From 1215c635a0e7ba72a8e94864d83f4def3480be73 Mon Sep 17 00:00:00 2001
From: teeverc <72298507+teeverc@users.noreply.github.com>
Date: Fri, 12 Dec 2025 00:14:19 -0800
Subject: [PATCH 1/4] fix: flush Claude SSE chunks immediately to match OpenAI
 behavior

- Write each SSE chunk directly to c.Writer and flush immediately
- Remove buffered writer and ticker-based flushing that caused delayed output
- Add 500ms timeout case for consistency with OpenAI/Gemini handlers
- Clean up unused bufio import

This fixes the 'not streaming' issue where small responses were held
in the buffer until timeout/threshold was reached.

Amp-Thread-ID: https://ampcode.com/threads/T-019b1186-164e-740c-96ab-856f64ee6bee
Co-authored-by: Amp <amp@ampcode.com>
---
 sdk/api/handlers/claude/code_handlers.go | 51 ++++++------------------
 1 file changed, 12 insertions(+), 39 deletions(-)

diff --git a/sdk/api/handlers/claude/code_handlers.go b/sdk/api/handlers/claude/code_handlers.go
index 8a57a0cc..5cca651a 100644
--- a/sdk/api/handlers/claude/code_handlers.go
+++ b/sdk/api/handlers/claude/code_handlers.go
@@ -7,7 +7,6 @@
 package claude
 
 import (
-	"bufio"
 	"bytes"
 	"compress/gzip"
 	"context"
@@ -219,52 +218,24 @@ func (h *ClaudeCodeAPIHandler) handleStreamingResponse(c *gin.Context, rawJSON [
 }
 
 func (h *ClaudeCodeAPIHandler) forwardClaudeStream(c *gin.Context, flusher http.Flusher, cancel func(error), data <-chan []byte, errs <-chan *interfaces.ErrorMessage) {
-	// v6.1: Intelligent Buffered Streamer strategy
-	// Enhanced buffering with larger buffer size (16KB) and longer flush interval (120ms).
-	// Smart flush only when buffer is sufficiently filled (≥50%), dramatically reducing
-	// flush frequency from ~12.5Hz to ~5-8Hz while maintaining low latency.
-	writer := bufio.NewWriterSize(c.Writer, 16*1024) // 4KB → 16KB
-	ticker := time.NewTicker(120 * time.Millisecond) // 80ms → 120ms
-	defer ticker.Stop()
-
-	var chunkIdx int
-
+	// OpenAI-style stream forwarding: write each SSE chunk and flush immediately.
+	// This guarantees clients see incremental output even for small responses.
 	for {
 		select {
 		case <-c.Request.Context().Done():
-			// Context cancelled, flush any remaining data before exit
-			_ = writer.Flush()
 			cancel(c.Request.Context().Err())
 			return
 
-		case <-ticker.C:
-			// Smart flush: only flush when buffer has sufficient data (≥50% full)
-			// This reduces flush frequency while ensuring data flows naturally
-			buffered := writer.Buffered()
-			if buffered >= 8*1024 { // At least 8KB (50% of 16KB buffer)
-				if err := writer.Flush(); err != nil {
-					// Error flushing, cancel and return
-					cancel(err)
-					return
-				}
-				flusher.Flush() // Also flush the underlying http.ResponseWriter
-			}
-
 		case chunk, ok := <-data:
 			if !ok {
-				// Stream ended, flush remaining data
-				_ = writer.Flush()
+				flusher.Flush()
 				cancel(nil)
 				return
 			}
-
-			// Forward the complete SSE event block directly (already formatted by the translator).
-			// The translator returns a complete SSE-compliant event block, including event:, data:, and separators.
-			// The handler just needs to forward it without reassembly.
 			if len(chunk) > 0 {
-				_, _ = writer.Write(chunk)
+				_, _ = c.Writer.Write(chunk)
+				flusher.Flush()
 			}
-			chunkIdx++
 
 		case errMsg, ok := <-errs:
 			if !ok {
@@ -276,21 +247,23 @@ func (h *ClaudeCodeAPIHandler) forwardClaudeStream(c *gin.Context, flusher http.
 					status = errMsg.StatusCode
 				}
 				c.Status(status)
+
 				// An error occurred: emit as a proper SSE error event
 				errorBytes, _ := json.Marshal(h.toClaudeError(errMsg))
-				_, _ = writer.WriteString("event: error\n")
-				_, _ = writer.WriteString("data: ")
-				_, _ = writer.Write(errorBytes)
-				_, _ = writer.WriteString("\n\n")
-				_ = writer.Flush()
+				_, _ = c.Writer.Write([]byte("event: error\n"))
+				_, _ = c.Writer.Write([]byte("data: "))
+				_, _ = c.Writer.Write(errorBytes)
+				_, _ = c.Writer.Write([]byte("\n\n"))
 				flusher.Flush()
 			}
+
 			var execErr error
 			if errMsg != nil {
 				execErr = errMsg.Error
 			}
 			cancel(execErr)
 			return
+		case <-time.After(500 * time.Millisecond):
 		}
 	}
 }

From 5ab30323354a4cff00a7ab9c0c1acacfe96dbb8d Mon Sep 17 00:00:00 2001
From: teeverc <72298507+teeverc@users.noreply.github.com>
Date: Fri, 12 Dec 2025 00:26:01 -0800
Subject: [PATCH 2/4] Update sdk/api/handlers/claude/code_handlers.go

thank you gemini

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
---
 sdk/api/handlers/claude/code_handlers.go | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)

diff --git a/sdk/api/handlers/claude/code_handlers.go b/sdk/api/handlers/claude/code_handlers.go
index 5cca651a..8a4c4806 100644
--- a/sdk/api/handlers/claude/code_handlers.go
+++ b/sdk/api/handlers/claude/code_handlers.go
@@ -250,10 +250,7 @@ func (h *ClaudeCodeAPIHandler) forwardClaudeStream(c *gin.Context, flusher http.
 
