fix: Default model selection (#533)

* fix: Default model selection

* cr

apps/open-swe/src/graphs/manager/nodes/classify-message/index.ts

@@ -15,7 +15,7 @@ import {
   loadModel,
   supportsParallelToolCallsParam,
   Task,
-} from "../../../../utils/load-model.js";
+} from "../../../../utils/llms/index.js";
 import { Command, END } from "@langchain/langgraph";
 import { getMessageContentString } from "@open-swe/shared/messages";
 import {

apps/open-swe/src/graphs/manager/utils/generate-issue-fields.ts

@@ -5,7 +5,7 @@ import {
   loadModel,
   supportsParallelToolCallsParam,
   Task,
-} from "../../../utils/load-model.js";
+} from "../../../utils/llms/index.js";
 import { getMessageString } from "../../../utils/message/content.js";
 
 export async function createIssueFieldsFromMessages(

apps/open-swe/src/graphs/planner/nodes/determine-needs-context.ts

@@ -9,7 +9,7 @@ import {
   loadModel,
   supportsParallelToolCallsParam,
   Task,
-} from "../../../utils/load-model.js";
+} from "../../../utils/llms/index.js";
 import { getMissingMessages } from "../../../utils/github/issue-messages.js";
 import { getMessageString } from "../../../utils/message/content.js";
 import { isHumanMessage } from "@langchain/core/messages";

apps/open-swe/src/graphs/planner/nodes/generate-message/index.ts

@@ -2,7 +2,7 @@ import {
   loadModel,
   supportsParallelToolCallsParam,
   Task,
-} from "../../../../utils/load-model.js";
+} from "../../../../utils/llms/index.js";
 import {
   createGetURLContentTool,
   createShellTool,

apps/open-swe/src/graphs/planner/nodes/generate-plan/index.ts

@@ -6,7 +6,7 @@ import {
   loadModel,
   supportsParallelToolCallsParam,
   Task,
-} from "../../../../utils/load-model.js";
+} from "../../../../utils/llms/index.js";
 import {
   PlannerGraphState,
   PlannerGraphUpdate,

apps/open-swe/src/graphs/planner/nodes/notetaker.ts

@@ -9,7 +9,7 @@ import {
   loadModel,
   supportsParallelToolCallsParam,
   Task,
-} from "../../../utils/load-model.js";
+} from "../../../utils/llms/index.js";
 import { getMessageString } from "../../../utils/message/content.js";
 import { formatUserRequestPrompt } from "../../../utils/user-request.js";
 import { formatCustomRulesPrompt } from "../../../utils/custom-rules.js";

apps/open-swe/src/graphs/planner/nodes/rewrite-plan.ts

@@ -17,7 +17,7 @@ import {
   loadModel,
   supportsParallelToolCallsParam,
   Task,
-} from "../../../utils/load-model.js";
+} from "../../../utils/llms/index.js";
 import { FallbackRunnable } from "../../../utils/runtime-fallback.js";
 
 const systemPromptIdentifyChanges = `You are operating as an agentic coding assistant built by LangChain. You've previously been given a task to generate a plan of action for, to address the user's initial request.

apps/open-swe/src/graphs/programmer/nodes/diagnose-error.ts

@@ -18,7 +18,7 @@ import {
   loadModel,
   supportsParallelToolCallsParam,
   Task,
-} from "../../../utils/load-model.js";
+} from "../../../utils/llms/index.js";
 import { z } from "zod";
 import { createLogger, LogLevel } from "../../../utils/logger.js";
 import {

apps/open-swe/src/graphs/programmer/nodes/generate-conclusion.ts

@@ -4,7 +4,7 @@ import {
   GraphUpdate,
   PlanItem,
 } from "@open-swe/shared/open-swe/types";
-import { loadModel, Task } from "../../../utils/load-model.js";
+import { loadModel, Task } from "../../../utils/llms/index.js";
 import { getMessageContentString } from "@open-swe/shared/messages";
 import { getMessageString } from "../../../utils/message/content.js";
 import { createLogger, LogLevel } from "../../../utils/logger.js";

apps/open-swe/src/graphs/programmer/nodes/generate-message/index.ts

@@ -8,7 +8,7 @@ import {
   loadModel,
   supportsParallelToolCallsParam,
   Task,
-} from "../../../../utils/load-model.js";
+} from "../../../../utils/llms/index.js";
 import {
   createShellTool,
   createApplyPatchTool,

