Decouple stream cancellation and clarify failover rules

apps/api/internal/httpapi/handlers.go

@@ -1,6 +1,7 @@
 package httpapi
 
 import (
+	"context"
 	"encoding/json"
 	"errors"
 	"fmt"
@@ -564,10 +565,15 @@ func (s *Server) createTask(kind string, compatible bool) http.Handler {
 			writeTaskAccepted(w, task)
 			return
 		}
+		runCtx, cancelRun := s.requestExecutionContext(r)
+		defer cancelRun()
 		if compatible {
 			if boolValue(body, "stream") {
 				flusher := prepareCompatibleStream(w)
-				result, runErr := s.runner.ExecuteStream(r.Context(), task, user, func(delta string) error {
+				result, runErr := s.runner.ExecuteStream(runCtx, task, user, func(delta string) error {
+					if !requestStillConnected(r) {
+						return nil
+					}
 					writeCompatibleDelta(w, kind, model, delta)
 					if flusher != nil {
 						flusher.Flush()
@@ -575,6 +581,9 @@ func (s *Server) createTask(kind string, compatible bool) http.Handler {
 					return nil
 				})
 				if runErr != nil {
+					if !requestStillConnected(r) {
+						return
+					}
 					status := statusFromRunError(runErr)
 					errorPayload := map[string]any{
 						"code":    runErrorCode(runErr),
@@ -593,29 +602,66 @@ func (s *Server) createTask(kind string, compatible bool) http.Handler {
 					}
 					return
 				}
+				if !requestStillConnected(r) {
+					return
+				}
 				writeCompatibleDone(w, kind, model, result.Output)
 				if flusher != nil {
 					flusher.Flush()
 				}
 				return
 			}
-			result, runErr := s.runner.Execute(r.Context(), task, user)
+			result, runErr := s.runner.Execute(runCtx, task, user)
 			if runErr != nil {
+				if !requestStillConnected(r) {
+					return
+				}
 				writeError(w, statusFromRunError(runErr), runErr.Error(), runErrorCode(runErr))
 				return
 			}
+			if !requestStillConnected(r) {
+				return
+			}
 			writeJSON(w, http.StatusOK, result.Output)
 			return
 		}
-		result, runErr := s.runner.Execute(r.Context(), task, user)
+		result, runErr := s.runner.Execute(runCtx, task, user)
 		if runErr != nil {
 			s.logger.Warn("task completed with failure", "kind", kind, "taskId", task.ID, "error", runErr)
 		}
 
+		if !requestStillConnected(r) {
+			return
+		}
 		writeTaskAccepted(w, result.Task)
 	})
 }
 
+func (s *Server) requestExecutionContext(r *http.Request) (context.Context, context.CancelFunc) {
+	base := context.WithoutCancel(r.Context())
+	if s.ctx == nil {
+		return base, func() {}
+	}
+	ctx, cancel := context.WithCancel(base)
+	go func() {
+		select {
+		case <-s.ctx.Done():
+			cancel()
+		case <-ctx.Done():
+		}
+	}()
+	return ctx, cancel
+}
+
+func requestStillConnected(r *http.Request) bool {
+	select {
+	case <-r.Context().Done():
+		return false
+	default:
+		return true
+	}
+}
+
 func asyncRequest(r *http.Request) bool {
 	value := strings.TrimSpace(strings.ToLower(r.Header.Get("x-async")))
 	return value == "1" || value == "true" || value == "yes" || value == "on"