get-benchmark-results

name	get-benchmark-results
description	Get benchmark results from a workflow run. Use when: benchmark results, get benchmarks, latency numbers, run results, performance results, compare benchmarks.
argument-hint	<run-id|run-url|branch> [comparison-run-id|run-url]
user-invocable	true
disable-model-invocation	false

Get Benchmark Results

Preflight

Complete the shared repository skill preflight before continuing.

Steps

1. Identify the run 1.0. Resolve the repository 1.1. User provides run ID or URL — extract the numeric run ID 1.2. User provides a branch name — find latest run for branch 1.3. If neither provided, list recent runs and ask which one
2. Verify benchmarks ran — get job details 2.1. Find the tune-and-benchmark job(s) — one per architecture 2.2. If job failed/skipped → report that and stop
3. Fetch benchmark logs — get benchmark logs
4. Extract results — parse the log output for: 4.1. Per-experiment latencies — lines matching Median latency for <name>: <seconds> seconds 4.2. Failed experiments — lines matching failed or Error 4.3. Optional wall-clock summary — include the Experiment / Time (s) table only when present
5. Present results — separate results by runner/job and show experiment, median latency, and status
6. Compare runs (only when the user provides two runs) 6.1. Fetch both runs independently and match by runner/job plus experiment name 6.2. Report current, comparison, delta seconds, and delta percent where both medians exist 6.3. Delta percent = (current - comparison) / comparison * 100; negative is faster 6.4. Report unmatched or failed experiments without manufacturing a numeric comparison