nvidia-omniverse/usd-content-agents/content-articulation-review

Articulation Candidate Review

Bind one complete ordered candidate decision to the exact request, source, dependency bundle, candidate document, usd-cli evidence, and checkpoint.

When to Use

Use after inspection/proposal and before any authoring or human review receipt.

Limitations

• Every candidate ID must appear exactly once and in canonical order.
• Evidence must be observation-bound and remain inside the run directory.
• An agent patch never bypasses a configured human review gate.
• A human revise decision never permits in-place graph edits or authoring.

Prerequisites

Load content-articulation-inspection and content-articulation-proposal. Read the current observation and schema.

Instructions

1. Copy every digest and checkpoint revision from the current observation.
2. Decide accept or reject for each exact candidate.
3. Complete-candidate edits may change only confidence, parent_hint,

child_hint, component_name, component_type, or the explanatory evidence text. Candidate identity, topology, axis, limits, readiness, and every Joint-owned declared or pass-through provenance/evidence field are immutable. Include concrete bound evidence paths.

1. Write the patch to the observation's decision_patch_path.
2. Apply it through the focused command selected by the active workflow and

respect needs_review when the requested policy requires human approval. In an Asset articulation.author.v1 leaf, the first opaque entrypoint call returns content-agent-workflows.articulation-author-leaf-progress.v1. After writing the patch to its prescribed path, call that same content-workflow-cli articulation agentic-leaf author --invocation ... entrypoint again with the exact immutable invocation bytes. Do not call _agent-apply, import an internal Python API, or create or execute a Python helper in this Asset-leaf pause.

Command Reference

For a standalone content-workflow-cli articulation run, use content-workflow-cli articulation _agent-apply --run-dir <run> --decision-patch <patch>.

For an Asset articulation.author.v1 leaf, re-run the descriptor's same opaque content-workflow-cli articulation agentic-leaf author --invocation <invocation> entrypoint. It validates and durably binds the prescribed outer patch inside the active leaf attempt.

Common Workflows

With review policy all or uncertain, stop after the agent ledger is bound; the asset owner supplies the separate human receipt.

For an embedded canonical graph, the human receipt may use revise for exact candidate IDs. Preserve the parent graph and review bytes, author a content-agent-workflows.embedded-articulation-graph-revision-patch.v1 that binds the parent state revision, identity/evidence/proposal digests, parent graph file and semantic digests, human decision and complete decisions file, reviewer, reason, timestamp, exact supported field changes, and revised graph. Supported changes are joint type, axis, lower/upper limit, limit unit, and frame policy; topology, membership, coverage, candidate identity, and evidence remain immutable. Apply it with content-workflow-cli articulation revise-graph, then stop for a complete new human decision over every ID at the revised graph digest. Never reuse the parent acceptance.

Output Format

Use content-agent-workflows.articulation-decision-patch.v1. The deterministic step publishes the patch, reviewed candidate document, and immutable ledger.

Troubleshooting

Stale revision, digest drift, missing IDs, unsafe edits, and unbound evidence fail before authoring. Rebuild from the latest observation; never edit state.