fix: use Fireworks reasoning_effort parameter for DeepSeek V3.1/V3.2

- Changed from supportsReasoningBinary to supportsReasoningEffort for
  DeepSeek V3.1 and V3.2 models on Fireworks
- Added default reasoningEffort: "medium" for both models
- Override createStream in FireworksHandler to use getModelParams
  which passes reasoning_effort parameter (Fireworks API style) instead
  of thinking parameter (Anthropic style)
- Added tests to verify reasoning_effort is passed correctly

packages/types/src/providers/fireworks.ts

@@ -101,6 +101,8 @@ export const fireworksModels = {
 		supportsImages: false,
 		supportsPromptCache: false,
 		supportsNativeTools: true,
+		supportsReasoningEffort: ["low", "medium", "high"],
+		reasoningEffort: "medium",
 		inputPrice: 0.56,
 		outputPrice: 1.68,
 		description:
@@ -112,7 +114,8 @@ export const fireworksModels = {
 		supportsImages: false,
 		supportsPromptCache: false,
 		supportsNativeTools: true,
-		supportsReasoningBinary: true,
+		supportsReasoningEffort: ["low", "medium", "high"],
+		reasoningEffort: "medium",
 		inputPrice: 0.56,
 		outputPrice: 1.68,
 		description:

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

@@ -214,6 +214,8 @@ describe("FireworksHandler", () => {
 				contextWindow: 163840,
 				supportsImages: false,
 				supportsPromptCache: false,
+				supportsReasoningEffort: ["low", "medium", "high"],
+				reasoningEffort: "medium",
 				inputPrice: 0.56,
 				outputPrice: 1.68,
 				description: expect.stringContaining("DeepSeek v3.1 is an improved version"),
@@ -235,7 +237,8 @@ describe("FireworksHandler", () => {
 				contextWindow: 163840,
 				supportsImages: false,
 				supportsPromptCache: false,
-				supportsReasoningBinary: true,
+				supportsReasoningEffort: ["low", "medium", "high"],
+				reasoningEffort: "medium",
 				inputPrice: 0.56,
 				outputPrice: 1.68,
 				description: expect.stringContaining("DeepSeek v3.2 is the latest version"),
@@ -243,6 +246,70 @@ describe("FireworksHandler", () => {
 		)
 	})
 
+	it("should include reasoning_effort parameter for DeepSeek V3.2 when reasoning is enabled", async () => {
+		const testModelId: FireworksModelId = "accounts/fireworks/models/deepseek-v3p2"
+		const handlerWithModel = new FireworksHandler({
+			apiModelId: testModelId,
+			fireworksApiKey: "test-fireworks-api-key",
+			enableReasoningEffort: true,
+			reasoningEffort: "high",
+		})
+
+		mockCreate.mockImplementationOnce(() => {
+			return {
+				[Symbol.asyncIterator]: () => ({
+					async next() {
+						return { done: true }
+					},
+				}),
+			}
+		})
+
+		const systemPrompt = "Test system prompt"
+		const messages: Anthropic.Messages.MessageParam[] = [{ role: "user", content: "Test message" }]
+
+		const messageGenerator = handlerWithModel.createMessage(systemPrompt, messages)
+		await messageGenerator.next()
+
+		// Check that reasoning_effort was included in the request
+		expect(mockCreate).toHaveBeenCalledWith(
+			expect.objectContaining({
+				model: testModelId,
+				reasoning_effort: "high",
+			}),
+			undefined,
+		)
+	})
+
+	it("should not include reasoning_effort parameter when reasoning is disabled", async () => {
+		const testModelId: FireworksModelId = "accounts/fireworks/models/deepseek-v3p2"
+		const handlerWithModel = new FireworksHandler({
+			apiModelId: testModelId,
+			fireworksApiKey: "test-fireworks-api-key",
+			enableReasoningEffort: false,
+		})
+
+		mockCreate.mockImplementationOnce(() => {
+			return {
+				[Symbol.asyncIterator]: () => ({
+					async next() {
+						return { done: true }
+					},
+				}),
+			}
+		})
+
+		const systemPrompt = "Test system prompt"
+		const messages: Anthropic.Messages.MessageParam[] = [{ role: "user", content: "Test message" }]
+
+		const messageGenerator = handlerWithModel.createMessage(systemPrompt, messages)
+		await messageGenerator.next()
+
+		// Check that reasoning_effort was NOT included in the request
+		const callArgs = mockCreate.mock.calls[0][0]
+		expect(callArgs.reasoning_effort).toBeUndefined()
+	})
+
 	it("should return GLM-4.5 model with correct configuration", () => {
 		const testModelId: FireworksModelId = "accounts/fireworks/models/glm-4p5"
 		const handlerWithModel = new FireworksHandler({

