more models

[ctate]

No description provided.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Preview	Comments	Updated (UTC)
workflow-builder	Ready	Preview	Comment	Nov 26, 2025 9:08am

[vercel]

Additional Suggestions:

1. The default text generation model is not updated to match the new default, and the provider detection logic doesn't handle new model formats like meta/llama-4-scout, google/*, or moonshotai/*, causing incorrect provider assignment.

View Details

📝 Patch Details

diff --git a/lib/workflow-codegen-sdk.ts b/lib/workflow-codegen-sdk.ts
index ce83d8f..7557b73 100644
--- a/lib/workflow-codegen-sdk.ts
+++ b/lib/workflow-codegen-sdk.ts
@@ -493,13 +493,26 @@ export function generateWorkflowSDKCode(
 
   function buildAITextParams(config: Record<string, unknown>): string[] {
     imports.add("import { generateText } from 'ai';");
-    const modelId = (config.aiModel as string) || "gpt-4o-mini";
-    const provider =
-      modelId.startsWith("gpt-") || modelId.startsWith("o1-")
-        ? "openai"
-        : "anthropic";
+    const modelId = (config.aiModel as string) || "meta/llama-4-scout";
+    
+    // Determine the full model string with provider
+    // If the model already contains a "/", it already has a provider prefix, so use as-is
+    let modelString: string;
+    if (modelId.includes("/")) {
+      modelString = modelId;
+    } else {
+      // Infer provider from model name for models without provider prefix
+      const provider =
+        modelId.startsWith("gpt-") || modelId.startsWith("o1-")
+          ? "openai"
+          : modelId.startsWith("claude-")
+            ? "anthropic"
+            : "openai"; // default to openai
+      modelString = `${provider}/${modelId}`;
+    }
+    
     return [
-      `model: "${provider}/${modelId}"`,
+      `model: "${modelString}"`,
       `prompt: \`${convertTemplateToJS((config.aiPrompt as string) || "")}\``,
       "apiKey: process.env.OPENAI_API_KEY!",
     ];

Analysis

Broken provider detection in code generation for text steps with multi-provider models

What fails: buildAITextParams() in lib/workflow-codegen-sdk.ts generates invalid model strings for models that already include a provider prefix, producing strings like "anthropic/meta/llama-4-scout" instead of "meta/llama-4-scout".

How to reproduce:

1. Create a workflow with "Generate Text" action
2. Use the default model (meta/llama-4-scout) or select any model with a provider prefix:
   • meta/llama-* (e.g., meta/llama-4-scout, meta/llama-3.3-70b)
   • google/gemini-* (e.g., google/gemini-2.5-flash)
   • moonshotai/kimi-* (e.g., moonshotai/kimi-k2-0905)
   • openai/gpt-oss-* (e.g., openai/gpt-oss-120b)
3. Generate the workflow code (export or view generated code)
4. The generated code will contain invalid model strings like "anthropic/meta/llama-4-scout"

Result: When the generated code runs with the AI SDK, it attempts to call the Anthropic provider with model name "meta/llama-4-scout", which is not a valid Anthropic model. This causes a runtime error since the AI SDK/AI Gateway cannot recognize the invalid provider/model combination.

Expected: Model strings should be generated in the format "creator/model-name" as expected by the AI SDK:

• Models with provider prefix should pass through unchanged: meta/llama-4-scout → meta/llama-4-scout
• Models without prefix should get the correct provider prepended based on naming: gpt-4o → openai/gpt-4o, claude-* → anthropic/claude-*

Root cause: The provider detection logic (lines 506-512) only checks for gpt-* and o1- prefixes to detect OpenAI models, defaulting all others to Anthropic. This fails for models that already have a provider prefix like meta/llama-4-scout, which don't start with gpt- and get incorrectly prefixed with anthropic/.

Affected model count: 7 broken model string patterns identified:

• Meta Llama models: meta/llama-*
• Google Gemini models: google/gemini-*
• Moonshot Kimi models: moonshotai/kimi-*
• OpenAI OSS models: openai/gpt-oss-*
• Any already-prefixed Anthropic models: anthropic/claude-*
• When explicitly using full models with provider: openai/gpt-4o

2. The default model used in `buildAITextParams` is `"gpt-4o-mini"`\, which is inconsistent with the default `"gpt-5"` at line 250 and conflicts with the codebase\-wide transition to `"meta/llama-4-scout"`\.

View Details

📝 Patch Details

diff --git a/lib/workflow-codegen-sdk.ts b/lib/workflow-codegen-sdk.ts
index ce83d8f..19ed536 100644
--- a/lib/workflow-codegen-sdk.ts
+++ b/lib/workflow-codegen-sdk.ts
@@ -247,11 +247,13 @@ function _generateGenerateTextStepBody(
 ): string {
   imports.add("import { generateText, generateObject } from 'ai';");
   imports.add("import { z } from 'zod';");
-  const modelId = (config.aiModel as string) || "gpt-5";
-  const provider =
-    modelId.startsWith("gpt-") || modelId.startsWith("o1")
-      ? "openai"
-      : "anthropic";
+  const modelId = (config.aiModel as string) || "meta/llama-4-scout";
+  // If model already has a provider prefix (contains "/"), use it as-is
+  const model = modelId.includes("/")
+    ? modelId
+    : modelId.startsWith("gpt-") || modelId.startsWith("o1")
+      ? `openai/${modelId}`
+      : `anthropic/${modelId}`;
   const aiPrompt = (config.aiPrompt as string) || "";
   const convertedPrompt = convertTemplateToJS(aiPrompt);
 
