fix(providers): handle model-invalid 400 errors with fallback and warning

Add FailoverModelInvalid reason for 400 errors where the model ID is
invalid or unavailable (e.g. 'not a valid model ID'). Previously, all
400 errors were classified as FailoverFormat (non-retriable), which
aborted the fallback chain immediately instead of trying the next model.

Model-invalid patterns are now checked BEFORE HTTP status code
classification, allowing fallback to proceed. When fallback succeeds
after a model-invalid error, a warning message is sent to the user's
channel so they know to fix the config.

pkg/agent/loop.go

@@ -557,6 +557,22 @@ func (al *AgentLoop) runLLMIteration(
 					logger.InfoCF("agent", fmt.Sprintf("Fallback: succeeded with %s/%s after %d attempts",
 						fbResult.Provider, fbResult.Model, len(fbResult.Attempts)+1),
 						map[string]any{"agent_id": agent.ID, "iteration": iteration})
+
+					// Send warning for model-invalid errors so user knows to fix config
+					if !constants.IsInternalChannel(opts.Channel) {
+						for _, attempt := range fbResult.Attempts {
+							if failErr, ok := attempt.Error.(*providers.FailoverError); ok && failErr.IsModelInvalid() {
+								al.bus.PublishOutbound(bus.OutboundMessage{
+									Channel: opts.Channel,
+									ChatID:  opts.ChatID,
+									Content: fmt.Sprintf("⚠️ Model %s/%s is invalid or unavailable: %v\nUsing fallback: %s/%s",
+										attempt.Provider, attempt.Model, failErr.Wrapped,
+										fbResult.Provider, fbResult.Model),
+								})
+								break // Only send one warning even if multiple model-invalid errors
+							}
+						}
+					}
 				}
 				return fbResult.Response, nil
 			}

pkg/providers/error_classifier.go

@@ -78,6 +78,21 @@ var (
 		substr("invalid request format"),
 	}
 
+	// Model invalid/not found patterns: these are 400 errors that should be
+	// retriable (fallback to next model), NOT treated as format errors.
+	modelInvalidPatterns = []errorPattern{
+		substr("not a valid model"),
+		substr("model not found"),
+		substr("model_not_found"),
+		substr("model not available"),
+		substr("does not exist"),
+		substr("no such model"),
+		substr("invalid model"),
+		rxp(`model.*not.*supported`),
+		rxp(`model.*is.*unavailable`),
+		rxp(`model.*is.*deprecated`),
+	}
+
 	imageDimensionPatterns = []errorPattern{
 		rxp(`image dimensions exceed max`),
 	}
@@ -128,7 +143,19 @@ func ClassifyError(err error, provider, model string) *FailoverError {
 		}
 	}
 
-	// Try HTTP status code extraction first.
+	// Model invalid/not found: retriable, should fallback to another model.
+	// This MUST run before HTTP status classification, because 400 + "not a valid model"
+	// would otherwise be classified as non-retriable FailoverFormat.
+	if matchesAny(msg, modelInvalidPatterns) {
+		return &FailoverError{
+			Reason:   FailoverModelInvalid,
+			Provider: provider,
+			Model:    model,
+			Wrapped:  err,
+		}
+	}
+
+	// Try HTTP status code extraction.
 	if status := extractHTTPStatus(msg); status > 0 {
 		if reason := classifyByStatus(status); reason != "" {
 			return &FailoverError{

pkg/providers/error_classifier_test.go

@@ -207,6 +207,59 @@ func TestClassifyError_FormatPatterns(t *testing.T) {
 	}
 }
 
