feat(util): implement dynamic thinking suffix normalization and refactor budget resolution logic

- Added support for parsing and normalizing dynamic thinking model suffixes. - Centralized budget resolution across executors and payload helpers. - Retired legacy Gemini-specific thinking handlers in favor of unified logic. - Updated executors to use metadata-based thinking configuration. - Added `ResolveOriginalModel` utility for resolving normalized upstream models using request metadata. - Updated executors (Gemini, Codex, iFlow, OpenAI, Qwen) to incorporate upstream model resolution and substitute model values in payloads and request URLs. - Ensured fallbacks handle cases with missing or malformed metadata to derive models robustly. - Refactored upstream model resolution to dynamically incorporate metadata for selecting and normalizing models. - Improved handling of thinking configurations and model overrides in executors. - Removed hardcoded thinking model entries and migrated logic to metadata-based resolution. - Updated payload mutations to always include the resolved model.
2026-02-03 13:00:52 +08:00 · 2025-12-11 01:16:56 +08:00
parent e717939edb
commit 423ce97665
13 changed files with 579 additions and 647 deletions
--- a/internal/runtime/executor/claude_executor.go
+++ b/internal/runtime/executor/claude_executor.go
@@ -54,15 +54,22 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	// Use streaming translation to preserve function calling, except for claude.
 	stream := from != to
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), stream)
-	modelForUpstream := req.Model
-	if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
-		body, _ = sjson.SetBytes(body, "model", modelOverride)
-		modelForUpstream = modelOverride
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+	if upstreamModel == "" {
+		upstreamModel = req.Model
 	}
-	// Inject thinking config based on model suffix for thinking variants
-	body = e.injectThinkingConfig(req.Model, body)
+	if modelOverride := e.resolveUpstreamModel(upstreamModel, auth); modelOverride != "" {
+		upstreamModel = modelOverride
+	} else if !strings.EqualFold(upstreamModel, req.Model) {
+		if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
+			upstreamModel = modelOverride
+		}
+	}
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	// Inject thinking config based on model metadata for thinking variants
+	body = e.injectThinkingConfig(req.Model, req.Metadata, body)

-	if !strings.HasPrefix(modelForUpstream, "claude-3-5-haiku") {
+	if !strings.HasPrefix(upstreamModel, "claude-3-5-haiku") {
 		body = checkSystemInstructions(body)
 	}
 	body = applyPayloadConfig(e.cfg, req.Model, body)
@@ -161,11 +168,20 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("claude")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
-	if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
-		body, _ = sjson.SetBytes(body, "model", modelOverride)
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+	if upstreamModel == "" {
+		upstreamModel = req.Model
 	}
-	// Inject thinking config based on model suffix for thinking variants
-	body = e.injectThinkingConfig(req.Model, body)
+	if modelOverride := e.resolveUpstreamModel(upstreamModel, auth); modelOverride != "" {
+		upstreamModel = modelOverride
+	} else if !strings.EqualFold(upstreamModel, req.Model) {
+		if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
+			upstreamModel = modelOverride
+		}
+	}
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	// Inject thinking config based on model metadata for thinking variants
+	body = e.injectThinkingConfig(req.Model, req.Metadata, body)
 	body = checkSystemInstructions(body)
 	body = applyPayloadConfig(e.cfg, req.Model, body)

@@ -295,13 +311,20 @@ func (e *ClaudeExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Aut
 	// Use streaming translation to preserve function calling, except for claude.
 	stream := from != to
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), stream)
-	modelForUpstream := req.Model
-	if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
-		body, _ = sjson.SetBytes(body, "model", modelOverride)
-		modelForUpstream = modelOverride
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+	if upstreamModel == "" {
+		upstreamModel = req.Model
 	}
+	if modelOverride := e.resolveUpstreamModel(upstreamModel, auth); modelOverride != "" {
+		upstreamModel = modelOverride
+	} else if !strings.EqualFold(upstreamModel, req.Model) {
+		if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
+			upstreamModel = modelOverride
+		}
+	}
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)

