Unify model integrations through AI SDK

README.md

@@ -39,15 +39,16 @@ Install the SDK next to the Restate SDKs it builds on, or clone the
 with one agent and one handler:
 
 ```sh
-npm install @restatedev/agentlib @restatedev/restate-sdk @restatedev/restate-sdk-gen
+npm install @restatedev/agentlib @restatedev/restate-sdk @restatedev/restate-sdk-gen @ai-sdk/openai
 ```
 
 ```ts
 import * as restate from "@restatedev/restate-sdk-gen";
 import {agent} from "@restatedev/agentlib";
-import {openAIModel} from "@restatedev/agentlib/openai";
+import {openai} from "@ai-sdk/openai";
+import {aiModel} from "@restatedev/agentlib/provider";
 
-const assistant = agent({model: openAIModel({model: "gpt-5.6-luna"})});
+const assistant = agent({model: aiModel({model: openai("gpt-5.6-luna")})});
 
 export const hello = restate.service({
   name: "Hello",
@@ -76,9 +77,8 @@ with live controls. Start a Restate server, for example with
 ```sh
 pnpm install --frozen-lockfile
 export OPENAI_API_KEY=...
-# Optional: export AGENT_MODEL=...  An OpenAI model (with OPENAI_BASE_URL, one on
-# any Chat Completions server), or anthropic:<model> or bedrock:<model> with that
-# provider's credentials. Lessons that need other keys say so.
+# Each lesson defines its model locally; edit that definition to change providers.
+# Lessons that need other keys say so.
 pnpm dev
 ```
 
@@ -163,7 +163,7 @@ repeat a completed model call.
 | Long runs                    | `agent({compaction, contextWindow})`, `ContextOverflowError`               | —                                                           |
 | Limits                       | `maxSteps`, `maxToolCalls`, `maxToolDepth`, `maxOutputChars`, `finalize`   | —                                                           |
 | Custom model adapters        | `journaledModel`                                                           | —                                                           |
-| Models                       | `openAIModel` (`/openai`), `aiModel`, `modelGateway` (`/provider`)         | [shared.ts](packages/examples/src/shared.ts)                |
+| Models                       | `aiModel`, `modelGateway` (`/provider`)                                    | [01_hello.ts](packages/examples/src/lessons/01_hello.ts)    |
 
 Subpaths are imports from `@restatedev/agentlib`, such as
 `@restatedev/agentlib/mcp`. `aiModel` accepts an OpenTelemetry `tracer`, and

packages/core/README.md

@@ -6,7 +6,7 @@ code. Model calls, tool results, and approvals are journaled, so a crash,
 deploy, or long wait does not mean starting the work over.
 
 ```sh
-npm install @restatedev/agentlib @restatedev/restate-sdk @restatedev/restate-sdk-gen
+npm install @restatedev/agentlib @restatedev/restate-sdk @restatedev/restate-sdk-gen @ai-sdk/openai
 ```
 
 Or start from the [project template](https://github.com/restatedev/agentlib-template),
@@ -15,9 +15,10 @@ a ready-to-run Restate service with one agent.
 ```ts
 import * as restate from "@restatedev/restate-sdk-gen";
 import {agent} from "@restatedev/agentlib";
-import {openAIModel} from "@restatedev/agentlib/openai";
+import {openai} from "@ai-sdk/openai";
+import {aiModel} from "@restatedev/agentlib/provider";
 
-const assistant = agent({model: openAIModel({model: "gpt-5.6-luna"})});
+const assistant = agent({model: aiModel({model: openai("gpt-5.6-luna")})});
 
 export const hello = restate.service({
   name: "Hello",
@@ -37,14 +38,13 @@ go from a first run to multi-agent workflows.
 