apps/open-swe/src/graphs/programmer/nodes/open-pr.ts

@@ -20,7 +20,7 @@ import {
   loadModel,
   supportsParallelToolCallsParam,
   Task,
-} from "../../../utils/load-model.js";
+} from "../../../utils/llms/index.js";
 import { formatPlanPromptWithSummaries } from "../../../utils/plan-prompt.js";
 import { formatUserRequestPrompt } from "../../../utils/user-request.js";
 import { AIMessage, BaseMessage, ToolMessage } from "@langchain/core/messages";

apps/open-swe/src/graphs/programmer/nodes/summarize-history.ts

@@ -5,7 +5,7 @@ import {
   GraphUpdate,
   PlanItem,
 } from "@open-swe/shared/open-swe/types";
-import { loadModel, Task } from "../../../utils/load-model.js";
+import { loadModel, Task } from "../../../utils/llms/index.js";
 import {
   AIMessage,
   BaseMessage,

apps/open-swe/src/graphs/programmer/nodes/update-plan.ts

@@ -10,7 +10,7 @@ import {
   loadModel,
   supportsParallelToolCallsParam,
   Task,
-} from "../../../utils/load-model.js";
+} from "../../../utils/llms/index.js";
 import { z } from "zod";
 import {
   getActiveTask,

apps/open-swe/src/graphs/reviewer/nodes/final-review.ts

@@ -18,7 +18,7 @@ import {
   loadModel,
   supportsParallelToolCallsParam,
   Task,
-} from "../../../utils/load-model.js";
+} from "../../../utils/llms/index.js";
 import { GraphConfig, PlanItem } from "@open-swe/shared/open-swe/types";
 import { z } from "zod";
 import { addTaskPlanToIssue } from "../../../utils/github/issue-task.js";

apps/open-swe/src/graphs/reviewer/nodes/generate-review-actions/index.ts

@@ -2,7 +2,7 @@ import {
   loadModel,
   supportsParallelToolCallsParam,
   Task,
-} from "../../../../utils/load-model.js";
+} from "../../../../utils/llms/index.js";
 import {
   ReviewerGraphState,
   ReviewerGraphUpdate,

apps/open-swe/src/graphs/shared/diagnose-error.ts

@@ -15,7 +15,7 @@ import {
   loadModel,
   supportsParallelToolCallsParam,
   Task,
-} from "../../utils/load-model.js";
+} from "../../utils/llms/index.js";
 import { trackCachePerformance } from "../../utils/caching.js";
 
 const logger = createLogger(LogLevel.INFO, "SharedDiagnoseError");

apps/open-swe/src/tools/search-documents-for/index.ts

@@ -2,7 +2,7 @@ import { tool } from "@langchain/core/tools";
 import { createLogger, LogLevel } from "../../utils/logger.js";
 import { createSearchDocumentForToolFields } from "@open-swe/shared/open-swe/tools";
 import { FireCrawlLoader } from "@langchain/community/document_loaders/web/firecrawl";
-import { loadModel, Task } from "../../utils/load-model.js";
+import { loadModel, Task } from "../../utils/llms/index.js";
 import { GraphConfig, GraphState } from "@open-swe/shared/open-swe/types";
 import { getMessageContentString } from "@open-swe/shared/messages";
 import { DOCUMENT_SEARCH_PROMPT } from "./prompt.js";

apps/open-swe/src/utils/llms/constants.ts

@@ -0,0 +1,33 @@
+export enum Task {
+  /**
+   * Used for programmer tasks. This includes: writing code,
+   * generating plans, taking context gathering actions, etc.
+   */
+  PROGRAMMER = "programmer",
+  /**
+   * Used for routing tasks. This includes: initial request
+   * routing to different agents.
+   */
+  ROUTER = "router",
+  /**
+   * Used for summarizing tasks. This includes: summarizing
+   * the conversation history, summarizing actions taken during
+   * a task execution, etc. Should be a slightly advanced model.
+   */
+  SUMMARIZER = "summarizer",
+}
+
+export const TASK_TO_CONFIG_DEFAULTS_MAP = {
+  [Task.PROGRAMMER]: {
+    modelName: "anthropic:claude-sonnet-4-0",
+    temperature: 0,
+  },
+  [Task.ROUTER]: {
+    modelName: "anthropic:claude-3-5-haiku-latest",
+    temperature: 0,
+  },
+  [Task.SUMMARIZER]: {
+    modelName: "openai:gpt-4.1-mini",
+    temperature: 0,
+  },
+};

apps/open-swe/src/utils/llms/index.ts

@@ -0,0 +1,2 @@
+export * from "./load-model.js";
+export * from "./constants.js";

