Merge pull request #1099 from router-for-me/claude

refactor(claude): move max_tokens constraint enforcement to Apply method

internal/registry/model_definitions.go

@@ -765,21 +765,23 @@ func GetIFlowModels() []*ModelInfo {
 type AntigravityModelConfig struct {
 	Thinking            *ThinkingSupport
 	MaxCompletionTokens int
-	Name                string
 }
 
 // GetAntigravityModelConfig returns static configuration for antigravity models.
 // Keys use upstream model names returned by the Antigravity models endpoint.
 func GetAntigravityModelConfig() map[string]*AntigravityModelConfig {
 	return map[string]*AntigravityModelConfig{
-		"gemini-2.5-flash":           {Thinking: &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true}, Name: "models/gemini-2.5-flash"},
-		"gemini-2.5-flash-lite":      {Thinking: &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true}, Name: "models/gemini-2.5-flash-lite"},
-		"rev19-uic3-1p":              {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true}, Name: "models/rev19-uic3-1p"},
-		"gemini-3-pro-high":          {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}, Name: "models/gemini-3-pro-high"},
-		"gemini-3-pro-image":         {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}, Name: "models/gemini-3-pro-image"},
-		"gemini-3-flash":             {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}}, Name: "models/gemini-3-flash"},
+		"gemini-2.5-flash":           {Thinking: &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true}},
+		"gemini-2.5-flash-lite":      {Thinking: &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true}},
+		"rev19-uic3-1p":              {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true}},
+		"gemini-3-pro-high":          {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}},
+		"gemini-3-pro-image":         {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}},
+		"gemini-3-flash":             {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}}},
 		"claude-sonnet-4-5-thinking": {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
 		"claude-opus-4-5-thinking":   {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
+		"claude-sonnet-4-5":          {MaxCompletionTokens: 64000},
+		"gpt-oss-120b-medium":        {},
+		"tab_flash_lite_preview":     {},
 	}
 }
 
@@ -809,10 +811,9 @@ func LookupStaticModelInfo(modelID string) *ModelInfo {
 	}
 
 	// Check Antigravity static config
-	if cfg := GetAntigravityModelConfig()[modelID]; cfg != nil && cfg.Thinking != nil {
+	if cfg := GetAntigravityModelConfig()[modelID]; cfg != nil {
 		return &ModelInfo{
 			ID:                  modelID,
-			Name:                cfg.Name,
 			Thinking:            cfg.Thinking,
 			MaxCompletionTokens: cfg.MaxCompletionTokens,
 		}

internal/runtime/executor/antigravity_executor.go

@@ -1005,9 +1005,6 @@ func FetchAntigravityModels(ctx context.Context, auth *cliproxyauth.Auth, cfg *c
 			}
 			modelCfg := modelConfig[modelID]
 			modelName := modelID
-			if modelCfg != nil && modelCfg.Name != "" {
-				modelName = modelCfg.Name
-			}
 			modelInfo := &registry.ModelInfo{
 				ID:          modelID,
 				Name:        modelName,
@@ -1410,13 +1407,6 @@ func geminiToAntigravity(modelName string, payload []byte, projectID string) []b
 	template, _ = sjson.Delete(template, "request.safetySettings")
 	template, _ = sjson.Set(template, "request.toolConfig.functionCallingConfig.mode", "VALIDATED")
 
-	if !strings.HasPrefix(modelName, "gemini-3-") {
-		if thinkingLevel := gjson.Get(template, "request.generationConfig.thinkingConfig.thinkingLevel"); thinkingLevel.Exists() {
-			template, _ = sjson.Delete(template, "request.generationConfig.thinkingConfig.thinkingLevel")
-			template, _ = sjson.Set(template, "request.generationConfig.thinkingConfig.thinkingBudget", -1)
-		}
-	}
-
 	if strings.Contains(modelName, "claude") {
 		gjson.Get(template, "request.tools").ForEach(func(key, tool gjson.Result) bool {
 			tool.Get("functionDeclarations").ForEach(func(funKey, funcDecl gjson.Result) bool {

internal/runtime/executor/claude_executor.go

@@ -17,7 +17,6 @@ import (
 	claudeauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/claude"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
@@ -119,9 +118,6 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	// Disable thinking if tool_choice forces tool use (Anthropic API constraint)
 	body = disableThinkingIfToolChoiceForced(body)
 
