ifbars/blender-agent-studio/blender-simulation-workflow

Blender Simulation Workflow

Read the shared execution guidance once per task for autonomous decisions, evidence cadence, and long-task continuity.

Simulation is a cache-producing build step. Preserve the setup and cache provenance; a visually plausible single frame is not proof that it is stable.

Define the simulation contract

Record the simulation type, solver, timeline, scale, gravity/forces, collision objects, emitters, domain bounds, desired behavior, cache location, resolution or substeps, and deliverable format. State the frames that establish onset, peak interaction, and settle. For fluids, explicitly distinguish liquid from gas/fire/smoke and whether a mesh, particles, or volume is the deliverable.

Use a project-local cache directory outside source control. Never overwrite an existing approved bake without creating a new output location or explicit user approval.

Build a controlled setup

1. Apply or account for object scale; use real-world scene scale and record

units. Mis-scaled colliders and domains invalidate tuning conclusions.

1. Name domain, flow/emitter, effector/collider, force, cache, and output

objects semantically. Keep a collection for inputs separate from generated mesh, volume, particles, or bake artifacts.

1. Start with a short, coarse preview cache to prove the causal setup,

collision direction, and timing. Do not begin a costly final bake before opening preview evidence.

1. For fluid/smoke/fire, bound the domain tightly enough to control cost while

leaving required motion margin. Use Modular or Final cache intentionally; document the ordered bake stages when applicable.

1. For rigid, cloth, soft-body, particle, or hair work, establish collider

thickness, substeps, damping/friction, self-collision, and pin/attachment constraints before aesthetic forces or turbulence.

1. Set seeds explicitly for stochastic effects. Preserve cache settings and

simulator version alongside the generation script.

Bake and validate

1. Bake the preview, then inspect the specified critical frames from at least

two useful cameras.

1. Check that emitters and colliders affect the result at the intended time;

that domains do not clip it; that collision does not tunnel, explode, or penetrate; and that mass, drag, settling, and scale read plausibly.

1. Change one causal parameter group at a time, invalidate the affected cache,

and rebake the smallest frame interval that proves the change.

1. Bake final data only after the preview passes. Verify every expected cache

stage/file is present and the timeline reads it rather than recomputing an implicit temporary state. Reopen the saved scene in a fresh Blender process and inspect the same frames with blender_inspect_motion. For liquid, use the evaluated domain mesh as a target; confirm nonempty changing surfaces and native data/mesh bake metadata. A fluid modifier, replay cache, or static proxy is not a completed bake. Final cache uses the Python enum ALL; UI labels are not necessarily enum identifiers. Cloth requires the baked point cache as well as visible deformation. These checks do not prove physical correctness.

1. Render the requested critical frames and, when timing matters, an encoded

preview or final sequence using $blender-agent-studio:blender-rendering-workflow.

Completion gate

Deliver the reproducible setup script or .blend, cache manifest (type, location, frames, resolution/substeps, seed, Blender version), final cache or exported simulation artifact as requested, and visually inspected evidence. Report the scope honestly: a short cache preview validates those frames only; it does not validate the full final timeline. Use asset validation after converting a simulation to a mesh or export format.