Forge Validate

Forge — Validate

The line between a finished Forge asset and a broken one. An asset is not done until it passes this gate. This skill is both a gate (run it before export) and the reference every other Forge skill cites for validation rules — when those skills say "check manifold" or "run the render QA pass", the full how lives here.

Project memory: if FORGE.md exists at the project root, read it first — it carries the coordinate system, scale unit, poly budget, target engine, and render engine for this project. Honour every constraint it specifies before running checks.

Suite position:

• Upstream craft skills: forge-model, forge-parametric, forge-procedural, forge-topology, forge-uv, forge-material, forge-texture, forge-light, forge-render
• Downstream pipeline: forge-export → forge-optimize → Skill(atelier-webgl)
• Called by ALL Forge agents: forge-director, forge-modeler, forge-lookdev, forge-rigtech, forge-pipeline
• Perf/a11y analogue (web): atelier-perf-a11y (handles CWV, WCAG; defer web-runtime gate there) — this is the analogue for Tiers 0–2 (the deterministic self-checklist)
• Adversarial analogue (web): atelier-review — Forge-validate Tier 3 mirrors the atelier-perf-a11y → atelier-review escalation; this is the 3D-asset counterpart of that fan-out, run on substantial/print/web-delivery assets (analogy, not a handoff — no Skill() call)
• Standards reference: forge-standards (units, budgets, naming — read it for threshold values)

Run = call the Skill tool with the exact name. Writing "run forge-validate" in prose runs nothing.

How to use this skill

Two modes:

1. Inline gate — other skills call Skill("forge-validate") after a construction step; this skill runs the appropriate ladder tier and reports pass/fail.
2. Full audit — invoked directly on a finished asset; runs all four tiers in order, escalates to adversarial reviewers on substantial/print/web-delivery assets.

The quality ladder

Four tiers, escalating rigor. Each tier builds on the last — do not skip.

Tier 0 — Deterministic pre-pass (always, seconds) Run scripts/detect.py over the target directory or file. Advisory regex + structural checks: missing poster, oversized GLB, EEVEE headless flag, // Blender paths, missing -- separator. Exit 0 always. Fix every [BLOCK]; weigh [WARN] against FORGE.md intent.

Tier 1 — Mesh & spec self-checklist (every build) Run scripts/validate.py with the target file. Checks: GLB magic + size, raw manifold / watertight / winding (measured on the UNTOUCHED mesh — the gate reports the asset's real state, it does not silently repair before grading), volume > 0, broken faces, face count vs budget, UV presence + range + texel-density CV, scale/units. If trimesh is installed, full mesh checks; otherwise degrades to GLB structural checks only. Two scoped-out items that need Blender, not trimesh: (a) bmesh-level topology fields (non-manifold edge count, flipped faces, degenerate count, high-valence poles, n-gon %) come from the §6 audit run as blender -b <file> -P topology_audit.py, not from validate.py; (b) full UV-island overlap (intersection) is not computed — the range check is a tiling/UDIM signal only. Full checklist: references/mesh-checklist.md.

Tier 2 — glTF-Validator + render-QA contact sheet (pre-export) Run Khronos glTF-Validator (CLI or npm) — zero errors policy, warnings reviewed. Fire Blender headless (BLENDER_WORKBENCH only on Windows — never EEVEE headless) to produce turntable + 6-view + diagnostic overlays, assemble contact sheet, call Read on the PNG. Visual inspection guide: references/render-qa-guide.md.

Tier 3 — Adversarial escalation (substantial / print / web-delivery assets) Spawn independent sub-agent reviewers, one per dimension, in parallel using the Task tool. Each reviewer gets the validate.py JSON + the contact-sheet path; the topology reviewer additionally gets the §6 topology_audit.py JSON (non-manifold edges, flipped faces, poles) and the printability reviewer gets the --print block — those fields do NOT come from the base validate.py run. Dimensions: topology integrity, material/PBR spec compliance, printability, render fidelity. Escalation protocol: references/adversarial-escalation.md.

The flow

0. Read project memory If FORGE.md exists, read it. Extract: coordinate system, scale unit, poly budget, target engine, render engine. These override any defaults below.

1. Run the deterministic pre-pass

python "$CLAUDE_CONFIG_DIR/skills/forge-validate/scripts/detect.py" \
  --target <path> [--json]

Fix every [BLOCK]. Note [WARN] items. Exit if [BLOCK] found and asset is not in repair mode.

