_intake/solve_camera_pose.py → referenceCamera), de-light the reference so it is free of baked lighting (stage1_intake/delight_albedo.py, hard requirement — this is what makes projection safe, not the flat-lit icon), then project the de-lit crop onto the mesh and bake it into UVs (stage3_build/bake_projected_texture.py --mesh-id <id>). For a CS2 skin the mesh is the procedural family-specific component tree you author in the spec, and the projected de-lit crop IS the finish (front + back from the two views) — no procedural Doppler material. For characters, first capture landmarks (stage1_intake/extract_landmarks.py --out anatomy.json), fill preSpecAssessment.anatomy, route grimoire/character/reconstruction.md. A single view cannot show hidden sides — report per-region confidence and request more views when it matters. Character sub-routes, in the order they are needed — decide what parts exist before shaping any of them, and shape the head before the hair that sits on it:
   • Parts — grimoire/character/structure_decomposition.md: which parts the figure is made of, and where each one's boundary falls.
   • Head — grimoire/character/head_construction.md: skull, face plane and feature placement, the sub-route the likeness gate reads against.
   • Hair — grimoire/character/stylized_hair_threejs.md, with the parameter contract in grimoire/character/threejs_hair_parameter_contract.json. Lock topology first: material tuning cannot repair wrong lock topology, so run it only after the silhouette review passes. 2d. Reference-free humanoid — when the request is a generic figure with no reference image ("a low-poly humanoid"), there is nothing to measure, so fill anatomy from public canon with forge/stage2_spec/humanoid_proportions.py <spec> --style-heads 8 --in-place. It writes anatomy.source: "canon-table" so canon is never mistaken for measured evidence, refuses to run at all when the spec names a reference image, and lists what the corpus does NOT supply under anatomy.unsourced. Only complete head counts are derivable (currently 8); anything else fails with the missing landmark named rather than being interpolated.
4. Author the spec from the assessment: forge/stage2_spec/new_sculpt_spec.py "Name" --image <img> --assessment assessment.json --manifest cs2-intake.json --out object-sculpt-spec.json. Replace generic starter featureReviewTargets with the object's real identity-defining systems (≤5 critical, ≤3 important per pass); for characters add anatomy-proportion, face-landmark-placement, pose-silhouette, outfit-and-palette. Use 3D-graphics terms only (grimoire/glossary/3d_vocabulary.md), never "nice/smooth/shiny". Classify every component's topologyClass/topologyRationale per grimoire/intake/surface_topology.md before picking a primitive — this is what prevents a continuous organic form from being picked as a box.
5. When material fidelity matters and a source image exists, analyze each material's finish then extract reference PBR evidence, both per crop (crop the correct region — verify the crop is on the part you think it is):
   • forge/stage1_intake/analyze_texture.py <crop> --spec spec.json --material-id <id> --in-place classifies the finish (gem-metal | gemstone | painted-metal | worn-composite | brushed-steel | plastic), extracts the gradient palette, and writes doc-grounded MeshPhysicalMaterial scalars (metalness/roughness/clearcoat/transmission/ior/anisotropy/envMapIntensity) onto the material. Recipes + Three.js texture/PBR rules (colorSpace, CanvasTexture/DataTexture, height→normal) live in grimoire/build/threejs_texture_reference.md. Rule of thumb: solid albedo for flat paint, real reference crop for patterned finishes (doppler/quartz/hydro-dip/camo).
   • forge/stage1_intake/extract_pbr_evidence.py <crop> --out-dir <dir> --material-id <id> --target-threshold 0.7. Confidence < 0.7 is a stop/refine-input signal, not a pass. It is inference, not inverse rendering.
   • For multiple named regions, use forge/stage1_intake/material_region_analysis.py --manifest regions.json --out-dir material-evidence --out material-analysis.json. Resolve each accepted assignment from docs/materials/material-reference.json, then wire it into the spec with forge/stage2_spec/apply_material_analysis.py.
