What problem does it solve? Turning a free-form image or video instruction into an executable multi-step ComfyUI workflow requires deciding which capabilities to invoke, in what order, and how outputs chain between steps. This Skill automates that planning: it converts a natural-language instruction into a structured Plan envelope (matching contracts/agent/plan.schema.json) of typed sub-goals that the host Orchestrator executes. ## Core Features & Use Cases - LLM-driven planning (v1.2): propose_plan.py wraps the host Planner to emit Plans with sub-goal kinds like workflow_search, model_resolve, comfyui_execute, llm_transform, evaluate, checkpoint, and wait_user, with cross-step state://, template://, and model:// references. - Deterministic fallback router: when no LLM provider is configured, legacy keyword-based decompose.py and emit_dag.py split instructions on connectives, map clauses to capability skills, and emit a validated DAG with provenance and trust tiers. - Tool discovery and plan validation: discover_tools.py lists executor kinds, models, templates, and MCP tools; validate_plan.py checks schema conformance, DAG acyclicity, and reference resolution. - Use Case: A user says "generate a rainy alley, then inpaint a detective, then apply noir style while preserving the face" — the Skill emits a four-step DAG with dependencies, identity-preservation evaluators, and per-step provenance. ## Quick Start Plan the instruction "generate a cyberpunk portrait, then replace the background with a neon street" into an executable ComfyUI workflow DAG.
