Skip to content

Blender Agent Kit v0.2.3

Latest

Choose a tag to compare

@github-actions github-actions released this 20 Nov 03:38
a328883

Blender Agent Kit v0.2.3

v0.2.3 is an agent-efficiency and architecture-hardening release driven by two sources of evidence: the independent SupplyCrate/Helios Conduit production trial and a later blind two-subagent quadcopter workflow comparison.

The release focuses on making production rules explicit, reducing agent discovery overhead, improving review/delivery automation, and replacing documentation-only batch claims with executable caching/telemetry.

Agent-first discovery

Two host-only commands now expose executable truth without locating or launching Blender:

python skills/core/blender-core/scripts/run.py capabilities --json
python skills/core/blender-core/scripts/run.py contract --target unity --asset-class solid_prop
  • capabilities reports executable commands, host-vs-Blender execution mode, installed targets/policies/project profiles, evidence surfaces, and guidance/configuration skills.
  • contract resolves engine target -> default production policy -> optional project profile and returns naming/classification, QA, LOD/material/pivot/UV/rigging rules, texture packing, export/round-trip behavior, and review styles/presets in one JSON contract.
  • Public JSON schemas cover both surfaces.
  • Invalid host arguments/errors remain machine-readable.
  • The 10 hot-path SKILL.md files were reduced from roughly 1,800 words to 1,229 words (~31.7% less) while retaining critical Blender/PBR safety guardrails.

Engine target vs production policy

Unity engine semantics are now separated from generic real-time asset policy.

The raw Unity target owns:

  • coordinate/export behavior;
  • hierarchy-safe static Y-up FBX handling;
  • standard URP/HDRP texture layouts;
  • engine-relevant round-trip settings.

The reusable realtime-default policy owns toolkit production defaults such as:

  • triangle/material budgets;
  • pivot/UV/texel-density expectations;
  • LOD/collider naming;
  • directional-empty naming;
  • LOD transition defaults.

This keeps statements such as a 15k triangle budget correctly attributable to Blender Agent Kit policy for the Unity target, not to Unity as a universal engine requirement.

LOD, collider, and directional-empty naming are configuration-driven. Blender tests prove alternate naming can be supplied without changing core code.

Configurable review styles

The default production policy now supplies reusable review styles:

  • technical — headless-safe low-sample Cycles, brighter neutral presentation, deterministic high-contrast lighting, and tighter inspection framing.
  • showcase — higher-sample Cycles with presentation-oriented background, light rig, lens, and framing.

Render engine, samples, background, camera factors, and light rig are policy/configuration data rather than hidden constants. Project review presets can select and overlay styles.

Eevee remains configurable when the runtime has a suitable graphics/EGL stack, but is intentionally not the headless CI default.

QA-gated delivery

New host command:

python skills/core/blender-core/scripts/run.py deliver \
  --blend asset.blend \
  --target unity --asset-class solid_prop \
  --review-style technical --out delivery/

Default flow:

inspect -> QA gate -> optional pack -> review -> export/bundle -> fresh-import verification
  • QA failures stop before expensive review/export by default.
  • An explicit diagnostic override can continue after QA failure.
  • Results record per-stage status and measured duration, QA-gate state, retry guidance, and generated evidence.
  • Delivery reports validate against a public JSON schema.

Executable batch processing and safe caching

blender-batch is no longer guidance-only. The public batch command supports:

  • qa
  • pack
  • export
  • bundle
  • deliver

Blender-backed cache misses remain process-isolated. Cache identity includes:

  • source bytes;
  • operation-relevant resolved configuration;
  • executable toolkit identity;
  • relevant operation options;
  • Blender executable/version identity.

Export/bundle cache hits additionally verify manifest build hashes and required generated files before a skip is accepted.

Reports include measured per-asset duration, cache hit/miss counts, processed/skipped/failed counts, and tool/Blender identity. The runtime explicitly does not fabricate model/API-token or agent-turn telemetry.

Tier-4 CI proves a real fresh SupplyCrate batch miss followed by an unchanged zero-processing cache hit.

Canonical reproducible demos

The independently authored source-bound demonstrations under demo/ are now the canonical v0.2.3+ showcase lineage.

python demo/SupplyCrate/run_demo.py --mode full --clean
python demo/SciFiKit/run_demo.py --mode full --clean

Both also provide bounded --mode smoke entrypoints used directly by Tier-4 CI.

The older top-level demos/ tree is retained as historical v0.2.1 provenance/backwards links rather than rewritten.

The current Sci-Fi full showcase also verifies and displays the optional source-bound Unity _EmissionMap, which the original v0.2.1 evidence predated.

Independent evidence context

The earlier independent production trial reported retrospective estimates of:

  • 74–80% lower pipeline-engineering effort;
  • 55–63% lower combined end-to-end effort.

Those figures remain explicitly labeled estimates, not controlled benchmarks.

A later blind two-subagent quadcopter workflow comparison observed:

  • 90 -> 0 Blender Agent Kit solid_prop policy errors for the Unity target;
  • 376.12 s -> 56.49 s recorded Blender compute across the two delivered workflows;
  • 150 -> 0 evaluator-classified custom pipeline-boilerplate lines;
  • 8 -> 13 agent turns;
  • kit export fresh-import parity at 10,240 -> 10,240 triangles.

The compute/render workloads and final assets were not identical, so the 85% compute difference is an observed workflow outcome, not an isolated library-speed claim or guaranteed saving. Raw subagent orchestration transcripts were not included in the submitted evidence package, so complete prompt isolation remains experimenter-reported.

Verification

The release candidate is gated by:

  • Ubuntu host/schema tests;
  • Windows host/schema tests;
  • Blender 5.2 Tier-2/3;
  • Tier-4 canonical demo smoke using the documented run_demo.py entrypoints;
  • fresh Blender FBX round-trip verification;
  • tag-specific CI before GitHub Release publication.

Known limits

  • No Unity Editor import/prefab/LODGroup verification is claimed.
  • Geometric UV1 island-overlap proof remains unimplemented; requested UV1 emits uv1_overlap_unverified when only presence can be proven.
  • Collider convexity/decomposition quality is not geometrically certified by core QA.
  • The original large independent-trial binary commit remains in Git history to preserve provenance.