@@ -279,7 +281,7 @@ function _generateGenerateTextStepBody(
       const zodSchema = z.object(schemaShape);
 
       const { object } = await generateObject({
-        model: '${provider}/${modelId}',
+        model: '${model}',
         prompt: finalPrompt,
         schema: zodSchema,
       });
@@ -291,7 +293,7 @@ function _generateGenerateTextStepBody(
   }
   
   const { text } = await generateText({
-    model: '${provider}/${modelId}',
+    model: '${model}',
     prompt: finalPrompt,
   });
   
@@ -493,13 +495,15 @@ export function generateWorkflowSDKCode(
 
   function buildAITextParams(config: Record<string, unknown>): string[] {
     imports.add("import { generateText } from 'ai';");
-    const modelId = (config.aiModel as string) || "gpt-4o-mini";
-    const provider =
-      modelId.startsWith("gpt-") || modelId.startsWith("o1-")
-        ? "openai"
-        : "anthropic";
+    const modelId = (config.aiModel as string) || "meta/llama-4-scout";
+    // If model already has a provider prefix (contains "/"), use it as-is
+    const model = modelId.includes("/")
+      ? modelId
+      : modelId.startsWith("gpt-") || modelId.startsWith("o1")
+        ? `openai/${modelId}`
+        : `anthropic/${modelId}`;
     return [
-      `model: "${provider}/${modelId}"`,
+      `model: "${model}"`,
       `prompt: \`${convertTemplateToJS((config.aiPrompt as string) || "")}\``,
       "apiKey: process.env.OPENAI_API_KEY!",
     ];

Analysis

Default model inconsistency and provider detection bug in buildAITextParams

What fails: The buildAITextParams function (lib/workflow-codegen-sdk.ts line 496) and related code uses incorrect defaults and cannot properly handle pre-prefixed model identifiers, causing generated code to have malformed model references.

How to reproduce:

1. When no aiModel is specified in a workflow configuration, buildAITextParams defaults to "gpt-4o-mini" instead of the codebase standard "meta/llama-4-scout" used everywhere else
2. When a user selects "meta/llama-4-scout" from the model dropdown, the provider detection logic doesn't recognize it has a provider prefix, incorrectly generating model: "anthropic/meta/llama-4-scout"
3. Similarly, models like "anthropic/claude-opus-4.5" become model: "anthropic/anthropic/claude-opus-4.5" (double-prefixed)

Result: Generated code uses non-existent or incorrect model identifiers, causing runtime failures when the workflow SDK generates code.

Expected:

• Default should match the codebase standard: "meta/llama-4-scout" (used in components/workflow/config/action-config.tsx line 438, lib/steps/generate-text.ts line 90)
• Pre-prefixed models like "meta/llama-4-scout", "anthropic/claude-opus-4.5", "google/gemini-2.5-pro" should be used as-is
• Models without prefixes like "gpt-5" and "o1-preview" should be prefixed with their provider

Fix applied:

• Changed default from "gpt-4o-mini" to "meta/llama-4-scout" in both buildAITextParams (line 496) and _generateGenerateTextStepBody (line 244)
• Updated provider detection logic to check if a model already contains "/" (provider prefix) and if so, use it as-is
• For models without a prefix, correctly detect provider: gpt- and o1 prefix → openai, otherwise → anthropic

3. The default text generation model is still `"gpt-5"` instead of the new default `"meta/llama-4-scout"`\, creating inconsistency with the UI and other code generation paths\.

View Details

📝 Patch Details

diff --git a/lib/workflow-codegen.ts b/lib/workflow-codegen.ts
index 6c7640d..3cc9a38 100644
--- a/lib/workflow-codegen.ts
+++ b/lib/workflow-codegen.ts
@@ -408,7 +408,7 @@ export function generateWorkflowCode(
 
     const config = node.data.config || {};
     const aiPrompt = (config.aiPrompt as string) || "Generate a summary";
-    const aiModel = (config.aiModel as string) || "gpt-5";
+    const aiModel = (config.aiModel as string) || "meta/llama-4-scout";
     const aiFormat = (config.aiFormat as string) || "text";
     const aiSchema = config.aiSchema as string | undefined;
 

Analysis

Inconsistent default AI model in workflow code generation

What fails: The _generateGenerateTextStepBody() function in lib/workflow-codegen.ts line 411 uses "gpt-5" as the default model when generating code, but the UI (components/workflow/config/action-config.tsx line 438), display layer (components/workflow/nodes/action-node.tsx line 292), and step execution function (lib/steps/generate-text.ts line 90) all default to "meta/llama-4-scout".

How to reproduce:

1. Create a new workflow with a "Generate Text" action node
2. Do not explicitly select a model in the node configuration
3. Generate the workflow code using the code generation function
4. The generated code will specify model: "gpt-5" instead of model: "meta/llama-4-scout"
5. When the workflow runs, it uses "gpt-5" despite the UI suggesting "meta/llama-4-scout" as the default

Result: Generated code uses model: "gpt-5" when aiModel config is undefined, causing inconsistency between the UI default and the actual code generation default.

Expected: Generated code should use model: "meta/llama-4-scout" to match the default shown in the UI and used elsewhere in the codebase.

Fix: Changed line 411 in lib/workflow-codegen.ts from "gpt-5" to "meta/llama-4-scout" to align with the defaults in action-config.tsx (line 438), action-node.tsx (line 292), and generate-text.ts (line 90).