Merge branch 'router-for-me:main' into main

feat(iflow): add model-specific thinking configs for GLM-4.7 and Mini…

.dockerignore

@@ -13,8 +13,6 @@ Dockerfile
 docs/*
 README.md
 README_CN.md
-MANAGEMENT_API.md
-MANAGEMENT_API_CN.md
 LICENSE
 
 # Runtime data folders (should be mounted as volumes)
@@ -32,3 +30,4 @@ bin/*
 .agent/*
 .bmad/*
 _bmad/*
+_bmad-output/*

.github/ISSUE_TEMPLATE/bug_report.md

@@ -7,6 +7,13 @@ assignees: ''
 
 ---
 
+**Is it a request payload issue?**
+[  ] Yes, this is a request payload issue. I am using a client/cURL to send a request payload, but I received an unexpected error.
+[  ] No, it's another issue.
+
+**If it's a request payload issue, you MUST know**
+Our team doesn't have any GODs or ORACLEs or MIND READERs. Please make sure to attach the request log or curl payload.
+
 **Describe the bug**
 A clear and concise description of what the bug is.
 

.gitignore

@@ -12,11 +12,15 @@ bin/*
 logs/*
 conv/*
 temp/*
+refs/*
+
+# Storage backends
 pgstore/*
 gitstore/*
 objectstore/*
+
+# Static assets
 static/*
-refs/*
 
 # Authentication data
 auths/*
@@ -36,6 +40,7 @@ GEMINI.md
 .agent/*
 .bmad/*
 _bmad/*
+_bmad-output/*
 .mcp/cache/
 
 # macOS

internal/config/config.go

@@ -879,8 +879,8 @@ func getOrCreateMapValue(mapNode *yaml.Node, key string) *yaml.Node {
 }
 
 // mergeMappingPreserve merges keys from src into dst mapping node while preserving
-// key order and comments of existing keys in dst. Unknown keys from src are appended
-// to dst at the end, copying their node structure from src.
+// key order and comments of existing keys in dst. New keys are only added if their
+// value is non-zero to avoid polluting the config with defaults.
 func mergeMappingPreserve(dst, src *yaml.Node) {
 	if dst == nil || src == nil {
 		return
@@ -891,20 +891,19 @@ func mergeMappingPreserve(dst, src *yaml.Node) {
 		copyNodeShallow(dst, src)
 		return
 	}
-	// Build a lookup of existing keys in dst
 	for i := 0; i+1 < len(src.Content); i += 2 {
 		sk := src.Content[i]
 		sv := src.Content[i+1]
 		idx := findMapKeyIndex(dst, sk.Value)
 		if idx >= 0 {
-			// Merge into existing value node
+			// Merge into existing value node (always update, even to zero values)
 			dv := dst.Content[idx+1]
 			mergeNodePreserve(dv, sv)
 		} else {
-			if shouldSkipEmptyCollectionOnPersist(sk.Value, sv) {
+			// New key: only add if value is non-zero to avoid polluting config with defaults
+			if isZeroValueNode(sv) {
 				continue
 			}
-			// Append new key/value pair by deep-copying from src
 			dst.Content = append(dst.Content, deepCopyNode(sk), deepCopyNode(sv))
 		}
 	}
@@ -987,32 +986,49 @@ func findMapKeyIndex(mapNode *yaml.Node, key string) int {
 	return -1
 }
 
-func shouldSkipEmptyCollectionOnPersist(key string, node *yaml.Node) bool {
-	switch key {
-	case "generative-language-api-key",
-		"gemini-api-key",
-		"vertex-api-key",
-		"claude-api-key",
-		"codex-api-key",
-		"openai-compatibility":
-		return isEmptyCollectionNode(node)
-	default:
-		return false
-	}
-}
-
-func isEmptyCollectionNode(node *yaml.Node) bool {
+// isZeroValueNode returns true if the YAML node represents a zero/default value
+// that should not be written as a new key to preserve config cleanliness.
+// For mappings and sequences, recursively checks if all children are zero values.
+func isZeroValueNode(node *yaml.Node) bool {
 	if node == nil {
 		return true
 	}
 	switch node.Kind {
-	case yaml.SequenceNode:
-		return len(node.Content) == 0
 	case yaml.ScalarNode:
-		return node.Tag == "!!null"
-	default:
-		return false
+		switch node.Tag {
+		case "!!bool":
+			return node.Value == "false"
+		case "!!int", "!!float":
+			return node.Value == "0" || node.Value == "0.0"
+		case "!!str":
+			return node.Value == ""
+		case "!!null":
+			return true
+		}
+	case yaml.SequenceNode:
+		if len(node.Content) == 0 {
+			return true
+		}
+		// Check if all elements are zero values
+		for _, child := range node.Content {
+			if !isZeroValueNode(child) {
+				return false
+			}
+		}
+		return true
+	case yaml.MappingNode:
+		if len(node.Content) == 0 {
+			return true
+		}
+		// Check if all values are zero values (values are at odd indices)
+		for i := 1; i < len(node.Content); i += 2 {
+			if !isZeroValueNode(node.Content[i]) {
+				return false
+			}
+		}
+		return true
 	}
+	return false
 }
 
