houdini-vex

houdini-vex

Typed VEX authoring and diagnosis for agents. All tools are affinity: main because they call hou. Prefer these over houdini-scripting.execute_python for VEX creation.

Hard constraint: VEX snippets are set via hou.Parm.set() on the "snippet" parameter of Wrangle nodes. This skill NEVER constructs or evaluates Python code from VEX strings. Every VEX snippet is validated client-side (allowlist and deny-list) before it touches Houdini.

Tool groups

• vex-create: create a Wrangle node with optional initial VEX snippet (create_wrangle).
• vex-edit: update the VEX snippet on an existing Wrangle (update_vex_snippet).
• vex-validate: pre-cook validation of VEX syntax, bindings, and parameters (validate_vex_syntax).
• vex-cook: cook a Wrangle and collect diagnostics (cook_wrangle, diagnose_wrangle).
• vex-query (read-only): inspect an existing Wrangle or list all wrangles under a path (get_vex_info, list_wrangles).

Tracer-bullet flow

1. validate_vex_snippet — check VEX syntax and bindings before committing
2. create_wrangle(parent_path="/obj/geo1", wrangle_type="pointwrangle", ...) — create a typed Wrangle with the validated snippet
3. cook_wrangle — cook the node and get initial diagnostics
4. get_vex_info — read back the wrangle metadata (snippet, run-over, cook state)
5. If errors: diagnose_wrangle — localize the failure to a specific line or attribute binding

For iterative work, use update_vex_snippet → cook_wrangle → diagnose_wrangle in a loop until the VEX produces the expected geometry.

Security model

• Client-side VEX allowlist: only known VEX builtins, attributes, and control flow are permitted.
• Deny list blocks: python, exec, eval, import, subprocess, os.*, sys.*, hou.*, and unicode escape sequences.
• Size limits: 64KB / 2000 lines per snippet.
• No Python execution: VEX code flows through hou.Parm.set() — never through exec() or eval() or any HOM Python execution path.