-	if !strings.HasPrefix(modelForUpstream, "claude-3-5-haiku") {
+	if !strings.HasPrefix(upstreamModel, "claude-3-5-haiku") {
 		body = checkSystemInstructions(body)
 	}

@@ -427,25 +450,15 @@ func extractAndRemoveBetas(body []byte) ([]string, []byte) {
 	return betas, body
 }

-// injectThinkingConfig adds thinking configuration based on model name suffix
-func (e *ClaudeExecutor) injectThinkingConfig(modelName string, body []byte) []byte {
+// injectThinkingConfig adds thinking configuration based on metadata or legacy suffixes.
+func (e *ClaudeExecutor) injectThinkingConfig(modelName string, metadata map[string]any, body []byte) []byte {
 	// Only inject if thinking config is not already present
 	if gjson.GetBytes(body, "thinking").Exists() {
 		return body
 	}

-	var budgetTokens int
-	switch {
-	case strings.HasSuffix(modelName, "-thinking-low"):
-		budgetTokens = 1024
-	case strings.HasSuffix(modelName, "-thinking-medium"):
-		budgetTokens = 8192
-	case strings.HasSuffix(modelName, "-thinking-high"):
-		budgetTokens = 24576
-	case strings.HasSuffix(modelName, "-thinking"):
-		// Default thinking without suffix uses medium budget
-		budgetTokens = 8192
-	default:
+	budgetTokens, ok := resolveClaudeThinkingBudget(modelName, metadata)
+	if !ok || budgetTokens <= 0 {
 		return body
 	}

@@ -454,6 +467,44 @@ func (e *ClaudeExecutor) injectThinkingConfig(modelName string, body []byte) []b
 	return body
 }

+func resolveClaudeThinkingBudget(modelName string, metadata map[string]any) (int, bool) {
+	budget, include, effort, matched := util.ThinkingFromMetadata(metadata)
+	if matched {
+		if include != nil && !*include {
+			return 0, false
+		}
+		if budget != nil {
+			normalized := util.NormalizeThinkingBudget(modelName, *budget)
+			if normalized > 0 {
+				return normalized, true
+			}
+			return 0, false
+		}
+		if effort != nil {
+			if derived, ok := util.ThinkingEffortToBudget(modelName, *effort); ok && derived > 0 {
+				return derived, true
+			}
+		}
+	}
+	return claudeBudgetFromSuffix(modelName)
+}
+
+func claudeBudgetFromSuffix(modelName string) (int, bool) {
+	lower := strings.ToLower(strings.TrimSpace(modelName))
+	switch {
+	case strings.HasSuffix(lower, "-thinking-low"):
+		return 1024, true
+	case strings.HasSuffix(lower, "-thinking-medium"):
+		return 8192, true
+	case strings.HasSuffix(lower, "-thinking-high"):
+		return 24576, true
+	case strings.HasSuffix(lower, "-thinking"):
+		return 8192, true
+	default:
+		return 0, false
+	}
+}
+
 // 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.
@@ -491,35 +542,45 @@ func ensureMaxTokensForThinking(modelName string, body []byte) []byte {
 }