-	// Ensure max_tokens > thinking.budget_tokens when thinking is enabled
-	body = ensureMaxTokensForThinking(baseModel, body)
-
 	// Extract betas from body and convert to header
 	var extraBetas []string
 	extraBetas, body = extractAndRemoveBetas(body)
@@ -250,9 +246,6 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	// Disable thinking if tool_choice forces tool use (Anthropic API constraint)
 	body = disableThinkingIfToolChoiceForced(body)
 
-	// Ensure max_tokens > thinking.budget_tokens when thinking is enabled
-	body = ensureMaxTokensForThinking(baseModel, body)
-
 	// Extract betas from body and convert to header
 	var extraBetas []string
 	extraBetas, body = extractAndRemoveBetas(body)
@@ -541,81 +534,6 @@ func disableThinkingIfToolChoiceForced(body []byte) []byte {
 	return body
 }
 
-// ensureMaxTokensForThinking ensures max_tokens > thinking.budget_tokens when thinking is enabled.
-// Anthropic API requires this constraint; violating it returns a 400 error.
-// This function should be called after all thinking configuration is finalized.
-// It looks up the model's MaxCompletionTokens from the registry to use as the cap.
-func ensureMaxTokensForThinking(modelName string, body []byte) []byte {
-	thinkingType := gjson.GetBytes(body, "thinking.type").String()
-	if thinkingType != "enabled" {
-		return body
-	}
-
-	budgetTokens := gjson.GetBytes(body, "thinking.budget_tokens").Int()
-	if budgetTokens <= 0 {
-		return body
-	}
-
-	maxTokens := gjson.GetBytes(body, "max_tokens").Int()
-
-	// Look up the model's max completion tokens from the registry
-	maxCompletionTokens := 0
-	if modelInfo := registry.LookupModelInfo(modelName); modelInfo != nil {
-		maxCompletionTokens = modelInfo.MaxCompletionTokens
-	}
-
-	// Fall back to budget + buffer if registry lookup fails or returns 0
-	const fallbackBuffer = 4000
-	requiredMaxTokens := budgetTokens + fallbackBuffer
-	if maxCompletionTokens > 0 {
-		requiredMaxTokens = int64(maxCompletionTokens)
-	}
-
-	if maxTokens < requiredMaxTokens {
-		body, _ = sjson.SetBytes(body, "max_tokens", requiredMaxTokens)
-	}
-	return body
-}
-
-func (e *ClaudeExecutor) resolveClaudeConfig(auth *cliproxyauth.Auth) *config.ClaudeKey {
-	if auth == nil || e.cfg == nil {
-		return nil
-	}
-	var attrKey, attrBase string
-	if auth.Attributes != nil {
-		attrKey = strings.TrimSpace(auth.Attributes["api_key"])
-		attrBase = strings.TrimSpace(auth.Attributes["base_url"])
-	}
-	for i := range e.cfg.ClaudeKey {
-		entry := &e.cfg.ClaudeKey[i]
-		cfgKey := strings.TrimSpace(entry.APIKey)
-		cfgBase := strings.TrimSpace(entry.BaseURL)
-		if attrKey != "" && attrBase != "" {
-			if strings.EqualFold(cfgKey, attrKey) && strings.EqualFold(cfgBase, attrBase) {
-				return entry
-			}
-			continue
-		}
-		if attrKey != "" && strings.EqualFold(cfgKey, attrKey) {
-			if cfgBase == "" || strings.EqualFold(cfgBase, attrBase) {
-				return entry
-			}
-		}
-		if attrKey == "" && attrBase != "" && strings.EqualFold(cfgBase, attrBase) {
-			return entry
-		}
-	}
-	if attrKey != "" {
-		for i := range e.cfg.ClaudeKey {
-			entry := &e.cfg.ClaudeKey[i]
-			if strings.EqualFold(strings.TrimSpace(entry.APIKey), attrKey) {
-				return entry
-			}
-		}
-	}
-	return nil
-}
-
 type compositeReadCloser struct {
 	io.Reader
 	closers []func() error

internal/thinking/provider/claude/apply.go

@@ -80,9 +80,66 @@ func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *
 
 	result, _ := sjson.SetBytes(body, "thinking.type", "enabled")
 	result, _ = sjson.SetBytes(result, "thinking.budget_tokens", config.Budget)
+
+	// Ensure max_tokens > thinking.budget_tokens (Anthropic API constraint)
+	result = a.normalizeClaudeBudget(result, config.Budget, modelInfo)
 	return result, nil
 }
 
