houdini-automation

houdini-automation

Higher-level repeatable automation for Houdini sessions. Prefer run_python_file for reviewed scripts on disk, and build_node_chain for compact graph recipes.

run_python_file bounds inline stdout, stderr, and result payloads. Its default output_mode=full keeps the legacy string result while applying the configured character caps. Use output_mode=structured to preserve JSON-safe result types, or output_mode=summary to return counts and artifact metadata without inline bodies. When spill_overflow_to_artifact=true, truncated channels are written as complete UTF-8 temporary artifacts with byte counts and SHA-256; summary mode persists every non-empty channel. Treat those files as potentially sensitive script output and delete them after collection.

build_node_chain is a structured atomic mutation surface, not an arbitrary code executor. It validates the parent, every node type/name/reference, and all connection references/port indices before opening an undo group. Use dry_run=true to inspect validated and predicted affected_paths with zero scene mutation.

On execution, the complete recipe uses one named Houdini undo group. Results include transaction_id, undo_label, validation evidence, and post-cook readback. If creation, parameter assignment, connection, layout, cook, or readback fails, the tool explicitly removes created nodes and restores any existing input connections and existing node positions touched by layout. Check rollback.complete and rollback.errors before retrying a failed recipe.

Successful responses expose a compact summary with the created nodes, readback-verified connections, submitted parameter values, and counts. This is the preferred proof for generated MaterialX networks.

MaterialX displacement acceptance

For each existing MaterialX builder, use one build_node_chain recipe with an mtlximage feeding mtlxrange, then mtlxdisplacement. Put texture paths, range bounds, clamping, and displacement scale in each node's parameters and connect nodes by their recipe-local ref or node_name. Four materials need four such calls because they have four different parent networks; each call is independently prevalidated and leaves no partial network when rollback is complete.