Improve resiliency with better retry defaults

Code execution in particular is sensitive to timeouts and retries, and requires smarter defaults. Users can still override the default behavior by providing custom channel options or channel options.HttpHandler.

Author: kzu

src/xAI/GrokClient.cs

@@ -1,6 +1,10 @@
 using System.Collections.Concurrent;
+using System.Diagnostics;
 using System.Net.Http.Headers;
 using Grpc.Net.Client;
+using Microsoft.Extensions.Http;
+using Polly;
+using Polly.Extensions.Http;
 using xAI.Protocol;
 
 namespace xAI;
@@ -47,12 +51,59 @@ public sealed class GrokClient(string apiKey, GrokClientOptions options) : IDisp
 
     internal GrpcChannel Channel => channels.GetOrAdd((Endpoint, ApiKey), key =>
     {
+        var inner = Options.ChannelOptions?.HttpHandler;
+        if (inner == null)
+        {
+            // If no custom HttpHandler is provided, we create one with Polly retry
+            // policies to handle transient errors, including gRPC-specific ones.
+            var retryPolicy = HttpPolicyExtensions
+                .HandleTransientHttpError()
+                .Or<Grpc.Core.RpcException>(ex =>
+                    ex.StatusCode is Grpc.Core.StatusCode.Unavailable or
+                                     Grpc.Core.StatusCode.DeadlineExceeded or
+                                     Grpc.Core.StatusCode.Internal &&
+                    ex.Status.Detail?.Contains("504") == true ||
+                    ex.Status.Detail?.Contains("INTERNAL_ERROR") == true)
+                .WaitAndRetryAsync(
+                    retryCount: 3,
+                    sleepDurationProvider: attempt => TimeSpan.FromSeconds(Math.Pow(2, attempt))
+#if DEBUG
+                    , onRetry: (outcome, delay, retryCount, ctx) =>
+                    {
+                        Debug.WriteLine($"[xAI Streaming Retry #{retryCount}] {outcome.Exception?.Message} — waiting {delay.TotalSeconds}s");
+                    }
+#endif
+                    );
+
+            inner = new PolicyHttpMessageHandler(retryPolicy)
+            {
+                InnerHandler = new SocketsHttpHandler
+                {
+                    PooledConnectionLifetime = TimeSpan.FromMinutes(20),
+                    KeepAlivePingDelay = TimeSpan.FromSeconds(30),
+                    KeepAlivePingTimeout = TimeSpan.FromSeconds(10),
+                    KeepAlivePingPolicy = HttpKeepAlivePingPolicy.Always,   // crucial for long streams
+                    EnableMultipleHttp2Connections = true,
+                    ConnectTimeout = TimeSpan.FromSeconds(60),
+                    MaxConnectionsPerServer = 10
+                }
+            };
+        }
+
         var handler = new AuthenticationHeaderHandler(ApiKey)
         {
-            InnerHandler = Options.ChannelOptions?.HttpHandler ?? new HttpClientHandler()
+            InnerHandler = inner
+        };
+
+        // Provide some sensible defaults for gRPC channel options, while allowing users to
+        // override them via GrokClientOptions.ChannelOptions if needed.
+        var options = Options.ChannelOptions ?? new GrpcChannelOptions
+        {
+            DisposeHttpClient = true,
+            MaxReceiveMessageSize = 128 * 1024 * 1024,   // large enough for tool output
+            MaxSendMessageSize = 16 * 1024 * 1024,
         };
 
-        var options = Options.ChannelOptions ?? new GrpcChannelOptions();
         options.HttpHandler = handler;
 
         return GrpcChannel.ForAddress(Endpoint, options);

src/xAI/xAI.csproj

@@ -14,6 +14,7 @@
   </PropertyGroup>
 
   <ItemGroup>
+    <PackageReference Include="Microsoft.Extensions.Http.Polly" Version="10.0.3" />
     <PackageReference Include="NuGetizer" Version="1.4.7" PrivateAssets="all" />
     <PackageReference Include="Google.Protobuf" Version="3.34.0" />
     <PackageReference Include="Grpc.Net.Client" Version="2.76.0" />