 				// An error occurred: emit as a proper SSE error event
 				errorBytes, _ := json.Marshal(h.toClaudeError(errMsg))
-				_, _ = c.Writer.Write([]byte("event: error\n"))
-				_, _ = c.Writer.Write([]byte("data: "))
-				_, _ = c.Writer.Write(errorBytes)
-				_, _ = c.Writer.Write([]byte("\n\n"))
+				_, _ = fmt.Fprintf(c.Writer, "event: error\ndata: %s\n\n", errorBytes)
 				flusher.Flush()
 			}
 

From 374faa264028530a7c813378d9a7889bdb358e0e Mon Sep 17 00:00:00 2001
From: hkfires <10558748+hkfires@users.noreply.github.com>
Date: Fri, 12 Dec 2025 21:33:20 +0800
Subject: [PATCH 3/4] fix(thinking): map budgets to effort levels

Ensure thinking settings translate correctly across providers:
- Only apply reasoning_effort to level-based models and derive it from numeric
  budget suffixes when present
- Strip effort string fields for budget-based models and skip Claude/Gemini
  budget resolution for level-based or unsupported models
- Default Gemini include_thoughts when a nonzero budget override is set
- Add cross-protocol conversion and budget range tests
---
 internal/runtime/executor/payload_helpers.go |  34 +-
 internal/util/claude_thinking.go             |   3 +
 internal/util/gemini_thinking.go             |  20 +-
 internal/util/openai_thinking.go             |  34 ++
 internal/util/thinking_suffix.go             |   5 +
 test/thinking_conversion_test.go             | 561 +++++++++++++++++++
 6 files changed, 645 insertions(+), 12 deletions(-)
 create mode 100644 internal/util/openai_thinking.go
 create mode 100644 test/thinking_conversion_test.go

diff --git a/internal/runtime/executor/payload_helpers.go b/internal/runtime/executor/payload_helpers.go
index 9bc82f1f..222c6e37 100644
--- a/internal/runtime/executor/payload_helpers.go
+++ b/internal/runtime/executor/payload_helpers.go
@@ -59,8 +59,20 @@ func applyReasoningEffortMetadata(payload []byte, metadata map[string]any, model
 		return payload
 	}
 	if effort, ok := util.ReasoningEffortFromMetadata(metadata); ok && effort != "" {
-		if updated, err := sjson.SetBytes(payload, field, effort); err == nil {
-			return updated
+		if util.ModelUsesThinkingLevels(model) {
+			if updated, err := sjson.SetBytes(payload, field, effort); err == nil {
+				return updated
+			}
+		}
+	}
+	// Fallback: numeric thinking_budget suffix for level-based (OpenAI-style) models.
+	if util.ModelUsesThinkingLevels(model) {
+		if budget, _, _, matched := util.ThinkingFromMetadata(metadata); matched && budget != nil {
+			if effort, ok := util.OpenAIThinkingBudgetToEffort(model, *budget); ok && effort != "" {
+				if updated, err := sjson.SetBytes(payload, field, effort); err == nil {
+					return updated
+				}
+			}
 		}
 	}
 	return payload
@@ -219,31 +231,37 @@ func matchModelPattern(pattern, model string) bool {
 // normalizeThinkingConfig normalizes thinking-related fields in the payload
 // based on model capabilities. For models without thinking support, it strips
 // reasoning fields. For models with level-based thinking, it validates and
-// normalizes the reasoning effort level.
+// normalizes the reasoning effort level. For models with numeric budget thinking,
+// it strips the effort string fields.
 func normalizeThinkingConfig(payload []byte, model string) []byte {
 	if len(payload) == 0 || model == "" {
 		return payload
 	}
 
 	if !util.ModelSupportsThinking(model) {
-		return stripThinkingFields(payload)
+		return stripThinkingFields(payload, false)
 	}
 
 	if util.ModelUsesThinkingLevels(model) {
 		return normalizeReasoningEffortLevel(payload, model)
 	}
 
-	return payload
+	// Model supports thinking but uses numeric budgets, not levels.
+	// Strip effort string fields since they are not applicable.
+	return stripThinkingFields(payload, true)
 }
 
