Fix Copilot client startup cleanup race

Clean up partially established CLI connections if startup fails during protocol negotiation, and avoid rethrowing failed startup tasks during force-stop cleanup.

Add a TCP fake-CLI regression test that drops during startup verification and verifies ForceStopAsync remains cleanup-safe.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: stephentoub

dotnet/src/Client.cs

@@ -244,12 +244,20 @@ async Task<Connection> StartCoreAsync(CancellationToken ct)
 
             var connection = await result;
 
-            // Verify protocol version compatibility
-            await VerifyProtocolVersionAsync(connection, ct);
-            await ConfigureSessionFsAsync(ct);
+            try
+            {
+                // Verify protocol version compatibility
+                await VerifyProtocolVersionAsync(connection, ct);
+                await ConfigureSessionFsAsync(ct);
 
-            _logger.LogInformation("Copilot client connected");
-            return connection;
+                _logger.LogInformation("Copilot client connected");
+                return connection;
+            }
+            catch
+            {
+                await CleanupConnectionAsync(connection, errors: null);
+                throw;
+            }
         }
     }
 
@@ -353,11 +361,27 @@ private async Task CleanupConnectionAsync(List<Exception>? errors)
             return;
         }
 
-        var ctx = await _connectionTask;
+        var connectionTask = _connectionTask;
         _connectionTask = null;
 
+        Connection ctx;
+        try
+        {
+            ctx = await connectionTask;
+        }
+        catch (Exception ex)
+        {
+            _logger.LogDebug(ex, "Ignoring failed Copilot client startup during cleanup");
+            return;
+        }
+
+        await CleanupConnectionAsync(ctx, errors);
+    }
+
+    private async Task CleanupConnectionAsync(Connection ctx, List<Exception>? errors)
+    {
         try { ctx.Rpc.Dispose(); }
-        catch (Exception ex) { errors?.Add(ex); }
+        catch (Exception ex) { AddCleanupError(errors, ex); }
 
         // Clear RPC and models cache
         _serverRpc = null;
@@ -366,7 +390,7 @@ private async Task CleanupConnectionAsync(List<Exception>? errors)
         if (ctx.NetworkStream is not null)
         {
             try { await ctx.NetworkStream.DisposeAsync(); }
-            catch (Exception ex) { errors?.Add(ex); }
+            catch (Exception ex) { AddCleanupError(errors, ex); }
         }
 
         if (ctx.CliProcess is { } childProcess)
@@ -380,7 +404,19 @@ private async Task CleanupConnectionAsync(List<Exception>? errors)
                 }
                 childProcess.Dispose();
             }
-            catch (Exception ex) { errors?.Add(ex); }
+            catch (Exception ex) { AddCleanupError(errors, ex); }
+        }
+    }
+
+    private void AddCleanupError(List<Exception>? errors, Exception ex)
+    {
+        if (errors is not null)
+        {
+            errors.Add(ex);
+        }
+        else
+        {
+            _logger.LogDebug(ex, "Error while cleaning up Copilot CLI connection");
         }
     }
 

dotnet/test/E2E/ClientOptionsE2ETests.cs

@@ -202,6 +202,28 @@ public async Task Should_Propagate_Activity_TraceContext_To_Session_Create_And_S
         await session.DisposeAsync();
     }
 
+    [Fact]
+    public async Task ForceStop_Does_Not_Rethrow_When_Tcp_Cli_Drops_During_Startup()
+    {
+        var cliPath = Path.Join(Ctx.WorkDir, $"fake-tcp-drop-cli-{Guid.NewGuid():N}.js");
+        await File.WriteAllTextAsync(cliPath, FakeTcpDropDuringStartupCliScript);
+
+        await using var client = Ctx.CreateClient(
+            useStdio: false,
+            options: new CopilotClientOptions
+            {
+                AutoStart = false,
+                CliPath = cliPath,
+                UseLoggedInUser = false,
+            });
+
+        var ex = await Assert.ThrowsAsync<IOException>(() => client.StartAsync());
+        Assert.Contains("Communication error", ex.Message, StringComparison.Ordinal);
+
+        await client.ForceStopAsync();
+        Assert.Equal(ConnectionState.Disconnected, client.State);
+    }
+
     [Fact]
     public async Task Should_Propagate_Activity_TraceContext_To_Session_Resume()
     {
@@ -362,6 +384,24 @@ private static JsonElement GetCapturedRequestParams(JsonElement captureRoot, str
             .GetProperty("params");
     }
 
+    private const string FakeTcpDropDuringStartupCliScript = """
+        const net = require("net");
+
+        const server = net.createServer(socket => {
+          socket.on("data", () => {
+            socket.destroy();
+            server.close(() => process.exit(0));
+          });
+        });
+
+        server.listen(0, "localhost", () => {
+          const address = server.address();
+          console.log(`listening on port ${address.port}`);
+        });
+
+        setTimeout(() => process.exit(2), 30000).unref();
+        """;
+
     private const string FakeStdioCliScript = """
         const fs = require("fs");