fix: fallback parallel tool calling fix (#628)

apps/open-swe/src/utils/llms/load-model.ts

@@ -29,7 +29,7 @@ export async function loadModel(
   return fallbackModel;
 }
 
-const MODELS_NO_PARALLEL_TOOL_CALLING = ["openai:o3", "openai:o3-mini"];
+export const MODELS_NO_PARALLEL_TOOL_CALLING = ["openai:o3", "openai:o3-mini"];
 
 export function supportsParallelToolCallsParam(
   config: GraphConfig,

apps/open-swe/src/utils/llms/model-manager.ts

@@ -42,9 +42,9 @@ export enum CircuitState {
 }
 
 export const PROVIDER_FALLBACK_ORDER = [
-  "google-genai",
-  "anthropic",
   "openai",
+  "anthropic",
+  "google-genai",
 ] as const;
 export type Provider = (typeof PROVIDER_FALLBACK_ORDER)[number];
 
@@ -238,10 +238,21 @@ export class ModelManager {
         (!selectedModelConfig ||
           fallbackModel.modelName !== selectedModelConfig.modelName)
       ) {
+        // Check if fallback model is a thinking model
+        const isThinkingModel =
+          (provider === "openai" && fallbackModel.modelName.startsWith("o")) ||
+          fallbackModel.modelName.includes("extended-thinking");
+
         const fallbackConfig = {
           ...fallbackModel,
-          temperature: baseConfig.temperature,
+          temperature: isThinkingModel ? undefined : baseConfig.temperature,
           maxTokens: baseConfig.maxTokens,
+          ...(isThinkingModel
+            ? {
+                thinkingModel: true,
+                thinkingBudgetTokens: THINKING_BUDGET_TOKENS,
+              }
+            : {}),
         };
         configs.push(fallbackConfig);
       }
@@ -348,9 +359,9 @@ export class ModelManager {
         [Task.SUMMARIZER]: "gemini-2.5-pro",
       },
       openai: {
-        [Task.PLANNER]: "gpt-4.1",
+        [Task.PLANNER]: "o3",
         [Task.PROGRAMMER]: "gpt-4.1",
-        [Task.REVIEWER]: "gpt-4.1",
+        [Task.REVIEWER]: "o3",
         [Task.ROUTER]: "gpt-4o-mini",
         [Task.SUMMARIZER]: "gpt-4.1-mini",
       },

apps/open-swe/src/utils/runtime-fallback.ts

@@ -18,6 +18,7 @@ import { BaseLanguageModelInput } from "@langchain/core/language_models/base";
 import { BindToolsInput } from "@langchain/core/language_models/chat_models";
 import { getMessageContentString } from "@open-swe/shared/messages";
 import { getConfig } from "@langchain/langgraph";
+import { MODELS_NO_PARALLEL_TOOL_CALLING } from "./llms/load-model.js";
 
 const logger = createLogger(LogLevel.DEBUG, "FallbackRunnable");
 
@@ -141,9 +142,19 @@ export class FallbackRunnable<
           "bindTools" in runnableToUse &&
           runnableToUse.bindTools
         ) {
+          const supportsParallelToolCall =
+            !MODELS_NO_PARALLEL_TOOL_CALLING.some(
+              (modelName) => modelKey === modelName,
+            );
+
+          const kwargs = { ...toolsToUse.kwargs };
+          if (!supportsParallelToolCall && "parallel_tool_calls" in kwargs) {
+            delete kwargs.parallel_tool_calls;
+          }
+
           runnableToUse = (runnableToUse as ConfigurableModel).bindTools(
             toolsToUse.tools,
-            toolsToUse.kwargs,
+            kwargs,
           );
         }