 // stripThinkingFields removes thinking-related fields from the payload for
-// models that do not support thinking.
-func stripThinkingFields(payload []byte) []byte {
+// models that do not support thinking. If effortOnly is true, only removes
+// effort string fields (for models using numeric budgets).
+func stripThinkingFields(payload []byte, effortOnly bool) []byte {
 	fieldsToRemove := []string{
-		"reasoning",
 		"reasoning_effort",
 		"reasoning.effort",
 	}
+	if !effortOnly {
+		fieldsToRemove = append([]string{"reasoning"}, fieldsToRemove...)
+	}
 	out := payload
 	for _, field := range fieldsToRemove {
 		if gjson.GetBytes(out, field).Exists() {
diff --git a/internal/util/claude_thinking.go b/internal/util/claude_thinking.go
index b0c5a0a2..6176f57d 100644
--- a/internal/util/claude_thinking.go
+++ b/internal/util/claude_thinking.go
@@ -28,6 +28,9 @@ func ApplyClaudeThinkingConfig(body []byte, budget *int) []byte {
 // It uses the unified ResolveThinkingConfigFromMetadata and normalizes the budget.
 // Returns the normalized budget (nil if thinking should not be enabled) and whether it matched.
 func ResolveClaudeThinkingConfig(modelName string, metadata map[string]any) (*int, bool) {
+	if !ModelSupportsThinking(modelName) {
+		return nil, false
+	}
 	budget, include, matched := ResolveThinkingConfigFromMetadata(modelName, metadata)
 	if !matched {
 		return nil, false
diff --git a/internal/util/gemini_thinking.go b/internal/util/gemini_thinking.go
index f33928b0..a89aba26 100644
--- a/internal/util/gemini_thinking.go
+++ b/internal/util/gemini_thinking.go
@@ -25,9 +25,15 @@ func ApplyGeminiThinkingConfig(body []byte, budget *int, includeThoughts *bool)
 			updated = rewritten
 		}
 	}
-	if includeThoughts != nil {
+	// Default to including thoughts when a budget override is present but no explicit include flag is provided.
+	incl := includeThoughts
+	if incl == nil && budget != nil && *budget != 0 {
+		defaultInclude := true
+		incl = &defaultInclude
+	}
+	if incl != nil {
 		valuePath := "generationConfig.thinkingConfig.include_thoughts"
-		rewritten, err := sjson.SetBytes(updated, valuePath, *includeThoughts)
+		rewritten, err := sjson.SetBytes(updated, valuePath, *incl)
 		if err == nil {
 			updated = rewritten
 		}
@@ -47,9 +53,15 @@ func ApplyGeminiCLIThinkingConfig(body []byte, budget *int, includeThoughts *boo
 			updated = rewritten
 		}
 	}
-	if includeThoughts != nil {
+	// Default to including thoughts when a budget override is present but no explicit include flag is provided.
+	incl := includeThoughts
+	if incl == nil && budget != nil && *budget != 0 {
+		defaultInclude := true
+		incl = &defaultInclude
+	}
+	if incl != nil {
 		valuePath := "request.generationConfig.thinkingConfig.include_thoughts"
-		rewritten, err := sjson.SetBytes(updated, valuePath, *includeThoughts)
+		rewritten, err := sjson.SetBytes(updated, valuePath, *incl)
 		if err == nil {
 			updated = rewritten
 		}
diff --git a/internal/util/openai_thinking.go b/internal/util/openai_thinking.go
new file mode 100644
index 00000000..4dda38f6
--- /dev/null
+++ b/internal/util/openai_thinking.go
@@ -0,0 +1,34 @@
+package util
+
+// OpenAIThinkingBudgetToEffort maps a numeric thinking budget (tokens)
+// into an OpenAI-style reasoning effort level for level-based models.
+//
+// Ranges:
+//   - 0            -> "none"
+//   - 1..1024      -> "low"
+//   - 1025..8192   -> "medium"
+//   - 8193..24576  -> "high"
+//   - 24577..      -> highest supported level for the model (defaults to "xhigh")
+//
+// Negative values (except the dynamic -1 handled elsewhere) are treated as unsupported.
+func OpenAIThinkingBudgetToEffort(model string, budget int) (string, bool) {
+	switch {
+	case budget < 0:
+		return "", false
+	case budget == 0:
+		return "none", true
+	case budget > 0 && budget <= 1024:
+		return "low", true
+	case budget <= 8192:
+		return "medium", true
+	case budget <= 24576:
+		return "high", true
+	case budget > 24576:
+		if levels := GetModelThinkingLevels(model); len(levels) > 0 {
+			return levels[len(levels)-1], true
+		}
+		return "xhigh", true
+	default:
+		return "", false
+	}
+}
diff --git a/internal/util/thinking_suffix.go b/internal/util/thinking_suffix.go
index 7851c580..b877e109 100644
--- a/internal/util/thinking_suffix.go
+++ b/internal/util/thinking_suffix.go
@@ -163,6 +163,11 @@ func ResolveThinkingConfigFromMetadata(model string, metadata map[string]any) (*
 	if !matched {
 		return nil, nil, false
 	}
+	// Level-based models (OpenAI-style) do not accept numeric thinking budgets in
+	// Claude/Gemini-style protocols, so we don't derive budgets for them here.
+	if ModelUsesThinkingLevels(model) {
+		return nil, nil, false
+	}
 