 func (e *ClaudeExecutor) resolveUpstreamModel(alias string, auth *cliproxyauth.Auth) string {
-	if alias == "" {
+	trimmed := strings.TrimSpace(alias)
+	if trimmed == "" {
 		return ""
 	}
-	// Hardcoded mappings for thinking models to actual Claude model names
-	switch alias {
-	case "claude-opus-4-5-thinking", "claude-opus-4-5-thinking-low", "claude-opus-4-5-thinking-medium", "claude-opus-4-5-thinking-high":
-		return "claude-opus-4-5-20251101"
-	case "claude-sonnet-4-5-thinking":
-		return "claude-sonnet-4-5-20250929"
-	}
+
 	entry := e.resolveClaudeConfig(auth)
 	if entry == nil {
 		return ""
 	}
+
+	normalizedModel, metadata := util.NormalizeThinkingModel(trimmed)
+
+	// Candidate names to match against configured aliases/names.
+	candidates := []string{strings.TrimSpace(normalizedModel)}
+	if !strings.EqualFold(normalizedModel, trimmed) {
+		candidates = append(candidates, trimmed)
+	}
+	if original := util.ResolveOriginalModel(normalizedModel, metadata); original != "" && !strings.EqualFold(original, normalizedModel) {
+		candidates = append(candidates, original)
+	}
+
 	for i := range entry.Models {
 		model := entry.Models[i]
 		name := strings.TrimSpace(model.Name)
 		modelAlias := strings.TrimSpace(model.Alias)
-		if modelAlias != "" {
-			if strings.EqualFold(modelAlias, alias) {
+
+		for _, candidate := range candidates {
+			if candidate == "" {
+				continue
+			}
+			if modelAlias != "" && strings.EqualFold(modelAlias, candidate) {
 				if name != "" {
 					return name
 				}
-				return alias
+				return candidate
+			}
+			if name != "" && strings.EqualFold(name, candidate) {
+				return name
 			}
-			continue
-		}
-		if name != "" && strings.EqualFold(name, alias) {
-			return name
 		}
 	}
 	return ""
--- a/internal/runtime/executor/codex_executor.go
+++ b/internal/runtime/executor/codex_executor.go
@@ -49,14 +49,14 @@ func (e *CodexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
 	defer reporter.trackFailure(ctx, &err)

+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("codex")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-
-	body = e.setReasoningEffortByAlias(req.Model, body)
-
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model)
 	body = applyPayloadConfig(e.cfg, req.Model, body)
-
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	body, _ = sjson.SetBytes(body, "stream", true)
 	body, _ = sjson.DeleteBytes(body, "previous_response_id")

@@ -142,13 +142,16 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
 	defer reporter.trackFailure(ctx, &err)

+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("codex")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)

-	body = e.setReasoningEffortByAlias(req.Model, body)
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model)
 	body = applyPayloadConfig(e.cfg, req.Model, body)
 	body, _ = sjson.DeleteBytes(body, "previous_response_id")
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)

 	url := strings.TrimSuffix(baseURL, "/") + "/responses"
 	httpReq, err := e.cacheHelper(ctx, from, url, req, body)
@@ -235,14 +238,16 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 }

 func (e *CodexExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("codex")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)

 	modelForCounting := req.Model

-	body = e.setReasoningEffortByAlias(req.Model, body)
-
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model)
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	body, _ = sjson.DeleteBytes(body, "previous_response_id")
 	body, _ = sjson.SetBytes(body, "stream", false)

@@ -261,83 +266,6 @@ func (e *CodexExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth
 	return cliproxyexecutor.Response{Payload: []byte(translated)}, nil
 }

