What problem does it solve? PyPTO operator performance tuning involves many phases, iterations, and rollback decisions that are easy to skip or execute out of order. This Skill enforces a rigid state-machine-driven workflow so every tuning round is verified, measured, recorded, and judged before proceeding. ## Core Features & Use Cases - Three-Level State Machine: Controls the main flow (INIT through S5_REPORT), S4_TUNE sub-phases (FRONTEND, SWIMLANE, INCORE), and per-iteration loops (ITER_START through ITER_JUDGE) with hard transition gates. - Todo List Enforcement: Mandates complete upfront Todo creation covering all three outer-loop rounds, with real-time updates after every measurement, rollback, and phase summary. - Phase Handoff Isolation: Generates structured phase-summary handoffs and launches isolated Task subagents per phase, preventing context bloat and cross-phase packing. - Use Case: When tuning a PyPTO operator on Ascend NPU, the orchestrator ensures each optimization point comes from the analysis-derived tuning list, each code change modifies only one parameter, and accuracy failures trigger automatic rollback before continuing. ## Quick Start Ask the assistant to orchestrate a complete PyPTO operator performance tuning run from environment check through final report, following the enforced state machine.