+// normalizeClaudeBudget applies Claude-specific constraints to ensure max_tokens > budget_tokens.
+// Anthropic API requires this constraint; violating it returns a 400 error.
+func (a *Applier) normalizeClaudeBudget(body []byte, budgetTokens int, modelInfo *registry.ModelInfo) []byte {
+	if budgetTokens <= 0 {
+		return body
+	}
+
+	// Ensure the request satisfies Claude constraints:
+	//  1) Determine effective max_tokens (request overrides model default)
+	//  2) If budget_tokens >= max_tokens, reduce budget_tokens to max_tokens-1
+	//  3) If the adjusted budget falls below the model minimum, leave the request unchanged
+	//  4) If max_tokens came from model default, write it back into the request
+
+	effectiveMax, setDefaultMax := a.effectiveMaxTokens(body, modelInfo)
+	if setDefaultMax && effectiveMax > 0 {
+		body, _ = sjson.SetBytes(body, "max_tokens", effectiveMax)
+	}
+
+	// Compute the budget we would apply after enforcing budget_tokens < max_tokens.
+	adjustedBudget := budgetTokens
+	if effectiveMax > 0 && adjustedBudget >= effectiveMax {
+		adjustedBudget = effectiveMax - 1
+	}
+
+	minBudget := 0
+	if modelInfo != nil && modelInfo.Thinking != nil {
+		minBudget = modelInfo.Thinking.Min
+	}
+	if minBudget > 0 && adjustedBudget > 0 && adjustedBudget < minBudget {
+		// If enforcing the max_tokens constraint would push the budget below the model minimum,
+		// leave the request unchanged.
+		return body
+	}
+
+	if adjustedBudget != budgetTokens {
+		body, _ = sjson.SetBytes(body, "thinking.budget_tokens", adjustedBudget)
+	}
+
+	return body
+}
+
+// effectiveMaxTokens returns the max tokens to cap thinking:
+// prefer request-provided max_tokens; otherwise fall back to model default.
+// The boolean indicates whether the value came from the model default (and thus should be written back).
+func (a *Applier) effectiveMaxTokens(body []byte, modelInfo *registry.ModelInfo) (max int, fromModel bool) {
+	if maxTok := gjson.GetBytes(body, "max_tokens"); maxTok.Exists() && maxTok.Int() > 0 {
+		return int(maxTok.Int()), false
+	}
+	if modelInfo != nil && modelInfo.MaxCompletionTokens > 0 {
+		return modelInfo.MaxCompletionTokens, true
+	}
+	return 0, false
+}
+
 func applyCompatibleClaude(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
 	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeNone && config.Mode != thinking.ModeAuto {
 		return body, nil

internal/translator/codex/claude/codex_claude_response.go

@@ -117,8 +117,12 @@ func ConvertCodexResponseToClaude(_ context.Context, _ string, originalRequestRa
 		} else {
 			template, _ = sjson.Set(template, "delta.stop_reason", "end_turn")
 		}
-		template, _ = sjson.Set(template, "usage.input_tokens", rootResult.Get("response.usage.input_tokens").Int())
-		template, _ = sjson.Set(template, "usage.output_tokens", rootResult.Get("response.usage.output_tokens").Int())
+		inputTokens, outputTokens, cachedTokens := extractResponsesUsage(rootResult.Get("response.usage"))
+		template, _ = sjson.Set(template, "usage.input_tokens", inputTokens)
+		template, _ = sjson.Set(template, "usage.output_tokens", outputTokens)
+		if cachedTokens > 0 {
+			template, _ = sjson.Set(template, "usage.cache_read_input_tokens", cachedTokens)
+		}
 
