feat(thinking): normalize effort levels in adaptive thinking requests to prevent validation errors

internal/translator/claude/gemini/claude_gemini_request.go

@@ -120,6 +120,8 @@ func ConvertGeminiRequestToClaude(modelName string, inputRawJSON []byte, stream
 			supportsAdaptive := mi != nil && mi.Thinking != nil && len(mi.Thinking.Levels) > 0
 			supportsMax := supportsAdaptive && thinking.HasLevel(mi.Thinking.Levels, string(thinking.LevelMax))
 
+			// MapToClaudeEffort normalizes levels (e.g. minimal→low, xhigh→high) to avoid
+			// validation errors since validate treats same-provider unsupported levels as errors.
 			thinkingLevel := thinkingConfig.Get("thinkingLevel")
 			if !thinkingLevel.Exists() {
 				thinkingLevel = thinkingConfig.Get("thinking_level")
@@ -134,12 +136,12 @@ func ConvertGeminiRequestToClaude(modelName string, inputRawJSON []byte, stream
 						out, _ = sjson.Delete(out, "thinking.budget_tokens")
 						out, _ = sjson.Delete(out, "output_config.effort")
 					default:
-						effort, ok := thinking.MapToClaudeEffort(level, supportsMax)
-						if ok {
-							out, _ = sjson.Set(out, "thinking.type", "adaptive")
-							out, _ = sjson.Delete(out, "thinking.budget_tokens")
-							out, _ = sjson.Set(out, "output_config.effort", effort)
+						if mapped, ok := thinking.MapToClaudeEffort(level, supportsMax); ok {
+							level = mapped
 						}
+						out, _ = sjson.Set(out, "thinking.type", "adaptive")
+						out, _ = sjson.Delete(out, "thinking.budget_tokens")
+						out, _ = sjson.Set(out, "output_config.effort", level)
 					}
 				} else {
 					switch level {
@@ -173,12 +175,12 @@ func ConvertGeminiRequestToClaude(modelName string, inputRawJSON []byte, stream
 						default:
 							level, ok := thinking.ConvertBudgetToLevel(budget)
 							if ok {
-								effort, ok := thinking.MapToClaudeEffort(level, supportsMax)
-								if ok {
-									out, _ = sjson.Set(out, "thinking.type", "adaptive")
-									out, _ = sjson.Delete(out, "thinking.budget_tokens")
-									out, _ = sjson.Set(out, "output_config.effort", effort)
+								if mapped, okM := thinking.MapToClaudeEffort(level, supportsMax); okM {
+									level = mapped
 								}
+								out, _ = sjson.Set(out, "thinking.type", "adaptive")
+								out, _ = sjson.Delete(out, "thinking.budget_tokens")
+								out, _ = sjson.Set(out, "output_config.effort", level)
 							}
 						}
 					} else {

internal/translator/claude/openai/chat-completions/claude_openai_request.go

@@ -74,6 +74,8 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 			supportsMax := supportsAdaptive && thinking.HasLevel(mi.Thinking.Levels, string(thinking.LevelMax))
 
 			// Claude 4.6 supports adaptive thinking with output_config.effort.
+			// MapToClaudeEffort normalizes levels (e.g. minimal→low, xhigh→high) to avoid
+			// validation errors since validate treats same-provider unsupported levels as errors.
 			if supportsAdaptive {
 				switch effort {
 				case "none":
@@ -85,7 +87,6 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 					out, _ = sjson.Delete(out, "thinking.budget_tokens")
 					out, _ = sjson.Delete(out, "output_config.effort")
 				default:
-					// Map non-Claude effort levels into Claude 4.6 effort vocabulary.
 					if mapped, ok := thinking.MapToClaudeEffort(effort, supportsMax); ok {
 						effort = mapped
 					}

internal/translator/claude/openai/responses/claude_openai-responses_request.go

@@ -62,6 +62,8 @@ func ConvertOpenAIResponsesRequestToClaude(modelName string, inputRawJSON []byte
 			supportsMax := supportsAdaptive && thinking.HasLevel(mi.Thinking.Levels, string(thinking.LevelMax))
 
 			// Claude 4.6 supports adaptive thinking with output_config.effort.
+			// MapToClaudeEffort normalizes levels (e.g. minimal→low, xhigh→high) to avoid
+			// validation errors since validate treats same-provider unsupported levels as errors.
 			if supportsAdaptive {
 				switch effort {
 				case "none":
@@ -73,7 +75,6 @@ func ConvertOpenAIResponsesRequestToClaude(modelName string, inputRawJSON []byte
 					out, _ = sjson.Delete(out, "thinking.budget_tokens")
 					out, _ = sjson.Delete(out, "output_config.effort")
 				default:
-					// Map non-Claude effort levels into Claude 4.6 effort vocabulary.
 					if mapped, ok := thinking.MapToClaudeEffort(effort, supportsMax); ok {
 						effort = mapped
 					}