-func (e *CodexExecutor) setReasoningEffortByAlias(modelName string, payload []byte) []byte {
-	if util.InArray([]string{"gpt-5", "gpt-5-minimal", "gpt-5-low", "gpt-5-medium", "gpt-5-high"}, modelName) {
-		payload, _ = sjson.SetBytes(payload, "model", "gpt-5")
-		switch modelName {
-		case "gpt-5-minimal":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "minimal")
-		case "gpt-5-low":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "low")
-		case "gpt-5-medium":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "medium")
-		case "gpt-5-high":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "high")
-		}
-	} else if util.InArray([]string{"gpt-5-codex", "gpt-5-codex-low", "gpt-5-codex-medium", "gpt-5-codex-high"}, modelName) {
-		payload, _ = sjson.SetBytes(payload, "model", "gpt-5-codex")
-		switch modelName {
-		case "gpt-5-codex-low":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "low")
-		case "gpt-5-codex-medium":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "medium")
-		case "gpt-5-codex-high":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "high")
-		}
-	} else if util.InArray([]string{"gpt-5-codex-mini", "gpt-5-codex-mini-medium", "gpt-5-codex-mini-high"}, modelName) {
-		payload, _ = sjson.SetBytes(payload, "model", "gpt-5-codex-mini")
-		switch modelName {
-		case "gpt-5-codex-mini-medium":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "medium")
-		case "gpt-5-codex-mini-high":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "high")
-		}
-	} else if util.InArray([]string{"gpt-5.1", "gpt-5.1-none", "gpt-5.1-low", "gpt-5.1-medium", "gpt-5.1-high"}, modelName) {
-		payload, _ = sjson.SetBytes(payload, "model", "gpt-5.1")
-		switch modelName {
-		case "gpt-5.1-none":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "none")
-		case "gpt-5.1-low":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "low")
-		case "gpt-5.1-medium":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "medium")
-		case "gpt-5.1-high":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "high")
-		}
-	} else if util.InArray([]string{"gpt-5.1-codex", "gpt-5.1-codex-low", "gpt-5.1-codex-medium", "gpt-5.1-codex-high"}, modelName) {
-		payload, _ = sjson.SetBytes(payload, "model", "gpt-5.1-codex")
-		switch modelName {
-		case "gpt-5.1-codex-low":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "low")
-		case "gpt-5.1-codex-medium":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "medium")
-		case "gpt-5.1-codex-high":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "high")
-		}
-	} else if util.InArray([]string{"gpt-5.1-codex-mini", "gpt-5.1-codex-mini-medium", "gpt-5.1-codex-mini-high"}, modelName) {
-		payload, _ = sjson.SetBytes(payload, "model", "gpt-5.1-codex-mini")
-		switch modelName {
-		case "gpt-5.1-codex-mini-medium":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "medium")
-		case "gpt-5.1-codex-mini-high":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "high")
-		}
-	} else if util.InArray([]string{"gpt-5.1-codex-max", "gpt-5.1-codex-max-low", "gpt-5.1-codex-max-medium", "gpt-5.1-codex-max-high", "gpt-5.1-codex-max-xhigh"}, modelName) {
-		payload, _ = sjson.SetBytes(payload, "model", "gpt-5.1-codex-max")
-		switch modelName {
-		case "gpt-5.1-codex-max-low":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "low")
-		case "gpt-5.1-codex-max-medium":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "medium")
-		case "gpt-5.1-codex-max-high":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "high")
-		case "gpt-5.1-codex-max-xhigh":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "xhigh")
-		}
-	}
-	return payload
-}
-
 func tokenizerForCodexModel(model string) (tokenizer.Codec, error) {
 	sanitized := strings.ToLower(strings.TrimSpace(model))
 	switch {
--- a/internal/runtime/executor/gemini_executor.go
+++ b/internal/runtime/executor/gemini_executor.go
@@ -75,6 +75,8 @@ func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
 	defer reporter.trackFailure(ctx, &err)

+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+
 	// Official Gemini API via API key or OAuth bearer
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
@@ -85,6 +87,7 @@ func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
 	body = fixGeminiImageAspectRatio(req.Model, body)
 	body = applyPayloadConfig(e.cfg, req.Model, body)
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)

 	action := "generateContent"
 	if req.Metadata != nil {
@@ -93,7 +96,7 @@ func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 		}
 	}
 	baseURL := resolveGeminiBaseURL(auth)
-	url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, req.Model, action)
+	url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, upstreamModel, action)
 	if opts.Alt != "" && action != "countTokens" {
 		url = url + fmt.Sprintf("?$alt=%s", opts.Alt)
 	}
@@ -167,6 +170,8 @@ func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
 	defer reporter.trackFailure(ctx, &err)

+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
@@ -176,9 +181,10 @@ func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
 	body = fixGeminiImageAspectRatio(req.Model, body)
 	body = applyPayloadConfig(e.cfg, req.Model, body)
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)

 	baseURL := resolveGeminiBaseURL(auth)
-	url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, req.Model, "streamGenerateContent")
+	url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, upstreamModel, "streamGenerateContent")
 	if opts.Alt == "" {
 		url = url + "?alt=sse"
 	} else {
--- a/internal/runtime/executor/gemini_vertex_executor.go
+++ b/internal/runtime/executor/gemini_vertex_executor.go
@@ -105,10 +105,12 @@ func (e *GeminiVertexExecutor) CountTokens(ctx context.Context, auth *cliproxyau

