fix(auth): skip downtime for request-scoped 404 errors in model state management

sdk/cliproxy/auth/conductor.go

@@ -1734,77 +1734,79 @@ func (m *Manager) MarkResult(ctx context.Context, result Result) {
 			}
 		} else {
 			if result.Model != "" {
-				state := ensureModelState(auth, result.Model)
-				state.Unavailable = true
-				state.Status = StatusError
-				state.UpdatedAt = now
-				if result.Error != nil {
-					state.LastError = cloneError(result.Error)
-					state.StatusMessage = result.Error.Message
-					auth.LastError = cloneError(result.Error)
-					auth.StatusMessage = result.Error.Message
-				}
+				if !isRequestScopedNotFoundResultError(result.Error) {
+					state := ensureModelState(auth, result.Model)
+					state.Unavailable = true
+					state.Status = StatusError
+					state.UpdatedAt = now
+					if result.Error != nil {
+						state.LastError = cloneError(result.Error)
+						state.StatusMessage = result.Error.Message
+						auth.LastError = cloneError(result.Error)
+						auth.StatusMessage = result.Error.Message
+					}
 
-				statusCode := statusCodeFromResult(result.Error)
-				if isModelSupportResultError(result.Error) {
-					next := now.Add(12 * time.Hour)
-					state.NextRetryAfter = next
-					suspendReason = "model_not_supported"
-					shouldSuspendModel = true
-				} else {
-					switch statusCode {
-					case 401:
-						next := now.Add(30 * time.Minute)
-						state.NextRetryAfter = next
-						suspendReason = "unauthorized"
-						shouldSuspendModel = true
-					case 402, 403:
-						next := now.Add(30 * time.Minute)
-						state.NextRetryAfter = next
-						suspendReason = "payment_required"
-						shouldSuspendModel = true
-					case 404:
+					statusCode := statusCodeFromResult(result.Error)
+					if isModelSupportResultError(result.Error) {
 						next := now.Add(12 * time.Hour)
 						state.NextRetryAfter = next
-						suspendReason = "not_found"
+						suspendReason = "model_not_supported"
 						shouldSuspendModel = true
-					case 429:
-						var next time.Time
-						backoffLevel := state.Quota.BackoffLevel
-						if result.RetryAfter != nil {
-							next = now.Add(*result.RetryAfter)
-						} else {
-							cooldown, nextLevel := nextQuotaCooldown(backoffLevel, quotaCooldownDisabledForAuth(auth))
-							if cooldown > 0 {
-								next = now.Add(cooldown)
-							}
-							backoffLevel = nextLevel
-						}
-						state.NextRetryAfter = next
-						state.Quota = QuotaState{
-							Exceeded:      true,
-							Reason:        "quota",
-							NextRecoverAt: next,
-							BackoffLevel:  backoffLevel,
-						}
-						suspendReason = "quota"
-						shouldSuspendModel = true
-						setModelQuota = true
-					case 408, 500, 502, 503, 504:
-						if quotaCooldownDisabledForAuth(auth) {
-							state.NextRetryAfter = time.Time{}
-						} else {
-							next := now.Add(1 * time.Minute)
+					} else {
+						switch statusCode {
+						case 401:
+							next := now.Add(30 * time.Minute)
 							state.NextRetryAfter = next
+							suspendReason = "unauthorized"
+							shouldSuspendModel = true
+						case 402, 403:
+							next := now.Add(30 * time.Minute)
+							state.NextRetryAfter = next
+							suspendReason = "payment_required"
+							shouldSuspendModel = true
+						case 404:
+							next := now.Add(12 * time.Hour)
+							state.NextRetryAfter = next
+							suspendReason = "not_found"
+							shouldSuspendModel = true
+						case 429:
+							var next time.Time
+							backoffLevel := state.Quota.BackoffLevel
+							if result.RetryAfter != nil {
+								next = now.Add(*result.RetryAfter)
+							} else {
+								cooldown, nextLevel := nextQuotaCooldown(backoffLevel, quotaCooldownDisabledForAuth(auth))
+								if cooldown > 0 {
+									next = now.Add(cooldown)
+								}
+								backoffLevel = nextLevel
+							}
+							state.NextRetryAfter = next
+							state.Quota = QuotaState{
+								Exceeded:      true,
+								Reason:        "quota",
+								NextRecoverAt: next,
+								BackoffLevel:  backoffLevel,
+							}
+							suspendReason = "quota"
+							shouldSuspendModel = true
+							setModelQuota = true
+						case 408, 500, 502, 503, 504:
+							if quotaCooldownDisabledForAuth(auth) {
+								state.NextRetryAfter = time.Time{}
+							} else {
+								next := now.Add(1 * time.Minute)
+								state.NextRetryAfter = next
+							}
+						default:
+							state.NextRetryAfter = time.Time{}
 						}
-					default:
-						state.NextRetryAfter = time.Time{}
 					}
-				}
 