   • Emit the controlled material camera/crop contract with forge/stage4_review/material_views.py, compare saved visible-footprint crops with forge/stage4_review/material_comparator.py, apply only bounded material-scoped corrections with forge/stage4_review/material_feedback.py, and record the blocking result with forge/stage4_review/material_gate.py.
6. Validate, then strict-validate before generating code: forge/stage2_spec/validate_sculpt_spec.py object-sculpt-spec.json then --strict-quality. Strict blocks shallow specs (a complex object with one root, no repetition systems, no local overrides, no micro groups is NOT implementation-ready even if JSON validates).
7. Locked build passes — only touch the currently unlocked pass: forge/stage3_build/orchestrate_passes.py status object-sculpt-spec.json forge/stage3_build/generate_threejs_factory.py object-sculpt-spec.json --out src/createObjectModel.ts (generator is fail-closed: strict-quality must pass before it can write any factory; a future --pass-id also fails until prior passes are reviewed continue). If blocked, preserve the BLOCKED artifact and refine the subject-specific spec; do not substitute a generic template. 6a. Hitting a triangle budget. performanceBudget.targetTriangles selects a tessellation tier for every primitive that has segment counts (low ≤6k, standard ≤60k, else hero), and caps implicit-surface sampling grids. Where that is not precise enough — an SDF's grid is quantised, so a tier can only get near a number — add geometryDescriptor.decimate: {"targetRatio": 0.4} to that component. It emits a Garland-Heckbert quadric collapse into the generated factory and runs before skin binding, so weights are computed on the surviving vertices and no skinning data is interpolated across a vertex merge. It keeps position only, recomputing normals, so it is refused on an authored/unwrapped uvStrategy. For offline LOD tiers from an exported mesh, forge/stage3_build/decimate.py <mesh.json> --ratio <r> --json is the same algorithm.
8. Render the current pass in a browser/preview, capture a screenshot at a review viewpoint. 7a. Off-axis and placement gates — a single review viewpoint is not evidence about the model. Capture a turntable, not one frame, and run all three. Each catches a defect class the older gates pass by construction; skipping them is how a hole through a skull, a hat at hip height and a charm floating below the ground plane survived eight front-only review rounds. forge/stage4_review/turntable_gate.py --capture 0=front.png --capture 90=right.png --capture 180=rear.png --capture 270=left.png --json node runtime/scripts/export_mesh_geometry.mjs --url <preview> --out meshes.json then forge/stage4_review/self_intersection.py meshes.json --json forge/stage4_review/attachment_anchor.py object-sculpt-spec.json --measured measured.json --json All three exit 0 clean / 1 gate failure / 2 error. A failure blocks continue for the pass even when the global fidelity score passes — the score is computed from one camera and a 64×64 luma grid, and neither can represent any of these defects. Read sampledVertexCount / unmeasuredAttachments / missingAzimuths before believing a clean verdict: each names the part of the model the gate did not actually look at.
9. Package one side-by-side sheet, then inspect it with agent vision: forge/stage4_review/make_comparison_sheet.py --reference <img> --render <shot> --out cmp.png --json.
10. Record the review (overall + per-layer + per-feature scores + decision): forge/stage4_review/append_review.py object-sculpt-spec.json --pass-id <pass> --fidelity <0-1> --action <continue|refine-spec|refine-code|request-input|stop> --summary "..." --render-screenshot <shot> --comparison-image cmp.png --ai-vision-score <0-1> --layer-scores-json '{...}' --feature-reviews-json <f.json> --in-place. For the CS2 family path, also attach the versioned report with --cs2-review-json cs2-review.json --review-scene-json forge/tests/fixtures/knife_review_scene.json. A failed family, painted-region, projection-coverage, critical-detail, or orbit gate blocks continue even when the global score passes. See docs/cs2/review-gates.md.
11. Sync pipeline state after manual review edits: forge/stage3_build/orchestrate_passes.py sync object-sculpt-spec.json --in-place.