 // countTokensWithServiceAccount handles token counting using service account credentials.
 func (e *GeminiVertexExecutor) countTokensWithServiceAccount(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, projectID, location string, saJSON []byte) (cliproxyexecutor.Response, error) {
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 	translatedReq := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	if budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
+	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
 		if budgetOverride != nil {
 			norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
 			budgetOverride = &norm
@@ -117,13 +119,14 @@ func (e *GeminiVertexExecutor) countTokensWithServiceAccount(ctx context.Context
 	}
 	translatedReq = util.StripThinkingConfigIfUnsupported(req.Model, translatedReq)
 	translatedReq = fixGeminiImageAspectRatio(req.Model, translatedReq)
+	translatedReq, _ = sjson.SetBytes(translatedReq, "model", upstreamModel)
 	respCtx := context.WithValue(ctx, "alt", opts.Alt)
 	translatedReq, _ = sjson.DeleteBytes(translatedReq, "tools")
 	translatedReq, _ = sjson.DeleteBytes(translatedReq, "generationConfig")
 	translatedReq, _ = sjson.DeleteBytes(translatedReq, "safetySettings")

 	baseURL := vertexBaseURL(location)
-	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, req.Model, "countTokens")
+	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, upstreamModel, "countTokens")

 	httpReq, errNewReq := http.NewRequestWithContext(respCtx, http.MethodPost, url, bytes.NewReader(translatedReq))
 	if errNewReq != nil {
@@ -191,10 +194,12 @@ func (e *GeminiVertexExecutor) countTokensWithServiceAccount(ctx context.Context

 // countTokensWithAPIKey handles token counting using API key credentials.
 func (e *GeminiVertexExecutor) countTokensWithAPIKey(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, apiKey, baseURL string) (cliproxyexecutor.Response, error) {
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 	translatedReq := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	if budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
+	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
 		if budgetOverride != nil {
 			norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
 			budgetOverride = &norm
@@ -203,6 +208,7 @@ func (e *GeminiVertexExecutor) countTokensWithAPIKey(ctx context.Context, auth *
 	}
 	translatedReq = util.StripThinkingConfigIfUnsupported(req.Model, translatedReq)
 	translatedReq = fixGeminiImageAspectRatio(req.Model, translatedReq)
+	translatedReq, _ = sjson.SetBytes(translatedReq, "model", upstreamModel)
 	respCtx := context.WithValue(ctx, "alt", opts.Alt)
 	translatedReq, _ = sjson.DeleteBytes(translatedReq, "tools")
 	translatedReq, _ = sjson.DeleteBytes(translatedReq, "generationConfig")
@@ -286,10 +292,12 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au
 	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
 	defer reporter.trackFailure(ctx, &err)

+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	if budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
+	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
 		if budgetOverride != nil {
 			norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
 			budgetOverride = &norm
@@ -301,6 +309,7 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au
 	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
 	body = fixGeminiImageAspectRatio(req.Model, body)
 	body = applyPayloadConfig(e.cfg, req.Model, body)
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)

 	action := "generateContent"
 	if req.Metadata != nil {
@@ -309,7 +318,7 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au
 		}
 	}
 	baseURL := vertexBaseURL(location)
-	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, req.Model, action)
+	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, upstreamModel, action)
 	if opts.Alt != "" && action != "countTokens" {
 		url = url + fmt.Sprintf("?$alt=%s", opts.Alt)
 	}
@@ -383,10 +392,12 @@ func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *clip
 	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
 	defer reporter.trackFailure(ctx, &err)

+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	if budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
+	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
 		if budgetOverride != nil {
 			norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
 			budgetOverride = &norm
@@ -398,6 +409,7 @@ func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *clip
 	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
 	body = fixGeminiImageAspectRatio(req.Model, body)
 	body = applyPayloadConfig(e.cfg, req.Model, body)
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)

 	action := "generateContent"
 	if req.Metadata != nil {
@@ -410,7 +422,7 @@ func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *clip
 	if baseURL == "" {
 		baseURL = "https://generativelanguage.googleapis.com"
 	}
-	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, req.Model, action)
+	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, upstreamModel, action)
 	if opts.Alt != "" && action != "countTokens" {
 		url = url + fmt.Sprintf("?$alt=%s", opts.Alt)
 	}
@@ -481,10 +493,12 @@ func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Conte
 	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
 	defer reporter.trackFailure(ctx, &err)