 // deepCopyNode creates a deep copy of a yaml.Node graph.

internal/config/sdk_config.go

@@ -30,13 +30,13 @@ type SDKConfig struct {
 // StreamingConfig holds server streaming behavior configuration.
 type StreamingConfig struct {
 	// KeepAliveSeconds controls how often the server emits SSE heartbeats (": keep-alive\n\n").
-	// nil means default (15 seconds). <= 0 disables keep-alives.
-	KeepAliveSeconds *int `yaml:"keepalive-seconds,omitempty" json:"keepalive-seconds,omitempty"`
+	// <= 0 disables keep-alives. Default is 0.
+	KeepAliveSeconds int `yaml:"keepalive-seconds,omitempty" json:"keepalive-seconds,omitempty"`
 
 	// BootstrapRetries controls how many times the server may retry a streaming request before any bytes are sent,
 	// to allow auth rotation / transient recovery.
-	// nil means default (2). 0 disables bootstrap retries.
-	BootstrapRetries *int `yaml:"bootstrap-retries,omitempty" json:"bootstrap-retries,omitempty"`
+	// <= 0 disables bootstrap retries. Default is 0.
+	BootstrapRetries int `yaml:"bootstrap-retries,omitempty" json:"bootstrap-retries,omitempty"`
 }
 
 // AccessConfig groups request authentication providers.

internal/registry/model_definitions.go

@@ -741,7 +741,7 @@ func GetIFlowModels() []*ModelInfo {
 		{ID: "qwen3-235b-a22b-instruct", DisplayName: "Qwen3-235B-A22B-Instruct", Description: "Qwen3 235B A22B Instruct", Created: 1753401600},
 		{ID: "qwen3-235b", DisplayName: "Qwen3-235B-A22B", Description: "Qwen3 235B A22B", Created: 1753401600},
 		{ID: "minimax-m2", DisplayName: "MiniMax-M2", Description: "MiniMax M2", Created: 1758672000},
-		{ID: "minimax-m2.1", DisplayName: "MiniMax-M2.1", Description: "MiniMax M2.1", Created: 1766448000},
+		{ID: "minimax-m2.1", DisplayName: "MiniMax-M2.1", Description: "MiniMax M2.1", Created: 1766448000, Thinking: iFlowThinkingSupport},
 	}
 	models := make([]*ModelInfo, 0, len(entries))
 	for _, entry := range entries {

internal/runtime/executor/iflow_executor.go

@@ -67,6 +67,7 @@ func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 		return resp, errValidate
 	}
 	body = applyIFlowThinkingConfig(body)
+	body = preserveReasoningContentInMessages(body)
 	body = applyPayloadConfig(e.cfg, req.Model, body)
 
 	endpoint := strings.TrimSuffix(baseURL, "/") + iflowDefaultEndpoint
@@ -159,6 +160,7 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 		return nil, errValidate
 	}
 	body = applyIFlowThinkingConfig(body)
+	body = preserveReasoningContentInMessages(body)
 	// Ensure tools array exists to avoid provider quirks similar to Qwen's behaviour.
 	toolsResult := gjson.GetBytes(body, "tools")
 	if toolsResult.Exists() && toolsResult.IsArray() && len(toolsResult.Array()) == 0 {
@@ -445,20 +447,98 @@ func ensureToolsArray(body []byte) []byte {
 	return updated
 }
 
