What problem does it solve? Before/after performance comparisons fail when traffic generation is not reproducible: ad-hoc curl loops and hand-written pollers produce different commands each run, so two runs measure two different things. This Skill drives a fixed, replayable request scenario against a running service and records it verbatim, so the telemetry from an observation run can be compared against a verification run after a fix. ## Core Features & Use Cases - Scripted scenario driving: The shipped drive_scenario.py script sends warmup and load requests with deterministic identity headers (User-Agent and W3C traceparent derived from the run slug), records every request row, and prints a verbatim run record. - Stored k6 benchmark replay: Replays benchmarks stored under .odd/benchmarks/<name>/ via the k6-guides replay script, with stage-boundary arithmetic, ramp segmentation, and rules for reading k6's exit codes and summaries as evidence rather than verdicts. - Run identity and isolation: References define how to qualify queries by run identity (service.instance.id, User-Agent, trace-id prefix), handle occupied ports, forbidden resets, remote targets, and long or non-deterministic scenarios. - Use Case: After fixing a slow endpoint, re-run the exact recorded command from the observation report's Commands: line and compare p95 latency measured from the same trace identity, counts, and concurrency. ## Quick Start Ask the agent to run the run-scenario skill to drive 30 requests each against GET /api/users and POST /api/orders on http://127.0.0.1:8080 with run slug "baseline-01" and record the scenario for later replay.
