fix: prevent DeepSeek 400 error on direct mode switch from reasoning-capable model

Author: Luis Carmona

packages/types/src/message.ts

@@ -170,6 +170,7 @@ export const clineSays = [
 	"codebase_search_result",
 	"user_edit_todos",
 	"too_many_tools_warning",
+	"mode_switch_compatibility_warning",
 	"tool",
 ] as const
 

src/api/providers/deepseek.ts

@@ -1,6 +1,8 @@
 import { Anthropic } from "@anthropic-ai/sdk"
 import OpenAI from "openai"
 
+import { logger } from "../../utils/logging"
+
 import {
 	deepSeekModels,
 	deepSeekDefaultModelId,
@@ -14,6 +16,8 @@ import type { ApiHandlerOptions } from "../../shared/api"
 import { ApiStream, ApiStreamUsageChunk } from "../transform/stream"
 import { getModelParams } from "../transform/model-params"
 import { convertToR1Format } from "../transform/r1-format"
+import { historyHasToolCallsWithoutReasoning } from "./utils/reasoning-history-guard"
+import { historyHasToolCallsWithoutReasoning } from "./utils/reasoning-history-guard"
 
 import { OpenAiHandler } from "./openai"
 import { extractReasoningFromDelta } from "./utils/extract-reasoning"
@@ -96,12 +100,6 @@ export class DeepSeekHandler extends OpenAiHandler {
 		const { info: modelInfo, temperature, reasoningEffort, maxTokens } = this.getModel()
 
 		const isThinkingModel = isDeepSeekThinkingEnabled(modelId, this.options)
-		const thinking = supportsDeepSeekThinkingToggle(modelId)
-			? ({ type: isThinkingModel ? "enabled" : "disabled" } as const)
-			: isThinkingModel
-				? ({ type: "enabled" } as const)
-				: undefined
-		const deepSeekReasoningEffort = isThinkingModel ? normalizeDeepSeekReasoningEffort(reasoningEffort) : undefined
 
 		// Convert messages to R1 format (merges consecutive same-role messages)
 		// This is required for DeepSeek which does not support successive messages with the same role
@@ -113,9 +111,30 @@ export class DeepSeekHandler extends OpenAiHandler {
 			mergeToolResultText: isThinkingModel,
 		})
 
+		// Layer 1 Guard: detect incompatible history (tool_calls without reasoning_content)
+		// and disable thinking mode for this request to prevent a 400 error from the API.
+		const hasIncompatibleHistory = historyHasToolCallsWithoutReasoning(convertedMessages)
+		const effectiveThinkingEnabled = isThinkingModel && !hasIncompatibleHistory
+
+		if (hasIncompatibleHistory) {
+			logger.warn("provider_reasoning_guard_triggered", {
+				ctx: "deepseek",
+				provider: "deepseek",
+				modelId,
+				taskId: metadata?.taskId,
+			})
+		}
+
+		const thinking = supportsDeepSeekThinkingToggle(modelId)
+			? ({ type: effectiveThinkingEnabled ? "enabled" : "disabled" } as const)
+			: effectiveThinkingEnabled
+				? ({ type: "enabled" } as const)
+				: undefined
+		const deepSeekReasoningEffort = effectiveThinkingEnabled ? normalizeDeepSeekReasoningEffort(reasoningEffort) : undefined
+
 		const requestOptions: DeepSeekChatCompletionParams = {
 			model: modelId,
-			...(!isThinkingModel && { temperature: temperature ?? DEEP_SEEK_DEFAULT_TEMPERATURE }),
+			...(!effectiveThinkingEnabled && { temperature: temperature ?? DEEP_SEEK_DEFAULT_TEMPERATURE }),
 			messages: convertedMessages,
 			stream: true as const,
 			stream_options: { include_usage: true },

src/api/providers/mimo.ts

@@ -14,6 +14,8 @@ import { extractReasoningFromDelta } from "./utils/extract-reasoning"
 import { OpenAiHandler } from "./openai"
 import type { ApiHandlerCreateMessageMetadata } from "../index"
 import { sanitizeOpenAiCallId } from "../../utils/tool-id"
+import { historyHasToolCallsWithoutReasoning } from "./utils/reasoning-history-guard"
+import { logger } from "../../utils/logging"
 