-// applyIFlowThinkingConfig converts normalized reasoning_effort to iFlow chat_template_kwargs.enable_thinking.
-// This should be called after NormalizeThinkingConfig has processed the payload.
-// iFlow only supports boolean enable_thinking, so any non-"none" effort enables thinking.
-func applyIFlowThinkingConfig(body []byte) []byte {
-	effort := gjson.GetBytes(body, "reasoning_effort")
-	if !effort.Exists() {
+// preserveReasoningContentInMessages ensures reasoning_content from assistant messages in the
+// conversation history is preserved when sending to iFlow models that support thinking.
+// This is critical for multi-turn conversations where the model needs to see its previous
+// reasoning to maintain coherent thought chains across tool calls and conversation turns.
+//
+// For GLM-4.7 and MiniMax-M2.1, the full assistant response (including reasoning) must be
+// appended back into message history before the next call.
+func preserveReasoningContentInMessages(body []byte) []byte {
+	model := strings.ToLower(gjson.GetBytes(body, "model").String())
+
+	// Only apply to models that support thinking with history preservation
+	needsPreservation := strings.HasPrefix(model, "glm-4.7") ||
+		strings.HasPrefix(model, "glm-4-7") ||
+		strings.HasPrefix(model, "minimax-m2.1") ||
+		strings.HasPrefix(model, "minimax-m2-1")
+
+	if !needsPreservation {
 		return body
 	}
 
-	val := strings.ToLower(strings.TrimSpace(effort.String()))
-	enableThinking := val != "none" && val != ""
+	messages := gjson.GetBytes(body, "messages")
+	if !messages.Exists() || !messages.IsArray() {
+		return body
+	}
 
-	body, _ = sjson.DeleteBytes(body, "reasoning_effort")
-	body, _ = sjson.SetBytes(body, "chat_template_kwargs.enable_thinking", enableThinking)
+	// Check if any assistant message already has reasoning_content preserved
+	hasReasoningContent := false
+	messages.ForEach(func(_, msg gjson.Result) bool {
+		role := msg.Get("role").String()
+		if role == "assistant" {
+			rc := msg.Get("reasoning_content")
+			if rc.Exists() && rc.String() != "" {
+				hasReasoningContent = true
+				return false // stop iteration
+			}
+		}
+		return true
+	})
+
+	// If reasoning content is already present, the messages are properly formatted
+	// No need to modify - the client has correctly preserved reasoning in history
+	if hasReasoningContent {
+		log.Debugf("iflow executor: reasoning_content found in message history for %s", model)
+	}
+
+	return body
+}
+
+// applyIFlowThinkingConfig converts normalized reasoning_effort to model-specific thinking configurations.
+// This should be called after NormalizeThinkingConfig has processed the payload.
+//
+// Model-specific handling:
+//   - GLM-4.7: Uses extra_body={"thinking": {"type": "enabled"}, "clear_thinking": false}
+//   - MiniMax-M2.1: Uses reasoning_split=true for OpenAI-style reasoning separation
+//   - Other iFlow models: Uses chat_template_kwargs.enable_thinking (boolean)
+func applyIFlowThinkingConfig(body []byte) []byte {
+	effort := gjson.GetBytes(body, "reasoning_effort")
+	model := strings.ToLower(gjson.GetBytes(body, "model").String())
+
+	// Check if thinking should be enabled
+	val := ""
+	if effort.Exists() {
+		val = strings.ToLower(strings.TrimSpace(effort.String()))
+	}
+	enableThinking := effort.Exists() && val != "none" && val != ""
+
+	// Remove reasoning_effort as we'll convert to model-specific format
+	if effort.Exists() {
+		body, _ = sjson.DeleteBytes(body, "reasoning_effort")
+	}
+
+	// GLM-4.7: Use extra_body with thinking config and clear_thinking: false
+	if strings.HasPrefix(model, "glm-4.7") || strings.HasPrefix(model, "glm-4-7") {
+		if enableThinking {
+			body, _ = sjson.SetBytes(body, "extra_body.thinking.type", "enabled")
+			body, _ = sjson.SetBytes(body, "extra_body.clear_thinking", false)
+		}
+		return body
+	}
+
+	// MiniMax-M2.1: Use reasoning_split=true for interleaved thinking
+	if strings.HasPrefix(model, "minimax-m2.1") || strings.HasPrefix(model, "minimax-m2-1") {
+		if enableThinking {
+			body, _ = sjson.SetBytes(body, "reasoning_split", true)
+		}
+		return body
+	}
+
+	// Other iFlow models (including GLM-4.6): Use chat_template_kwargs.enable_thinking
+	if effort.Exists() {
+		body, _ = sjson.SetBytes(body, "chat_template_kwargs.enable_thinking", enableThinking)
+	}
 
 	return body
 }

internal/runtime/executor/usage_helpers.go

@@ -482,12 +482,16 @@ func StripUsageMetadataFromJSON(rawJSON []byte) ([]byte, bool) {
 	cleaned := jsonBytes
 	var changed bool
 