Forge Runtime Contracts

Subdivision runtime tests compile generated TypeScript against the showcase checkout. Set IMG2THREEJS_SHOWCASE_ROOT to that checkout; without it, local runtime-only tests skip with an actionable message while static contracts still run. CI should set IMG2THREEJS_REQUIRE_SHOWCASE=1 to turn a missing showcase checkout into a test failure.

IMG2THREEJS_SHOWCASE_ROOT=/path/to/img2threejs-showcase python3 forge/tests/test_subdivision.py
IMG2THREEJS_SHOWCASE_ROOT=/path/to/img2threejs-showcase python3 -m unittest discover -s forge/tests
IMG2THREEJS_SHOWCASE_ROOT=/path/to/img2threejs-showcase python3 forge/tests/test_showcase_tsc_smoke.py

(generator is pass-gated: a future --pass-id fails until prior passes are reviewed continue). The local state adds --force only for a new pass or refine-spec; refine-code edits the current artifact without regenerating it. Before overwriting, carry valid hand refinement back into the spec; generated code must not be the only copy of reconstruction decisions. 7. Render the current pass in a browser/preview, capture a screenshot at a review viewpoint. 8. Run deterministic gates before AI vision. MUST read grimoire/review/gates_reference.md and grimoire/review/self_correction.md completely. Run forge/stage4_review/diagnose_render.py and record the passing Tier 1 result with --spec object-sculpt-spec.json --pass-id <pass> --in-place; for non-planar forms also run forge/stage4_review/diagnose_render_multi_angle.py with the fixed view and at least two meaningful orbit views. Then run forge/stage3_build/orchestrate_passes.py check object-sculpt-spec.json --pass-id <pass>. 9. Package one side-by-side sheet, then inspect it with agent vision: forge/stage4_review/make_comparison_sheet.py --reference <img> --render <shot> --out cmp.png --json. 10. Record the review (overall + per-layer + per-feature scores + decision): forge/stage4_review/append_review.py object-sculpt-spec.json --pass-id <pass> --fidelity <0-1> --action <continue|refine-spec|refine-code|request-input|stop> --summary "..." --render-screenshot <shot> --comparison-image cmp.png --ai-vision-score <0-1> --layer-scores-json '{...}' --feature-reviews-json <f.json> --in-place. For the CS2 knife path, also attach the versioned report with --cs2-review-json cs2-review.json --review-scene-json forge/tests/fixtures/knife_review_scene.json. Produce that report first with forge/stage4_review/cs2_review.py --manifest cs2-intake.json --metrics cs2-review-inputs.json --scene forge/tests/fixtures/knife_review_scene.json --out cs2-review.json. A failed family, painted-region, projection-coverage, critical-detail, or orbit gate blocks continue even when the global score passes. See docs/cs2/review-gates.md. 11. Sync pipeline state after manual review edits, record checklist evidence, then re-run the local state gate before another correction or pass: forge/stage3_build/orchestrate_passes.py sync object-sculpt-spec.json --in-place python3 forge/next.py --state .img2threejs/state.json object-sculpt-spec.json. 12. Before declaring completion, run forge/stage4_review/check_part_coverage.py --spec object-sculpt-spec.json --manifest parts.json and verify the action-ready hierarchy. Mark part-coverage and action-ready only with evidence.

CS2 image-matched rule

For a CS2 item, the target is observable agreement between the supplied image and the rendered item: silhouette, proportions, edge profile, hardware layout, coating colour, pattern placement, wear, roughness response, and camera framing. Every decision must be traceable to evidence or be labelled as an approximation.

The initial CS2 family boundary covers supported knife subtypes and the Glock-18 pistol adapter. Rifle, SMG, sniper, heavy, glove, unsupported pistol, and unknown knife subtypes must stop with unsupported-family or unsupported-subtype; they must not receive another family's component tree as a generic fallback.

Layer contract

Pass these records between layers. Do not copy an informal vision description into the next stage:

The canonical hand-off is cs2-intake.json (schemaVersion: 1). Its state is one of proceed, request-input, fallback, rejected, unsupported-family, or unsupported-subtype. Write it atomically and preserve unknown provider fields under extensions; a fallback must never erase prior evidence.

CS2 intake order

1. Admit and technically probe every view. Reject undecodable, empty, tiny, fragmented, or duplicate references before classification.
2. Record the heuristic CS2 signal only as a routing hint. detect_cs2.py is never authoritative identity evidence.
3. Require a classification record before selecting a family adapter. If classification is absent, timed out, or contradicts a high-confidence objectness result, return request-input.
4. Resolve identity in this order: explicit user metadata, uniquely resolved metadata, then the authoritative classification record. Preserve ambiguity rather than guessing.
5. Select route and exactness independently:
   • reference-projection: default for matching a specific patterned image;
   • authored-texture: only when independent texture maps are supplied or legally acquired;
   • procedural-finish: fallback when projection evidence is unavailable or live response is the stated priority. Exactness is image-only, metadata-assisted, or exact-texture; changing route must not silently upgrade or downgrade the evidence tier.
6. Select the family adapter only after family/subtype validation. Record painted regions, unpainted substrate, visible hardware, hidden-region confidence, and every approximation in the spec.
7. For projection, solve the camera and de-light the source first. Projected pixels provide colour evidence, not automatic geometry truth; geometry still comes from the adapter and silhouette review.

