api-load-tester

Run hey with -n for total requests, -c for concurrency, and -t for timeout, adding -H flags for auth headers and -d for request bodies. The skill checks for hey first, then wrk, then ab, falling back to curl with background bash processes if none are installed.

The breaking point is the concurrency level where the error rate exceeds 1%, p95 latency exceeds 5x baseline, throughput drops versus the previous stage, or more than 5% of connections are refused. The report names which condition triggered it.

Yes, POST, PUT, and DELETE endpoints with Bearer tokens and JSON bodies are supported, but the skill asks for confirmation that the endpoint is idempotent or side-effect free. Auth tokens are masked in all output and reports.

The skill never load-tests production without explicit user confirmation and warns when the URL appears to be a production domain. It also respects 429 rate limits by reducing concurrency instead of continuing to hammer the endpoint.

Connection errors at high concurrency usually indicate OS-level file descriptor limits on the client machine or connection pool exhaustion on the server. The skill detects these errors, suggests increasing ulimit -n, and classifies the bottleneck in the report.

The report includes p50 through p99 latency percentiles per concurrency stage, throughput in requests per second, error breakdowns by status code and type, ASCII throughput and latency charts, bottleneck classification, and capacity scaling projections.