 ## Subpaths
 
-| Import                           | Purpose                                       |
-| -------------------------------- | --------------------------------------------- |
-| `@restatedev/agentlib`           | Agents, tools, conversations, guardrails      |
-| `@restatedev/agentlib/openai`    | OpenAI and Chat Completions-compatible models |
-| `@restatedev/agentlib/provider`  | Models through the Vercel AI SDK providers    |
-| `@restatedev/agentlib/mcp`       | Tools from MCP servers                        |
-| `@restatedev/agentlib/program`   | Programmatic tool calling in a sandbox        |
-| `@restatedev/agentlib/discovery` | Expose agents and tools for discovery         |
+| Import                           | Purpose                                    |
+| -------------------------------- | ------------------------------------------ |
+| `@restatedev/agentlib`           | Agents, tools, conversations, guardrails   |
+| `@restatedev/agentlib/provider`  | Models through the Vercel AI SDK providers |
+| `@restatedev/agentlib/mcp`       | Tools from MCP servers                     |
+| `@restatedev/agentlib/program`   | Programmatic tool calling in a sandbox     |
+| `@restatedev/agentlib/discovery` | Expose agents and tools for discovery      |
 
 ## License
 

packages/core/package.json

@@ -29,10 +29,6 @@
       "types": "./dist/index.d.ts",
       "import": "./dist/index.js"
     },
-    "./openai": {
-      "types": "./dist/openai.d.ts",
-      "import": "./dist/openai.js"
-    },
     "./program": {
       "types": "./dist/program.d.ts",
       "import": "./dist/program.js"
@@ -64,7 +60,6 @@
     "@opentelemetry/api": "1.9.0",
     "ai": "7.0.26",
     "minisearch": "7.2.0",
-    "openai": "^7.17.0",
     "quickjs-emscripten-core": "0.32.0",
     "zod": "^4.6.5"
   },

packages/core/src/provider.ts

@@ -54,6 +54,19 @@ function finished(reason: string | undefined) {
   if (reason !== "stop" && reason !== "tool-calls")
     throw new ModelProtocolError(`Incomplete model response: ${reason ?? "unknown"}`);
 }
+
+function contextOverflow(error: APICallError): boolean {
+  if (error.statusCode === 413) return true;
+  const data: unknown = error.data;
+  if (!data || typeof data !== "object" || !("error" in data)) return false;
+  const detail = data.error;
+  return (
+    !!detail &&
+    typeof detail === "object" &&
+    "code" in detail &&
+    detail.code === "context_length_exceeded"
+  );
+}
 // Namespaced: each provider reads its own key, so no provider detection is needed.
 const CACHE_POINT = {
   anthropic: {cacheControl: {type: "ephemeral"}},
@@ -253,9 +266,9 @@ export function aiModel(options: ProviderOptions): Model {
       // throttling and transient provider failures stay retryable.
       if (LoadAPIKeyError.isInstance(error) || UnsupportedFunctionalityError.isInstance(error))
         throw new TerminalError(error.message);
-      // Providers word an overflow differently and give it no code of its own,
-      // so only a 413 counts here; contextWindow avoids the rest up front.
-      if (APICallError.isInstance(error) && error.statusCode === 413)
+      // Use the SDK's parsed error code or HTTP 413, never provider-specific
+      // message text. contextWindow avoids uncoded overflows up front.
+      if (APICallError.isInstance(error) && contextOverflow(error))
         throw new ContextOverflowError(error.message);
       if (APICallError.isInstance(error) && !error.isRetryable)
         throw new TerminalError(error.message, {errorCode: error.statusCode});

packages/examples/README.md

@@ -61,8 +61,7 @@ from the workspace root:
 ```sh
 pnpm install --frozen-lockfile
 export OPENAI_API_KEY=...
-# Optional: export AGENT_MODEL=...  For example anthropic:claude-sonnet-5-5,
-# with ANTHROPIC_API_KEY; see "Model configuration" below.
+# To change providers, edit the model in the lesson; see "Model configuration" below.
 pnpm dev
 ```
 
@@ -516,28 +515,44 @@ agent with a sandbox, as in lesson 11, can copy a skill in and run them.
 
