feat(llm): improve model-specific patches

src/llms/OpenAIClient.ts

@@ -3,13 +3,7 @@
  */
 import { InvokeError, InvokeErrorType } from './errors'
 import type { InvokeResult, LLMClient, Message, OpenAIClientConfig, Tool } from './types'
-import { zodToOpenAITool } from './utils'
+import { modelPatch, zodToOpenAITool } from './utils'
-
-// Claude's openAI-API has different format for some fields
-const CLAUDE_PATCH = {
-	tool_choice: { type: 'tool', name: 'AgentOutput' },
-	thinking: { type: 'disabled' },
-}
 
 export class OpenAIClient implements LLMClient {
 	config: OpenAIClientConfig
@@ -26,9 +20,10 @@ export class OpenAIClient implements LLMClient {
 		// 1. Convert tools to OpenAI format
 		const openaiTools = Object.entries(tools).map(([name, tool]) => zodToOpenAITool(name, tool))
 
-		// 2. Detect if Claude (auto-compatibility)
+		// 2. Detect patch (auto-compatibility)
 		// TODO: Gemini also uses slightly different format than OpenAI
 		const isClaude = this.config.model.toLowerCase().startsWith('claude')
+		const isGrok = this.config.model.toLowerCase().startsWith('grok')
 
 		// 3. Call API
 		let response: Response
@@ -39,7 +34,8 @@ export class OpenAIClient implements LLMClient {
 					'Content-Type': 'application/json',
 					Authorization: `Bearer ${this.config.apiKey}`,
 				},
-				body: JSON.stringify({
+				body: JSON.stringify(
+					modelPatch({
 						model: this.config.model,
 						temperature: this.config.temperature,
 						max_tokens: this.config.maxTokens,
@@ -54,9 +50,8 @@ export class OpenAIClient implements LLMClient {
 						// reasoning_effort: 'minimal',
 						// verbosity: 'low',
 						parallel_tool_calls: false,
-
+					})
-					...(isClaude ? CLAUDE_PATCH : {}),
+				),
-				}),
 				signal: abortSignal,
 			})
 		} catch (error: unknown) {

src/llms/OpenAILenientClient.ts

@@ -5,13 +5,7 @@ import type { MacroToolInput } from '@/PageAgent'
 
 import { InvokeError, InvokeErrorType } from './errors'
 import type { InvokeResult, LLMClient, Message, OpenAIClientConfig, Tool } from './types'
-import { lenientParseMacroToolCall, zodToOpenAITool } from './utils'
+import { lenientParseMacroToolCall, modelPatch, zodToOpenAITool } from './utils'
-
-// Claude's openAI-API has different format for some fields
-const CLAUDE_PATCH = {
-	tool_choice: { type: 'tool', name: 'AgentOutput' },
-	thinking: { type: 'disabled' },
-}
 
 export class OpenAIClient implements LLMClient {
 	config: OpenAIClientConfig
@@ -41,7 +35,8 @@ export class OpenAIClient implements LLMClient {
 					'Content-Type': 'application/json',
 					Authorization: `Bearer ${this.config.apiKey}`,
 				},
-				body: JSON.stringify({
+				body: JSON.stringify(
+					modelPatch({
 						model: this.config.model,
 						temperature: this.config.temperature,
 						max_tokens: this.config.maxTokens,
@@ -56,9 +51,8 @@ export class OpenAIClient implements LLMClient {
 						// reasoning_effort: 'minimal',
 						// verbosity: 'low',
 						parallel_tool_calls: false,
-
+					})
-					...(isClaude ? CLAUDE_PATCH : {}),
+				),
-				}),
 				signal: abortSignal,
 			})
 		} catch (error: unknown) {

src/llms/utils.ts

@@ -192,3 +192,23 @@ export function lenientParseMacroToolCall(
 		)
 	}
 }
+
+export function modelPatch(body: Record<string, any>) {
+	const model: string = body.model || ''
+
+	if (model.toLowerCase().startsWith('claude')) {
+		body.tool_choice = { type: 'tool', name: 'AgentOutput' }
+		body.thinking = { type: 'disabled' }
+		// body.reasoning = { enabled: 'disabled' }
+	}
+
+	if (model.toLowerCase().includes('grok')) {
+		console.log('Applying Grok patch: removing tool_choice')
+		delete body.tool_choice
+		console.log('Applying Grok patch: disable reasoning and thinking')
+		body.thinking = { type: 'disabled', effort: 'minimal' }
+		body.reasoning = { enabled: false, effort: 'low' }
+	}
+
+	return body
+}