 	if budget == nil && effort != nil {
 		if derived, ok := ThinkingEffortToBudget(model, *effort); ok {
diff --git a/test/thinking_conversion_test.go b/test/thinking_conversion_test.go
new file mode 100644
index 00000000..c2f4aa8d
--- /dev/null
+++ b/test/thinking_conversion_test.go
@@ -0,0 +1,561 @@
+package test
+
+import (
+	"fmt"
+	"net/http"
+	"strings"
+	"testing"
+	"time"
+
+	_ "github.com/router-for-me/CLIProxyAPI/v6/internal/translator"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
+	sdktranslator "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// statusErr mirrors executor.statusErr to keep validation behavior aligned.
+type statusErr struct {
+	code int
+	msg  string
+}
+
+func (e statusErr) Error() string { return e.msg }
+
+// registerCoreModels loads representative models across providers into the registry
+// so NormalizeThinkingBudget and level validation use real ranges.
+func registerCoreModels(t *testing.T) func() {
+	t.Helper()
+	reg := registry.GetGlobalRegistry()
+	uid := fmt.Sprintf("thinking-core-%d", time.Now().UnixNano())
+	reg.RegisterClient(uid+"-gemini", "gemini", registry.GetGeminiModels())
+	reg.RegisterClient(uid+"-claude", "claude", registry.GetClaudeModels())
+	reg.RegisterClient(uid+"-openai", "codex", registry.GetOpenAIModels())
+	reg.RegisterClient(uid+"-qwen", "qwen", registry.GetQwenModels())
+	return func() {
+		reg.UnregisterClient(uid + "-gemini")
+		reg.UnregisterClient(uid + "-claude")
+		reg.UnregisterClient(uid + "-openai")
+		reg.UnregisterClient(uid + "-qwen")
+	}
+}
+
+func buildRawPayload(fromProtocol, modelWithSuffix string) []byte {
+	switch fromProtocol {
+	case "gemini":
+		return []byte(fmt.Sprintf(`{"model":"%s","contents":[{"role":"user","parts":[{"text":"hi"}]}]}`, modelWithSuffix))
+	case "openai-response":
+		return []byte(fmt.Sprintf(`{"model":"%s","input":[{"role":"user","content":[{"type":"text","text":"hi"}]}]}`, modelWithSuffix))
+	default: // openai / claude and other chat-style payloads
+		return []byte(fmt.Sprintf(`{"model":"%s","messages":[{"role":"user","content":"hi"}]}`, modelWithSuffix))
+	}
+}
+
+// applyThinkingMetadataLocal mirrors executor.applyThinkingMetadata.
+func applyThinkingMetadataLocal(payload []byte, metadata map[string]any, model string) []byte {
+	budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(model, metadata)
+	if !ok || (budgetOverride == nil && includeOverride == nil) {
+		return payload
+	}
+	if !util.ModelSupportsThinking(model) {
+		return payload
+	}
+	if budgetOverride != nil {
+		norm := util.NormalizeThinkingBudget(model, *budgetOverride)
+		budgetOverride = &norm
+	}
+	return util.ApplyGeminiThinkingConfig(payload, budgetOverride, includeOverride)
+}
+
+// applyReasoningEffortMetadataLocal mirrors executor.applyReasoningEffortMetadata.
+func applyReasoningEffortMetadataLocal(payload []byte, metadata map[string]any, model, field string) []byte {
+	if len(metadata) == 0 {
+		return payload
+	}
+	if !util.ModelSupportsThinking(model) {
+		return payload
+	}
+	if field == "" {
+		return payload
+	}
+	if effort, ok := util.ReasoningEffortFromMetadata(metadata); ok && effort != "" {
+		if updated, err := sjson.SetBytes(payload, field, effort); err == nil {
+			return updated
+		}
+	}
+	if util.ModelUsesThinkingLevels(model) {
+		if budget, _, _, matched := util.ThinkingFromMetadata(metadata); matched && budget != nil {
+			if effort, ok := util.OpenAIThinkingBudgetToEffort(model, *budget); ok && effort != "" {
+				if updated, err := sjson.SetBytes(payload, field, effort); err == nil {
+					return updated
+				}
+			}
+		}
+	}
+	return payload
+}
+
+// normalizeThinkingConfigLocal mirrors executor.normalizeThinkingConfig.
+func normalizeThinkingConfigLocal(payload []byte, model string) []byte {
+	if len(payload) == 0 || model == "" {
+		return payload
+	}
+
+	if !util.ModelSupportsThinking(model) {
+		return stripThinkingFieldsLocal(payload, false)
+	}
+
+	if util.ModelUsesThinkingLevels(model) {
+		return normalizeReasoningEffortLevelLocal(payload, model)
+	}
+
+	// Model supports thinking but uses numeric budgets, not levels.
+	// Strip effort string fields since they are not applicable.
+	return stripThinkingFieldsLocal(payload, true)
+}
+
+// stripThinkingFieldsLocal mirrors executor.stripThinkingFields.
+func stripThinkingFieldsLocal(payload []byte, effortOnly bool) []byte {
+	fieldsToRemove := []string{
+		"reasoning_effort",
+		"reasoning.effort",
+	}
+	if !effortOnly {
+		fieldsToRemove = append([]string{"reasoning"}, fieldsToRemove...)
+	}
+	out := payload
+	for _, field := range fieldsToRemove {
+		if gjson.GetBytes(out, field).Exists() {
+			out, _ = sjson.DeleteBytes(out, field)
+		}
+	}
+	return out
+}
+
+// normalizeReasoningEffortLevelLocal mirrors executor.normalizeReasoningEffortLevel.
+func normalizeReasoningEffortLevelLocal(payload []byte, model string) []byte {
+	out := payload
+
+	if effort := gjson.GetBytes(out, "reasoning_effort"); effort.Exists() {
+		if normalized, ok := util.NormalizeReasoningEffortLevel(model, effort.String()); ok {
+			out, _ = sjson.SetBytes(out, "reasoning_effort", normalized)
+		}
+	}
+
+	if effort := gjson.GetBytes(out, "reasoning.effort"); effort.Exists() {
+		if normalized, ok := util.NormalizeReasoningEffortLevel(model, effort.String()); ok {
+			out, _ = sjson.SetBytes(out, "reasoning.effort", normalized)
+		}
+	}
+
+	return out
+}
+
+// validateThinkingConfigLocal mirrors executor.validateThinkingConfig.
+func validateThinkingConfigLocal(payload []byte, model string) error {
+	if len(payload) == 0 || model == "" {
+		return nil
+	}
+	if !util.ModelSupportsThinking(model) || !util.ModelUsesThinkingLevels(model) {
+		return nil
+	}
+
+	levels := util.GetModelThinkingLevels(model)
+	checkField := func(path string) error {
+		if effort := gjson.GetBytes(payload, path); effort.Exists() {
+			if _, ok := util.NormalizeReasoningEffortLevel(model, effort.String()); !ok {
+				return statusErr{
+					code: http.StatusBadRequest,
+					msg:  fmt.Sprintf("unsupported reasoning effort level %q for model %s (supported: %s)", effort.String(), model, strings.Join(levels, ", ")),
+				}
+			}
+		}
+		return nil
+	}
+
+	if err := checkField("reasoning_effort"); err != nil {
+		return err
+	}
+	if err := checkField("reasoning.effort"); err != nil {
+		return err
+	}
+	return nil
+}
+
+// normalizeCodexPayload mirrors codex_executor's reasoning + streaming tweaks.
+func normalizeCodexPayload(body []byte, upstreamModel string) ([]byte, error) {
+	body = normalizeThinkingConfigLocal(body, upstreamModel)
+	if err := validateThinkingConfigLocal(body, upstreamModel); err != nil {
+		return body, err
+	}
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	body, _ = sjson.SetBytes(body, "stream", true)
+	body, _ = sjson.DeleteBytes(body, "previous_response_id")
+	return body, nil
+}
+
+// buildBodyForProtocol runs a minimal request through the same translation and
+// thinking pipeline used in executors for the given target protocol.
+func buildBodyForProtocol(t *testing.T, fromProtocol, toProtocol, modelWithSuffix string) ([]byte, error) {
+	t.Helper()
