DongyunZou

Clarify a loose product or engineering idea through focused user discovery, repository research, and distinct design options, then write a concise spec. Use for brainstorming, requirement discovery, 头脑风暴, 需求澄清, or before gen-plan.

Turn an approved spec or sufficiently clear Markdown draft into a concise, repo-grounded implementation plan with verifiable deliverables, scope, milestones, and checks. Use for generate a plan, implementation planning, 生成计划, or after gen-idea.

Orchestrate deep idea discovery, implementation planning, independent Architect and Engineer reviews, and a GO/NO-GO gate for large features or refactors. Use for superman flow, superman mode, or a reviewed launch-ready plan; this skill remains plan-only.

Profile a warp-specialized GPU kernel (CuTe-DSL or CUDA C++) by stamping the per-SM clock() register at each cross-warp synchronization point, then reconstruct a per-warp pipeline timeline (Gantt chart). Reveals where warps STALL, the cross-warp DATA DEPENDENCIES (which warp's output unblocks which), and how compute/memory phases OVERLAP. Use when the user wants to understand or visualize a warp-specialized kernel's runtime pipeline, find where warps wait on each other, or verify that producer/consumer warps actually run concurrently. Triggers include "profile the warp specialization", "pipeline timeline", "where does it stall", "how do the warps overlap", "visualize the producer/consumer pipeline". NOT applicable to Triton (cannot express warp specialization nor write timestamps to shared memory) and NOT a substitute for Nsight Compute/Systems for occupancy/throughput/memory-bandwidth analysis.

Coordinate multiple Codex agents or tasks sharing the same NVIDIA GPU by acquiring a local GPU lease before tests, benchmarks, profilers, or any command that may execute CUDA work; includes stale lease timeout cleanup, soft GPU hiding, and zombie/unknown GPU process checks.