What problem does it solve? Developing code for the MindIE-SD repository (patterns, operators, graph dispatch, tests) involves many sub-workflows—writing tests first, deploying to Ascend NPU environments, running pytest, profiling, and retrospectives—and it is easy to skip gates or route tasks to the wrong sub-skill. This Skill acts as the L1 orchestration entry point that routes every MindIE-SD code change to the right capability skill and enforces order and deliverable contracts. ## Core Features & Use Cases - Entry Routing: Splits work between MindIE-SD repo changes (this entry), third-party framework changes (framework-integration), and environment/deployment tasks (env-install, remote-access), based on which repository the modified files live in. - Test-First Closed Loop: Enforces the cycle of writing a failing test, implementing, deploying to the remote Ascend environment, and verifying with pytest, with evidence (commands, exit codes, results) recorded per feature point. - Parallel Development & Retrospective: Runs independent modules in parallel on separate NPU cards, merges shared files last, and closes each phase with a retrospective checklist, rework-lessons updates, and case backfilling. - Use Case: When asked to add a new fusion pattern for a model, this Skill routes pattern authoring to pattern-dev, deployment to env-install, validation to dummy-run/framework-integration, and ensures the Test-First gate and retrospective are completed before claiming done. ## Quick Start Ask the assistant to add a new fusion pattern to MindIE-SD following the full development workflow with tests written first and remote NPU verification.