Surface and review rule

For a specific CS2 reference, preserve the reference's own colour/pattern pixels whenever legal and technically possible. Procedural Doppler/Fade/Gamma/Marble patterns are not equivalent to the input image and may only be used with an explicit procedural-finish route and approximation warning. Keep albedo, roughness, metalness, normal/height, AO, mask, and wear as independent channels. Record channel source, colour space, UV orientation, dimensions, packed-channel decoding, and missing-channel derivation. A low-confidence PBR inference is a refine-input signal, not proof of exact material.

Single-view reconstruction may proceed only when visible identity features are sufficiently covered; hidden blade sides, underside, and back hardware must carry inference confidence and may trigger request-input. Review the fixed camera plus two meaningful orbit views. Report what changed, which evidence caused it, what still differs, and choose exactly one next action: continue, refine-spec, refine-code, request-input, or stop.

Gates (do not skip)

• Suitability + reference integrity: pass / conditional / reject before any planning (grimoire/intake/validation_rubric.md), AND every reference admitted via forge/stage1_intake/check_reference_admission.py (rejects empty/fragmented/tiny/duplicate/ undecodable refs with a reason). Intake understanding cross-checked by forge/stage1_intake/check_intake_correctness.py (halts on a confident class contradiction).
• Divine Eye (the harness heart) — deterministic-first, model-last: the render evaluator is forge/stage4_review/divine_eye.py — a zero-token multi-signal ensemble (IoU/scale HARD gates; proportion/symmetry-parity/pHash/SSIM/edge/blowout/flat/tonal-parity soft) with self-uncertainty (probe on signal disagreement) and deterministic routing (continue/refine-spec/refine-code/ probe). The VLM (forge/stage4_review/vlm_gate.py) is a gated, calibrated, cross-checked last layer: never consulted on a hard-gate failure, multi-sample-voted, and can rescue a soft near-threshold reject but never grant past a hard geometric failure.
• Multi-angle or it didn't happen: a non-planar form must hold from ≥2 camera angles. forge/stage4_review/diagnose_render_multi_angle.py flags degenerate-view when an orbited silhouette collapses (a flat plane faking a volume). Orbit angles use reference-free self-consistency — never scored against a reference angle the photo doesn't cover.
• CS2 review contract: forge/stage4_review/cs2_review.py consumes the manifest and versioned scene fixture, then blocks wrong family identity, missing projection coverage, painted-region mismatch, critical identity-detail failure, finish/material response failure, and degenerate orbit form. It records exactness tier, hidden-region confidence, per-region confidence, approximation notes, camera, environment hash, exposure, tone mapping, resolution, background, and renderer version.
• Bounded correction loop (token-burn safety): forge/stage4_review/correction_loop.py guarantees termination: hard gates route to refine-code; repeated defects and oscillation route to refine-spec; plateau and the hard ceiling route to request-input — never a silent infinite burn. Deterministic analysis-by-synthesis parameter fitting and Divine Eye provenance are documented in grimoire/build/analysis_by_synthesis_fitting.md.
• Executable Divine Eye fitting: fit_against_divine_eye() in forge/stage4_review/fit_params.py connects deterministic parameter-to-render callbacks to bounded gate-aware Divine Eye optimization. Clean candidates use raw fidelity; hard-gated candidates score below all clean results while retaining original fidelity and provenance. The returned objective and optional selected raw fidelity are explicit, and each copied record has candidate/reference/render provenance. It lazily loads the default evaluator and returns normalized raw-fidelity correction-loop provenance without mutating sources.
