fix(runtime): unify reasoning effort metadata overrides

internal/runtime/executor/codex_executor.go

@@ -54,7 +54,7 @@ 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)
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort")
 	body = normalizeThinkingConfig(body, upstreamModel)
 	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
 		return resp, errValidate
@@ -152,7 +152,7 @@ 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)
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort")
 	body = normalizeThinkingConfig(body, upstreamModel)
 	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)
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort")
 	body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	body, _ = sjson.DeleteBytes(body, "previous_response_id")
 	body, _ = sjson.SetBytes(body, "stream", false)

internal/runtime/executor/iflow_executor.go

@@ -57,7 +57,7 @@ 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)
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort")
 	if upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata); upstreamModel != "" {
 		body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	}
@@ -143,7 +143,7 @@ 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)
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort")
 	if upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata); upstreamModel != "" {
 		body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	}

internal/runtime/executor/openai_compat_executor.go

@@ -58,7 +58,7 @@ 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)
+	translated = applyReasoningEffortMetadata(translated, req.Metadata, req.Model, "reasoning_effort")
 	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
 	if upstreamModel != "" {
 		translated, _ = sjson.SetBytes(translated, "model", upstreamModel)
@@ -152,7 +152,7 @@ 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)
+	translated = applyReasoningEffortMetadata(translated, req.Metadata, req.Model, "reasoning_effort")
 	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
 	if upstreamModel != "" {
 		translated, _ = sjson.SetBytes(translated, "model", upstreamModel)

internal/runtime/executor/payload_helpers.go

@@ -45,40 +45,20 @@ func applyThinkingMetadataCLI(payload []byte, metadata map[string]any, model str
 	return util.ApplyGeminiCLIThinkingConfig(payload, budgetOverride, includeOverride)
 }
 
-// applyReasoningEffortMetadata applies reasoning effort overrides (reasoning.effort) when present in metadata.
+// applyReasoningEffortMetadata applies reasoning effort overrides from metadata to the given JSON path.
-// It avoids overwriting an existing reasoning.effort field and only applies to models that support thinking.
+// Metadata values take precedence over any existing field when the model supports thinking.
-func applyReasoningEffortMetadata(payload []byte, metadata map[string]any, model string) []byte {
+func applyReasoningEffortMetadata(payload []byte, metadata map[string]any, model, field string) []byte {
 	if len(metadata) == 0 {
 		return payload
 	}
 	if !util.ModelSupportsThinking(model) {
 		return payload
 	}
-	if gjson.GetBytes(payload, "reasoning.effort").Exists() {
+	if field == "" {
 		return payload
 	}
 	if effort, ok := util.ReasoningEffortFromMetadata(metadata); ok && effort != "" {
-		if updated, err := sjson.SetBytes(payload, "reasoning.effort", effort); err == nil {
+		if updated, err := sjson.SetBytes(payload, field, 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
 		}
 	}

internal/runtime/executor/qwen_executor.go

@@ -51,7 +51,7 @@ 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)
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort")
 	if upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata); upstreamModel != "" {
 		body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	}
@@ -126,7 +126,7 @@ 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)
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort")
 	if upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata); upstreamModel != "" {
 		body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	}

internal/util/thinking_suffix.go

@@ -55,16 +55,42 @@ func NormalizeThinkingModel(modelName string) (string, map[string]any) {
 			value := modelName[idx+len("-thinking-"):]
 			if value != "" {
 				if parsed, ok := parseIntPrefix(value); ok {
-					baseModel = modelName[:idx]
+					candidateBase := modelName[:idx]
-					budgetOverride = &parsed
+					if ModelUsesThinkingLevels(candidateBase) {
-					matched = true
+						baseModel = candidateBase
+						// Numeric suffix on level-aware models should still surface as reasoning effort metadata.
+						raw := strings.ToLower(strings.TrimSpace(value))
+						if raw != "" {
+							reasoningEffort = &raw
+						}
+						matched = true
+					} else {
+						baseModel = candidateBase
+						budgetOverride = &parsed
+						matched = true
+					}
 				} else {
 					baseModel = modelName[:idx]
 					if normalized, ok := NormalizeReasoningEffortLevel(baseModel, value); ok {
 						reasoningEffort = &normalized
 						matched = true
+					} else if !ModelUsesThinkingLevels(baseModel) {
+						// Keep unknown effort tokens so callers can honor user intent even without normalization.
+						raw := strings.ToLower(strings.TrimSpace(value))
+						if raw != "" {
+							reasoningEffort = &raw
+							matched = true
+						} else {
+							baseModel = modelName
+						}
 					} else {
-						baseModel = modelName
+						raw := strings.ToLower(strings.TrimSpace(value))
+						if raw != "" {
+							reasoningEffort = &raw
+							matched = true
+						} else {
+							baseModel = modelName
+						}
 					}
 				}
 			}