 /**
  * MiMoHandler extends OpenAiHandler with MiMo-specific adaptations.
@@ -81,6 +83,20 @@ export class MimoHandler extends OpenAiHandler {
 
 		const tools = metadata?.tools
 
+		// Layer 1 Guard: detect incompatible history (tool_calls without reasoning_content)
+		// and disable thinking mode for this request to prevent a 400 error from the API.
+		// MiMo previously had NO disable path — this closes that gap.
+		const hasIncompatibleHistory = historyHasToolCallsWithoutReasoning(convertedMessages)
+
+		if (hasIncompatibleHistory) {
+			logger.warn("provider_reasoning_guard_triggered", {
+				ctx: "mimo",
+				provider: "mimo",
+				modelId,
+				taskId: metadata?.taskId,
+			})
+		}
+
 		// Build request per MiMo's OpenAI-compatible API
 		// https://developer.puter.com/ai/xiaomi/mimo-v2.5-pro/
 		// Note: temperature is omitted because MiMo forces it to 1.0 when thinking mode
@@ -91,7 +107,7 @@ export class MimoHandler extends OpenAiHandler {
 			stream: true,
 			stream_options: { include_usage: true },
 			// MiMo requires thinking to be enabled via extra_body
-			extra_body: { thinking: { type: "enabled" } },
+			extra_body: { thinking: { type: hasIncompatibleHistory ? "disabled" : "enabled" } },
 		}
 
 		if (tools && tools.length > 0) {

src/api/providers/utils/__tests__/reasoning-history-guard.spec.ts

@@ -0,0 +1,107 @@
+// npx vitest run api/providers/utils/__tests__/reasoning-history-guard.spec.ts
+
+import { historyHasToolCallsWithoutReasoning } from "../reasoning-history-guard"
+
+describe("historyHasToolCallsWithoutReasoning", () => {
+	it("returns false for empty messages", () => {
+		expect(historyHasToolCallsWithoutReasoning([])).toBe(false)
+	})
+
+	it("returns false when no assistant messages have tool_calls", () => {
+		const messages = [
+			{ role: "user", content: "hello" },
+			{ role: "assistant", content: "hi there" },
+		]
+		expect(historyHasToolCallsWithoutReasoning(messages)).toBe(false)
+	})
+
+	it("returns false when assistant messages have tool_calls with reasoning_content", () => {
+		const messages = [
+			{
+				role: "assistant",
+				content: null,
+				tool_calls: [{ id: "call_1", function: { name: "test" } }],
+				reasoning_content: "I should call test because...",
+			},
+		]
+		expect(historyHasToolCallsWithoutReasoning(messages)).toBe(false)
+	})
+
+	it("returns true when assistant messages have tool_calls but no reasoning_content field", () => {
+		const messages = [
+			{
+				role: "assistant",
+				content: null,
+				tool_calls: [{ id: "call_1", function: { name: "test" } }],
+			},
+		]
+		expect(historyHasToolCallsWithoutReasoning(messages)).toBe(true)
+	})
+
+	it("returns true when assistant messages have tool_calls with empty reasoning_content", () => {
+		const messages = [
+			{
+				role: "assistant",
+				content: null,
+				tool_calls: [{ id: "call_1", function: { name: "test" } }],
+				reasoning_content: "",
+			},
+		]
+		expect(historyHasToolCallsWithoutReasoning(messages)).toBe(true)
+	})
+
+	it("returns true when assistant messages have empty tool_calls array", () => {
+		const messages = [
+			{
+				role: "assistant",
+				content: "hello",
+				tool_calls: [],
+			},
+		]
+		expect(historyHasToolCallsWithoutReasoning(messages)).toBe(false)
+	})
+
+	it("returns false for non-assistant messages with tool_calls", () => {
+		const messages = [
+			{
+				role: "user",
+				content: "hello",
+				tool_calls: [{ id: "call_1" }],
+			},
+		]
+		expect(historyHasToolCallsWithoutReasoning(messages)).toBe(false)
+	})
+
+	it("returns true when at least one assistant message is missing reasoning_content despite tool_calls", () => {
+		const messages = [
+			{
+				role: "assistant",
+				content: "Let me think...",
+				reasoning_content: "thinking step 1",
+			},
+			{
+				role: "assistant",
+				content: null,
+				tool_calls: [{ id: "call_1", function: { name: "read_file" } }],
+				// no reasoning_content — this is the problematic message
+			},
+			{
+				role: "tool",
+				content: "file content",
+				tool_call_id: "call_1",
+			},
+		]
+		expect(historyHasToolCallsWithoutReasoning(messages)).toBe(true)
+	})
+
+	it("handles non-array tool_calls gracefully", () => {
+		const messages = [
+			{
+				role: "assistant",
+				content: null,
+				tool_calls: "not-an-array" as unknown as unknown[],
+			},
+		]
+		expect(historyHasToolCallsWithoutReasoning(messages)).toBe(false)
+	})
+})

