Glina

Sculpt job

A sculpt job (pipeline/llm_blender.js, sculptWithLlm) is the LLM-driven build loop: prompt → the model writes bpy code → execute through the Blender session → iterate → export GLB → run the verification gate. The model is the brain; every execution still goes through the Blender MCP layer, and the model key comes from Skarbiec like every other secret.

Entry points: glina sculpt <prompt>, MCP tool glina_sculpt, HTTP POST /v1/sculpt.

Definition and shape

Job inputs: prompt, outDir (default assets/models), filename (default: slugified prompt + .glb), maxRounds (default llm.maxRounds, 12). Result: { outPath, verification, rounds } — the CLI strips the transcript from what it prints.

Each model reply must be one JSON object:

{ "thought": "one short sentence", "code": "python bpy code", "done": false }

The reply parser tolerates fences and surrounding prose (first balanced JSON object wins) but a round with no usable JSON fails the job: model reply not usable in round N: <reason>.

The system prompt's execution contract

The model is told, verbatim in spirit (see SYSTEM_PROMPT):

  • each block executes in a fresh namespace containing only bpy — nothing persists between calls except Blender data itself,
  • never use bpy.context for object access; work through bpy.data,
  • never call bpy.ops.wm.read_factory_settings — it kills the bridge,
  • flat shading, material colors only, ~6000-triangle budget (Thronefall-style chunky low-poly),
  • when animations are required: one armature, automatic weights, at least two named actions (idle plus one characteristic motion) keyframed over 24–64 frames,
  • reply done: true with empty code when finished.

Animations become required in the job when verify.requireAnimations is true or verify.minAnimationClips > 0 — the requirement is injected into the first user message, since the gate will refuse static meshes anyway.

Round loop

  1. Ask the model (llm.maxTokens, default 8192).
  2. Execute step.code if non-empty. Execution errors are fed back to the model as ERROR: <message> — the model gets a chance to repair its own bug.
  3. Bridge-down rule: three consecutive transport failures end the run — Blender bridge is down after 3 consecutive failures: <reason> — because the model cannot heal a dead bridge; probing through the LLM only burns rounds.
  4. If the Blender MCP server exposes get_viewport_screenshot, a screenshot is attached to the next round (best-effort; the loop works text-only).
  5. done: true breaks the loop. Running out of rounds fails the job: model did not finish within N rounds.

Export and gate

The scene is exported with exportGlb and then, unless verify.enabled: false, gated with throwOnFail — a failing asset raises asset failed verification: <reasons> and the job fails. The Blender session is closed in a finally either way.

Proven runs

Generated end-to-end and gate-verified on live runs (per the README): kamien.glb (granite boulder, 12 bpy rounds, Brama → Blender MCP, 2026-07-27) and krasnolud-wojownik.glb (dwarven warrior, local Brama, 2026-07-28). Both files live in assets/models/.