Merge pull request #1886 from rumpl/retries

Add retries even for models without fallbacks

Author: dgageot

pkg/runtime/fallback.go

@@ -90,6 +90,7 @@ func extractHTTPStatusCode(err error) int {
 //
 // Retryable status codes:
 // - 5xx (server errors): 500, 502, 503, 504
+// - 529 (Anthropic overloaded)
 // - 408 (request timeout)
 //
 // Non-retryable status codes (skip to next model immediately):
@@ -99,6 +100,8 @@ func isRetryableStatusCode(statusCode int) bool {
 	switch statusCode {
 	case 500, 502, 503, 504: // Server errors
 		return true
+	case 529: // Anthropic overloaded
+		return true
 	case 408: // Request timeout
 		return true
 	case 429: // Rate limit - NOT retryable, skip to next model
@@ -119,6 +122,7 @@ func isRetryableStatusCode(statusCode int) bool {
 // - Network timeouts
 // - Temporary network errors
 // - HTTP 5xx errors (server errors)
+// - HTTP 529 (Anthropic overloaded)
 // - HTTP 408 (request timeout)
 //
 // Non-retryable errors (skip to next model in chain immediately):
@@ -180,6 +184,10 @@ func isRetryableModelError(err error) bool {
 		"gateway timeout",       // Gateway timeout
 		"overloaded",            // Server overloaded
 		"overloaded_error",      // Server overloaded
+		"other side closed",     // Connection closed by peer
+		"fetch failed",          // Network fetch failure
+		"reset before headers",  // Connection reset before headers received
+		"upstream connect",      // Upstream connection error
 	}
 
 	for _, pattern := range retryablePatterns {
@@ -370,9 +378,10 @@ func getEffectiveCooldown(a *agent.Agent) time.Duration {
 }
 
 // getEffectiveRetries returns the number of retries to use for the agent.
-// If no retries are explicitly configured (retries == 0) and fallback models
-// are configured, returns DefaultFallbackRetries to provide sensible retry
-// behavior out of the box.
+// If no retries are explicitly configured (retries == 0), returns
+// DefaultFallbackRetries to provide sensible retry behavior out of the box.
+// This ensures that transient errors (e.g., Anthropic 529 overloaded) are
+// retried even when no fallback models are configured.
 //
 // Note: Users who explicitly want 0 retries can set retries: -1 in their config
 // (though this is an edge case - most users want some retries for resilience).
@@ -382,8 +391,8 @@ func getEffectiveRetries(a *agent.Agent) int {
 	if retries < 0 {
 		return 0
 	}
-	// 0 means "use default" when fallback models are configured
-	if retries == 0 && len(a.FallbackModels()) > 0 {
+	// 0 means "use default" - always provide retries for transient error resilience
+	if retries == 0 {
 		return DefaultFallbackRetries
 	}
 	return retries

pkg/runtime/fallback_test.go

@@ -176,6 +176,31 @@ func TestIsRetryableModelError(t *testing.T) {
 			err:      errors.New("authentication failed"),
 			expected: false,
 		},
+		{
+			name:     "anthropic overloaded 529",
+			err:      errors.New("529 overloaded"),
+			expected: true,
+		},
+		{
+			name:     "other side closed",
+			err:      errors.New("other side closed the connection"),
+			expected: true,
+		},
+		{
+			name:     "fetch failed",
+			err:      errors.New("fetch failed"),
+			expected: true,
+		},
+		{
+			name:     "reset before headers",
+			err:      errors.New("reset before headers"),
+			expected: true,
+		},
+		{
+			name:     "upstream connect error",
+			err:      errors.New("upstream connect error"),
+			expected: true,
+		},
 		{
 			name:     "unknown error",
 			err:      errors.New("something weird happened"),
@@ -591,6 +616,7 @@ func TestIsRetryableStatusCode(t *testing.T) {
 		{503, true},  // Service unavailable - retryable
 		{504, true},  // Gateway timeout - retryable
 		{408, true},  // Request timeout - retryable
+		{529, true},  // Anthropic overloaded - retryable
 		{429, false}, // Rate limit - NOT retryable (skip to next model)
 		{400, false}, // Bad request - not retryable
 		{401, false}, // Unauthorized - not retryable
@@ -719,12 +745,13 @@ func TestGetEffectiveRetries(t *testing.T) {
 	mockModel := &mockProvider{id: "test/model", stream: newStreamBuilder().AddContent("ok").AddStopWithUsage(1, 1).Build()}
 	mockFallback := &mockProvider{id: "test/fallback", stream: newStreamBuilder().AddContent("ok").AddStopWithUsage(1, 1).Build()}
 
-	// Agent with no retries configured and no fallback models should return 0
+	// Agent with no retries configured and no fallback models should use default
+	// retries for transient error resilience (e.g., Anthropic 529 overloaded)
 	agentNoFallback := agent.New("no-fallback", "test",
 		agent.WithModel(mockModel),
 	)
 	retries := getEffectiveRetries(agentNoFallback)
-	assert.Equal(t, 0, retries, "no fallback models = no retries (nothing to retry to)")
+	assert.Equal(t, DefaultFallbackRetries, retries, "should use default retries even without fallback models")
 
 	// Agent with no retries configured but with fallback models should use default
 	agentWithFallback := agent.New("with-fallback", "test",