+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
-	if budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
+	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
 		if budgetOverride != nil {
 			norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
 			budgetOverride = &norm
@@ -496,9 +510,10 @@ func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Conte
 	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
 	body = fixGeminiImageAspectRatio(req.Model, body)
 	body = applyPayloadConfig(e.cfg, req.Model, body)
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)

 	baseURL := vertexBaseURL(location)
-	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, req.Model, "streamGenerateContent")
+	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, upstreamModel, "streamGenerateContent")
 	if opts.Alt == "" {
 		url = url + "?alt=sse"
 	} else {
@@ -595,10 +610,12 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth
 	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
 	defer reporter.trackFailure(ctx, &err)

+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
-	if budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
+	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
 		if budgetOverride != nil {
 			norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
 			budgetOverride = &norm
@@ -610,12 +627,13 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth
 	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
 	body = fixGeminiImageAspectRatio(req.Model, body)
 	body = applyPayloadConfig(e.cfg, req.Model, body)
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)

 	// For API key auth, use simpler URL format without project/location
 	if baseURL == "" {
 		baseURL = "https://generativelanguage.googleapis.com"
 	}
-	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, req.Model, "streamGenerateContent")
+	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, upstreamModel, "streamGenerateContent")
 	if opts.Alt == "" {
 		url = url + "?alt=sse"
 	} else {
--- a/internal/runtime/executor/iflow_executor.go
+++ b/internal/runtime/executor/iflow_executor.go
@@ -57,6 +57,10 @@ func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
+	body = applyReasoningEffortMetadataChatCompletions(body, req.Metadata, req.Model)
+	if upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata); upstreamModel != "" {
+		body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	}
 	body = applyPayloadConfig(e.cfg, req.Model, body)

 	endpoint := strings.TrimSuffix(baseURL, "/") + iflowDefaultEndpoint
@@ -139,6 +143,10 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 	to := sdktranslator.FromString("openai")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)

+	body = applyReasoningEffortMetadataChatCompletions(body, req.Metadata, req.Model)
+	if upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata); upstreamModel != "" {
+		body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	}
 	// 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 {
--- a/internal/runtime/executor/openai_compat_executor.go
+++ b/internal/runtime/executor/openai_compat_executor.go
@@ -58,6 +58,10 @@ func (e *OpenAICompatExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 		translated = e.overrideModel(translated, modelOverride)
 	}
 	translated = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", translated)
+	translated = applyReasoningEffortMetadataChatCompletions(translated, req.Metadata, req.Model)
+	if upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata); upstreamModel != "" {
+		translated, _ = sjson.SetBytes(translated, "model", upstreamModel)
+	}

 	url := strings.TrimSuffix(baseURL, "/") + "/chat/completions"
 	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(translated))
@@ -143,6 +147,10 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 		translated = e.overrideModel(translated, modelOverride)
 	}
 	translated = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", translated)
+	translated = applyReasoningEffortMetadataChatCompletions(translated, req.Metadata, req.Model)
+	if upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata); upstreamModel != "" {
+		translated, _ = sjson.SetBytes(translated, "model", upstreamModel)
+	}