 ## Model configuration
 
-[shared.ts](src/shared.ts) configures the example model from `AGENT_MODEL`:
-`anthropic:<model>` (reads `ANTHROPIC_API_KEY`), `bedrock:<model>` (reads the
-AWS credentials), or an OpenAI model, optionally as `openai:<model>`. The
-OpenAI adapter speaks Chat Completions, so with `OPENAI_BASE_URL` it reaches
-compatible servers such as DeepSeek, xAI, OpenRouter, Ollama or vLLM.
-`openAIModel` defaults
-GPT-5.6 function-tool requests to `reasoning_effort: "none"` to avoid the default
-Luna model's Chat Completions tool/reasoning compatibility error. Explicit model
-settings are preserved. Change an agent's `model` field to use another adapter
-or gateway.
+Each lesson defines its own model with `aiModel`, so the provider and retry
+policy are visible alongside the agent. The examples use OpenAI Responses
+and read `OPENAI_API_KEY` when the first call is made; importing a lesson
+requires no credentials.
+
+To choose another model or provider, edit that lesson's definition. For
+example, use Anthropic with `ANTHROPIC_API_KEY`:
+
+```ts
+import {anthropic} from "@ai-sdk/anthropic";
+import {aiModel} from "@restatedev/agentlib/provider";
+
+const model = aiModel({
+  model: anthropic("claude-sonnet-5-5"),
+  retry: {maxAttempts: 3},
+});
+```
+
+For Bedrock, use `createAmazonBedrock({region: "us-east-1"})("<model id>")`
+from `@ai-sdk/amazon-bedrock` with AWS credentials. For a Chat Completions-compatible
+server, use `createOpenAI({baseURL: "https://your-server/v1", apiKey: "..."}).chat("<model id>")`
+from `@ai-sdk/openai`. Pass provider-specific settings through
+`aiModel({providerOptions, ...})`.
+
+There is no shared `AGENT_MODEL` selector; each example's model configuration
+is independent.
 
 ## Retry policies
 
 Configure retries at the operation that can fail; there is no agent-wide retry
 default. `maxSteps` limits the model/tool loop, not retries.
 
-| Operation           | Where to configure it                                  | Example                                |
-| ------------------- | ------------------------------------------------------ | -------------------------------------- |
-| Ordinary async tool | `asyncTool({retry, ...})`                              | [02_tools.ts](src/lessons/02_tools.ts) |
-| Model call          | `openAIModel({retry, ...})` or `aiModel({retry, ...})` | [shared.ts](src/shared.ts)             |
-| Individual action   | `restate.run(action, {name, retry})`                   | Below                                  |
-| Handler-backed tool | `retryPolicy` on the called service or handler         | —                                      |
+| Operation           | Where to configure it                          | Example                                |
+| ------------------- | ---------------------------------------------- | -------------------------------------- |
+| Ordinary async tool | `asyncTool({retry, ...})`                      | [02_tools.ts](src/lessons/02_tools.ts) |
+| Model call          | `aiModel({retry, ...})`                        | [01_hello.ts](src/lessons/01_hello.ts) |
+| Individual action   | `restate.run(action, {name, retry})`           | Below                                  |
+| Handler-backed tool | `retryPolicy` on the called service or handler | —                                      |
 
 Lesson 02 uses this policy; the same shape works for model adapters and
 individual actions:

packages/examples/package.json

@@ -91,6 +91,7 @@
     "@ag-ui/core": "1.0.0",
     "@ai-sdk/amazon-bedrock": "5.0.20",
     "@ai-sdk/anthropic": "4.0.14",
+    "@ai-sdk/openai": "4.0.13",
     "@restatedev/agentlib": "workspace:*",
     "@restatedev/restate-sdk": "^1.18.2",
     "@restatedev/restate-sdk-clients": "^1.18.2",