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

@@ -0,0 +1,28 @@
+import { GraphConfig } from "@open-swe/shared/open-swe/types";
+import { getModelManager } from "./model-manager.js";
+import { FallbackRunnable } from "../runtime-fallback.js";
+import { Task, TASK_TO_CONFIG_DEFAULTS_MAP } from "./constants.js";
+
+export async function loadModel(config: GraphConfig, task: Task) {
+  const modelManager = getModelManager();
+
+  const model = await modelManager.loadModel(config, task);
+  if (!model) {
+    throw new Error(`Model loading returned undefined for task: ${task}`);
+  }
+  const fallbackModel = new FallbackRunnable(model, config, task, modelManager);
+  return fallbackModel;
+}
+
+const MODELS_NO_PARALLEL_TOOL_CALLING = ["openai:o3", "openai:o3-mini"];
+
+export function supportsParallelToolCallsParam(
+  config: GraphConfig,
+  task: Task,
+): boolean {
+  const modelStr =
+    config.configurable?.[`${task}ModelName`] ??
+    TASK_TO_CONFIG_DEFAULTS_MAP[task].modelName;
+
+  return !MODELS_NO_PARALLEL_TOOL_CALLING.some((model) => modelStr === model);
+}

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

@@ -3,10 +3,11 @@ import {
   initChatModel,
 } from "langchain/chat_models/universal";
 import { GraphConfig } from "@open-swe/shared/open-swe/types";
-import { createLogger, LogLevel } from "./logger.js";
-import { Task } from "./load-model.js";
+import { createLogger, LogLevel } from "../logger.js";
+import { Task } from "./constants.js";
 import { isAllowedUser } from "@open-swe/shared/github/allowed-users";
 import { decryptSecret } from "@open-swe/shared/crypto";
+import { TASK_TO_CONFIG_DEFAULTS_MAP } from "./constants.js";
 
 const logger = createLogger(LogLevel.INFO, "ModelManager");
 
@@ -242,19 +243,19 @@ export class ModelManager {
       [Task.PROGRAMMER]: {
         modelName:
           config.configurable?.[`${task}ModelName`] ??
-          "anthropic:claude-sonnet-4-0",
+          TASK_TO_CONFIG_DEFAULTS_MAP[task].modelName,
         temperature: config.configurable?.[`${task}Temperature`] ?? 0,
       },
       [Task.ROUTER]: {
         modelName:
           config.configurable?.[`${task}ModelName`] ??
-          "anthropic:claude-3-5-haiku-latest",
+          TASK_TO_CONFIG_DEFAULTS_MAP[task].modelName,
         temperature: config.configurable?.[`${task}Temperature`] ?? 0,
       },
       [Task.SUMMARIZER]: {
         modelName:
           config.configurable?.[`${task}ModelName`] ??
-          "google-genai:gemini-2.5-pro",
+          TASK_TO_CONFIG_DEFAULTS_MAP[task].modelName,
         temperature: config.configurable?.[`${task}Temperature`] ?? 0,
       },
     };
@@ -307,7 +308,7 @@ export class ModelManager {
       openai: {
         [Task.PROGRAMMER]: "gpt-4o",
         [Task.ROUTER]: "gpt-4o-mini",
-        [Task.SUMMARIZER]: "gpt-4o",
+        [Task.SUMMARIZER]: "gpt-4.1-mini",
       },
     };
 

