RecStore

Use when setting up RecStore end-to-end DLRM/TorchRec benchmarks with explicit client placement, PS placement/type, sharding, runnable commands, and Chinese performance summaries.

Run RecStore KVEngine correctness, YCSB, and storage-only batch lookup benchmark workflows, including read_mode=batch_get_flat for aligning KVEngine limits with PS RDMA batch GET. Use when Codex needs to validate src/test/test_kvengine.cpp, run tools/benchmarks/run_kvengine_compare.py, prompt for thread count, SSD benchmark path, results directory, read mode, and batch keys, then generate a Chinese summary.md.

Configure, run, validate, and report RecStore Parameter Server benchmarks with tools/benchmarks/run_benchmark_ps.py, covering local and cross-host topologies from explicit IP lists, RDMA/GRPC/BRPC comparison runs, RDMA module correctness checks, SSH/container execution, optimized DRAM_PET_HASH RDMA GET auto/staging-copy mode, and Chinese summary.md output.

Run matched RecStore and TorchRec DLRM jobs, align their embedding and dense-model initialization, and compare per-rank per-step loss values. Use when validating RecStore/TorchRec numerical equivalence, diagnosing loss divergence, checking sparse update visibility or ordering, or verifying changes to rs_demo training semantics.

Quick-reference card for all ponytail modes, skills, and commands. One-shot display, not a persistent mode. Trigger: /ponytail-help, "ponytail help", "what ponytail commands", "how do I use ponytail".