-	if gjson.GetBytes(cleaned, "usageMetadata").Exists() {
+	if usageMetadata = gjson.GetBytes(cleaned, "usageMetadata"); usageMetadata.Exists() {
+		// Rename usageMetadata to cpaUsageMetadata in the message_start event of Claude
+		cleaned, _ = sjson.SetRawBytes(cleaned, "cpaUsageMetadata", []byte(usageMetadata.Raw))
 		cleaned, _ = sjson.DeleteBytes(cleaned, "usageMetadata")
 		changed = true
 	}
 
-	if gjson.GetBytes(cleaned, "response.usageMetadata").Exists() {
+	if usageMetadata = gjson.GetBytes(cleaned, "response.usageMetadata"); usageMetadata.Exists() {
+		// Rename usageMetadata to cpaUsageMetadata in the message_start event of Claude
+		cleaned, _ = sjson.SetRawBytes(cleaned, "response.cpaUsageMetadata", []byte(usageMetadata.Raw))
 		cleaned, _ = sjson.DeleteBytes(cleaned, "response.usageMetadata")
 		changed = true
 	}

internal/translator/antigravity/claude/antigravity_claude_response.go

@@ -99,6 +99,14 @@ func ConvertAntigravityResponseToClaude(_ context.Context, _ string, originalReq
 		// This follows the Claude Code API specification for streaming message initialization
 		messageStartTemplate := `{"type": "message_start", "message": {"id": "msg_1nZdL29xx5MUA1yADyHTEsnR8uuvGzszyY", "type": "message", "role": "assistant", "content": [], "model": "claude-3-5-sonnet-20241022", "stop_reason": null, "stop_sequence": null, "usage": {"input_tokens": 0, "output_tokens": 0}}}`
 
+		// Use cpaUsageMetadata within the message_start event for Claude.
+		if promptTokenCount := gjson.GetBytes(rawJSON, "response.cpaUsageMetadata.promptTokenCount"); promptTokenCount.Exists() {
+			messageStartTemplate, _ = sjson.Set(messageStartTemplate, "message.usage.input_tokens", promptTokenCount.Int())
+		}
+		if candidatesTokenCount := gjson.GetBytes(rawJSON, "response.cpaUsageMetadata.candidatesTokenCount"); candidatesTokenCount.Exists() {
+			messageStartTemplate, _ = sjson.Set(messageStartTemplate, "message.usage.output_tokens", candidatesTokenCount.Int())
+		}
+
 		// Override default values with actual response metadata if available from the Gemini CLI response
 		if modelVersionResult := gjson.GetBytes(rawJSON, "response.modelVersion"); modelVersionResult.Exists() {
 			messageStartTemplate, _ = sjson.Set(messageStartTemplate, "message.model", modelVersionResult.String())

internal/translator/antigravity/openai/chat-completions/antigravity_openai_request.go

@@ -247,7 +247,7 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 			} else if role == "assistant" {
 				node := []byte(`{"role":"model","parts":[]}`)
 				p := 0
-				if content.Type == gjson.String {
+				if content.Type == gjson.String && content.String() != "" {
 					node, _ = sjson.SetBytes(node, "parts.-1.text", content.String())
 					p++
 				} else if content.IsArray() {

internal/translator/claude/openai/chat-completions/claude_openai_response.go

@@ -209,9 +209,12 @@ func ConvertClaudeResponseToOpenAI(_ context.Context, modelName string, original
 		if usage := root.Get("usage"); usage.Exists() {
 			inputTokens := usage.Get("input_tokens").Int()
 			outputTokens := usage.Get("output_tokens").Int()
-			template, _ = sjson.Set(template, "usage.prompt_tokens", inputTokens)
+			cacheReadInputTokens := usage.Get("cache_read_input_tokens").Int()
+			cacheCreationInputTokens := usage.Get("cache_creation_input_tokens").Int()
+			template, _ = sjson.Set(template, "usage.prompt_tokens", inputTokens+cacheCreationInputTokens)
 			template, _ = sjson.Set(template, "usage.completion_tokens", outputTokens)
 			template, _ = sjson.Set(template, "usage.total_tokens", inputTokens+outputTokens)
+			template, _ = sjson.Set(template, "usage.prompt_tokens_details.cached_tokens", cacheReadInputTokens)
 		}
 		return []string{template}
 