src/api/providers/fireworks.ts

@@ -1,8 +1,16 @@
+import { Anthropic } from "@anthropic-ai/sdk"
+import OpenAI from "openai"
+
 import { type FireworksModelId, fireworksDefaultModelId, fireworksModels } from "@roo-code/types"
 
 import type { ApiHandlerOptions } from "../../shared/api"
+import { getModelParams } from "../transform/model-params"
+import { getModelMaxOutputTokens } from "../../shared/api"
+import { convertToOpenAiMessages } from "../transform/openai-format"
 
+import type { ApiHandlerCreateMessageMetadata } from "../index"
 import { BaseOpenAiCompatibleProvider } from "./base-openai-compatible-provider"
+import { handleOpenAIError } from "./utils/openai-error-handler"
 
 export class FireworksHandler extends BaseOpenAiCompatibleProvider<FireworksModelId> {
 	constructor(options: ApiHandlerOptions) {
@@ -16,4 +24,57 @@ export class FireworksHandler extends BaseOpenAiCompatibleProvider<FireworksMode
 			defaultTemperature: 0.5,
 		})
 	}
+
+	override getModel() {
+		const id =
+			this.options.apiModelId && this.options.apiModelId in this.providerModels
+				? (this.options.apiModelId as FireworksModelId)
+				: this.defaultProviderModelId
+
+		const info = this.providerModels[id]
+		const params = getModelParams({ format: "openai", modelId: id, model: info, settings: this.options })
+		return { id, info, ...params }
+	}
+
+	protected override createStream(
+		systemPrompt: string,
+		messages: Anthropic.Messages.MessageParam[],
+		metadata?: ApiHandlerCreateMessageMetadata,
+		requestOptions?: OpenAI.RequestOptions,
+	) {
+		const { id: model, info, reasoning } = this.getModel()
+
+		// Centralized cap: clamp to 20% of the context window (unless provider-specific exceptions apply)
+		const max_tokens =
+			getModelMaxOutputTokens({
+				modelId: model,
+				model: info,
+				settings: this.options,
+				format: "openai",
+			}) ?? undefined
+
+		const temperature = this.options.modelTemperature ?? this.defaultTemperature
+
+		const params: OpenAI.Chat.Completions.ChatCompletionCreateParamsStreaming = {
+			model,
+			max_tokens,
+			temperature,
+			messages: [{ role: "system", content: systemPrompt }, ...convertToOpenAiMessages(messages)],
+			stream: true,
+			stream_options: { include_usage: true },
+			...(metadata?.tools && { tools: this.convertToolsForOpenAI(metadata.tools) }),
+			...(metadata?.tool_choice && { tool_choice: metadata.tool_choice }),
+			...(metadata?.toolProtocol === "native" && {
+				parallel_tool_calls: metadata.parallelToolCalls ?? false,
+			}),
+			// Use Fireworks-style reasoning_effort parameter instead of Anthropic-style "thinking"
+			...(reasoning && reasoning),
+		}
+
+		try {
+			return this.client.chat.completions.create(params, requestOptions)
+		} catch (error) {
+			throw handleOpenAIError(error, "Fireworks")
+		}
+	}
 }