Startup Performance

The startup gate records three wrapper-controlled metrics:

• publicImportMs: a fresh .NET child process times loading the public SDK assembly from inside the process, excluding .NET runtime boot.
• sdkStartReturnMs: a ready benchmark runtime times the public StartAsync call against a deterministic native fixture.
• fixtureSpawnToFirstRpcMs: the same runtime times fixture spawn through a successful GetStateAsync response.

Every metric uses 5 warmups and 50 measured samples. The benchmark exits nonzero if any p95 is not below the 50 ms budget.

dotnet run --project benchmarks/Autohand.CodeAgentSdk.StartupBenchmark/Autohand.CodeAgentSdk.StartupBenchmark.csproj --configuration Release

Measured on 2026-07-20 on the development macOS host:

{
  "language": "csharp",
  "budgetMs": 50,
  "metrics": {
    "publicImportMs": { "samples": 50, "medianMs": 0.996, "p95Ms": 1.309, "maxMs": 1.703, "passed": true },
    "sdkStartReturnMs": { "samples": 50, "medianMs": 10.133, "p95Ms": 12.92, "maxMs": 18.512, "passed": true },
    "fixtureSpawnToFirstRpcMs": { "samples": 50, "medianMs": 11.462, "p95Ms": 17.45, "maxMs": 28.253, "passed": true }
  },
  "passed": true
}

The native fixture excludes network and provider latency. A real Autohand CLI session still depends on local CLI initialization, authentication, provider availability, and network conditions; this benchmark does not claim those environment-dependent paths are below 50 ms.