+	normalizedModel, metadata := util.NormalizeThinkingModel(modelWithSuffix)
+	upstreamModel := util.ResolveOriginalModel(normalizedModel, metadata)
+	raw := buildRawPayload(fromProtocol, modelWithSuffix)
+	stream := fromProtocol != toProtocol
+
+	body := sdktranslator.TranslateRequest(
+		sdktranslator.FromString(fromProtocol),
+		sdktranslator.FromString(toProtocol),
+		normalizedModel,
+		raw,
+		stream,
+	)
+
+	var err error
+	switch toProtocol {
+	case "gemini":
+		body = applyThinkingMetadataLocal(body, metadata, normalizedModel)
+		body = util.ApplyDefaultThinkingIfNeeded(normalizedModel, body)
+		body = util.NormalizeGeminiThinkingBudget(normalizedModel, body)
+		body = util.StripThinkingConfigIfUnsupported(normalizedModel, body)
+	case "claude":
+		if budget, ok := util.ResolveClaudeThinkingConfig(normalizedModel, metadata); ok {
+			body = util.ApplyClaudeThinkingConfig(body, budget)
+		}
+	case "openai":
+		body = applyReasoningEffortMetadataLocal(body, metadata, normalizedModel, "reasoning_effort")
+		body = normalizeThinkingConfigLocal(body, upstreamModel)
+		err = validateThinkingConfigLocal(body, upstreamModel)
+	case "codex": // OpenAI responses / codex
+		body = applyReasoningEffortMetadataLocal(body, metadata, normalizedModel, "reasoning.effort")
+		// Mirror CodexExecutor final normalization and model override so tests log the final body.
+		body, err = normalizeCodexPayload(body, upstreamModel)
+	default:
+	}
+
+	// Mirror executor behavior: final payload uses the upstream (base) model name.
+	if upstreamModel != "" {
+		body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	}
+
+	// For tests we only keep model + thinking-related fields to avoid noise.
+	body = filterThinkingBody(toProtocol, body, upstreamModel, normalizedModel)
+	return body, err
+}
+
+// filterThinkingBody projects the translated payload down to only model and
+// thinking-related fields for the given target protocol.
+func filterThinkingBody(toProtocol string, body []byte, upstreamModel, normalizedModel string) []byte {
+	if len(body) == 0 {
+		return body
+	}
+	out := []byte(`{}`)
+
+	// Preserve model if present, otherwise fall back to upstream/normalized model.
+	if m := gjson.GetBytes(body, "model"); m.Exists() {
+		out, _ = sjson.SetBytes(out, "model", m.Value())
+	} else if upstreamModel != "" {
+		out, _ = sjson.SetBytes(out, "model", upstreamModel)
+	} else if normalizedModel != "" {
+		out, _ = sjson.SetBytes(out, "model", normalizedModel)
+	}
+
+	switch toProtocol {
+	case "gemini":
+		if tc := gjson.GetBytes(body, "generationConfig.thinkingConfig"); tc.Exists() {
+			out, _ = sjson.SetRawBytes(out, "generationConfig.thinkingConfig", []byte(tc.Raw))
+		}
+	case "claude":
+		if tcfg := gjson.GetBytes(body, "thinking"); tcfg.Exists() {
+			out, _ = sjson.SetRawBytes(out, "thinking", []byte(tcfg.Raw))
+		}
+	case "openai":
+		if re := gjson.GetBytes(body, "reasoning_effort"); re.Exists() {
+			out, _ = sjson.SetBytes(out, "reasoning_effort", re.Value())
+		}
+	case "codex":
+		if re := gjson.GetBytes(body, "reasoning.effort"); re.Exists() {
+			out, _ = sjson.SetBytes(out, "reasoning.effort", re.Value())
+		}
+	}
+	return out
+}
+
+func TestThinkingConversionsAcrossProtocolsAndModels(t *testing.T) {
+	cleanup := registerCoreModels(t)
+	defer cleanup()
+
+	models := []string{
+		"gpt-5",             // supports levels (low/medium/high)
+		"gemini-2.5-pro",    // supports numeric budget
+		"qwen3-coder-flash", // no thinking support
+	}
+	fromProtocols := []string{"openai", "claude", "gemini", "openai-response"}
+	toProtocols := []string{"gemini", "claude", "openai", "codex"}
+
+	type scenario struct {
+		name        string
+		modelSuffix string
+		expectFn    func(info *registry.ModelInfo) (present bool, budget int64)
+	}
+
+	buildBudgetFn := func(raw int) func(info *registry.ModelInfo) (bool, int64) {
+		return func(info *registry.ModelInfo) (bool, int64) {
+			if info == nil || info.Thinking == nil {
+				return false, 0
+			}
+			return true, int64(util.NormalizeThinkingBudget(info.ID, raw))
+		}
+	}
+
+	levelBudgetFn := func(level string) func(info *registry.ModelInfo) (bool, int64) {
+		return func(info *registry.ModelInfo) (bool, int64) {
+			if info == nil || info.Thinking == nil {
+				return false, 0
+			}
+			if b, ok := util.ThinkingEffortToBudget(info.ID, level); ok {
+				return true, int64(b)
+			}
+			return false, 0
+		}
+	}
+
+	for _, model := range models {
+		info := registry.GetGlobalRegistry().GetModelInfo(model)
+		min, max := 0, 0
+		if info != nil && info.Thinking != nil {
+			min = info.Thinking.Min
+			max = info.Thinking.Max
+		}
+
+		for _, from := range fromProtocols {
+			// Scenario selection follows protocol semantics:
+			// - OpenAI-style protocols (openai/openai-response) express thinking as levels.
+			// - Claude/Gemini-style protocols express thinking as numeric budgets.
+			cases := []scenario{
+				{name: "no-suffix", modelSuffix: model, expectFn: func(_ *registry.ModelInfo) (bool, int64) { return false, 0 }},
+			}
+			if from == "openai" || from == "openai-response" {
+				cases = append(cases,
+					scenario{name: "level-low", modelSuffix: fmt.Sprintf("%s(low)", model), expectFn: levelBudgetFn("low")},
+					scenario{name: "level-high", modelSuffix: fmt.Sprintf("%s(high)", model), expectFn: levelBudgetFn("high")},
+					scenario{name: "level-auto", modelSuffix: fmt.Sprintf("%s(auto)", model), expectFn: levelBudgetFn("auto")},
+				)
+			} else { // claude or gemini
+				if util.ModelUsesThinkingLevels(model) {
+					// Numeric budgets for level-based models are mapped into levels when needed.
+					cases = append(cases,
+						scenario{name: "numeric-0", modelSuffix: fmt.Sprintf("%s(0)", model), expectFn: buildBudgetFn(0)},
+						scenario{name: "numeric-1024", modelSuffix: fmt.Sprintf("%s(1024)", model), expectFn: buildBudgetFn(1024)},
+						scenario{name: "numeric-1025", modelSuffix: fmt.Sprintf("%s(1025)", model), expectFn: buildBudgetFn(1025)},
+						scenario{name: "numeric-8192", modelSuffix: fmt.Sprintf("%s(8192)", model), expectFn: buildBudgetFn(8192)},
+						scenario{name: "numeric-8193", modelSuffix: fmt.Sprintf("%s(8193)", model), expectFn: buildBudgetFn(8193)},
+						scenario{name: "numeric-24576", modelSuffix: fmt.Sprintf("%s(24576)", model), expectFn: buildBudgetFn(24576)},
+						scenario{name: "numeric-24577", modelSuffix: fmt.Sprintf("%s(24577)", model), expectFn: buildBudgetFn(24577)},
+					)
+				} else {
+					cases = append(cases,
+						scenario{name: "numeric-below-min", modelSuffix: fmt.Sprintf("%s(%d)", model, min-10), expectFn: buildBudgetFn(min - 10)},
+						scenario{name: "numeric-above-max", modelSuffix: fmt.Sprintf("%s(%d)", model, max+10), expectFn: buildBudgetFn(max + 10)},
+					)
+				}
+			}
+
+			for _, to := range toProtocols {
+				if from == to {
+					continue
+				}
+				t.Logf("─────────────────────────────────────────────────────────────────────────────────")
+				t.Logf("  %s -> %s | model: %s", from, to, model)
+				t.Logf("─────────────────────────────────────────────────────────────────────────────────")
+				for _, cs := range cases {
+					from := from
+					to := to
+					cs := cs