 	url := strings.TrimSuffix(baseURL, "/") + "/chat/completions"
 	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(translated))
--- a/internal/runtime/executor/payload_helpers.go
+++ b/internal/runtime/executor/payload_helpers.go
@@ -12,8 +12,8 @@ import (
 // applyThinkingMetadata applies thinking config from model suffix metadata (e.g., -reasoning, -thinking-N)
 // for standard Gemini format payloads. It normalizes the budget when the model supports thinking.
 func applyThinkingMetadata(payload []byte, metadata map[string]any, model string) []byte {
-	budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(metadata)
-	if !ok {
+	budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(model, metadata)
+	if !ok || (budgetOverride == nil && includeOverride == nil) {
 		return payload
 	}
 	if !util.ModelSupportsThinking(model) {
@@ -29,17 +29,60 @@ func applyThinkingMetadata(payload []byte, metadata map[string]any, model string
 // applyThinkingMetadataCLI applies thinking config from model suffix metadata (e.g., -reasoning, -thinking-N)
 // for Gemini CLI format payloads (nested under "request"). It normalizes the budget when the model supports thinking.
 func applyThinkingMetadataCLI(payload []byte, metadata map[string]any, model string) []byte {
-	budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(metadata)
-	if !ok {
+	budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(model, metadata)
+	if !ok || (budgetOverride == nil && includeOverride == nil) {
 		return payload
 	}
-	if budgetOverride != nil && util.ModelSupportsThinking(model) {
+	if !util.ModelSupportsThinking(model) {
+		return payload
+	}
+	if budgetOverride != nil {
 		norm := util.NormalizeThinkingBudget(model, *budgetOverride)
 		budgetOverride = &norm
 	}
 	return util.ApplyGeminiCLIThinkingConfig(payload, budgetOverride, includeOverride)
 }

+// applyReasoningEffortMetadata applies reasoning effort overrides (reasoning.effort) when present in metadata.
+// It avoids overwriting an existing reasoning.effort field and only applies to models that support thinking.
+func applyReasoningEffortMetadata(payload []byte, metadata map[string]any, model string) []byte {
+	if len(metadata) == 0 {
+		return payload
+	}
+	if !util.ModelSupportsThinking(model) {
+		return payload
+	}
+	if gjson.GetBytes(payload, "reasoning.effort").Exists() {
+		return payload
+	}
+	if effort, ok := util.ReasoningEffortFromMetadata(metadata); ok && effort != "" {
+		if updated, err := sjson.SetBytes(payload, "reasoning.effort", effort); err == nil {
+			return updated
+		}
+	}
+	return payload
+}
+
+// applyReasoningEffortMetadataChatCompletions applies reasoning_effort (OpenAI chat completions field)
+// when present in metadata. It avoids overwriting an existing reasoning_effort field.
+func applyReasoningEffortMetadataChatCompletions(payload []byte, metadata map[string]any, model string) []byte {
+	if len(metadata) == 0 {
+		return payload
+	}
+	if !util.ModelSupportsThinking(model) {
+		return payload
+	}
+	if gjson.GetBytes(payload, "reasoning_effort").Exists() {
+		return payload
+	}
+	if effort, ok := util.ReasoningEffortFromMetadata(metadata); ok && effort != "" {
+		if updated, err := sjson.SetBytes(payload, "reasoning_effort", effort); err == nil {
+			return updated
+		}
+	}
+	return payload
+}
+
 // applyPayloadConfig applies payload default and override rules from configuration
 // to the given JSON payload for the specified model.
 // Defaults only fill missing fields, while overrides always overwrite existing values.
--- a/internal/runtime/executor/qwen_executor.go
+++ b/internal/runtime/executor/qwen_executor.go
@@ -12,6 +12,7 @@ import (

 	qwenauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/qwen"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
 	sdktranslator "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
@@ -50,6 +51,10 @@ func (e *QwenExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
+	body = applyReasoningEffortMetadataChatCompletions(body, req.Metadata, req.Model)
+	if upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata); upstreamModel != "" {
+		body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	}
 	body = applyPayloadConfig(e.cfg, req.Model, body)

 	url := strings.TrimSuffix(baseURL, "/") + "/chat/completions"
@@ -121,6 +126,10 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 	to := sdktranslator.FromString("openai")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)

+	body = applyReasoningEffortMetadataChatCompletions(body, req.Metadata, req.Model)
+	if upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata); upstreamModel != "" {
+		body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	}
 	toolsResult := gjson.GetBytes(body, "tools")
 	// I'm addressing the Qwen3 "poisoning" issue, which is caused by the model needing a tool to be defined. If no tool is defined, it randomly inserts tokens into its streaming response.
 	// This will have no real consequences. It's just to scare Qwen3.