# Models for pi-ai

`agents/models/pi-ai` is the [pi-ai](https://github.com/earendil-works/pi/tree/main/packages/ai)
twin of [`agents/models/ai-sdk`](./models.md): the same `createAI` factory, the same options,
the same gateway, returning pi-ai `Model` objects instead of AI SDK ones. Use it with pi-ai's
`stream`/`complete`, with [pi-durable](https://github.com/earendil-works/pi/tree/main/packages/durable)'s
`Harness`, or with any framework built on a pi-ai `Models` registry.

## The rule

**Workers AI is the catalog Cloudflare keeps up with. Every other vendor is pi-ai's.**

`agents` ships no vendor ids, wire formats, thinking rules or effort tables for any model
that is not `@cf/…`. A third-party model comes from pi-ai's own registry — the package that
tracks that vendor — and `agents` only routes it through AI Gateway. So there are two call
forms:

```ts
import { createAI } from "agents/models/pi-ai";
import { getBuiltinModel } from "@earendil-works/pi-ai/providers/all";

const ai = createAI({ binding: env.AI, id: "prod", cacheTtl: 60 });

ai("@cf/zai-org/glm-5.3-flash"); // Workers AI: our catalog, our compat layer
ai(getBuiltinModel("anthropic", "claude-opus-4-8")); // their model, our transport
```

For convenience, pi-ai also ships a generated catalog of the models Cloudflare's gateway
serves, so a `<provider>/<id>` string resolves without importing anything:

```ts
ai("anthropic/claude-opus-4.8"); // the same model, as AI Gateway's catalog names it
ai("openai/gpt-5.4");
```

Any other string is a `TypeError` that names the way in. There is no prefix guessing here by
design: a model this package cannot name exactly is yours to construct.

1. **Workers AI ids start with `@cf/`.** Everything else is a vendor's, and comes from a
   vendor registry pi-ai maintains.
2. **The gateway is an option, not a product.** Caching, logging, metadata, timeouts and
   retries are options on `createAI`, on the model, or on the call; the `default` gateway is
   created on first use.
3. **No provider keys.** AI Gateway holds the credential — unified billing, or a key you
   store on the gateway (BYOK) — so nothing vendor-shaped needs a secret in your Worker.
4. **The binding, or an account id and token.** Workers AI and catalog models go through
   `env.AI.run`, other model objects through the binding's `fetch` to AI Gateway's
   provider-native path. With [an account id and API token](#account-id-and-api-token) the
   same requests go over HTTP instead.

## Quick start

```jsonc
// wrangler.jsonc
{ "ai": { "binding": "AI" } }
```

```sh
npm install agents @earendil-works/pi-ai
```

```ts
import { createAI } from "agents/models/pi-ai";

export default {
  async fetch(_request: Request, env: { AI: Ai }) {
    const ai = createAI({ binding: env.AI });
    const reply = await ai.complete(ai("@cf/zai-org/glm-5.3-flash"), {
      messages: [{ role: "user", content: "Hello", timestamp: Date.now() }]
    });
    // `content` is an array of parts — text, thinking, tool calls.
    const text = reply.content
      .filter((part) => part.type === "text")
      .map((part) => part.text)
      .join("");
    return Response.json({ text });
  }
};
```

### Account id and API token

To use a gateway in another account, or to call from outside a Worker, pass `accountId` and
`apiToken` instead of `binding`. Everything else is the same, and the gateway still defaults to
`default`:

```ts
const ai = createAI({
  accountId: "27b146402af2103944379f33841b6234",
  apiToken: env.CLOUDFLARE_API_TOKEN,
  gateway: "project-gateway"
});
```

Workers AI and catalog models go to AI Gateway's
[REST API](https://developers.cloudflare.com/ai-gateway/usage/rest-api/), which runs them on
`/run` as `env.AI.run` does. Other model objects go to the provider-native path on
`gateway.ai.cloudflare.com`, with the token as `cf-aig-authorization`. Create the token with
**Workers AI Read** and **AI Gateway Run**; it is sent to those two hosts only.

## Where a model comes from

```ts
// 1. Workers AI — ours.
ai("@cf/zai-org/glm-5.3-flash");

// 2. pi-ai's Cloudflare AI Gateway catalog — a string, resolved by pi, not us.
ai("anthropic/claude-opus-4.8");

// 2b. An AI Gateway dynamic route — the route on the gateway picks the model.
ai("dynamic/support");

// 3. Any pi-ai model object, from any registry you imported.
import { anthropicProvider } from "@earendil-works/pi-ai/providers/anthropic";
const claude = anthropicProvider()
  .getModels()
  .find((m) => m.id === "claude-opus-4-8")!;
ai(claude);
```

Forms 2 and 3 keep the model's own metadata verbatim — `api`, `compat`, `thinkingLevelMap`,
cost, context window, max tokens. That metadata is why Claude Opus gets
`thinking: { type: "adaptive" }` and GPT-5 gets `reasoning.effort`: the model's author said
so, and this package does not second-guess it. Form 3 never mutates the object you
passed; the options ride on a copy.

A `dynamic/<route>` id runs on `env.AI.run` with the Workers AI wire, under the gateway the
route was created on. It answers in the chat-completions shape whichever model the route
picked, and takes the same metadata defaults and overrides as a Workers AI id.