-				auth.Status = StatusError
-				auth.UpdatedAt = now
-				updateAggregatedAvailability(auth, now)
+					auth.Status = StatusError
+					auth.UpdatedAt = now
+					updateAggregatedAvailability(auth, now)
+				}
 			} else {
 				applyAuthFailureState(auth, result.Error, result.RetryAfter, now)
 			}
@@ -2056,11 +2058,29 @@ func isModelSupportResultError(err *Error) bool {
 	return isModelSupportErrorMessage(err.Message)
 }
 
+func isRequestScopedNotFoundMessage(message string) bool {
+	if message == "" {
+		return false
+	}
+	lower := strings.ToLower(message)
+	return strings.Contains(lower, "item with id") &&
+		strings.Contains(lower, "not found") &&
+		strings.Contains(lower, "items are not persisted when `store` is set to false")
+}
+
+func isRequestScopedNotFoundResultError(err *Error) bool {
+	if err == nil || statusCodeFromResult(err) != http.StatusNotFound {
+		return false
+	}
+	return isRequestScopedNotFoundMessage(err.Message)
+}
+
 // isRequestInvalidError returns true if the error represents a client request
 // error that should not be retried. Specifically, it treats 400 responses with
-// "invalid_request_error" and all 422 responses as request-shape failures,
-// where switching auths or pooled upstream models will not help. Model-support
-// errors are excluded so routing can fall through to another auth or upstream.
+// "invalid_request_error", request-scoped 404 item misses caused by `store=false`,
+// and all 422 responses as request-shape failures, where switching auths or
+// pooled upstream models will not help. Model-support errors are excluded so
+// routing can fall through to another auth or upstream.
 func isRequestInvalidError(err error) bool {
 	if err == nil {
 		return false
@@ -2072,6 +2092,8 @@ func isRequestInvalidError(err error) bool {
 	switch status {
 	case http.StatusBadRequest:
 		return strings.Contains(err.Error(), "invalid_request_error")
+	case http.StatusNotFound:
+		return isRequestScopedNotFoundMessage(err.Error())
 	case http.StatusUnprocessableEntity:
 		return true
 	default:
@@ -2083,6 +2105,9 @@ func applyAuthFailureState(auth *Auth, resultErr *Error, retryAfter *time.Durati
 	if auth == nil {
 		return
 	}
+	if isRequestScopedNotFoundResultError(resultErr) {
+		return
+	}
 	auth.Unavailable = true
 	auth.Status = StatusError
 	auth.UpdatedAt = now

sdk/cliproxy/auth/conductor_overrides_test.go

@@ -12,6 +12,8 @@ import (
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
 )
 