@@ -285,8 +288,6 @@ func ConvertClaudeResponseToOpenAINonStream(_ context.Context, _ string, origina
 	var messageID string
 	var model string
 	var createdAt int64
-	var inputTokens, outputTokens int64
-	var reasoningTokens int64
 	var stopReason string
 	var contentParts []string
 	var reasoningParts []string
@@ -303,9 +304,6 @@ func ConvertClaudeResponseToOpenAINonStream(_ context.Context, _ string, origina
 				messageID = message.Get("id").String()
 				model = message.Get("model").String()
 				createdAt = time.Now().Unix()
-				if usage := message.Get("usage"); usage.Exists() {
-					inputTokens = usage.Get("input_tokens").Int()
-				}
 			}
 
 		case "content_block_start":
@@ -368,11 +366,14 @@ func ConvertClaudeResponseToOpenAINonStream(_ context.Context, _ string, origina
 				}
 			}
 			if usage := root.Get("usage"); usage.Exists() {
-				outputTokens = usage.Get("output_tokens").Int()
-				// Estimate reasoning tokens from accumulated thinking content
-				if len(reasoningParts) > 0 {
-					reasoningTokens = int64(len(strings.Join(reasoningParts, "")) / 4) // Rough estimation
-				}
+				inputTokens := usage.Get("input_tokens").Int()
+				outputTokens := usage.Get("output_tokens").Int()
+				cacheReadInputTokens := usage.Get("cache_read_input_tokens").Int()
+				cacheCreationInputTokens := usage.Get("cache_creation_input_tokens").Int()
+				out, _ = sjson.Set(out, "usage.prompt_tokens", inputTokens+cacheCreationInputTokens)
+				out, _ = sjson.Set(out, "usage.completion_tokens", outputTokens)
+				out, _ = sjson.Set(out, "usage.total_tokens", inputTokens+outputTokens)
+				out, _ = sjson.Set(out, "usage.prompt_tokens_details.cached_tokens", cacheReadInputTokens)
 			}
 		}
 	}
@@ -431,16 +432,5 @@ func ConvertClaudeResponseToOpenAINonStream(_ context.Context, _ string, origina
 		out, _ = sjson.Set(out, "choices.0.finish_reason", mapAnthropicStopReasonToOpenAI(stopReason))
 	}
 
-	// Set usage information including prompt tokens, completion tokens, and total tokens
-	totalTokens := inputTokens + outputTokens
-	out, _ = sjson.Set(out, "usage.prompt_tokens", inputTokens)
-	out, _ = sjson.Set(out, "usage.completion_tokens", outputTokens)
-	out, _ = sjson.Set(out, "usage.total_tokens", totalTokens)
-
-	// Add reasoning tokens to usage details if any reasoning content was processed
-	if reasoningTokens > 0 {
-		out, _ = sjson.Set(out, "usage.completion_tokens_details.reasoning_tokens", reasoningTokens)
-	}
-
 	return out
 }

internal/translator/claude/openai/responses/claude_openai-responses_request.go

@@ -114,13 +114,16 @@ func ConvertOpenAIResponsesRequestToClaude(modelName string, inputRawJSON []byte
 					var builder strings.Builder
 					if parts := item.Get("content"); parts.Exists() && parts.IsArray() {
 						parts.ForEach(func(_, part gjson.Result) bool {
-							text := part.Get("text").String()
+							textResult := part.Get("text")
+							text := textResult.String()
 							if builder.Len() > 0 && text != "" {
 								builder.WriteByte('\n')
 							}
 							builder.WriteString(text)
 							return true
 						})
+					} else if parts.Type == gjson.String {
+						builder.WriteString(parts.String())
 					}
 					instructionsText = builder.String()
 					if instructionsText != "" {
@@ -207,6 +210,8 @@ func ConvertOpenAIResponsesRequestToClaude(modelName string, inputRawJSON []byte
 						}
 						return true
 					})
+				} else if parts.Type == gjson.String {
+					textAggregate.WriteString(parts.String())
 				}
 
