fix: reasoning_effort parameter silently dropped for OpenAI-compatible providers

Root cause: BaseOpenAiCompatibleProvider and OpenAICompatibleHandler
never called getModelParams() and never read options.reasoningEffort,
causing reasoning_effort to be silently dropped from all API requests
regardless of user settings.

Changes:
- BaseOpenAiCompatibleProvider.getModel() now calls getModelParams()
  (same pattern as OpenAiHandler and DeepSeekHandler)
- Added openAiCustomModelInfo fallback for custom/unknown models
- createStream() and completePrompt() now spread reasoning from getModel()
- OpenAICompatibleHandler now passes reasoning_effort via AI SDK's
  providerOptions.openaiCompatible.reasoningEffort

Affected providers: Baseten, Fireworks, SambaNova, ZAi, Moonshot

Author: bravos2k5

src/api/providers/base-openai-compatible-provider.ts

@@ -7,6 +7,7 @@ import { type ApiHandlerOptions, getModelMaxOutputTokens } from "../../shared/ap
 import { TagMatcher } from "../../utils/tag-matcher"
 import { ApiStream, ApiStreamUsageChunk } from "../transform/stream"
 import { convertToOpenAiMessages } from "../transform/openai-format"
+import { getModelParams } from "../transform/model-params"
 
 import type { SingleCompletionHandler, ApiHandlerCreateMessageMetadata } from "../index"
 import { DEFAULT_HEADERS } from "./constants"
@@ -73,7 +74,7 @@ export abstract class BaseOpenAiCompatibleProvider<ModelName extends string>
 		metadata?: ApiHandlerCreateMessageMetadata,
 		requestOptions?: OpenAI.RequestOptions,
 	) {
-		const { id: model, info } = this.getModel()
+		const { id: model, info, reasoning } = this.getModel()
 
 		// Centralized cap: clamp to 20% of the context window (unless provider-specific exceptions apply)
 		const max_tokens =
@@ -98,8 +99,14 @@ export abstract class BaseOpenAiCompatibleProvider<ModelName extends string>
 			parallel_tool_calls: metadata?.parallelToolCalls ?? true,
 		}
 
-		// Add thinking parameter if reasoning is enabled and model supports it
-		if (this.options.enableReasoningEffort && info.supportsReasoningBinary) {
+		// Add reasoning_effort from centralized model params (handled by getModelParams)
+		if (reasoning) {
+			Object.assign(params, reasoning)
+		}
+
+		// Fallback: Add binary thinking parameter when reasoning_effort not used but
+		// reasoning is still enabled and model supports simple on/off thinking
+		if (!reasoning && this.options.enableReasoningEffort && info.supportsReasoningBinary) {
 			;(params as any).thinking = { type: "enabled" }
 		}
 
@@ -220,15 +227,20 @@ export abstract class BaseOpenAiCompatibleProvider<ModelName extends string>
 	}
 
 	async completePrompt(prompt: string): Promise<string> {
-		const { id: modelId, info: modelInfo } = this.getModel()
+		const { id: modelId, info: modelInfo, reasoning } = this.getModel()
 
 		const params: OpenAI.Chat.Completions.ChatCompletionCreateParams = {
 			model: modelId,
 			messages: [{ role: "user", content: prompt }],
 		}
 
-		// Add thinking parameter if reasoning is enabled and model supports it
-		if (this.options.enableReasoningEffort && modelInfo.supportsReasoningBinary) {
+		// Add reasoning_effort from centralized model params
+		if (reasoning) {
+			Object.assign(params, reasoning)
+		}
+
+		// Fallback: Add binary thinking parameter when reasoning_effort not used
+		if (!reasoning && this.options.enableReasoningEffort && modelInfo.supportsReasoningBinary) {
 			;(params as any).thinking = { type: "enabled" }
 		}
 
@@ -250,11 +262,23 @@ export abstract class BaseOpenAiCompatibleProvider<ModelName extends string>
 	}
 
 	override getModel() {
-		const id =
+		const providerId =
 			this.options.apiModelId && this.options.apiModelId in this.providerModels
 				? (this.options.apiModelId as ModelName)
 				: this.defaultProviderModelId
 
-		return { id, info: this.providerModels[id] }
+		// Allow user to override model info via openAiCustomModelInfo (like OpenAiHandler),
+		// enabling support for custom/unknown models with reasoning effort capability
+		const info: ModelInfo = this.options.openAiCustomModelInfo ?? this.providerModels[providerId]
+
+		const params = getModelParams({
+			format: "openai",
+			modelId: providerId,
+			model: info,
+			settings: this.options,
+			defaultTemperature: this.defaultTemperature,
+		})
+
+		return { id: providerId as string, info, ...params }
 	}
 }

src/api/providers/openai-compatible.ts

@@ -165,6 +165,12 @@ export abstract class OpenAICompatibleHandler extends BaseProvider implements Si
 		const openAiTools = this.convertToolsForOpenAI(metadata?.tools)
 		const aiSdkTools = convertToolsForAiSdk(openAiTools) as ToolSet | undefined
 
+		// Build provider options for reasoning_effort (supported by @ai-sdk/openai-compatible)
+		const modelReasoning = (model as any).reasoning as { reasoning_effort?: string } | undefined
+		const openaiCompatibleOptions = modelReasoning?.reasoning_effort
+			? { reasoningEffort: modelReasoning.reasoning_effort }
+			: undefined
+
 		// Build the request options
 		const requestOptions: Parameters<typeof streamText>[0] = {
 			model: languageModel,
@@ -174,6 +180,9 @@ export abstract class OpenAICompatibleHandler extends BaseProvider implements Si
 			maxOutputTokens: this.getMaxOutputTokens(),
 			tools: aiSdkTools,
 			toolChoice: this.mapToolChoice(metadata?.tool_choice),
+			...(openaiCompatibleOptions
+				? { providerOptions: { openaiCompatible: openaiCompatibleOptions } as any }
+				: {}),
 		}
 
 		// Use streamText for streaming responses
@@ -199,12 +208,22 @@ export abstract class OpenAICompatibleHandler extends BaseProvider implements Si
 	 */
 	async completePrompt(prompt: string): Promise<string> {
 		const languageModel = this.getLanguageModel()
+		const model = this.getModel()
+
+		// Build provider options for reasoning_effort
+		const modelReasoning = (model as any).reasoning as { reasoning_effort?: string } | undefined
+		const openaiCompatibleOptions = modelReasoning?.reasoning_effort
+			? { reasoningEffort: modelReasoning.reasoning_effort }
+			: undefined
 
 		const { text } = await generateText({
 			model: languageModel,
 			prompt,
 			maxOutputTokens: this.getMaxOutputTokens(),
 			temperature: this.config.temperature ?? 0,
+			...(openaiCompatibleOptions
+				? { providerOptions: { openaiCompatible: openaiCompatibleOptions } as any }
+				: {}),
 		})
 
 		return text