fix(executor): properly handle thinking application errors

internal/runtime/executor/aistudio_executor.go

@@ -393,7 +393,10 @@ func (e *AIStudioExecutor) translateRequest(req cliproxyexecutor.Request, opts c
 	}
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, stream)
 	payload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), stream)
-	payload, _ = thinking.ApplyThinking(payload, req.Model, "gemini")
+	payload, err := thinking.ApplyThinking(payload, req.Model, "gemini")
+	if err != nil {
+		return nil, translatedPayload{}, err
+	}
 	payload = fixGeminiImageAspectRatio(baseModel, payload)
 	payload = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", payload, originalTranslated)
 	payload, _ = sjson.DeleteBytes(payload, "generationConfig.maxOutputTokens")

internal/runtime/executor/antigravity_executor.go

@@ -136,7 +136,10 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
 	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	translated, _ = thinking.ApplyThinking(translated, req.Model, "antigravity")
+	translated, err = thinking.ApplyThinking(translated, req.Model, "antigravity")
+	if err != nil {
+		return resp, err
+	}
 
 	// Preserve Claude special handling (use baseModel for registry lookups)
 	translated = normalizeAntigravityThinking(baseModel, translated, isClaude)
@@ -254,7 +257,10 @@ func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
 	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 
-	translated, _ = thinking.ApplyThinking(translated, req.Model, "antigravity")
+	translated, err = thinking.ApplyThinking(translated, req.Model, "antigravity")
+	if err != nil {
+		return resp, err
+	}
 
 	// Preserve Claude special handling (use baseModel for registry lookups)
 	translated = normalizeAntigravityThinking(baseModel, translated, true)
@@ -620,7 +626,10 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
 	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 
-	translated, _ = thinking.ApplyThinking(translated, req.Model, "antigravity")
+	translated, err = thinking.ApplyThinking(translated, req.Model, "antigravity")
+	if err != nil {
+		return nil, err
+	}
 
 	// Preserve Claude special handling (use baseModel for registry lookups)
 	translated = normalizeAntigravityThinking(baseModel, translated, isClaude)
@@ -801,7 +810,10 @@ func (e *AntigravityExecutor) CountTokens(ctx context.Context, auth *cliproxyaut
 	// Prepare payload once (doesn't depend on baseURL)
 	payload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	payload, _ = thinking.ApplyThinking(payload, req.Model, "antigravity")
+	payload, err := thinking.ApplyThinking(payload, req.Model, "antigravity")
+	if err != nil {
+		return cliproxyexecutor.Response{}, err
+	}
 
 	// Preserve Claude special handling (use baseModel for registry lookups)
 	payload = normalizeAntigravityThinking(baseModel, payload, isClaude)

internal/runtime/executor/claude_executor.go

@@ -106,7 +106,10 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), stream)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	body, _ = thinking.ApplyThinking(body, req.Model, "claude")
+	body, err = thinking.ApplyThinking(body, req.Model, "claude")
+	if err != nil {
+		return resp, err
+	}
 
 	if !strings.HasPrefix(baseModel, "claude-3-5-haiku") {
 		body = checkSystemInstructions(body)
@@ -236,7 +239,10 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	body, _ = thinking.ApplyThinking(body, req.Model, "claude")
+	body, err = thinking.ApplyThinking(body, req.Model, "claude")
+	if err != nil {
+		return nil, err
+	}
 
 	body = checkSystemInstructions(body)
 	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)

internal/runtime/executor/codex_executor.go

@@ -96,7 +96,10 @@ func (e *CodexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	body = sdktranslator.TranslateRequest(from, to, baseModel, body, false)
 	body = misc.StripCodexUserAgent(body)
 
-	body, _ = thinking.ApplyThinking(body, req.Model, "codex")
+	body, err = thinking.ApplyThinking(body, req.Model, "codex")
+	if err != nil {
+		return resp, err
+	}
 
 	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
@@ -201,7 +204,10 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 	body = sdktranslator.TranslateRequest(from, to, baseModel, body, true)
 	body = misc.StripCodexUserAgent(body)
 
-	body, _ = thinking.ApplyThinking(body, req.Model, "codex")
+	body, err = thinking.ApplyThinking(body, req.Model, "codex")
+	if err != nil {
+		return nil, err
+	}
 
 	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
 	body, _ = sjson.DeleteBytes(body, "previous_response_id")
@@ -302,7 +308,10 @@ func (e *CodexExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth
 	body = sdktranslator.TranslateRequest(from, to, baseModel, body, false)
 	body = misc.StripCodexUserAgent(body)
 