+func TestClassifyError_ModelInvalidPatterns(t *testing.T) {
+	patterns := []string{
+		"nemotron-3-nano-30b-a3b:free is not a valid model ID",
+		"model not found",
+		"model_not_found: the requested model does not exist",
+		"model not available in this region",
+		"the model does not exist or you do not have access",
+		"no such model: gpt-5-turbo",
+		"invalid model specified",
+		"model llama-3-8b is not supported",
+		"model gpt-4o-mini is unavailable",
+		"model codellama is deprecated",
+	}
+
+	for _, msg := range patterns {
+		err := errors.New(msg)
+		result := ClassifyError(err, "nvidia", "test-model")
+		if result == nil {
+			t.Errorf("pattern %q: expected non-nil", msg)
+			continue
+		}
+		if result.Reason != FailoverModelInvalid {
+			t.Errorf("pattern %q: reason = %q, want model_invalid", msg, result.Reason)
+		}
+		if !result.IsRetriable() {
+			t.Errorf("pattern %q: should be retriable to allow fallback", msg)
+		}
+		if !result.IsModelInvalid() {
+			t.Errorf("pattern %q: should be classified as model invalid", msg)
+		}
+	}
+}
+
+func TestClassifyError_ModelInvalid_OverridesStatus400(t *testing.T) {
+	// This is the exact production error: status 400 + "not a valid model".
+	// Before the fix, status 400 was classified as FailoverFormat (non-retriable),
+	// which prevented fallback to other models.
+	err := fmt.Errorf("API request failed:\n  Status: 400\n  Body:   {\"error\":{\"message\":\"nemotron-3-nano-30b-a3b:free is not a valid model ID\",\"code\":400}}")
+	result := ClassifyError(err, "nvidia", "nemotron-3-nano-30b-a3b:free")
+	if result == nil {
+		t.Fatal("expected non-nil for model-invalid 400 error")
+	}
+	if result.Reason != FailoverModelInvalid {
+		t.Errorf("reason = %q, want model_invalid (should override status 400)", result.Reason)
+	}
+	if !result.IsRetriable() {
+		t.Error("model-invalid error should be retriable to allow fallback to next model")
+	}
+	if !result.IsModelInvalid() {
+		t.Error("should be classified as model invalid")
+	}
+}
+
 func TestClassifyError_ImageDimensionError(t *testing.T) {
 	err := errors.New("image dimensions exceed max allowed 2048x2048")
 	result := ClassifyError(err, "openai", "gpt-4o")
@@ -264,6 +317,7 @@ func TestFailoverError_IsRetriable(t *testing.T) {
 		{FailoverBilling, true},
 		{FailoverTimeout, true},
 		{FailoverOverloaded, true},
+		{FailoverModelInvalid, true},
 		{FailoverFormat, false},
 		{FailoverUnknown, true},
 	}

pkg/providers/types.go

@@ -39,13 +39,14 @@ type StatefulProvider interface {
 type FailoverReason string
 
 const (
-	FailoverAuth       FailoverReason = "auth"
-	FailoverRateLimit  FailoverReason = "rate_limit"
-	FailoverBilling    FailoverReason = "billing"
-	FailoverTimeout    FailoverReason = "timeout"
-	FailoverFormat     FailoverReason = "format"
-	FailoverOverloaded FailoverReason = "overloaded"
-	FailoverUnknown    FailoverReason = "unknown"
+	FailoverAuth         FailoverReason = "auth"
+	FailoverRateLimit    FailoverReason = "rate_limit"
+	FailoverBilling      FailoverReason = "billing"
+	FailoverTimeout      FailoverReason = "timeout"
+	FailoverFormat       FailoverReason = "format"
+	FailoverOverloaded   FailoverReason = "overloaded"
+	FailoverModelInvalid FailoverReason = "model_invalid"
+	FailoverUnknown      FailoverReason = "unknown"
 )
 
 // FailoverError wraps an LLM provider error with classification metadata.
@@ -72,6 +73,11 @@ func (e *FailoverError) IsRetriable() bool {
 	return e.Reason != FailoverFormat
 }
 
+// IsModelInvalid returns true if this error is due to an invalid/unavailable model.
+func (e *FailoverError) IsModelInvalid() bool {
+	return e.Reason == FailoverModelInvalid
+}
+
 // ModelConfig holds primary model and fallback list.
 type ModelConfig struct {
 	Primary   string