+const requestScopedNotFoundMessage = "Item with id 'rs_0b5f3eb6f51f175c0169ca74e4a85881998539920821603a74' not found. Items are not persisted when `store` is set to false. Try again with `store` set to true, or remove this item from your input."
+
 func TestManager_ShouldRetryAfterError_RespectsAuthRequestRetryOverride(t *testing.T) {
 	m := NewManager(nil, nil, nil)
 	m.SetRetryConfig(3, 30*time.Second, 0)
@@ -447,3 +449,114 @@ func TestManager_MarkResult_RespectsAuthDisableCoolingOverride(t *testing.T) {
 		t.Fatalf("expected NextRetryAfter to be zero when disable_cooling=true, got %v", state.NextRetryAfter)
 	}
 }
+
+func TestManager_MarkResult_RequestScopedNotFoundDoesNotCooldownAuth(t *testing.T) {
+	m := NewManager(nil, nil, nil)
+
+	auth := &Auth{
+		ID:       "auth-1",
+		Provider: "openai",
+	}
+	if _, errRegister := m.Register(context.Background(), auth); errRegister != nil {
+		t.Fatalf("register auth: %v", errRegister)
+	}
+
+	model := "gpt-4.1"
+	m.MarkResult(context.Background(), Result{
+		AuthID:   auth.ID,
+		Provider: auth.Provider,
+		Model:    model,
+		Success:  false,
+		Error: &Error{
+			HTTPStatus: http.StatusNotFound,
+			Message:    requestScopedNotFoundMessage,
+		},
+	})
+
+	updated, ok := m.GetByID(auth.ID)
+	if !ok || updated == nil {
+		t.Fatalf("expected auth to be present")
+	}
+	if updated.Unavailable {
+		t.Fatalf("expected request-scoped 404 to keep auth available")
+	}
+	if !updated.NextRetryAfter.IsZero() {
+		t.Fatalf("expected request-scoped 404 to keep auth cooldown unset, got %v", updated.NextRetryAfter)
+	}
+	if state := updated.ModelStates[model]; state != nil {
+		t.Fatalf("expected request-scoped 404 to avoid model cooldown state, got %#v", state)
+	}
+}
+
+func TestManager_RequestScopedNotFoundStopsRetryWithoutSuspendingAuth(t *testing.T) {
+	m := NewManager(nil, nil, nil)
+	executor := &authFallbackExecutor{
+		id: "openai",
+		executeErrors: map[string]error{
+			"aa-bad-auth": &Error{
+				HTTPStatus: http.StatusNotFound,
+				Message:    requestScopedNotFoundMessage,
+			},
+		},
+	}
+	m.RegisterExecutor(executor)
+
+	model := "gpt-4.1"
+	badAuth := &Auth{ID: "aa-bad-auth", Provider: "openai"}
+	goodAuth := &Auth{ID: "bb-good-auth", Provider: "openai"}
+
+	reg := registry.GetGlobalRegistry()
+	reg.RegisterClient(badAuth.ID, "openai", []*registry.ModelInfo{{ID: model}})
+	reg.RegisterClient(goodAuth.ID, "openai", []*registry.ModelInfo{{ID: model}})
+	t.Cleanup(func() {
+		reg.UnregisterClient(badAuth.ID)
+		reg.UnregisterClient(goodAuth.ID)
+	})
+
+	if _, errRegister := m.Register(context.Background(), badAuth); errRegister != nil {
+		t.Fatalf("register bad auth: %v", errRegister)
+	}
+	if _, errRegister := m.Register(context.Background(), goodAuth); errRegister != nil {
+		t.Fatalf("register good auth: %v", errRegister)
+	}
+
+	_, errExecute := m.Execute(context.Background(), []string{"openai"}, cliproxyexecutor.Request{Model: model}, cliproxyexecutor.Options{})
+	if errExecute == nil {
+		t.Fatal("expected request-scoped not-found error")
+	}
+	errResult, ok := errExecute.(*Error)
+	if !ok {
+		t.Fatalf("expected *Error, got %T", errExecute)
+	}
+	if errResult.HTTPStatus != http.StatusNotFound {
+		t.Fatalf("status = %d, want %d", errResult.HTTPStatus, http.StatusNotFound)
+	}
+	if errResult.Message != requestScopedNotFoundMessage {
+		t.Fatalf("message = %q, want %q", errResult.Message, requestScopedNotFoundMessage)
+	}
+
+	got := executor.ExecuteCalls()
+	want := []string{badAuth.ID}
+	if len(got) != len(want) {
+		t.Fatalf("execute calls = %v, want %v", got, want)
+	}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("execute call %d auth = %q, want %q", i, got[i], want[i])
+		}
+	}
+
+	updatedBad, ok := m.GetByID(badAuth.ID)
+	if !ok || updatedBad == nil {
+		t.Fatalf("expected bad auth to remain registered")
+	}
+	if updatedBad.Unavailable {
+		t.Fatalf("expected request-scoped 404 to keep bad auth available")
+	}
+	if !updatedBad.NextRetryAfter.IsZero() {
+		t.Fatalf("expected request-scoped 404 to keep bad auth cooldown unset, got %v", updatedBad.NextRetryAfter)
+	}
+	if state := updatedBad.ModelStates[model]; state != nil {
+		t.Fatalf("expected request-scoped 404 to avoid bad auth model cooldown state, got %#v", state)
+	}
+}