2. Run the mesh + spec validator

python "$CLAUDE_CONFIG_DIR/skills/forge-validate/scripts/validate.py" \
  --input <file.glb|.stl|.obj> [--budget-faces N] [--units mm|m] [--json]

Parse the JSON. Any "status": "FAIL" item in the results is a hard gate — fix before proceeding. Repair guide: references/mesh-repair.md.

3. Run glTF-Validator (GLB/GLTF assets only)

# PowerShell — via npm-installed CLI
gltf_validator.exe --stdout "model.glb" | ConvertFrom-Json | Select-Object -ExpandProperty issues

Gate: numErrors == 0. Warnings reviewed against FORGE.md intent. Graceful degradation: if gltf_validator is not on PATH and neither npx @gltf-transform/cli validate nor a global gltf-validator is available, do NOT stall — emit WARN: "glTF spec validation skipped (validator not installed; see references/gltf-validator-guide.md for install)" and fall back to validate.py's stdlib GLB header/chunk + size checks as the floor. Full severity table and common error codes: references/gltf-validator-guide.md.

4. Render QA contact sheet Use BLENDER_WORKBENCH (not EEVEE) for all QA renders on Windows. Invoke via PowerShell using the absolute Blender path. Produce turntable (12 angles) + 6-view ortho + diagnostic variants (matcap, wireframe, normal, UV-checker). Pin the seed so two QA renders of an unchanged asset are byte-comparable (see render-qa-guide.md §1.5). Assemble contact sheet.

4a. Verify the PNG before reading it. A failed Cycles/Workbench render on Windows commonly writes a 0-byte or tiny black PNG while Blender still exits 0. Confirm the file exists and is >= 1024 bytes before trusting it:

python "$CLAUDE_CONFIG_DIR/skills/forge-validate/scripts/validate.py" \
  --verify-png <contact_sheet.png> [--json]

If missing/tiny, the render failed silently (EEVEE-on-Windows or missing write_still=True) → inspect .forge-build/blender.log, switch to CYCLES/Workbench, re-render. Never Read a 0-byte PNG and report PASS.

4b. Then call Read on the verified PNG path — visually inspect for:

• Black patches (inverted normals)
• Silhouette discontinuities (floating geo, scale error)
• Elongated checker squares (UV stretch > 2:1)
• Non-uniform wireframe density (poles, ngons) Visual inspection guide: references/render-qa-guide.md.

5. Printability gate (print-target assets only) If FORGE.md specifies Target: print or process: fdm|sla|sls|dmls, run the printability sub-pass of validate.py — it emits the printability block the Tier-3 reviewer reads (volume_mm3, thin_sample_frac, overhang_face_count, drain_hole_count, body_count):

python "$CLAUDE_CONFIG_DIR/skills/forge-validate/scripts/validate.py" \
  --input <file.stl> --print --process fdm|sla|sls|dmls --units mm [--json]

Wall thickness (inward ray-cast sampling) and overhang need trimesh; absence degrades to WARN, same as the structural trimesh pass. Thresholds, drain-hole rules and repair patterns: references/printability.md.

6. Adversarial escalation (substantial builds) If the asset is a hero asset, a print deliverable, or a web-delivery GLB, spawn parallel sub-agent reviewers via the Task tool. See references/adversarial-escalation.md for the exact delegation prompt template and required JSON output contract.

7. Report and gate decision Summarise all tier results in a structured report. Gate:

• BLOCK: any Tier 0 [BLOCK] or Tier 1 FAIL or glTF numErrors > 0 → fix before merge
• WARN: glTF warnings or Tier 3 advisory findings → review vs FORGE.md intent
• PASS: all tiers green

Operating principles

• Gates block, advisories inform. A BLOCK finding stops the pipeline. A WARN is evidence to weigh against the project brief, not a mandatory fix.
• Workbench only for headless QA renders. EEVEE Next is unsupported headless on Windows. Use BLENDER_WORKBENCH; switch to CYCLES (CPU) only for the UV-checker and normal-pass variants which require shader nodes.
• Absolute forward-slash paths in Blender Python. Never // relative paths in bpy filepath fields; never backslash literals in Python path strings — use pathlib.Path or forward slashes.
• Scripts, not inline code. validate.py and detect.py carry the real logic. Code blocks in this skill body are illustrations only — the model does NOT run $() expressions from SKILL.md.
• Run = call the Skill tool. When handing off to another Forge skill (e.g., forge-topology for repair), use Skill("forge-topology"). Narrating the handoff in prose does nothing.