 		output = "event: message_delta\n"
 		output += fmt.Sprintf("data: %s\n\n", template)
@@ -204,8 +208,12 @@ func ConvertCodexResponseToClaudeNonStream(_ context.Context, _ string, original
 	out := `{"id":"","type":"message","role":"assistant","model":"","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":0,"output_tokens":0}}`
 	out, _ = sjson.Set(out, "id", responseData.Get("id").String())
 	out, _ = sjson.Set(out, "model", responseData.Get("model").String())
-	out, _ = sjson.Set(out, "usage.input_tokens", responseData.Get("usage.input_tokens").Int())
-	out, _ = sjson.Set(out, "usage.output_tokens", responseData.Get("usage.output_tokens").Int())
+	inputTokens, outputTokens, cachedTokens := extractResponsesUsage(responseData.Get("usage"))
+	out, _ = sjson.Set(out, "usage.input_tokens", inputTokens)
+	out, _ = sjson.Set(out, "usage.output_tokens", outputTokens)
+	if cachedTokens > 0 {
+		out, _ = sjson.Set(out, "usage.cache_read_input_tokens", cachedTokens)
+	}
 
 	hasToolCall := false
 
@@ -308,12 +316,27 @@ func ConvertCodexResponseToClaudeNonStream(_ context.Context, _ string, original
 		out, _ = sjson.SetRaw(out, "stop_sequence", stopSequence.Raw)
 	}
 
-	if responseData.Get("usage.input_tokens").Exists() || responseData.Get("usage.output_tokens").Exists() {
-		out, _ = sjson.Set(out, "usage.input_tokens", responseData.Get("usage.input_tokens").Int())
-		out, _ = sjson.Set(out, "usage.output_tokens", responseData.Get("usage.output_tokens").Int())
+	return out
+}
+
+func extractResponsesUsage(usage gjson.Result) (int64, int64, int64) {
+	if !usage.Exists() || usage.Type == gjson.Null {
+		return 0, 0, 0
 	}
 
-	return out
+	inputTokens := usage.Get("input_tokens").Int()
+	outputTokens := usage.Get("output_tokens").Int()
+	cachedTokens := usage.Get("input_tokens_details.cached_tokens").Int()
+
+	if cachedTokens > 0 {
+		if inputTokens >= cachedTokens {
+			inputTokens -= cachedTokens
+		} else {
+			inputTokens = 0
+		}
+	}
+
+	return inputTokens, outputTokens, cachedTokens
 }
 
 // buildReverseMapFromClaudeOriginalShortToOriginal builds a map[short]original from original Claude request tools.

internal/translator/openai/claude/openai_claude_request.go

@@ -88,7 +88,7 @@ func ConvertClaudeRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 	var messagesJSON = "[]"
 
 	// Handle system message first
-	systemMsgJSON := `{"role":"system","content":[{"type":"text","text":"Use ANY tool, the parameters MUST accord with RFC 8259 (The JavaScript Object Notation (JSON) Data Interchange Format), the keys and value MUST be enclosed in double quotes."}]}`
+	systemMsgJSON := `{"role":"system","content":[]}`
 	if system := root.Get("system"); system.Exists() {
 		if system.Type == gjson.String {
 			if system.String() != "" {

internal/translator/openai/claude/openai_claude_response.go

@@ -289,21 +289,17 @@ func convertOpenAIStreamingChunkToAnthropic(rawJSON []byte, param *ConvertOpenAI
 	// Only process if usage has actual values (not null)
 	if param.FinishReason != "" {
 		usage := root.Get("usage")
-		var inputTokens, outputTokens int64
+		var inputTokens, outputTokens, cachedTokens int64
 		if usage.Exists() && usage.Type != gjson.Null {
-			// Check if usage has actual token counts
-			promptTokens := usage.Get("prompt_tokens")
-			completionTokens := usage.Get("completion_tokens")
-
-			if promptTokens.Exists() && completionTokens.Exists() {
-				inputTokens = promptTokens.Int()
-				outputTokens = completionTokens.Int()
-			}
+			inputTokens, outputTokens, cachedTokens = extractOpenAIUsage(usage)
 			// Send message_delta with usage
 			messageDeltaJSON := `{"type":"message_delta","delta":{"stop_reason":"","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
 			messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "delta.stop_reason", mapOpenAIFinishReasonToAnthropic(param.FinishReason))
 			messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "usage.input_tokens", inputTokens)
 			messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "usage.output_tokens", outputTokens)
+			if cachedTokens > 0 {
+				messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "usage.cache_read_input_tokens", cachedTokens)
+			}
 			results = append(results, "event: message_delta\ndata: "+messageDeltaJSON+"\n\n")
 			param.MessageDeltaSent = true
 
