Vercel AI SDK Migration

For Claude: REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.

Goal: Replace raw Anthropic/OpenAI SDKs with the Vercel AI SDK's unified streamText API, aligned with OpenCode Zen's recommended provider packages.

Architecture: Create 4 provider instances (Anthropic, OpenAI, Google, OpenAI-compatible) each pointing at the Zen gateway. Replace 3 separate streaming generator functions with a single streamText call that picks the right provider based on model.provider.npm. Keep the existing SSE protocol to the frontend unchanged.

Tech Stack: ai@4, @ai-sdk/anthropic@1, @ai-sdk/openai@1, @ai-sdk/google@1, Hono SSE streaming (note: @ai-sdk/openai-compatible dropped due to esm.sh/Deno zod 4 incompatibility — openai-compatible models use createOpenAI with provider.chat() instead)

Files:

• Modify: backend/routes/api.ts:1-10 (imports)
• Modify: backend/routes/api.ts:68-90 (client factories)

Step 1: Replace imports

Remove:

Create val
import Anthropic from "https://esm.sh/@anthropic-ai/sdk@0.39.0";
import OpenAI from "https://esm.sh/openai@4.96.0";

Add:

Create val
import { streamText } from "https://esm.sh/ai@4";
import { createAnthropic } from "https://esm.sh/@ai-sdk/anthropic@1";
import { createOpenAI } from "https://esm.sh/@ai-sdk/openai@1";
import { createGoogleGenerativeAI } from "https://esm.sh/@ai-sdk/google@1";
import { createOpenAICompatible } from "https://esm.sh/@ai-sdk/openai-compatible@1";

Note: use major-version-only pins on esm.sh — it resolves to latest within that major.

Step 2: Replace client factories

Remove getAnthropicClient(), getOpenAIClient(), and their cached variables (lines 68-90).

Add a single getProvider function:

Create val
function getProviderModel(model: any) {
  const apiKey = Deno.env.get("OPENCODE_API_KEY") || "";
  const npm = model.provider?.npm || "";

  if (npm === "@ai-sdk/anthropic") {
    const provider = createAnthropic({
      baseURL: "https://opencode.ai/zen/v1",
      apiKey,
    });
    return provider(model.id);
  }

  if (npm === "@ai-sdk/openai") {
    const provider = createOpenAI({
      baseURL: "https://opencode.ai/zen/v1",
      apiKey,
    });
    return provider(model.id);
  }

  if (npm === "@ai-sdk/google") {
    const provider = createGoogleGenerativeAI({
      baseURL: "https://opencode.ai/zen/v1",
      apiKey,
    });
    return provider(model.id);
  }

  // Default: openai-compatible (GLM, Qwen, Kimi, MiniMax, etc.)
  const provider = createOpenAICompatible({
    name: "opencode",
    baseURL: "https://opencode.ai/zen/v1/chat/completions",
    apiKey,
  });
  return provider.chatModel(model.id);
}

Step 3: Commit

git add backend/routes/api.ts
git commit -m "refactor: replace raw SDKs with Vercel AI SDK provider instances"

Files:

• Modify: backend/routes/api.ts:92-208 (streaming functions)

Step 1: Remove old streaming functions

Delete streamAnthropic, streamOpenAIResponses, streamChatCompletions, and streamModel (lines 92-208).

Step 2: Write new unified streamModel

Create val
async function* streamModel(
  model: any,
  messages: { role: string; content: string }[],
): AsyncGenerator<StreamEvent> {
  try {
    const result = streamText({
      model: getProviderModel(model),
      messages: messages.map((m) => ({
        role: m.role as "user" | "assistant" | "system",
        content: m.content,
      })),
      maxTokens: Math.min(4096, model.limit.output),
    });

    let fullContent = "";
    for await (const chunk of result.textStream) {
      fullContent += chunk;
      yield { type: "chunk", data: { model_id: model.id, content: chunk } };
    }

    yield {
      type: "done",
      data: { model_id: model.id, full_content: fullContent },
    };
  } catch (err: any) {
    console.error(`[streamModel] ${model.id} error:`, err);
    yield { type: "error", data: { model_id: model.id, error: err.message } };
  }
}

This preserves the exact same StreamEvent interface and SSE protocol — no frontend changes needed.

Step 3: Commit

git add backend/routes/api.ts
git commit -m "refactor: replace 3 streaming functions with unified streamText"

Step 1: Push to Val Town and test

Open the app, create a conversation selecting models from different providers:

• A Claude model (Anthropic provider)
• A GPT model (OpenAI provider)
• A GLM or Kimi model (openai-compatible provider)

Send a message and verify:

• All models stream text chunks in real-time
• SSE events parse correctly (no frontend errors in console)
• Final messages are saved to the database
• Conversation history works on follow-up messages

Step 2: Check for Google models if any are selected

If Gemini models are in your model list, test one to verify the Google provider path works.

Step 3: Commit any fixes