What problem does it solve? When running multi-stage AI pipelines for Ascend C operator generation (cake, cake-evo, cake-partial modes), there is no visibility into which skills were called, how long each took, or how skill usage correlates with final compilation success and speedup. This Skill provides a standardized JSON tracing mechanism to answer those questions. ## Core Features & Use Cases - Lifecycle Tracing: TRACE-INIT, TRACE-START, TRACE-END, and TRACE-FINALIZE operations record each skill's start time, duration, status, retry count, inputs, and outputs into skill_trace.json. - Metadata Attachment: TRACE-META appends extra context such as optimization strategies applied during dsl-lowering. - Multi-Variant Aggregation: TRACE-AGGREGATE merges traces across cake-evo variants to compute skill frequency, average duration, success rate, and best/worst variant correlation analysis. - Use Case: After running a cake-evo pipeline with multiple parallel variants, aggregate all skill_trace.json files to identify which skills and strategies the best-speedup variant used compared to the worst. ## Quick Start Initialize a skill trace file for this operator task, then record the start and completion of each skill invocation and finalize the trace with the evaluation results.