src/api/providers/utils/reasoning-history-guard.ts

@@ -0,0 +1,26 @@
+/**
+ * Detects whether a converted OpenAI-format message history contains any
+ * assistant message with `tool_calls` but no non-empty `reasoning_content`.
+ * Used as a guard before enabling strict provider "thinking" modes that
+ * require reasoning_content to accompany every tool-call turn.
+ *
+ * This function operates on messages *after* conversion to the provider's
+ * OpenAI-compatible format (e.g. `convertToR1Format`, `convertToZAiFormat`),
+ * because it is only in the converted format that `reasoning_content`
+ * presence can be reliably determined.
+ *
+ * @param messages - Array of converted messages in OpenAI-compatible format
+ * @returns `true` if any assistant message has tool_calls but lacks
+ *          non-empty reasoning_content
+ */
+export function historyHasToolCallsWithoutReasoning(
+	messages: Array<{ role?: string; tool_calls?: unknown[]; reasoning_content?: unknown }>,
+): boolean {
+	return messages.some(
+		(m) =>
+			m.role === "assistant" &&
+			Array.isArray(m.tool_calls) &&
+			m.tool_calls.length > 0 &&
+			(typeof m.reasoning_content !== "string" || m.reasoning_content.length === 0),
+	)
+}

src/api/providers/zai.ts

@@ -13,10 +13,12 @@ import {
 
 import { type ApiHandlerOptions, getModelMaxOutputTokens } from "../../shared/api"
 import { convertToZAiFormat } from "../transform/zai-format"
+import { historyHasToolCallsWithoutReasoning } from "./utils/reasoning-history-guard"
 
 import type { ApiHandlerCreateMessageMetadata } from "../index"
 import { BaseOpenAiCompatibleProvider } from "./base-openai-compatible-provider"
 import { handleOpenAIError } from "./utils/error-handler"
+import { logger } from "../../utils/logging"
 
 // Custom interface for Z.ai params to support thinking mode and reasoning effort tiers.
 // Z.ai accepts the standard `reasoning_effort` ladder (none/minimal/low/medium/high/xhigh/max)
@@ -107,6 +109,22 @@ export class ZAiHandler extends BaseOpenAiCompatibleProvider<string> {
 		// Use Z.ai format to preserve reasoning_content and merge post-tool text into tool messages
 		const convertedMessages = convertToZAiFormat(messages, { mergeToolResultText: true })
 
+		// Layer 1 Guard: detect incompatible history (tool_calls without reasoning_content)
+		// and disable thinking mode for this request to prevent a 400 error from the API.
+		const hasIncompatibleHistory = historyHasToolCallsWithoutReasoning(convertedMessages)
+
+		if (hasIncompatibleHistory) {
+			logger.warn("provider_reasoning_guard_triggered", {
+				ctx: "zai",
+				provider: "zai",
+				model,
+				taskId: metadata?.taskId,
+			})
+		}
+
+		// When incompatible history is detected, force reasoning off regardless of user settings.
+		const effectiveReasoning = hasIncompatibleHistory ? false : useReasoning
+
 		const params: ZAiChatCompletionParams = {
 			model,
 			max_tokens,
@@ -115,8 +133,8 @@ export class ZAiHandler extends BaseOpenAiCompatibleProvider<string> {
 			stream: true,
 			stream_options: { include_usage: true },
 			// Thinking is ON by default for these models, so explicitly disable it when needed.
-			thinking: useReasoning ? { type: "enabled" } : { type: "disabled" },
-			reasoning_effort: reasoningEffort,
+			thinking: effectiveReasoning ? { type: "enabled" } : { type: "disabled" },
+			reasoning_effort: effectiveReasoning ? reasoningEffort : undefined,
 			tools: this.convertToolsForOpenAI(metadata?.tools),
 			tool_choice: metadata?.tool_choice,
 			parallel_tool_calls: metadata?.parallelToolCalls ?? true,

src/core/tools/SwitchModeTool.ts

@@ -56,7 +56,8 @@ export class SwitchModeTool extends BaseTool<"switch_mode"> {
 			}
 
