Merge pull request #563 from Szpadel/r1-display-reason

Display reasoning for supported openrouter models

src/api/providers/openrouter.ts

@@ -10,6 +10,7 @@ import delay from "delay"
 // Add custom interface for OpenRouter params
 type OpenRouterChatCompletionParams = OpenAI.Chat.ChatCompletionCreateParams & {
 	transforms?: string[]
+	include_reasoning?: boolean
 }
 
 // Add custom interface for OpenRouter usage chunk
@@ -110,14 +111,23 @@ export class OpenRouterHandler implements ApiHandler, SingleCompletionHandler {
 				maxTokens = 8_192
 				break
 		}
+
+		let temperature = 0
+		switch (this.getModel().id) {
+			case "deepseek/deepseek-r1":
+				// Recommended temperature for DeepSeek reasoning models
+				temperature = 0.6
+		}
+
 		// https://openrouter.ai/docs/transforms
 		let fullResponseText = ""
 		const stream = await this.client.chat.completions.create({
 			model: this.getModel().id,
 			max_tokens: maxTokens,
-			temperature: 0,
+			temperature: temperature,
 			messages: openAiMessages,
 			stream: true,
+			include_reasoning: true,
 			// This way, the transforms field will only be included in the parameters when openRouterUseMiddleOutTransform is true.
 			...(this.options.openRouterUseMiddleOutTransform && { transforms: ["middle-out"] }),
 		} as OpenRouterChatCompletionParams)
@@ -137,6 +147,12 @@ export class OpenRouterHandler implements ApiHandler, SingleCompletionHandler {
 			}
 
 			const delta = chunk.choices[0]?.delta
+			if ("reasoning" in delta && delta.reasoning) {
+				yield {
+					type: "reasoning",
+					text: delta.reasoning,
+				} as ApiStreamChunk
+			}
 			if (delta?.content) {
 				fullResponseText += delta.content
 				yield {

src/api/transform/stream.ts

@@ -1,11 +1,16 @@
 export type ApiStream = AsyncGenerator<ApiStreamChunk>
-export type ApiStreamChunk = ApiStreamTextChunk | ApiStreamUsageChunk
+export type ApiStreamChunk = ApiStreamTextChunk | ApiStreamUsageChunk | ApiStreamReasoningChunk
 
 export interface ApiStreamTextChunk {
 	type: "text"
 	text: string
 }
 
+export interface ApiStreamReasoningChunk {
+	type: "reasoning"
+	text: string
+}
+
 export interface ApiStreamUsageChunk {
 	type: "usage"
 	inputTokens: number

src/core/Cline.ts

@@ -2219,7 +2219,7 @@ export class Cline {
 		}
 
 		/*
-		Seeing out of bounds is fine, it means that the next too call is being built up and ready to add to assistantMessageContent to present. 
+		Seeing out of bounds is fine, it means that the next too call is being built up and ready to add to assistantMessageContent to present.
 		When you see the UI inactive during this, it means that a tool is breaking without presenting any UI. For example the write_to_file tool was breaking when relpath was undefined, and for invalid relpath it never presented UI.
 		*/
 		this.presentAssistantMessageLocked = false // this needs to be placed here, if not then calling this.presentAssistantMessage below would fail (sometimes) since it's locked
@@ -2391,9 +2391,14 @@ export class Cline {
 
 			const stream = this.attemptApiRequest(previousApiReqIndex) // yields only if the first chunk is successful, otherwise will allow the user to retry the request (most likely due to rate limit error, which gets thrown on the first chunk)
 			let assistantMessage = ""
+			let reasoningMessage = ""
 			try {
 				for await (const chunk of stream) {
 					switch (chunk.type) {
+						case "reasoning":
+							reasoningMessage += chunk.text
+							await this.say("reasoning", reasoningMessage, undefined, true)
+							break
 						case "usage":
 							inputTokens += chunk.inputTokens
 							outputTokens += chunk.outputTokens

src/shared/ExtensionMessage.ts

@@ -121,6 +121,7 @@ export interface ClineMessage {
 	text?: string
 	images?: string[]
 	partial?: boolean
+	reasoning?: string
 }
 
 export type ClineAsk =
@@ -142,6 +143,7 @@ export type ClineSay =
 	| "api_req_started"
 	| "api_req_finished"
 	| "text"
+	| "reasoning"
 	| "completion_result"
 	| "user_feedback"
 	| "user_feedback_diff"