• Tier 1 (legacy, still valid): "Tier 2 (AI-vision) never runs against a render that has not passed Tier 1." Run forge/stage4_review/diagnose_render.py (silhouette IoU/proportion/symmetry/per-part color) and record it (--spec ... --in-place) before requesting a comparison sheet; orchestrate_passes.py check refuses otherwise.
• Pre-spec / strict-quality: blocks code gen until the spec is deep enough for its contract.
• Screenshot feedback: continue is allowed only with a render + comparison sheet + global AI-vision score ≥ threshold (default 0.7) AND every critical feature ≥ its own threshold. Details + per-layer scorecard: grimoire/feedback/render_capture.md.
• Action-ready: build a runtime hierarchy (pivots, sockets, colliders, destruction groups), never an inert lump; expose root.userData.sculptRuntime. grimoire/readiness/action_rigging.md.
• Procedural rig contract (1.5-alpha): for humanoid/character builds, validate the authored joints/parents/names/matrix_local/packed skin payload with forge/stage5_rig/validate_rig_payload.py before binding THREE.Skeleton. The gate proves structural payload integrity only; pose stress, dynamic bounds, readable screenshots, and visual likeness remain separate gates. Payload ownership and non-goals: grimoire/readiness/procedural_rigging_contract.md.
• Assembly gate (structure, not pixels) — every model ships explodable AND clickable: this is a build requirement, not a per-project extra. Name every mesh; flag surface relief userData.explodeWithParent so it rides its shell; let a named group of anonymous meshes be one part while a named group of named parts stays a container. Explode and part-picking must share one definition of "a part" — if they disagree, both are wrong. Separate parts by SCALING the layout about the model centre, never by pushing every part the same distance (that translates the arrangement without opening any gap). Then run forge/stage4_review/check_part_coverage.py --spec <spec> --manifest <parts.json>: it FAILS on a specified component that was never built and on two components fused onto one mesh; it warns on inventoried details that never reached the spec and on meshes belonging to no named part. This is the only gate that scores STRUCTURE — every other one scores pixels, and a single fused mesh wearing a projected photo passes all of those. Its limit is honest and must be stated when reporting: it proves you built what you specified, never that you specified enough. Full contract + the two rules it took a wrong pass to learn: grimoire/build/geometry_patterns.md.
• Attachment: child appendages (branches/limbs/handles/tubes) need attachment.parentSocket, localStart, localEnd, contactType, embedDepth/overlap, gapTolerance — no mid-air parts. grimoire/readiness/joint_attachment.md.
• Material/lighting: grimoire/feedback/shading_realism.md — independent PBR channels (never alias albedo into roughness/normal/AO), macro/meso/micro frequency bands, real lights.
• Detail inventory: for moderate+ subjects strict-quality blocks code gen until the detailInventory reaches targetMinDetails and every detail maps to a real component/material entry (gloss needs low-roughness/clearcoat; fasteners need instancing/micro parts).
• Character track: when primaryDomain is character/hybrid (or --character), the spec author auto-builds a stylized humanoid template (head/neck/torso/arms + hair, glasses, headphones, face features), flattened to world space under a hidden root, with per-part character materials and character build passes (proportion-lock, feature-placement). strict-quality requires a filled anatomy block (head-units, proportions, face landmarks) and character feature targets. Suitability routing for humans: grimoire/intake/validation_rubric.md (stylized vs maximum-likeness). Stylized bust, not a face-copy; refine positions per reference. The initial CS2 family boundary is knife only. Pistol, rifle, SMG, sniper, heavy, glove, and unknown knife subtypes must stop with unsupported-family or unsupported-subtype; they must not receive the knife component tree as a generic fallback.

For every CS2 reconstruction, MUST read the full layer contract, intake order, and surface/review rule in grimoire/intake/cs2_intake_contract.md before intake state can advance.