+					testName := fmt.Sprintf("%s->%s/%s/%s", from, to, model, cs.name)
+					t.Run(testName, func(t *testing.T) {
+						normalizedModel, metadata := util.NormalizeThinkingModel(cs.modelSuffix)
+						expectPresent, expectValue, expectErr := func() (bool, string, bool) {
+							switch to {
+							case "gemini":
+								budget, include, ok := util.ResolveThinkingConfigFromMetadata(normalizedModel, metadata)
+								if !ok || !util.ModelSupportsThinking(normalizedModel) {
+									return false, "", false
+								}
+								if include != nil && !*include {
+									return false, "", false
+								}
+								if budget == nil {
+									return false, "", false
+								}
+								norm := util.NormalizeThinkingBudget(normalizedModel, *budget)
+								return true, fmt.Sprintf("%d", norm), false
+							case "claude":
+								if !util.ModelSupportsThinking(normalizedModel) {
+									return false, "", false
+								}
+								budget, ok := util.ResolveClaudeThinkingConfig(normalizedModel, metadata)
+								if !ok || budget == nil {
+									return false, "", false
+								}
+								return true, fmt.Sprintf("%d", *budget), false
+							case "openai":
+								if !util.ModelSupportsThinking(normalizedModel) {
+									return false, "", false
+								}
+								if !util.ModelUsesThinkingLevels(normalizedModel) {
+									// Non-levels models don't support effort strings in openai
+									return false, "", false
+								}
+								effort, ok := util.ReasoningEffortFromMetadata(metadata)
+								if !ok || strings.TrimSpace(effort) == "" {
+									if budget, _, _, matched := util.ThinkingFromMetadata(metadata); matched && budget != nil {
+										if mapped, okMap := util.OpenAIThinkingBudgetToEffort(normalizedModel, *budget); okMap {
+											effort = mapped
+											ok = true
+										}
+									}
+								}
+								if !ok || strings.TrimSpace(effort) == "" {
+									return false, "", false
+								}
+								effort = strings.ToLower(strings.TrimSpace(effort))
+								if normalized, okLevel := util.NormalizeReasoningEffortLevel(normalizedModel, effort); okLevel {
+									return true, normalized, false
+								}
+								return false, "", true // validation would fail
+							case "codex":
+								if !util.ModelSupportsThinking(normalizedModel) {
+									return false, "", false
+								}
+								if !util.ModelUsesThinkingLevels(normalizedModel) {
+									// Non-levels models don't support effort strings in codex
+									if from != "openai-response" {
+										return false, "", false
+									}
+									return false, "", false
+								}
+								effort, ok := util.ReasoningEffortFromMetadata(metadata)
+								if ok && strings.TrimSpace(effort) != "" {
+									effort = strings.ToLower(strings.TrimSpace(effort))
+									if normalized, okLevel := util.NormalizeReasoningEffortLevel(normalizedModel, effort); okLevel {
+										return true, normalized, false
+									}
+									return false, "", true
+								}
+								if budget, _, _, matched := util.ThinkingFromMetadata(metadata); matched && budget != nil {
+									if mapped, okMap := util.OpenAIThinkingBudgetToEffort(normalizedModel, *budget); okMap && mapped != "" {
+										mapped = strings.ToLower(strings.TrimSpace(mapped))
+										if normalized, okLevel := util.NormalizeReasoningEffortLevel(normalizedModel, mapped); okLevel {
+											return true, normalized, false
+										}
+										return false, "", true
+									}
+								}
+								if from != "openai-response" {
+									// Codex translators default reasoning.effort to "medium" when
+									// no explicit thinking suffix/metadata is provided.
+									return true, "medium", false
+								}
+								return false, "", false
+							default:
+								return false, "", false
+							}
+						}()
+
+						body, err := buildBodyForProtocol(t, from, to, cs.modelSuffix)
+						actualPresent, actualValue := func() (bool, string) {
+							path := ""
+							switch to {
+							case "gemini":
+								path = "generationConfig.thinkingConfig.thinkingBudget"
+							case "claude":
+								path = "thinking.budget_tokens"
+							case "openai":
+								path = "reasoning_effort"
+							case "codex":
+								path = "reasoning.effort"
+							}
+							if path == "" {
+								return false, ""
+							}
+							val := gjson.GetBytes(body, path)
+							if to == "codex" && !val.Exists() {
+								reasoning := gjson.GetBytes(body, "reasoning")
+								if reasoning.Exists() {
+									val = reasoning.Get("effort")
+								}
+							}
+							if !val.Exists() {
+								return false, ""
+							}
+							if val.Type == gjson.Number {
+								return true, fmt.Sprintf("%d", val.Int())
+							}
+							return true, val.String()
+						}()
+
+						t.Logf("from=%s to=%s model=%s suffix=%s present(expect=%v got=%v) value(expect=%s got=%s) err(expect=%v got=%v) body=%s",
+							from, to, model, cs.modelSuffix, expectPresent, actualPresent, expectValue, actualValue, expectErr, err != nil, string(body))
+
+						if expectErr {
+							if err == nil {
+								t.Fatalf("expected validation error but got none, body=%s", string(body))
+							}
+							return
+						}
+						if err != nil {
+							t.Fatalf("unexpected error: %v body=%s", err, string(body))
+						}
+
+						if expectPresent != actualPresent {
+							t.Fatalf("presence mismatch: expect %v got %v body=%s", expectPresent, actualPresent, string(body))
+						}
+						if expectPresent && expectValue != actualValue {
+							t.Fatalf("value mismatch: expect %s got %s body=%s", expectValue, actualValue, string(body))
+						}
+					})
+				}
+			}
+		}
+	}
+}
+
+func TestOpenAIThinkingBudgetToEffortRanges(t *testing.T) {
+	cleanup := registerCoreModels(t)
+	defer cleanup()
+
+	cases := []struct {
+		name   string
+		model  string
+		budget int
+		want   string
+		ok     bool
+	}{
+		{name: "zero-none", model: "gpt-5", budget: 0, want: "none", ok: true},
+		{name: "low-min", model: "gpt-5", budget: 1, want: "low", ok: true},
+		{name: "low-max", model: "gpt-5", budget: 1024, want: "low", ok: true},
+		{name: "medium-min", model: "gpt-5", budget: 1025, want: "medium", ok: true},
+		{name: "medium-max", model: "gpt-5", budget: 8192, want: "medium", ok: true},
+		{name: "high-min", model: "gpt-5", budget: 8193, want: "high", ok: true},
+		{name: "high-max", model: "gpt-5", budget: 24576, want: "high", ok: true},
+		{name: "over-max-clamps-to-highest", model: "gpt-5", budget: 24577, want: "high", ok: true},
+		{name: "over-max-xhigh-model", model: "gpt-5.2", budget: 50000, want: "xhigh", ok: true},
+		{name: "negative-unsupported", model: "gpt-5", budget: -5, want: "", ok: false},
+	}
+
+	for _, cs := range cases {
+		cs := cs
+		t.Run(cs.name, func(t *testing.T) {
+			got, ok := util.OpenAIThinkingBudgetToEffort(cs.model, cs.budget)
+			if ok != cs.ok {
+				t.Fatalf("ok mismatch for model=%s budget=%d: expect %v got %v", cs.model, cs.budget, cs.ok, ok)
+			}
+			if got != cs.want {
+				t.Fatalf("value mismatch for model=%s budget=%d: expect %q got %q", cs.model, cs.budget, cs.want, got)
+			}
+		})
+	}
+}

