Browse reusable Agent Skills, each with a clear purpose and practical guidance.
Use when evaluating a newly adapted model's accuracy/eval score. Runs evalscope eval (or the test/srt/run_eval.py fallback) against an OpenAI-compatible endpoint on datasets like gsm8k / aime, producing per-dataset reports and a score. Accuracy only — throughput/latency benchmarks belong to model-speed-benchmark. Triggers on accuracy, eval, evalscope, gsm8k, aime, run_eval — not throughput.
Use when running throughput/latency (serving speed) benchmarks against an already-running model server. Uses sgl_jax.bench_serving to sweep batch-size/concurrency and measure throughput and latency, producing summary.csv and per-point bs_<N>/result.jsonl. Speed only — accuracy/eval belongs to model-accuracy-benchmark. Triggers on throughput, latency, bench_serving, batch-size sweep, tok/s, TTFT, TPOT — not accuracy/eval.
Use when you need to CAPTURE / PRODUCE XProf (jax.profiler) profiling artifacts from sgl-jax (SGLang on JAX/TPU) for later performance analysis: driving a live server (/start_profile, stage-separated prefill/decode), offline bench_one_batch --profile, or kernel microbench. This skill produces the trace files; it does not analyze them.
Use when a user wants to report, troubleshoot, or collect context for a WispTerm issue, including crashes, rendering/DPI glitches, high CPU, keyboard/input bugs, selection/copy/scrolling, SSH/SCP failures, SSH image preview failures, HTML preview/browser panel failures, SSH disconnects such as ssh_packet_write_poll/eother, file explorer behavior, updater failures, or remote console behavior.
Concurrency analysis for tractor's trio-based async primitives. Trace task scheduling across checkpoint boundaries, identify race windows in shared mutable state, and verify synchronization correctness. Invoke on code segments the user points at, OR proactively when reviewing/writing concurrent cache, lock, or multi-task acm code.
Review design specification documents for Go programs before implementation begins. Use this skill when a user has a Go spec, design doc, RFC, or PRD to review, wants feedback on a planned feature in a Go codebase, is about to start implementing from a spec, or asks "is this spec ready?" or "review this before I build it". Applies Go philosophy — simplicity, composition, explicit errors, context propagation — plus Cobra/Viper CLI conventions where applicable.
Record a hard-won engineering lesson after a bug or mistake is resolved. Use when a bug resisted ATLAS until Boss's guidance cracked it, or when a long debugging hunt ended in an aha moment worth never relearning.
Use when triaging a live Area Occupancy Detection symptom report — occupancy stuck on/won't clear, occupancy won't turn on, probability pinned at 0.99 or 0.01, wrong or spammy repair issues, a config setting that seems to be "ignored" (especially decay half-life or min_prior_override), the database growing/slow, config-flow errors, or entities showing 100% occupied right after a restart. Also load this before touching timezone/DST datetime handling, decay half-life resolution, or prior/global_prior calculation code — these are the project's three most expensive historical bug classes and this skill has the exact traps.
Use when you need to MEASURE what Area Occupancy Detection is actually doing instead of guessing — downloading and reading the HA diagnostics export, interpreting a health/repair issue, reading hourly-analysis debug logs, or using the simulator to compare a predicted probability against the live sensor. Load this before answering "why is this area stuck occupied", "is the prior learned correctly", "what does this repair mean", or "how do I reproduce this probability by hand".