nvidia-omniverse/usd-content-agents/content-workflow-asset

Agentic Asset Composition

Use this skill from a repository-root call to content_workflow_cli.asset_runner.run_interactive_asset_workflow or resume_interactive_asset_workflow, or from content-workflow-cli asset run and asset resume. The public launchers expose one AssetCoordinatorSession to exactly one outer reasoner for one prompt while owning the sole parent usd-cli lifecycle. There is no selector, domain planner, nested coordinator, provider-owned controller, or built-in total domain/stage order.

Public Agentic Contract

• New public runs are selected_mode=agentic and request schema

content-agents.asset-composition-request.v3. They bind the staged source, prompt/source/config/reference digests, the deterministic repository-owned content-agents.asset-leaf-catalog.v3, and one durable asset-sole-coordinator-identity.v1. Never author or substitute run-local descriptors; --leaf-catalog accepts only an exact repository-catalog readback.

• The sole outer reasoner interprets the prompt and supplies one immutable

content-agents.asset-execution-graph.v2. Use stable catalog leaf IDs only. List the prompt-specific selected set, every explicit dependency, required/optional requirement, terminal-output leaves, and every omitted ID. Nodes, dependency IDs, and omissions use canonical lexical storage only; that storage order is never execution policy.

• Bind each selected node to its catalog descriptor digest. Static catalog

requirements must appear in depends_on; incompatible selections are invalid. A descriptor's required_dependents are conditional producer constraints: if both leaves are selected, each named dependent node must include the descriptor leaf in depends_on. Omitting the dependent imposes no global order. Unknown, missing, duplicate, cyclic, incompatible, mixed-mode, or stale graph identities fail closed.

• State validates and follows the frozen graph. It never chooses, inserts,

removes, or reorders a leaf. No agentic transition can enter the historical STAGE_ORDER, fixed pipeline, classic mode, or a fallback path.

• Omitted catalog leaves are durably not_requested. A selected required leaf

must pass. A selected optional leaf may be not_evaluated only with native result/evidence and only when no selected dependent requires it.

• If articulation.proposal-provider.v1 is selected, its predecessor

preparation must use proposal_status=not_evaluated with no proposal. Use proposal_status=not_requested only when that proposal leaf is omitted.

• Every leaf receipt binds the graph/catalog/coordinator identities, dependency

receipts, invocation, native result and terminal receipt, operation/evidence indexes, evidence, saved-stage readbacks, native disposition, timing, resource claims/releases, and supersession. A failed or cancelled opaque result stays failed or cancelled with those exact artifacts; composition must never upgrade it to passed or not_evaluated. The graph terminal receipt binds all selected receipts, all omissions, terminal artifacts, total timing, and parent/leaf resource release.

Discover the capability catalog with content-workflow-cli asset catalog and use the identical frozen catalog in request.json. Each descriptor resolves to repository-owned invocation/result schemas and a declared deterministic projector; catalog discovery advertises capabilities but never selects or orders them. Create the prompt-specific AssetExecutionNode list and AssetExecutionGraph.create(...) inside the sole reasoning invocation. The factory canonicalizes storage lexically; only explicit dependency edges govern readiness.

Agentic Execution

1. Read and verify the frozen request.json and asset_run.json. Never edit

either. Confirm the request coordinator identity matches this invocation and next_action is freeze_graph for a fresh run. A public launcher session also supplies the immutable typed ParentUsdCliSessionIdentity and its exact artifact binding. Reuse that parent session; never start, replace, or stop a daemon from the reasoning loop.

1. Inspect the prompt, references, staged-source identity, and frozen public

catalog. Read back content-workflow-cli asset catalog and require its exact descriptor and catalog digests to match the request. Author the exact graph JSON below the run root. Include every catalog ID in exactly one of selected_leaf_ids or omitted_leaf_ids; freeze it once with freeze-graph. Absence of a graph is not permission to use a fixed order. When the request carries required_leaf_ids, select every one as required; mark exactly required_terminal_leaf_ids as terminal outputs; and include every edge in required_leaf_dependencies. Missing catalog capabilities or any partial graph contract must fail before provider, renderer, authoring, or runtime work begins.

