Startup Performance

The SDK has a deterministic startup budget for work controlled by this wrapper. TestStartupBudgets uses a local JSON-RPC fixture, five warmup iterations, and 50 measured samples for each metric. It reports the median and p95 and fails if any p95 reaches 50 ms.

The metrics have deliberately narrow boundaries:

• publicImportMs: the package-initialization interval recorded from the earliest package variable initializer through package init, read from a fresh Go test process. Go runtime process boot is outside the timer.
• sdkStartReturnMs: elapsed time for the public SDK.Start API, including its readiness getState request.
• fixtureSpawnToFirstRpcMs: elapsed time to start the deterministic CLI fixture and complete a successful getState request through RPCClient.

Baseline captured on 2026-07-20:

Metric	Median	p95	Budget
publicImportMs	0.001000 ms	0.001208 ms	< 50 ms
sdkStartReturnMs	6.405417 ms	7.070250 ms	< 50 ms
fixtureSpawnToFirstRpcMs	6.243750 ms	6.837833 ms	< 50 ms

Run the benchmark directly:

go test -run TestStartupBudgets -count=1 -v

The test emits one machine-readable JSON object with top-level language, budgetMs, metrics, and passed fields. Each metric contains samples, medianMs, p95Ms, maxMs, and passed. The fixed protocol is five warmups and 50 measured samples.

These measurements isolate wrapper overhead. Startup of a real Autohand CLI, provider authentication, network access, model loading, and provider readiness depend on the host environment and are intentionally reported separately from the deterministic 50 ms gate.