From 660aabc437585b42825522fd1250f403a0f24516 Mon Sep 17 00:00:00 2001
From: Luis Pater <webmaster@idotorg.org>
Date: Sat, 13 Dec 2025 04:06:02 +0800
Subject: [PATCH 4/4] fix(executor): add `allowCompat` support for reasoning
 effort normalization

Introduced `allowCompat` parameter to improve compatibility handling for reasoning effort in payloads across OpenAI and similar models.
---
 internal/runtime/executor/codex_executor.go   | 10 +++---
 internal/runtime/executor/iflow_executor.go   |  8 ++---
 .../executor/openai_compat_executor.go        | 31 ++++++++++++++++---
 internal/runtime/executor/payload_helpers.go  | 14 ++++-----
 internal/runtime/executor/qwen_executor.go    |  8 ++---
 5 files changed, 47 insertions(+), 24 deletions(-)

diff --git a/internal/runtime/executor/codex_executor.go b/internal/runtime/executor/codex_executor.go
index 866f1bdc..2e966237 100644
--- a/internal/runtime/executor/codex_executor.go
+++ b/internal/runtime/executor/codex_executor.go
@@ -54,8 +54,8 @@ func (e *CodexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("codex")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort")
-	body = normalizeThinkingConfig(body, upstreamModel)
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort", false)
+	body = normalizeThinkingConfig(body, upstreamModel, false)
 	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
 		return resp, errValidate
 	}
@@ -152,8 +152,8 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 	to := sdktranslator.FromString("codex")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
 
-	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort")
-	body = normalizeThinkingConfig(body, upstreamModel)
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort", false)
+	body = normalizeThinkingConfig(body, upstreamModel, false)
 	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
 		return nil, errValidate
 	}
@@ -254,7 +254,7 @@ func (e *CodexExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth
 
 	modelForCounting := req.Model
 
-	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort")
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort", false)
 	body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	body, _ = sjson.DeleteBytes(body, "previous_response_id")
 	body, _ = sjson.SetBytes(body, "stream", false)
diff --git a/internal/runtime/executor/iflow_executor.go b/internal/runtime/executor/iflow_executor.go
index 4424ea0f..03df1be0 100644
--- a/internal/runtime/executor/iflow_executor.go
+++ b/internal/runtime/executor/iflow_executor.go
@@ -57,12 +57,12 @@ 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 = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort")
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
 	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
 	if upstreamModel != "" {
 		body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	}
-	body = normalizeThinkingConfig(body, upstreamModel)
+	body = normalizeThinkingConfig(body, upstreamModel, false)
 	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
 		return resp, errValidate
 	}
@@ -148,12 +148,12 @@ 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 = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort")
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
 	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
 	if upstreamModel != "" {
 		body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	}
-	body = normalizeThinkingConfig(body, upstreamModel)
+	body = normalizeThinkingConfig(body, upstreamModel, false)
 	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
 		return nil, errValidate
 	}
diff --git a/internal/runtime/executor/openai_compat_executor.go b/internal/runtime/executor/openai_compat_executor.go
index d4b0afcb..1bbd0c8e 100644
--- a/internal/runtime/executor/openai_compat_executor.go
+++ b/internal/runtime/executor/openai_compat_executor.go
@@ -59,12 +59,13 @@ 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 = applyReasoningEffortMetadata(translated, req.Metadata, req.Model, "reasoning_effort")
+	allowCompat := e.allowCompatReasoningEffort(req.Model, auth)
+	translated = applyReasoningEffortMetadata(translated, req.Metadata, req.Model, "reasoning_effort", allowCompat)
 	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
 	if upstreamModel != "" && modelOverride == "" {
 		translated, _ = sjson.SetBytes(translated, "model", upstreamModel)
 	}
-	translated = normalizeThinkingConfig(translated, upstreamModel)
+	translated = normalizeThinkingConfig(translated, upstreamModel, allowCompat)
 	if errValidate := validateThinkingConfig(translated, upstreamModel); errValidate != nil {
 		return resp, errValidate
 	}
@@ -154,12 +155,13 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 		translated = e.overrideModel(translated, modelOverride)
 	}
 	translated = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", translated)
-	translated = applyReasoningEffortMetadata(translated, req.Metadata, req.Model, "reasoning_effort")
+	allowCompat := e.allowCompatReasoningEffort(req.Model, auth)
+	translated = applyReasoningEffortMetadata(translated, req.Metadata, req.Model, "reasoning_effort", allowCompat)
 	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
 	if upstreamModel != "" && modelOverride == "" {
 		translated, _ = sjson.SetBytes(translated, "model", upstreamModel)
 	}
-	translated = normalizeThinkingConfig(translated, upstreamModel)
+	translated = normalizeThinkingConfig(translated, upstreamModel, allowCompat)
 	if errValidate := validateThinkingConfig(translated, upstreamModel); errValidate != nil {
 		return nil, errValidate
 	}
@@ -325,6 +327,27 @@ func (e *OpenAICompatExecutor) resolveUpstreamModel(alias string, auth *cliproxy
 	return ""
 }
 
