feat(provider): add low reasoning effort for DeepSeek V4 Flash

README.md

@@ -18,7 +18,7 @@
 **Pick DeepSeek V4 from the Copilot Chat model picker — and keep everything else Copilot already gives you.**
 
 <p align="center">
-  <img src="resources/screenshots/01-picker.png" alt="DeepSeek V4 Pro and Flash in the Copilot Chat model picker, with the per-model Thinking Effort dropdown (None / High / Max)" width="800">
+  <img src="resources/screenshots/01-picker.png" alt="DeepSeek V4 Pro and Flash in the Copilot Chat model picker, with a per-model Thinking Effort dropdown" width="800">
 </p>
 
 Love DeepSeek's price-performance but don't want to give up GitHub Copilot's agent mode, tool calling, and polished UI? This extension drops **DeepSeek V4 Pro & Flash** straight into the Copilot Chat model selector — with **vision**, **thinking mode**, and your own API key.
@@ -45,7 +45,7 @@ This proxy is a compatibility bridge; if DeepSeek native vision becomes availabl
 </p>
 
 ### Thinking Mode with Reasoning Effort Control
-Full support for DeepSeek V4's `reasoning_content`. Use Copilot Chat's native model picker menu to choose `none` (off), `high` (balanced, default), or `max` (deep reasoning for hard agent tasks).
+Full support for DeepSeek V4's `reasoning_content`. DeepSeek V4 Flash offers `none` (off), `low` (light reasoning), `high` (balanced, default), and `max` (deep reasoning for hard agent tasks). DeepSeek V4 Pro currently offers `none`, `high`, and `max`, matching the effort levels implemented by the official API.
 
 ### Inherits Every Copilot Capability
 Because this plugs into Copilot's native provider API, you get the full stack for free:
@@ -88,10 +88,10 @@ Install from the registry used by your editor:
 
 ## Models
 
-| Model | Best For |
-|---|---|
-| **DeepSeek V4 Flash** | Fast everyday coding, quick edits, cheap iteration |
-| **DeepSeek V4 Pro** | Complex refactors, agent tasks, deep reasoning |
+| Model | Thinking Effort | Best For |
+|---|---|---|
+| **DeepSeek V4 Flash** | `none` / `low` / `high` / `max` | Fast everyday coding, quick edits, cheap iteration |
+| **DeepSeek V4 Pro** | `none` / `high` / `max` | Complex refactors, agent tasks, deep reasoning |
 
 Both support optional thinking mode, tool calling, and 1M token context.
 

README.zh-cn.md

@@ -18,7 +18,7 @@
 **在 Copilot Chat 模型选择器中直接使用 DeepSeek V4——无需离开你熟悉的 Copilot 工作流。**
 
 <p align="center">
-  <img src="resources/screenshots/01-picker.png" alt="DeepSeek V4 Pro 和 Flash 出现在 Copilot Chat 模型选择器中，带有可按模型独立设置的思考深度下拉菜单（停用 / 标准 / 深度）" width="800">
+  <img src="resources/screenshots/01-picker.png" alt="DeepSeek V4 Pro 和 Flash 出现在 Copilot Chat 模型选择器中，带有可按模型独立设置的思考深度下拉菜单" width="800">
 </p>
 
 喜欢 DeepSeek 的性价比，但不想放弃 GitHub Copilot 的 Agent 模式、工具调用和成熟的交互体验？本扩展将 **DeepSeek V4 Pro 和 Flash** 直接接入 Copilot Chat 模型选择器，支持**视觉识别**、**思考模式**，使用你自己的 API Key。
@@ -45,7 +45,7 @@ DeepSeek V4 是纯文本模型。将截图拖入聊天，本扩展会自动将
 </p>
 
 ### 思考模式与推理深度控制
-完整支持 DeepSeek V4 的 `reasoning_content`。通过 Copilot Chat 模型选择器的菜单选择 `停用`、`标准`（均衡，默认）或 `深度`（适用于复杂 Agent 任务）。
+完整支持 DeepSeek V4 的 `reasoning_content`。DeepSeek V4 Flash 可选择 `停用`、`轻量`、`标准`（均衡，默认）或 `深度`（适用于复杂 Agent 任务）；DeepSeek V4 Pro 目前提供 `停用`、`标准` 和 `深度`，与官方 API 已实现的推理档位保持一致。
 
 ### 继承全部 Copilot 能力
 由于本扩展接入的是 Copilot 的原生 provider API，你免费获得完整能力栈：
@@ -88,10 +88,10 @@ API Key 存储在 VS Code 的 `SecretStorage` 中（macOS 钥匙串 / Windows 
 