@@ -423,13 +419,12 @@ func convertOpenAINonStreamingToAnthropic(rawJSON []byte) []string {
 
 	// Set usage information
 	if usage := root.Get("usage"); usage.Exists() {
-		out, _ = sjson.Set(out, "usage.input_tokens", usage.Get("prompt_tokens").Int())
-		out, _ = sjson.Set(out, "usage.output_tokens", usage.Get("completion_tokens").Int())
-		reasoningTokens := int64(0)
-		if v := usage.Get("completion_tokens_details.reasoning_tokens"); v.Exists() {
-			reasoningTokens = v.Int()
+		inputTokens, outputTokens, cachedTokens := extractOpenAIUsage(usage)
+		out, _ = sjson.Set(out, "usage.input_tokens", inputTokens)
+		out, _ = sjson.Set(out, "usage.output_tokens", outputTokens)
+		if cachedTokens > 0 {
+			out, _ = sjson.Set(out, "usage.cache_read_input_tokens", cachedTokens)
 		}
-		out, _ = sjson.Set(out, "usage.reasoning_tokens", reasoningTokens)
 	}
 
 	return []string{out}
@@ -674,8 +669,12 @@ func ConvertOpenAIResponseToClaudeNonStream(_ context.Context, _ string, origina
 	}
 
 	if respUsage := root.Get("usage"); respUsage.Exists() {
-		out, _ = sjson.Set(out, "usage.input_tokens", respUsage.Get("prompt_tokens").Int())
-		out, _ = sjson.Set(out, "usage.output_tokens", respUsage.Get("completion_tokens").Int())
+		inputTokens, outputTokens, cachedTokens := extractOpenAIUsage(respUsage)
+		out, _ = sjson.Set(out, "usage.input_tokens", inputTokens)
+		out, _ = sjson.Set(out, "usage.output_tokens", outputTokens)
+		if cachedTokens > 0 {
+			out, _ = sjson.Set(out, "usage.cache_read_input_tokens", cachedTokens)
+		}
 	}
 
 	if !stopReasonSet {
@@ -692,3 +691,23 @@ func ConvertOpenAIResponseToClaudeNonStream(_ context.Context, _ string, origina
 func ClaudeTokenCount(ctx context.Context, count int64) string {
 	return fmt.Sprintf(`{"input_tokens":%d}`, count)
 }
+
+func extractOpenAIUsage(usage gjson.Result) (int64, int64, int64) {
+	if !usage.Exists() || usage.Type == gjson.Null {
+		return 0, 0, 0
+	}
+
+	inputTokens := usage.Get("prompt_tokens").Int()
+	outputTokens := usage.Get("completion_tokens").Int()
+	cachedTokens := usage.Get("prompt_tokens_details.cached_tokens").Int()
+
+	if cachedTokens > 0 {
+		if inputTokens >= cachedTokens {
+			inputTokens -= cachedTokens
+		} else {
+			inputTokens = 0
+		}
+	}
+
+	return inputTokens, outputTokens, cachedTokens
+}

test/thinking_conversion_test.go

@@ -20,6 +20,7 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	sdktranslator "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
 	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
 )
 
 // thinkingTestCase represents a common test case structure for both suffix and body tests.
@@ -2707,6 +2708,9 @@ func runThinkingTests(t *testing.T, cases []thinkingTestCase) {
 				[]byte(tc.inputJSON),
 				true,
 			)
+			if applyTo == "claude" {
+				body, _ = sjson.SetBytes(body, "max_tokens", 200000)
+			}
 
 			body, err := thinking.ApplyThinking(body, tc.model, tc.from, applyTo)