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

@@ -1,61 +0,0 @@
-import { GraphConfig } from "@open-swe/shared/open-swe/types";
-import { getModelManager } from "./model-manager.js";
-import { FallbackRunnable } from "./runtime-fallback.js";
-
-export enum Task {
-  /**
-   * Used for programmer tasks. This includes: writing code,
-   * generating plans, taking context gathering actions, etc.
-   */
-  PROGRAMMER = "programmer",
-  /**
-   * Used for routing tasks. This includes: initial request
-   * routing to different agents.
-   */
-  ROUTER = "router",
-  /**
-   * Used for summarizing tasks. This includes: summarizing
-   * the conversation history, summarizing actions taken during
-   * a task execution, etc. Should be a slightly advanced model.
-   */
-  SUMMARIZER = "summarizer",
-}
-
-const TASK_TO_CONFIG_DEFAULTS_MAP = {
-  [Task.PROGRAMMER]: {
-    modelName: "anthropic:claude-sonnet-4-0",
-    temperature: 0,
-  },
-  [Task.ROUTER]: {
-    modelName: "anthropic:claude-3-5-haiku-latest",
-    temperature: 0,
-  },
-  [Task.SUMMARIZER]: {
-    modelName: "openai:gpt-4.1-mini",
-    temperature: 0,
-  },
-};
-
-export async function loadModel(config: GraphConfig, task: Task) {
-  const modelManager = getModelManager();
-
-  const model = await modelManager.loadModel(config, task);
-  if (!model) {
-    throw new Error(`Model loading returned undefined for task: ${task}`);
-  }
-  const fallbackModel = new FallbackRunnable(model, config, task, modelManager);
-  return fallbackModel;
-}
-
-const MODELS_NO_PARALLEL_TOOL_CALLING = ["openai:o3", "openai:o3-mini"];
-
-export function supportsParallelToolCallsParam(
-  config: GraphConfig,
-  task: Task,
-): boolean {
-  const modelStr =
-    config.configurable?.[`${task}ModelName`] ??
-    TASK_TO_CONFIG_DEFAULTS_MAP[task].modelName;
-
-  return !MODELS_NO_PARALLEL_TOOL_CALLING.some((model) => modelStr === model);
-}

apps/open-swe/src/utils/mcp-output/index.ts

@@ -1,5 +1,5 @@
 import { GraphConfig } from "@open-swe/shared/open-swe/types";
-import { loadModel, Task } from "../load-model.js";
+import { loadModel, Task } from "../llms/index.js";
 import { createLogger, LogLevel } from "../logger.js";
 import { DOCUMENT_TOC_GENERATION_PROMPT } from "./prompt.js";
 import { getMessageContentString } from "@open-swe/shared/messages";

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

@@ -1,6 +1,6 @@
 import { GraphConfig } from "@open-swe/shared/open-swe/types";
-import { Task } from "./load-model.js";
-import { ModelManager } from "./model-manager.js";
+import { Task } from "./llms/index.js";
+import { ModelManager } from "./llms/model-manager.js";
 import { createLogger, LogLevel } from "./logger.js";
 import { Runnable, RunnableConfig } from "@langchain/core/runnables";
 import { StructuredToolInterface } from "@langchain/core/tools";

packages/shared/src/open-swe/models.ts

@@ -1,20 +1,20 @@
 export const MODEL_OPTIONS = [
-  // {
-  //   label: "Claude Sonnet 4 (Extended Thinking)",
-  //   value: "anthropic:extended-thinking:claude-sonnet-4-0",
-  // },
-  // {
-  //   label: "Claude Opus 4 (Extended Thinking)",
-  //   value: "anthropic:extended-thinking:claude-opus-4-0",
-  // },
+  {
+    label: "Claude Sonnet 4 (Extended Thinking)",
+    value: "anthropic:extended-thinking:claude-sonnet-4-0",
+  },
+  {
+    label: "Claude Opus 4 (Extended Thinking)",
+    value: "anthropic:extended-thinking:claude-opus-4-0",
+  },
   {
     label: "Claude Sonnet 4",
     value: "anthropic:claude-sonnet-4-0",
   },
-  // {
-  //   label: "Claude Opus 4",
-  //   value: "anthropic:claude-opus-4-0",
-  // },
+  {
+    label: "Claude Opus 4",
+    value: "anthropic:claude-opus-4-0",
+  },
   {
     label: "Claude 3.7 Sonnet",
     value: "anthropic:claude-3-7-sonnet-latest",
@@ -27,10 +27,10 @@ export const MODEL_OPTIONS = [
     label: "Claude 3.5 Haiku",
     value: "anthropic:claude-3-5-haiku-latest",
   },
-  // {
-  //   label: "o4",
-  //   value: "openai:o4",
-  // },
+  {
+    label: "o4",
+    value: "openai:o4",
+  },
   {
     label: "o4 mini",
     value: "openai:o4-mini",
@@ -60,12 +60,12 @@ export const MODEL_OPTIONS = [
     value: "openai:gpt-4.1-mini",
   },
   {
-    label: "Gemini 2.5 Pro Preview",
-    value: "google-genai:gemini-2.5-pro-preview-05-06",
+    label: "Gemini 2.5 Pro",
+    value: "google-genai:gemini-2.5-pro",
   },
   {
-    label: "Gemini 2.5 Flash Preview",
-    value: "google-genai:gemini-2.5-flash-preview-05-20",
+    label: "Gemini 2.5 Flash",
+    value: "google-genai:gemini-2.5-flash",
   },
 ];