1. When next_action=begin_leaf, select only a dependency-ready leaf already

frozen in the graph, call begin-leaf, then leaf-dir; put every invocation/result/evidence/readback/release file below that attempt directory. On managed Windows, invoke the descriptor through content-workflow-asset-state invoke-leaf with the exact run state, active leaf ID, absolute invocation path, and a new absolute result JSON path in the attempt directory. Do not invoke a domain executable or usd-cli-tel directly. Use the launcher-owned readiness artifact, command journal/checkpoint, staged source, and exact session identity for retained pre-native preparation facts; the trusted domain leaf owns authenticated scene commands. Other hosts invoke the descriptor's opaque public entrypoint directly. The entrypoint retains its own native validation and cannot have a conditional or failed disposition upgraded by composition state. For every invocation artifact inherited from a predecessor leaf, copy its absolute path, sha256, and size_bytes verbatim from that predecessor's verified leaf_receipt.json; never reconstruct, abbreviate, or relocate the path. Before calling the opaque entrypoint, require that each copied path is an existing regular file below the declared run directory and rehashes to the copied identity. Correct an invocation-construction error before the native call; after the native call, preserve any failed result exactly. When articulation.preparation-publisher.v1 feeds the standalone articulation.author.v1 leaf, require its retained inspector configuration to set capabilities.canonical_output_evidence_required=true before publication. That capability declares the mandatory post-author evidence policy; keep the author invocation's canonical_visual_envelope null until validation.canonical-ovrtx-evidence.v1 runs after authoring. Load the exact bound preparation publication before constructing the author invocation, and copy its complete source binding into the invocation. Never reuse the request's staged-source binding: the retained publication source is authoritative even when both paths contain byte-identical files. Author its preparation readback as content-agent-workflows.articulation-preparation-readback-draft.v2 and omit both source_dependency_bundle_sha256 and saved_stage_dependency_bundle_sha256. Never calculate, guess, or recover those package-aware identities from a publisher error. The trusted preparation publisher derives and seals them; v1 readbacks remain only for trusted producers that already own both exact identities. A typed failed preparation result is terminal for that attempt and must be passed unchanged to fail-leaf, not repaired by retrying with a guessed digest. articulation.author.v1 has one explicit two-call pause inside the same active leaf attempt. Its immutable invocation prescribes articulation_decision_patch.json. The first entrypoint call returns typed awaiting_decision progress with the exact identity and observation. Write only the complete decision patch at that prescribed path, then call the same content-workflow-cli articulation agentic-leaf author --invocation ... entrypoint with the same invocation bytes. Never substitute the standalone articulation _agent-apply command, import an internal Python API, or create or execute a Python helper at this pause. Never create, pre-create, or write the entrypoint-owned native/ directory. Only the second, terminal result, which releases its normalized, ledger-bound native patch, may be passed to complete-leaf. In that patch, set the top-level evidence_requirements to a non-empty list and set non-empty evidence_ids on every candidate_decisions entry and every entry in canonical_graph.groups, canonical_graph.memberships, canonical_graph.joints, and canonical_graph.rigid_link_operations. Those lists may contain only exact evidence_id values from the current articulation_author_observation.json.evidence_records. Verify every required list before the first native call; never submit an empty list and try to repair it after a native failure, and never put future post-author validation goals in these fields. When canonical OVRTX evidence is selected with Articulation authoring, the canonical evidence node must depend on articulation.author.v1; do not render the pre-author stage. This is the sole outer coordinator's decision pause; it does not launch a Joint child or another coordinator. material.assignment.v1 has two typed pauses inside one active attempt. Keep the invocation bytes unchanged. The first call seals preparation and returns awaiting_decision; write only the prescribed native/raw/material_decision_patch.json, then call the same entrypoint. The second call seals the applied result and exact post-apply OVRTX evidence and returns awaiting_review; visually review every prescribed binding, write only native/raw/material_post_apply_review.json, and call the same entrypoint a third time. Only a terminal passed result with exact output bytes and a resource-release receipt may reach complete-leaf; a conditional Material result is a failed required leaf. Keep Articulation-review and final-package visual identities separate. validation.canonical-ovrtx-evidence.v1 stays bound to the authored Articulation bytes. Carry the accepted publication through Material, Texture, Physics inspection/apply, SimReady conformance, and portable packaging. Run asset.final-ovrtx-evidence.v1 and SimReady validation on the exact portable package. asset.combined-result.v1 must receive every required same-run predecessor receipt and can pass only when its projector proves the complete cross-domain byte chain.

1. Call complete-leaf with only the exact typed invocation and result. The

repository binding executes its declared projector API v2 with an immutable content-agents.asset-leaf-projection-context.v1 that binds the selected invocation/result path, digest, and size. The domain projector must compare that context with its native receipt chain before returning every exact digest/size-qualified terminal, index, evidence, readback, and release artifact. State exposes newly dependency-ready leaves without inventing a successor order. On native interruption or failure, use cancel-leaf or fail-leaf with the exact typed invocation and result; their declared projectors retain native terminal, evidence, readback, and release artifacts. Recovery uses recover-leaf, retains and supersedes the prior immutable receipt, and retries the same leaf without changing the graph. A required failed or cancelled leaf cannot aggregate or finalize successfully. Selected-leaf path fields must be absolute so projector replay is independe