webview-ui/src/components/chat/ChatRow.tsx

@@ -15,6 +15,7 @@ import { vscode } from "../../utils/vscode"
 import CodeAccordian, { removeLeadingNonAlphanumeric } from "../common/CodeAccordian"
 import CodeBlock, { CODE_BLOCK_BG_COLOR } from "../common/CodeBlock"
 import MarkdownBlock from "../common/MarkdownBlock"
+import ReasoningBlock from "./ReasoningBlock"
 import Thumbnails from "../common/Thumbnails"
 import McpResourceRow from "../mcp/McpResourceRow"
 import McpToolRow from "../mcp/McpToolRow"
@@ -79,6 +80,14 @@ export const ChatRowContent = ({
 	isStreaming,
 }: ChatRowContentProps) => {
 	const { mcpServers, alwaysAllowMcp } = useExtensionState()
+	const [reasoningCollapsed, setReasoningCollapsed] = useState(false)
+
+	// Auto-collapse reasoning when new messages arrive
+	useEffect(() => {
+		if (!isLast && message.say === "reasoning") {
+			setReasoningCollapsed(true)
+		}
+	}, [isLast, message.say])
 	const [cost, apiReqCancelReason, apiReqStreamingFailedMessage] = useMemo(() => {
 		if (message.text != null && message.say === "api_req_started") {
 			const info: ClineApiReqInfo = JSON.parse(message.text)
@@ -472,6 +481,14 @@ export const ChatRowContent = ({
 	switch (message.type) {
 		case "say":
 			switch (message.say) {
+				case "reasoning":
+					return (
+						<ReasoningBlock
+							content={message.text || ""}
+							isCollapsed={reasoningCollapsed}
+							onToggleCollapse={() => setReasoningCollapsed(!reasoningCollapsed)}
+						/>
+					)
 				case "api_req_started":
 					return (
 						<>

webview-ui/src/components/chat/ReasoningBlock.tsx

@@ -0,0 +1,70 @@
+import React, { useEffect, useRef } from "react"
+import { CODE_BLOCK_BG_COLOR } from "../common/CodeBlock"
+import MarkdownBlock from "../common/MarkdownBlock"
+
+interface ReasoningBlockProps {
+	content: string
+	isCollapsed?: boolean
+	onToggleCollapse?: () => void
+	autoHeight?: boolean
+}
+
+const ReasoningBlock: React.FC<ReasoningBlockProps> = ({
+	content,
+	isCollapsed = false,
+	onToggleCollapse,
+	autoHeight = false,
+}) => {
+	const contentRef = useRef<HTMLDivElement>(null)
+
+	// Scroll to bottom when content updates
+	useEffect(() => {
+		if (contentRef.current && !isCollapsed) {
+			contentRef.current.scrollTop = contentRef.current.scrollHeight
+		}
+	}, [content, isCollapsed])
+
+	return (
+		<div
+			style={{
+				backgroundColor: CODE_BLOCK_BG_COLOR,
+				border: "1px solid var(--vscode-editorGroup-border)",
+				borderRadius: "3px",
+				overflow: "hidden",
+			}}>
+			<div
+				onClick={onToggleCollapse}
+				style={{
+					padding: "8px 12px",
+					cursor: "pointer",
+					userSelect: "none",
+					display: "flex",
+					alignItems: "center",
+					justifyContent: "space-between",
+					borderBottom: isCollapsed ? "none" : "1px solid var(--vscode-editorGroup-border)",
+				}}>
+				<span style={{ fontWeight: "bold" }}>Reasoning</span>
+				<span className={`codicon codicon-chevron-${isCollapsed ? "right" : "down"}`}></span>
+			</div>
+			{!isCollapsed && (
+				<div
+					ref={contentRef}
+					style={{
+						padding: "8px 12px",
+						maxHeight: autoHeight ? "none" : "160px",
+						overflowY: "auto",
+					}}>
+					<div
+						style={{
+							fontSize: "13px",
+							opacity: 0.9,
+						}}>
+						<MarkdownBlock markdown={content} />
+					</div>
+				</div>
+			)}
+		</div>
+	)
+}
+
+export default ReasoningBlock