-	body, _ = thinking.ApplyThinking(body, req.Model, "codex")
+	body, err := thinking.ApplyThinking(body, req.Model, "codex")
+	if err != nil {
+		return cliproxyexecutor.Response{}, err
+	}
 
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 	body, _ = sjson.DeleteBytes(body, "previous_response_id")

internal/runtime/executor/gemini_cli_executor.go

@@ -123,7 +123,10 @@ func (e *GeminiCLIExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
 	basePayload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	basePayload, _ = thinking.ApplyThinking(basePayload, req.Model, "gemini-cli")
+	basePayload, err = thinking.ApplyThinking(basePayload, req.Model, "gemini-cli")
+	if err != nil {
+		return resp, err
+	}
 
 	basePayload = fixGeminiCLIImageAspectRatio(baseModel, basePayload)
 	basePayload = applyPayloadConfigWithRoot(e.cfg, baseModel, "gemini", "request", basePayload, originalTranslated)
@@ -269,7 +272,10 @@ func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyaut
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
 	basePayload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 
-	basePayload, _ = thinking.ApplyThinking(basePayload, req.Model, "gemini-cli")
+	basePayload, err = thinking.ApplyThinking(basePayload, req.Model, "gemini-cli")
+	if err != nil {
+		return nil, err
+	}
 
 	basePayload = fixGeminiCLIImageAspectRatio(baseModel, basePayload)
 	basePayload = applyPayloadConfigWithRoot(e.cfg, baseModel, "gemini", "request", basePayload, originalTranslated)
@@ -473,7 +479,10 @@ func (e *GeminiCLIExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.
 	for range models {
 		payload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-		payload, _ = thinking.ApplyThinking(payload, req.Model, "gemini-cli")
+		payload, err = thinking.ApplyThinking(payload, req.Model, "gemini-cli")
+		if err != nil {
+			return cliproxyexecutor.Response{}, err
+		}
 
 		payload = deleteJSONField(payload, "project")
 		payload = deleteJSONField(payload, "model")

internal/runtime/executor/gemini_executor.go

@@ -120,7 +120,10 @@ func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	body, _ = thinking.ApplyThinking(body, req.Model, "gemini")
+	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
+	if err != nil {
+		return resp, err
+	}
 
 	body = fixGeminiImageAspectRatio(baseModel, body)
 	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
@@ -219,7 +222,10 @@ func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 
-	body, _ = thinking.ApplyThinking(body, req.Model, "gemini")
+	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
+	if err != nil {
+		return nil, err
+	}
 
 	body = fixGeminiImageAspectRatio(baseModel, body)
 	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
@@ -332,7 +338,10 @@ func (e *GeminiExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Aut
 	to := sdktranslator.FromString("gemini")
 	translatedReq := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	translatedReq, _ = thinking.ApplyThinking(translatedReq, req.Model, "gemini")
+	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, "gemini")
+	if err != nil {
+		return cliproxyexecutor.Response{}, err
+	}
 
 	translatedReq = fixGeminiImageAspectRatio(baseModel, translatedReq)
 	respCtx := context.WithValue(ctx, "alt", opts.Alt)

internal/runtime/executor/gemini_vertex_executor.go

@@ -170,7 +170,10 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	body, _ = thinking.ApplyThinking(body, req.Model, "gemini")
+	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
+	if err != nil {
+		return resp, err
+	}
 
 	body = fixGeminiImageAspectRatio(baseModel, body)
 	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
@@ -269,7 +272,10 @@ func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *clip
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	body, _ = thinking.ApplyThinking(body, req.Model, "gemini")
+	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
+	if err != nil {
+		return resp, err
+	}
 
 	body = fixGeminiImageAspectRatio(baseModel, body)
 	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
@@ -369,7 +375,10 @@ func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Conte
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 
-	body, _ = thinking.ApplyThinking(body, req.Model, "gemini")
+	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
+	if err != nil {
+		return nil, err
+	}
 
 	body = fixGeminiImageAspectRatio(baseModel, body)
 	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
@@ -485,7 +494,10 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 
-	body, _ = thinking.ApplyThinking(body, req.Model, "gemini")
+	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
+	if err != nil {
+		return nil, err
+	}
 
 	body = fixGeminiImageAspectRatio(baseModel, body)
 	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
@@ -593,7 +605,10 @@ func (e *GeminiVertexExecutor) countTokensWithServiceAccount(ctx context.Context
 
 	translatedReq := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	translatedReq, _ = thinking.ApplyThinking(translatedReq, req.Model, "gemini")
+	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, "gemini")
+	if err != nil {
+		return cliproxyexecutor.Response{}, err
+	}
 