 ## 模型
 
-| 模型 | 适用场景 |
-|---|---|
-| **DeepSeek V4 Flash** | 日常快速编码、小改动、低成本迭代 |
-| **DeepSeek V4 Pro** | 复杂重构、Agent 任务、深度推理 |
+| 模型 | 思考深度 | 适用场景 |
+|---|---|---|
+| **DeepSeek V4 Flash** | `停用` / `轻量` / `标准` / `深度` | 日常快速编码、小改动、低成本迭代 |
+| **DeepSeek V4 Pro** | `停用` / `标准` / `深度` | 复杂重构、Agent 任务、深度推理 |
 
 两者均支持可选的思考模式、工具调用和 1M Token 上下文。
 

src/consts.ts

@@ -59,7 +59,11 @@ export const MODELS: ModelDefinition[] = [
 		capabilities: {
 			toolCalling: DEEPSEEK_TOOLS_LIMIT,
 			imageInput: true,
-			thinking: true,
+			thinking: {
+				supportedEfforts: ['low', 'high', 'max'],
+				defaultEffort: 'high',
+				canDisable: true,
+			},
 		},
 		requiresThinkingParam: true,
 		pricing: {
@@ -79,7 +83,11 @@ export const MODELS: ModelDefinition[] = [
 		capabilities: {
 			toolCalling: DEEPSEEK_TOOLS_LIMIT,
 			imageInput: true,
-			thinking: true,
+			thinking: {
+				supportedEfforts: ['high', 'max'],
+				defaultEffort: 'high',
+				canDisable: true,
+			},
 		},
 		requiresThinkingParam: true,
 		pricing: {

src/i18n.ts

@@ -37,6 +37,8 @@ const zh: Translations = {
 	'status.thinking': '思考模式',
 	'thinking.none': '停用',
 	'thinking.none.desc': '停用思考，响应更快',
+	'thinking.low': '轻量',
+	'thinking.low.desc': '轻量推理，适合快速编辑和简单任务',
 	'thinking.high': '标准',
 	'thinking.high.desc': '推荐日常使用',
 	'thinking.max': '深度',
@@ -227,6 +229,8 @@ const en: Translations = {
 	'status.thinking': 'Thinking Effort',
 	'thinking.none': 'None',
 	'thinking.none.desc': 'Disable thinking for faster responses',
+	'thinking.low': 'Low',
+	'thinking.low.desc': 'Light reasoning for quick edits and simple tasks',
 	'thinking.high': 'High',
 	'thinking.high.desc': 'Recommended for most tasks',
 	'thinking.max': 'Max',

src/provider/models.ts

@@ -1,6 +1,11 @@
 import vscode from 'vscode';
 import { t } from '../i18n';
-import type { ModelDefinition, PricingCurrency } from '../types';
+import type {
+	ModelDefinition,
+	PricingCurrency,
+	ReasoningEffort,
+	ThinkingCapability,
+} from '../types';
 import { toModelCostInfo, type ModelCostInformation } from './pricing/costs';
 
 /**
@@ -14,7 +19,7 @@ import { toModelCostInfo, type ModelCostInformation } from './pricing/costs';
  * metadata and per-model configuration controls.
  */
 
-export type ThinkingEffort = 'none' | 'high' | 'max';
+export type ThinkingEffort = 'none' | ReasoningEffort;
 
 export type ModelConfigurationOptions = vscode.ProvideLanguageModelChatResponseOptions & {
 	readonly modelConfiguration?: Record<string, unknown>;
@@ -38,6 +43,7 @@ export function toChatInfo(
 ): ModelPickerChatInformation {
 	const modelDetail = resolveModelText(m, 'detail') ?? m.detail;
 	const modelTooltip = resolveModelText(m, 'tooltip');
+	const thinkingCapability = m.capabilities.thinking;
 	return {
 		id: m.id,
 		name: m.name,
@@ -55,45 +61,58 @@ export function toChatInfo(
 			imageInput: m.capabilities.imageInput,
 		},
 		...toModelCostInfo(m, pricingCurrency),
-		...(m.capabilities.thinking ? { configurationSchema: buildThinkingEffortSchema() } : {}),
+		...(thinkingCapability
+			? { configurationSchema: buildThinkingEffortSchema(thinkingCapability) }
+			: {}),
 	};
 }
 
-export function getConfiguredThinkingEffort(options: ModelConfigurationOptions): ThinkingEffort {
+export function getConfiguredThinkingEffort(
+	options: ModelConfigurationOptions,
+	thinkingCapability: ThinkingCapability,
+): ThinkingEffort {
 	const configuredEffort =
 		options.modelConfiguration?.reasoningEffort ?? options.configuration?.reasoningEffort;
 
-	if (configuredEffort === 'none') {
+	if (configuredEffort === 'none' && thinkingCapability.canDisable) {
 		return 'none';
 	}
 