internal/translator/codex/claude/codex_claude_request.go

@@ -232,19 +232,14 @@ func ConvertClaudeRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 			}
 		case "adaptive", "auto":
 			// Adaptive thinking can carry an explicit effort in output_config.effort (Claude 4.6).
-			// Preserve it when present; otherwise keep the previous "max capacity" sentinel.
+			// Pass through directly; ApplyThinking handles clamping to target model's levels.
 			effort := ""
 			if v := rootResult.Get("output_config.effort"); v.Exists() && v.Type == gjson.String {
 				effort = strings.ToLower(strings.TrimSpace(v.String()))
 			}
-			switch effort {
-			case "minimal", "low", "medium", "high":
+			if effort != "" {
 				reasoningEffort = effort
-			case "max":
-				reasoningEffort = string(thinking.LevelXHigh)
-			default:
-				// Keep adaptive/auto as a high level sentinel; ApplyThinking resolves it
-				// to model-specific max capability.
+			} else {
 				reasoningEffort = string(thinking.LevelXHigh)
 			}
 		case "disabled":

internal/translator/gemini-cli/claude/gemini-cli_claude_request.go

@@ -171,7 +171,8 @@ func ConvertClaudeRequestToCLI(modelName string, inputRawJSON []byte, _ bool) []
 		}
 	}
 
-	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when type==enabled
+	// Map Anthropic thinking -> Gemini CLI thinkingConfig when enabled
+	// Translator only does format conversion, ApplyThinking handles model capability validation.
 	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() {
 		switch t.Get("type").String() {
 		case "enabled":
@@ -181,9 +182,19 @@ func ConvertClaudeRequestToCLI(modelName string, inputRawJSON []byte, _ bool) []
 				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 			}
 		case "adaptive", "auto":
-			// Keep adaptive/auto as a high level sentinel; ApplyThinking resolves it
-			// to model-specific max capability.
-			out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingLevel", "high")
+			// For adaptive thinking:
+			// - If output_config.effort is explicitly present, pass through as thinkingLevel.
+			// - Otherwise, treat it as "enabled with target-model maximum" and emit high.
+			// ApplyThinking handles clamping to target model's supported levels.
+			effort := ""
+			if v := gjson.GetBytes(rawJSON, "output_config.effort"); v.Exists() && v.Type == gjson.String {
+				effort = strings.ToLower(strings.TrimSpace(v.String()))
+			}
+			if effort != "" {
+				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingLevel", effort)
+			} else {
+				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingLevel", "high")
+			}
 			out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 		}
 	}

internal/translator/gemini/claude/gemini_claude_request.go

@@ -164,19 +164,15 @@ func ConvertClaudeRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 			}
 		case "adaptive", "auto":
 			// For adaptive thinking:
-			// - If output_config.effort is explicitly present, map it to thinkingLevel.
+			// - If output_config.effort is explicitly present, pass through as thinkingLevel.
 			// - Otherwise, treat it as "enabled with target-model maximum" and emit thinkingBudget=max.
+			// ApplyThinking handles clamping to target model's supported levels.
 			effort := ""
 			if v := gjson.GetBytes(rawJSON, "output_config.effort"); v.Exists() && v.Type == gjson.String {
 				effort = strings.ToLower(strings.TrimSpace(v.String()))
 			}
 			if effort != "" {
-				level := effort
-				switch level {
-				case "xhigh", "max":
-					level = "high"
-				}
-				out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingLevel", level)
+				out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingLevel", effort)
 			} else {
 				maxBudget := 0
 				if mi := registry.LookupModelInfo(modelName, "gemini"); mi != nil && mi.Thinking != nil {

internal/translator/openai/claude/openai_claude_request.go

@@ -77,19 +77,14 @@ func ConvertClaudeRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 				}
 			case "adaptive", "auto":
 				// Adaptive thinking can carry an explicit effort in output_config.effort (Claude 4.6).
-				// Preserve it when present; otherwise keep the previous "max capacity" sentinel.
+				// Pass through directly; ApplyThinking handles clamping to target model's levels.
 				effort := ""
 				if v := root.Get("output_config.effort"); v.Exists() && v.Type == gjson.String {
 					effort = strings.ToLower(strings.TrimSpace(v.String()))
 				}
-				switch effort {
-				case "minimal", "low", "medium", "high":
+				if effort != "" {
 					out, _ = sjson.Set(out, "reasoning_effort", effort)
-				case "max":
-					out, _ = sjson.Set(out, "reasoning_effort", string(thinking.LevelXHigh))
-				default:
-					// Keep adaptive/auto as a high level sentinel; ApplyThinking resolves it
-					// to model-specific max capability.
+				} else {
 					out, _ = sjson.Set(out, "reasoning_effort", string(thinking.LevelXHigh))
 				}
 			case "disabled":