Skip to content

Toolchain

afb and the repository scripts form the factory's operating surface.

Which commands should operators use?

afb is the primary interface for planning, running, validating and packaging Asset Factory work. Repository scripts supply bounded backend adapters, diagram generation and specialised verification helpers. Operators should begin with afb info and afb capabilities, then use the documented command whose service owns the intended state change.

Primary surfaces

  • afb is the operator and automation CLI. In addition to the core workflow, it exposes graph validation, semantic migration, exact Profile validation, task-fitness evidence, bound governance decisions, durable tool serving, single-use tool approvals, reference capsules and release evidence.
  • scripts/discover_capabilities.py probes available capabilities and plans on-demand installs.
  • scripts/generate_diagrams.py owns documentation diagrams.
  • scripts/reconstruction/ holds backend adapters, run, provision and install helpers, PartCrafter USD stage authoring and GLB preview rendering.
  • scripts/segmentation/ holds USD mesh segment conditioning.
  • scripts/texturing/ holds rollout PBR texture creation, PBR semantics calibration, ambientCG import, live texture generation and probing and visual set pass composition.
  • scripts/simready/ holds the bounded official Profile bridge, Isaac load and behavioural check and Isaac visual set render.
  • scripts/ci/ holds process-level service smoke checks that also run from the extracted source appliance and deployment image.
  • scripts/afb and scripts/afb-agent-launchable are the CLI and agent tool-server launchers.

Release workflow

  1. Verify the requirements and install the locked source appliance with uv sync --frozen --all-extras.
  2. Generate or validate the run plan.
  3. Run the agentic workflow in dry-run or live mode.
  4. Validate manifests, skills, diagrams and the complete record graph.
  5. Build the source archive, inspect its appliance contents, extract it into a clean directory and rerun the locked workflow there before creating reusable reference-run evidence.

Pytest suites and the reproducibility benchmark live in this repository under tests/ and benchmarks/; .github/workflows/ci.yml runs them on every push.

Commands

uv sync --frozen --all-extras
uv run afb --help
uv run python scripts/generate_diagrams.py --check
uv run afb skill-audit --root . --output artifacts/skill-audit.json

Recorded execution

Commands leave reports and checksums so reviewers can reconstruct execution after the process exits.