FusionSparse

FusionSparse is a generator-driven Python layer for the Autodesk Fusion API.

The project keeps the full raw adsk.* surface reachable while generating a smaller, cleaner, more token-sparse interface for common Design workspace workflows.

Example

This is a real Autodesk sample workflow from extrudeFeatures_addSimple_Sample.htm, compared with the FusionSparse remake that produces the same result.

Official Autodesk style:

import adsk.core
import adsk.fusion

def run(context):
    app = adsk.core.Application.get()
    app.documents.add(adsk.core.DocumentTypes.FusionDesignDocumentType)
    design = adsk.fusion.Design.cast(app.activeProduct)

    root_comp = design.rootComponent
    sketch = root_comp.sketches.add(root_comp.xYConstructionPlane)
    lines = sketch.sketchCurves.sketchLines
    lines.addTwoPointRectangle(adsk.core.Point3D.create(0, 0, 0), adsk.core.Point3D.create(3, 2, 0))

    profile = sketch.profiles.item(0)
    distance = adsk.core.ValueInput.createByString("100 mm")
    operation = adsk.fusion.FeatureOperations.NewBodyFeatureOperation
    extrude_features = root_comp.features.extrudeFeatures
    extrude_features.addSimple(profile, distance, operation)

FusionSparse:

import fusion_sparse as fx

def run(context):
    design = fx.new_design()
    root = design.root
    sketch = root.sketch("xy")
    sketch.rect((0, 0), (3, 2))
    profile = sketch.profile()
    root.extrude(profile, "100 mm")

Measured on this exact validated pair:

• lines: 23 -> 13 (43.5% smaller)
• characters: 871 -> 342 (60.7% smaller)
• estimated tokens: 176 -> 82 (53.4% smaller)

Benefits:

• less boilerplate around Application, document creation, casting, and feature collections
• much fewer API ceremony tokens for agents to read and write
• the code describes the modeling intent directly: sketch rectangle, get profile, extrude
• raw escape hatch is still available through .raw when the compact layer is not enough

Goals

• Keep raw Fusion access available at all times.
• Generate as much surface area as possible from Autodesk's official API corpus.
• Make common modeling code much shorter and easier for agents to read and write.
• Keep updates mechanical: refresh corpus, rebuild IR, regenerate artifacts, review diffs.

Current scope

FusionSparse is focused on the Design workspace, not the whole Fusion app.

Current compact coverage includes:

• context/bootstrap
• sketches
• construction planes / axes / points
• extrude / revolve / sweep / loft / patch
• shell / draft / hole / fillet / chamfer
• combine / mirror / circular pattern / rectangular pattern
• move / offset / replace face / scale / split body / thread / trim

Everything else remains available through raw adsk.* access and .raw.

Status

• generator pipeline is working end to end
• Autodesk corpus is pinned as a submodule
• IR build, diff, and regeneration are in place
• Fusion deployment exists for both scripts and add-ins
• official-vs-compact validation runs in real Fusion through MCP

Current proof points:

• 53/53 official-vs-compact sample pairs matched in real Fusion
• paired samples are 61.9% smaller by characters and 52.7% smaller by estimated tokens
• Design-workspace validated families: 41

Repo shape

• src/fusion_sparse: shipped library
• tools: parsers, generators, reporting, validation
• rules: handwritten policy
• corpus/FusionAPIReference: pinned Autodesk source corpus
• fusion: Fusion script/add-in deployment assets
• tests: unit and integration coverage

Useful commands

python -m tools.cli build-ir
python -m tools.cli generate --skip-build-ir
python -m tools.cli run-sample-pairs
python -m tools.cli map-coverage
python -m unittest discover -s tests/unit -p 'test_*.py'

Generated docs and reports

• compact_reference.md
• raw_mapping.md
• update_workflow.md
• build/reports/api_coverage_map.md
• build/reports/design_workspace_backlog.md
• build/reports/sample_pairs/sample_pairs_report.md