 				// Fallback to given role if content types not decisive

sdk/api/handlers/handlers.go

@@ -104,8 +104,8 @@ func BuildErrorResponseBody(status int, errText string) []byte {
 // Returning 0 disables keep-alives (default when unset).
 func StreamingKeepAliveInterval(cfg *config.SDKConfig) time.Duration {
 	seconds := defaultStreamingKeepAliveSeconds
-	if cfg != nil && cfg.Streaming.KeepAliveSeconds != nil {
-		seconds = *cfg.Streaming.KeepAliveSeconds
+	if cfg != nil {
+		seconds = cfg.Streaming.KeepAliveSeconds
 	}
 	if seconds <= 0 {
 		return 0
@@ -116,8 +116,8 @@ func StreamingKeepAliveInterval(cfg *config.SDKConfig) time.Duration {
 // StreamingBootstrapRetries returns how many times a streaming request may be retried before any bytes are sent.
 func StreamingBootstrapRetries(cfg *config.SDKConfig) int {
 	retries := defaultStreamingBootstrapRetries
-	if cfg != nil && cfg.Streaming.BootstrapRetries != nil {
-		retries = *cfg.Streaming.BootstrapRetries
+	if cfg != nil {
+		retries = cfg.Streaming.BootstrapRetries
 	}
 	if retries < 0 {
 		retries = 0

sdk/api/handlers/handlers_stream_bootstrap_test.go

@@ -94,10 +94,9 @@ func TestExecuteStreamWithAuthManager_RetriesBeforeFirstByte(t *testing.T) {
 		registry.GetGlobalRegistry().UnregisterClient(auth2.ID)
 	})
 
-	bootstrapRetries := 1
 	handler := NewBaseAPIHandlers(&sdkconfig.SDKConfig{
 		Streaming: sdkconfig.StreamingConfig{
-			BootstrapRetries: &bootstrapRetries,
+			BootstrapRetries: 1,
 		},
 	}, manager)
 	dataChan, errChan := handler.ExecuteStreamWithAuthManager(context.Background(), "openai", "test-model", []byte(`{"model":"test-model"}`), "")

sdk/cliproxy/auth/conductor.go

@@ -263,7 +263,6 @@ func (m *Manager) Execute(ctx context.Context, providers []string, req cliproxye
 		return cliproxyexecutor.Response{}, &Error{Code: "provider_not_found", Message: "no provider supplied"}
 	}
 	rotated := m.rotateProviders(req.Model, normalized)
-	defer m.advanceProviderCursor(req.Model, normalized)
 
 	retryTimes, maxWait := m.retrySettings()
 	attempts := retryTimes + 1
@@ -302,7 +301,6 @@ func (m *Manager) ExecuteCount(ctx context.Context, providers []string, req clip
 		return cliproxyexecutor.Response{}, &Error{Code: "provider_not_found", Message: "no provider supplied"}
 	}
 	rotated := m.rotateProviders(req.Model, normalized)
-	defer m.advanceProviderCursor(req.Model, normalized)
 
 	retryTimes, maxWait := m.retrySettings()
 	attempts := retryTimes + 1
@@ -341,7 +339,6 @@ func (m *Manager) ExecuteStream(ctx context.Context, providers []string, req cli
 		return nil, &Error{Code: "provider_not_found", Message: "no provider supplied"}
 	}
 	rotated := m.rotateProviders(req.Model, normalized)
-	defer m.advanceProviderCursor(req.Model, normalized)
 
 	retryTimes, maxWait := m.retrySettings()
 	attempts := retryTimes + 1
@@ -640,13 +637,20 @@ func (m *Manager) normalizeProviders(providers []string) []string {
 	return result
 }
 
+// rotateProviders returns a rotated view of the providers list starting from the
+// current offset for the model, and atomically increments the offset for the next call.
+// This ensures concurrent requests get different starting providers.
 func (m *Manager) rotateProviders(model string, providers []string) []string {
 	if len(providers) == 0 {
 		return nil
 	}
-	m.mu.RLock()
+
+	// Atomic read-and-increment: get current offset and advance cursor in one lock
+	m.mu.Lock()
 	offset := m.providerOffsets[model]
-	m.mu.RUnlock()
+	m.providerOffsets[model] = (offset + 1) % len(providers)
+	m.mu.Unlock()
+
 	if len(providers) > 0 {
 		offset %= len(providers)
 	}
@@ -662,19 +666,6 @@ func (m *Manager) rotateProviders(model string, providers []string) []string {
 	return rotated
 }
 
-func (m *Manager) advanceProviderCursor(model string, providers []string) {
-	if len(providers) == 0 {
-		m.mu.Lock()
-		delete(m.providerOffsets, model)
-		m.mu.Unlock()
-		return
-	}
-	m.mu.Lock()
-	current := m.providerOffsets[model]
-	m.providerOffsets[model] = (current + 1) % len(providers)
-	m.mu.Unlock()
-}
-
 func (m *Manager) retrySettings() (int, time.Duration) {
 	if m == nil {
 		return 0, 0