fix: add proper support for DeepSeek V3 models in OpenRouter

- DeepSeek V3 models (v3, 3.2, 3.2-exp) now correctly use standard message format instead of R1 format
- Added explicit model detection to differentiate between DeepSeek R1 and V3 models
- Ensured V3 models have proper tool calling support and reasonable token limits
- Added comprehensive tests to verify correct behavior for both V3 and R1 models

Fixes #9779

src/api/providers/__tests__/openrouter.spec.ts

@@ -338,4 +338,129 @@ describe("OpenRouterHandler", () => {
 			await expect(handler.completePrompt("test prompt")).rejects.toThrow("Unexpected error")
 		})
 	})
+
+	describe("DeepSeek V3 Model Handling", () => {
+		it("should NOT use R1 format for DeepSeek V3 models", async () => {
+			const deepseekV3Handler = new OpenRouterHandler({
+				...mockOptions,
+				openRouterModelId: "deepseek/deepseek-v3",
+			})
+
+			const mockStream = {
+				[Symbol.asyncIterator]: async function* () {
+					yield {
+						choices: [{ delta: { content: "test" }, finish_reason: null }],
+						usage: null,
+					}
+					yield {
+						choices: [{ delta: {}, finish_reason: "stop" }],
+						usage: { prompt_tokens: 10, completion_tokens: 5 },
+					}
+				},
+			}
+
+			const mockCreate = vitest.fn().mockResolvedValue(mockStream)
+			;(OpenAI as any).prototype.chat = {
+				completions: { create: mockCreate },
+			} as any
+
+			const generator = deepseekV3Handler.createMessage("system prompt", [])
+			const chunks = []
+			for await (const chunk of generator) {
+				chunks.push(chunk)
+			}
+
+			// Verify that the messages were NOT converted to R1 format
+			expect(mockCreate).toHaveBeenCalledWith(
+				expect.objectContaining({
+					messages: expect.arrayContaining([
+						expect.objectContaining({ role: "system", content: expect.anything() }),
+					]),
+				}),
+				undefined,
+			)
+		})
+
+		it("should NOT use R1 format for DeepSeek 3.2 models", async () => {
+			const deepseek32Handler = new OpenRouterHandler({
+				...mockOptions,
+				openRouterModelId: "deepseek/deepseek-3.2",
+			})
+
+			const mockStream = {
+				[Symbol.asyncIterator]: async function* () {
+					yield {
+						choices: [{ delta: { content: "test" }, finish_reason: null }],
+						usage: null,
+					}
+					yield {
+						choices: [{ delta: {}, finish_reason: "stop" }],
+						usage: { prompt_tokens: 10, completion_tokens: 5 },
+					}
+				},
+			}
+
+			const mockCreate = vitest.fn().mockResolvedValue(mockStream)
+			;(OpenAI as any).prototype.chat = {
+				completions: { create: mockCreate },
+			} as any
+
+			const generator = deepseek32Handler.createMessage("system prompt", [])
+			const chunks = []
+			for await (const chunk of generator) {
+				chunks.push(chunk)
+			}
+
+			// Verify that the messages were NOT converted to R1 format
+			expect(mockCreate).toHaveBeenCalledWith(
+				expect.objectContaining({
+					messages: expect.arrayContaining([
+						expect.objectContaining({ role: "system", content: expect.anything() }),
+					]),
+				}),
+				undefined,
+			)
+		})
+
+		it("should still use R1 format for DeepSeek R1 models", async () => {
+			const deepseekR1Handler = new OpenRouterHandler({
+				...mockOptions,
+				openRouterModelId: "deepseek/deepseek-r1",
+			})
+
+			const mockStream = {
+				[Symbol.asyncIterator]: async function* () {
+					yield {
+						choices: [{ delta: { content: "test" }, finish_reason: null }],
+						usage: null,
+					}
+					yield {
+						choices: [{ delta: {}, finish_reason: "stop" }],
+						usage: { prompt_tokens: 10, completion_tokens: 5 },
+					}
+				},
+			}
+
+			const mockCreate = vitest.fn().mockResolvedValue(mockStream)
+			;(OpenAI as any).prototype.chat = {
+				completions: { create: mockCreate },
+			} as any
+
+			const generator = deepseekR1Handler.createMessage("system prompt", [])
+			const chunks = []
+			for await (const chunk of generator) {
+				chunks.push(chunk)
+			}
+
+			// Verify that the messages WERE converted to R1 format (user role instead of system)
+			expect(mockCreate).toHaveBeenCalledWith(
+				expect.objectContaining({
+					messages: expect.arrayContaining([
+						expect.objectContaining({ role: "user", content: expect.anything() }),
+					]),
+				}),
+				undefined,
+			)
+		})
+	})
 })

src/api/providers/fetchers/openrouter.ts

@@ -266,5 +266,23 @@ export const parseOpenRouterModel = ({
 		modelInfo.maxTokens = 32768
 	}
 
+	// Configure DeepSeek V3 models properly
+	// These models support standard tool calling but NOT R1 format
+	if (
+		id.startsWith("deepseek/deepseek-v3") ||
+		id.startsWith("deepseek/deepseek-3") ||
+		id === "deepseek/deepseek-chat"
+	) {
+		// Ensure these models are marked as supporting native tools
+		// but NOT reasoning format (they're not R1 models)
+		if (modelInfo.supportsNativeTools === undefined) {
+			modelInfo.supportsNativeTools = true
+		}
+		// Ensure reasonable max tokens if not set
+		if (!modelInfo.maxTokens || modelInfo.maxTokens < 8192) {
+			modelInfo.maxTokens = 8192
+		}
+	}
+
 	return modelInfo
 }

src/api/providers/openrouter.ts

@@ -135,7 +135,8 @@ export class OpenRouterHandler extends BaseProvider implements SingleCompletionH
 			...convertToOpenAiMessages(messages),
 		]
 
-		// DeepSeek highly recommends using user instead of system role.
+		// DeepSeek R1 models require user instead of system role.
+		// Note: DeepSeek V3 models (deepseek-v3, deepseek-3.2, etc.) do NOT use R1 format
 		if (modelId.startsWith("deepseek/deepseek-r1") || modelId === "perplexity/sonar-reasoning") {
 			openAiMessages = convertToR1Format([{ role: "user", content: systemPrompt }, ...messages])
 		}
@@ -388,7 +389,13 @@ export class OpenRouterHandler extends BaseProvider implements SingleCompletionH
 			info = this.endpoints[this.options.openRouterSpecificProvider]
 		}
 
+		// Only DeepSeek R1 models use special temperature and topP settings
+		// DeepSeek V3 models (v3, 3.2, 3.2-exp) use standard settings
 		const isDeepSeekR1 = id.startsWith("deepseek/deepseek-r1") || id === "perplexity/sonar-reasoning"
+		const isDeepSeekV3 =
+			id.startsWith("deepseek/deepseek-v3") ||
+			id.startsWith("deepseek/deepseek-3") ||
+			id === "deepseek/deepseek-chat"
 
 		const params = getModelParams({
 			format: "openrouter",