+func (e *OpenAICompatExecutor) allowCompatReasoningEffort(model string, auth *cliproxyauth.Auth) bool {
+	trimmed := strings.TrimSpace(model)
+	if trimmed == "" || e == nil || e.cfg == nil {
+		return false
+	}
+	compat := e.resolveCompatConfig(auth)
+	if compat == nil || len(compat.Models) == 0 {
+		return false
+	}
+	for i := range compat.Models {
+		entry := compat.Models[i]
+		if strings.EqualFold(strings.TrimSpace(entry.Alias), trimmed) {
+			return true
+		}
+		if strings.EqualFold(strings.TrimSpace(entry.Name), trimmed) {
+			return true
+		}
+	}
+	return false
+}
+
 func (e *OpenAICompatExecutor) resolveCompatConfig(auth *cliproxyauth.Auth) *config.OpenAICompatibility {
 	if auth == nil || e.cfg == nil {
 		return nil
diff --git a/internal/runtime/executor/payload_helpers.go b/internal/runtime/executor/payload_helpers.go
index b791dac7..6e352c51 100644
--- a/internal/runtime/executor/payload_helpers.go
+++ b/internal/runtime/executor/payload_helpers.go
@@ -48,7 +48,7 @@ func applyThinkingMetadataCLI(payload []byte, metadata map[string]any, model str
 // applyReasoningEffortMetadata applies reasoning effort overrides from metadata to the given JSON path.
 // Metadata values take precedence over any existing field when the model supports thinking, intentionally
 // overwriting caller-provided values to honor suffix/default metadata priority.
-func applyReasoningEffortMetadata(payload []byte, metadata map[string]any, model, field string) []byte {
+func applyReasoningEffortMetadata(payload []byte, metadata map[string]any, model, field string, allowCompat bool) []byte {
 	if len(metadata) == 0 {
 		return payload
 	}
@@ -59,20 +59,20 @@ func applyReasoningEffortMetadata(payload []byte, metadata map[string]any, model
 	if baseModel == "" {
 		baseModel = model
 	}
-	if !util.ModelSupportsThinking(baseModel) && !util.IsOpenAICompatibilityModel(baseModel) {
+	if !util.ModelSupportsThinking(baseModel) && !allowCompat {
 		return payload
 	}
 	if effort, ok := util.ReasoningEffortFromMetadata(metadata); ok && effort != "" {
-		if util.ModelUsesThinkingLevels(model) {
+		if util.ModelUsesThinkingLevels(baseModel) || allowCompat {
 			if updated, err := sjson.SetBytes(payload, field, effort); err == nil {
 				return updated
 			}
 		}
 	}
 	// Fallback: numeric thinking_budget suffix for level-based (OpenAI-style) models.
-	if util.ModelUsesThinkingLevels(model) {
+	if util.ModelUsesThinkingLevels(baseModel) || allowCompat {
 		if budget, _, _, matched := util.ThinkingFromMetadata(metadata); matched && budget != nil {
-			if effort, ok := util.OpenAIThinkingBudgetToEffort(model, *budget); ok && effort != "" {
+			if effort, ok := util.OpenAIThinkingBudgetToEffort(baseModel, *budget); ok && effort != "" {
 				if updated, err := sjson.SetBytes(payload, field, effort); err == nil {
 					return updated
 				}
@@ -237,13 +237,13 @@ func matchModelPattern(pattern, model string) bool {
 // reasoning fields. For models with level-based thinking, it validates and
 // normalizes the reasoning effort level. For models with numeric budget thinking,
 // it strips the effort string fields.
-func normalizeThinkingConfig(payload []byte, model string) []byte {
+func normalizeThinkingConfig(payload []byte, model string, allowCompat bool) []byte {
 	if len(payload) == 0 || model == "" {
 		return payload
 	}
 
 	if !util.ModelSupportsThinking(model) {
-		if util.IsOpenAICompatibilityModel(model) {
+		if allowCompat {
 			return payload
 		}
 		return stripThinkingFields(payload, false)
diff --git a/internal/runtime/executor/qwen_executor.go b/internal/runtime/executor/qwen_executor.go
index c8e2974a..3bd61021 100644
--- a/internal/runtime/executor/qwen_executor.go
+++ b/internal/runtime/executor/qwen_executor.go
@@ -51,12 +51,12 @@ 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 = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort")
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
 	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
 	if upstreamModel != "" {
 		body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	}
-	body = normalizeThinkingConfig(body, upstreamModel)
+	body = normalizeThinkingConfig(body, upstreamModel, false)
 	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
 		return resp, errValidate
 	}
@@ -131,12 +131,12 @@ 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 = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort")
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
 	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
 	if upstreamModel != "" {
 		body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	}
-	body = normalizeThinkingConfig(body, upstreamModel)
+	body = normalizeThinkingConfig(body, upstreamModel, false)
 	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
 		return nil, errValidate
 	}