-	if (configuredEffort === 'high') {
-		return 'high';
+	if (isSupportedReasoningEffort(configuredEffort, thinkingCapability)) {
+		return configuredEffort;
 	}
 
-	return configuredEffort === 'max' ? 'max' : 'high';
+	return thinkingCapability.defaultEffort;
 }
 
-function buildThinkingEffortSchema() {
+function buildThinkingEffortSchema(thinkingCapability: ThinkingCapability) {
+	const efforts: ThinkingEffort[] = [
+		...(thinkingCapability.canDisable ? (['none'] as const) : []),
+		...thinkingCapability.supportedEfforts,
+	];
+
 	return {
 		properties: {
 			reasoningEffort: {
 				type: 'string',
 				title: t('status.thinking'),
-				enum: ['none', 'high', 'max'],
-				enumItemLabels: [t('thinking.none'), t('thinking.high'), t('thinking.max')],
-				enumDescriptions: [
-					t('thinking.none.desc'),
-					t('thinking.high.desc'),
-					t('thinking.max.desc'),
-				],
-				default: 'high',
+				enum: efforts,
+				enumItemLabels: efforts.map((effort) => t(`thinking.${effort}`)),
+				enumDescriptions: efforts.map((effort) => t(`thinking.${effort}.desc`)),
+				default: thinkingCapability.defaultEffort,
 				group: 'navigation',
 			},
 		},
 	} as const;
 }
 
+function isSupportedReasoningEffort(
+	value: unknown,
+	thinkingCapability: ThinkingCapability,
+): value is ReasoningEffort {
+	return thinkingCapability.supportedEfforts.some((effort) => effort === value);
+}
+
 function resolveModelText(m: ModelDefinition, field: 'detail' | 'tooltip'): string | undefined {
 	const suffix = m.id.startsWith('deepseek-v4-') ? m.id.slice('deepseek-v4-'.length) : m.id;
 	const key = `model.${suffix}.${field}`;

src/provider/request.ts

@@ -64,7 +64,8 @@ export async function prepareChatRequest({
 	const baseUrl = getBaseUrl();
 	const client = new DeepSeekClient(baseUrl, apiKey);
 	const modelDef = MODELS.find((m) => m.id === modelInfo.id);
-	const isThinkingModel = modelDef?.capabilities.thinking ?? false;
+	const thinkingCapability = modelDef?.capabilities.thinking;
+	const isThinkingModel = Boolean(thinkingCapability);
 	const maxTokens = getMaxTokens();
 
 	const visionResolution = await resolveImageMessages(messages, token, getVisionDescriber);
@@ -85,9 +86,9 @@ export async function prepareChatRequest({
 		request: baseRequest,
 		inputMessages: messages,
 	});
-	const configuredThinkingEffort = getConfiguredThinkingEffort(
-		options as ModelConfigurationOptions,
-	);
+	const configuredThinkingEffort = thinkingCapability
+		? getConfiguredThinkingEffort(options as ModelConfigurationOptions, thinkingCapability)
+		: 'none';
 	// Only force helper requests into disabled thinking on the official API.
 	// Custom endpoints keep their configured effort to preserve pre-#137 request shape.
 	const forceNoneThinking =

src/types.ts

@@ -4,6 +4,9 @@
 
 // ---- API request/response types ----
 
+/** API-level reasoning efforts. Disabling thinking is modeled separately. */
+export type ReasoningEffort = 'low' | 'high' | 'max';
+
 export interface DeepSeekMessage {
 	role: 'system' | 'user' | 'assistant' | 'tool';
 	content: string;
@@ -48,7 +51,7 @@ export interface DeepSeekRequest {
 	tools?: DeepSeekTool[];
 	tool_choice?: 'none' | 'auto' | 'required';
 	thinking?: { type: 'enabled' | 'disabled' };
-	reasoning_effort?: 'high' | 'max';
+	reasoning_effort?: ReasoningEffort;
 	stream_options?: {
 		include_usage: boolean;
 	};
@@ -103,6 +106,13 @@ export interface ModelPricing {
 	output: number;
 }
 
+export interface ThinkingCapability {
+	/** Effort values this model implements and may receive in API requests. */
+	supportedEfforts: readonly ReasoningEffort[];
+	defaultEffort: ReasoningEffort;
+	canDisable: boolean;
+}
+
 export interface ModelDefinition {
 	id: string;
 	name: string;
@@ -114,7 +124,7 @@ export interface ModelDefinition {
 	capabilities: {
 		toolCalling: boolean | number;
 		imageInput: boolean;
-		thinking: boolean;
+		thinking: ThinkingCapability | false;
 	};
 	requiresThinkingParam: boolean;
 	pricing?: Readonly<Record<PricingCurrency, ModelPricing>>;