 			// Switch the mode using shared handler
-			await task.providerRef.deref()?.handleModeSwitch(mode_slug)
+			// via: "switch_mode" — explicit tool call
+			await task.providerRef.deref()?.handleModeSwitch(mode_slug, "switch_mode")
 
 			pushToolResult(
 				`Successfully switched from ${getModeBySlug(currentMode)?.name ?? currentMode} mode to ${

src/core/webview/ClineProvider.ts

@@ -95,6 +95,8 @@ import { t } from "../../i18n"
 import { buildApiHandler } from "../../api"
 import { forceFullModelDetailsLoad, hasLoadedFullDetails } from "../../api/providers/fetchers/lmstudio"
 
+import { logger } from "../../utils/logging"
+import { modeSwitchRisksReasoningIncompatibility } from "../../shared/reasoning-mode-compatibility"
 import { ContextProxy } from "../config/ContextProxy"
 import { ProviderSettingsManager } from "../config/ProviderSettingsManager"
 import { CustomModesManager } from "../config/CustomModesManager"
@@ -1481,14 +1483,68 @@ export class ClineProvider
 	/**
 	 * Handle switching to a new mode, including updating the associated API configuration
 	 * @param newMode The mode to switch to
+	 * @param via - The origin of the mode switch:
+	 *              "switch_mode" — explicit switch_mode tool call
+	 *              "new_task_delegation" — delegation via new_task
+	 *              "unknown_bypass" — cannot be attributed to explicit tool call (default)
 	 */
-	public async handleModeSwitch(newMode: Mode) {
+	public async handleModeSwitch(newMode: Mode, via: "switch_mode" | "new_task_delegation" | "unknown_bypass" = "unknown_bypass") {
 		const task = this.getCurrentTask()
+		const fromMode = (await this.getState())?.mode ?? "unknown"
 
 		if (task) {
 			TelemetryService.instance.captureModeSwitch(task.taskId, newMode)
 			task.emit(RooCodeEventName.TaskModeSwitched, task.taskId, newMode)
 
+			// Layer 2 — Orchestration safety: check for reasoning-mode incompatibility
+			// when switching modes without proper delegation (only relevant for bypass switches).
+			if (via !== "new_task_delegation") {
+				try {
+					const fromProviderName = task.apiConfiguration?.apiProvider
+					const toConfigId = await this.providerSettingsManager.getModeConfigId(newMode)
+					const listApiConfig = await this.providerSettingsManager.listConfig()
+					const toProfile = toConfigId
+						? listApiConfig.find((c) => c.id === toConfigId)
+						: undefined
+					const toProviderName = toProfile
+						? (await this.providerSettingsManager.getProfile({ name: toProfile.name })).apiProvider
+						: fromProviderName
+
+					if (
+						toProviderName &&
+						modeSwitchRisksReasoningIncompatibility(fromProviderName, toProviderName)
+					) {
+						// Behavior A — always show visible warning in chat
+						await task.say(
+							"mode_switch_compatibility_warning",
+							JSON.stringify({
+								fromMode,
+								toMode: newMode,
+								fromProvider: fromProviderName,
+								toProvider: toProviderName,
+								via,
+							}),
+							undefined,
+							undefined,
+							undefined,
+							undefined,
+							{ isNonInteractive: true },
+						)
+
+						// Behavior B — optional auto-condense when setting is enabled
+						const autoCondense = this.context.workspaceState.get("autoCondenseOnRiskyModeSwitch", false)
+						if (autoCondense) {
+							await task.condenseContext()
+						}
+					}
+				} catch (innerError) {
+					// Non-fatal: log but don't block the mode switch if the check fails.
+					this.log(
+						\`Mode-switch compatibility check failed: ${innerError instanceof Error ? innerError.message : String(innerError)}\`,
+					)
+				}
+			}
+
 			try {
 				// Update the task history with the new mode first.
 				const taskHistoryItem =