Gates (do not skip)

Before any visual review or continue decision, MUST read the full gate-by-gate contract in grimoire/review/gates_reference.md (Divine Eye, VLM rescue, multi-angle, CS2 review, bounded correction, screenshot feedback, assembly, attachment, material, detail inventory, character track). In short:

• Validate references first (grimoire/intake/validation_rubric.md, check_reference_admission.py).
• divine_eye.py is deterministic-first; the VLM (vlm_gate.py) is a gated last layer, never consulted on a hard-gate failure.
• A non-planar form must hold from ≥2 angles (diagnose_render_multi_angle.py).
• CS2 knife builds also run cs2_review.py against the versioned scene fixture.
• Local state enforces 3 corrections per pass and 6 total by default; reaching either limit is a hard stop. correction_loop.py may stop earlier on repeated defects, oscillation, or plateau.
• continue requires a render + comparison sheet + AI-vision score ≥ threshold, every critical feature ≥ its own threshold (grimoire/feedback/render_capture.md).
• Every model ships explodable AND clickable — a structure gate, not pixels (check_part_coverage.py, grimoire/build/geometry_patterns.md).
• Action-ready, attachment, material/lighting, detail inventory, and character-track requirements: grimoire/readiness/action_rigging.md, grimoire/readiness/joint_attachment.md, grimoire/feedback/shading_realism.md, grimoire/intake/quality_contract.md, grimoire/intake/validation_rubric.md.

Self-Correction

After every pass, decide exactly one: continue | refine-spec | refine-code | request-input | stop. refine-spec fixes a wrong/missing/shallow spec (re-validate, don't patch code around it); refine-code fixes geometry/material/lighting that doesn't match a sound spec. Before making the decision, MUST read the root-cause guide + fidelity scale in grimoire/review/self_correction.md, record the decision, and re-run the local state gate.

Small features need a different instrument. Divine Eye's SSIM/tonal/edge signals run on a 64×64 luma grid and a 96×96 edge grid, so a detail a few pixels wide in the reference is not scored badly — it is absent before any comparison happens, and per_feature.py cannot compensate because it consumes a scores dict and never opens an image. When fidelity depends on individual tears, spars, fangs or eyes, use the four-tier microscope: grimoire/review/divine_eye_microscope.md. Two rules there are empirically established rather than proposed, and both produced confident false findings before being understood: measure fidelity on a component's visible footprint (the full frame minus a component-hidden frame), never on an isolation render, which reveals geometry the reference cannot see; and never colour-gate a concave feature, where a dark ratio captures cavity shading rather than material.

Implementation Rules (brief)

TypeScript + plain Three.js unless the project uses a wrapper. Group factory createObjectNameModel(spec, options), reconstruction data kept separate from renderer objects, deterministic seeds for all procedural noise. Prefer primitives / Shape extrude / curve+tube / instancing / displacement / generated canvas textures before any external art. Full geometry & material recipes + hard-won failure patterns: grimoire/build/geometry_patterns.md.

Optional Python ↔ Three.js render bridge

When Python is requested for character rendering, use it as a deterministic job/evidence layer around the browser Three.js runtime: camera-batch manifests, source/output hashes, readiness and settle checks, screenshot persistence, masks, diagnostics, and comparison packaging. The target Three.js browser route remains the rendering authority. Do not silently replace the procedural TypeScript factory with Blender/VRM/GLB output. Full routing, manifest fields, and failure rules: grimoire/build/python_threejs_render_bridge.md.

Standard character pipeline (merged 1.5 beta + alpha)

Use grimoire/readiness/standard_character_pipeline.md for character work. Beta owns the strict sculpt/build/review gates; alpha owns deterministic camera manifests, browser screenshot evidence and UniRig-shaped rig validation. CharacterGen, Tripo, VRM and other neural/asset systems are opt-in adapters with source, checkpoint, license, coordinate conversion and output hashes. They never silently replace the procedural TypeScript factory. Image-to-mesh systems emit a static mesh with no skeleton, so their output is never animation-ready however good it looks; neural riggers are offline inference whose value here is the serialized payload, not a browser dependency.

Executable entry points are forge/stage4_review/render_bridge.py and scripts/capture_threejs_playwright.py. Run init → browser capture → validate → diagnose; capture must operate on the real showcase/browser route and leave readable PNGs in the workspace.

Output

• Analysis-only: suitability verdict + scores, object extraction, macro→micro hierarchy, geometry strategy, material/lighting recipe, animation/destruction feasibility, plan + risks.
• Implementation: the above briefly, then edit code; verify with typecheck/build + a screenshot.
• Not feasible: name the blocker, ask for more views / cleaner image / accepted stylization / a narrower target. "This cannot reach the requested fidelity from this image" is a valid result.