 	translatedReq = fixGeminiImageAspectRatio(baseModel, translatedReq)
 	translatedReq, _ = sjson.SetBytes(translatedReq, "model", baseModel)
@@ -674,7 +689,10 @@ func (e *GeminiVertexExecutor) countTokensWithAPIKey(ctx context.Context, auth *
 
 	translatedReq := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	translatedReq, _ = thinking.ApplyThinking(translatedReq, req.Model, "gemini")
+	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, "gemini")
+	if err != nil {
+		return cliproxyexecutor.Response{}, err
+	}
 
 	translatedReq = fixGeminiImageAspectRatio(baseModel, translatedReq)
 	translatedReq, _ = sjson.SetBytes(translatedReq, "model", baseModel)

internal/runtime/executor/iflow_executor.go

@@ -92,7 +92,10 @@ func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	body, _ = thinking.ApplyThinking(body, req.Model, "iflow")
+	body, err = thinking.ApplyThinking(body, req.Model, "iflow")
+	if err != nil {
+		return resp, err
+	}
 
 	body = preserveReasoningContentInMessages(body)
 	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
@@ -187,7 +190,10 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	body, _ = thinking.ApplyThinking(body, req.Model, "iflow")
+	body, err = thinking.ApplyThinking(body, req.Model, "iflow")
+	if err != nil {
+		return nil, err
+	}
 
 	body = preserveReasoningContentInMessages(body)
 	// Ensure tools array exists to avoid provider quirks similar to Qwen's behaviour.

internal/runtime/executor/openai_compat_executor.go

@@ -92,7 +92,10 @@ func (e *OpenAICompatExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), opts.Stream)
 	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", translated, originalTranslated)
 
-	translated, _ = thinking.ApplyThinking(translated, req.Model, "openai")
+	translated, err = thinking.ApplyThinking(translated, req.Model, "openai")
+	if err != nil {
+		return resp, err
+	}
 
 	url := strings.TrimSuffix(baseURL, "/") + "/chat/completions"
 	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(translated))
@@ -184,7 +187,10 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", translated, originalTranslated)
 
-	translated, _ = thinking.ApplyThinking(translated, req.Model, "openai")
+	translated, err = thinking.ApplyThinking(translated, req.Model, "openai")
+	if err != nil {
+		return nil, err
+	}
 
 	url := strings.TrimSuffix(baseURL, "/") + "/chat/completions"
 	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(translated))
@@ -291,7 +297,10 @@ func (e *OpenAICompatExecutor) CountTokens(ctx context.Context, auth *cliproxyau
 
 	modelForCounting := baseModel
 
-	translated, _ = thinking.ApplyThinking(translated, req.Model, "openai")
+	translated, err := thinking.ApplyThinking(translated, req.Model, "openai")
+	if err != nil {
+		return cliproxyexecutor.Response{}, err
+	}
 
 	enc, err := tokenizerForModel(modelForCounting)
 	if err != nil {

internal/runtime/executor/qwen_executor.go

@@ -86,7 +86,10 @@ func (e *QwenExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	body, _ = thinking.ApplyThinking(body, req.Model, "openai")
+	body, err = thinking.ApplyThinking(body, req.Model, "openai")
+	if err != nil {
+		return resp, err
+	}
 
 	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
 
@@ -169,7 +172,10 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	body, _ = thinking.ApplyThinking(body, req.Model, "openai")
+	body, err = thinking.ApplyThinking(body, req.Model, "openai")
+	if err != nil {
+		return nil, err
+	}
 
 	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.

internal/thinking/errors.go

@@ -1,6 +1,8 @@
 // Package thinking provides unified thinking configuration processing logic.
 package thinking
 
+import "net/http"
+
 // ErrorCode represents the type of thinking configuration error.
 type ErrorCode string
 
@@ -69,3 +71,8 @@ func NewThinkingErrorWithModel(code ErrorCode, message, model string) *ThinkingE
 		Model:   model,
 	}
 }
+
+// StatusCode implements a portable status code interface for HTTP handlers.
+func (e *ThinkingError) StatusCode() int {
+	return http.StatusBadRequest
+}

internal/thinking/provider/geminicli/apply.go

@@ -55,6 +55,9 @@ func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *
 	if config.Mode == thinking.ModeAuto {
 		return a.applyBudgetFormat(body, config)
 	}
+	if config.Mode == thinking.ModeBudget {
+		return a.applyBudgetFormat(body, config)
+	}
 
 	// For non-auto modes, choose format based on model capabilities
 	support := modelInfo.Thinking