## Running a model

The instance carries pi-ai's four entry points, so nothing global is registered:

```ts
const model = ai("anthropic/claude-opus-4.8");
const context = {
  messages: [{ role: "user", content: "Hi", timestamp: Date.now() }]
};

const message = await ai.complete(model, context); // AssistantMessage
for await (const event of ai.stream(model, context)) {
  // AssistantMessageEvent: text_delta, thinking_delta, toolcall_*, done, error
}
await ai.completeSimple(model, context, { reasoning: "low" }); // SimpleStreamOptions
ai.streamSimple(model, context, { reasoning: "high", maxTokens: 2048 });
```

### With pi-durable

pi-durable's `Harness` takes a `Models` registry. Register the provider on one and name the
model by id:

```ts
import { Harness } from "@earendil-works/pi-durable";
import { createModels } from "@earendil-works/pi-ai/models";

const models = createModels();
models.setProvider(ai.provider);
const harness = await Harness.open(storage, { models, registry }, context);
// a session's model: { provider: "cloudflare", modelId: "anthropic/claude-opus-4.8" }
```

A session stores its model as `{ provider, modelId }`, not as a model object, so use the
`@cf/…` and `<provider>/<id>` forms there. Options given to `ai(model, options)` do not reach a
session; set them on `createAI` instead.

A model stream that is cut off, or that sends nothing for `streamIdleTimeoutMs`, ends in an error
pi-ai's `isRetryableAssistantError()` treats as transient, so the harness retries the turn under
its retry settings rather than failing it.

`ai.streamFn` is `streamSimple`, bound, for frameworks that take a stream function, such as
pi-agent-core's `Agent({ streamFn })`.

### With a pi-ai `Models` registry

```ts
import { createModels } from "@earendil-works/pi-ai/models";

const models = createModels();
models.setProvider(ai.provider); // provider id: "cloudflare"
models.getModel("cloudflare", "anthropic/claude-opus-4.8");
models.streamSimple(ai("@cf/zai-org/glm-5.3-flash"), context);
```

The provider's static list is every Workers AI id pi-ai knows plus every `<provider>/<id>` in
its Cloudflare AI Gateway catalog, so `cloudflare/<id>` specifiers resolve. A model object
you pass to `ai(model)` streams whether or not it is on that list.

## Options

Gateway options may be given flat or nested, on any of the three layers:

```ts
const ai = createAI({ binding: env.AI, id: "prod", cacheTtl: 60 });

ai("anthropic/claude-opus-4.8", {
  gateway: { id: "my-gateway" }, // or a string id
  cacheTtl: 3600, // the same keys, flat
  skipCache: false,
  cacheKey: "greeting-v1",
  collectLog: true,
  metadata: { team: "growth" },
  eventId: "turn-17",
  requestTimeoutMs: 30_000,
  retries: { maxAttempts: 3, backoff: "exponential", retryDelayMs: 500 },
  byokAlias: "team-a", // which stored BYOK key, by alias
  fallback: ["@cf/zai-org/glm-5.3-flash"], // ids or models, in any mix
  headers: { "x-request-source": "docs" },
  sessionAffinity: "chat-42", // Workers AI prefix-cache routing
  reasoningEffort: "low", // "low" | "medium" | "high" | null
  chatTemplateKwargs: { enable_thinking: false } // Workers AI only
});
```

`reasoningEffort` and `chatTemplateKwargs` are Workers AI settings. On a model object or a
`<slug>/<id>` model they are not sent; the drop is recorded on the message as a
`cloudflare-compat` diagnostic, and the pi `reasoning` level (mapped by the model's own
metadata) is the way to ask a vendor model to think. A string in `fallback` follows the same
rule as `ai(string)`: a `@cf/` id or a registry-known `<slug>/<id>`, never a raw vendor id.

The `@cf/` form also takes pi-ai metadata overrides — `name`, `contextWindow`, `maxTokens`,
`reasoning`, `input`, `cost`, `streamIdleTimeoutMs` — because the Workers AI catalog moves
faster than any generated registry. A model object brings its own; nothing here overrides it.

A pi-ai model is plain data, and these per-model options ride on it under symbol keys, which
`JSON.stringify` and `structuredClone` drop. Routing survives that — a model stored in `Agent`
state and read back still reaches its vendor under the vendor's own id — but the options do
not, so re-apply them with `ai(model, options)` after loading if the model carried any.

Per call, pi-ai's own options map onto the same layers: `headers` are merged last, `metadata`
(scalars) joins the gateway metadata, `sessionId` sets session affinity, and `reasoning` (a pi
thinking level) is remapped through the model's own `thinkingLevelMap` — as far as the wire
can carry it; see [thinking shapes](#thinking-shapes) for the one limit. Precedence is call,
then model, then `createAI` settings.

## How a request travels

| Model                                   | `api`                | Through the binding                                       | With an account id and token                           |
| --------------------------------------- | -------------------- | --------------------------------------------------------- | ------------------------------------------------------ |
| `@cf/…`, `dynamic/…`                    | `cloudflare-ai`      | `env.AI.run` + Cloudflare's compat layer                  | the REST API + Cloudflare's compat layer               |
| `anthropic/…` from pi's gateway catalog | `anthropic-messages` | `env.AI.run` → AI Gateway `/run`, Messages in and out     | the REST API's `/ai/v1/messages` → AI Gateway `/run`   |
| `openai/…` from pi's gateway catalog    | `openai-responses`   | `env.AI.run` → AI Gateway `/run`, Responses in and out    | the REST API's `/ai/v1/responses` → AI Gateway `/run`  |
| Any other model object                  | its own `api`        | the binding's `fetch` → AI Gateway's provider-native path | `gateway.ai.cloudflare.com` → the provider-native path |

A model from pi's Cloudflare AI Gateway catalog runs on `env.AI.run` under the gateway's
catalog id (`anthropic/claude-opus-4.8`). pi-ai spells Claude ids as Anthropic does
(`claude-opus-4-8`), so `agents` lists and resolves them with the catalog's dot instead. The body is pi-ai's own,
built by pi-ai's converter for the model's `api`, and it is sent in that format: `/run`
converts nothing, so the answer, streams included, is the shape pi-ai's stream processors
read. `/run` picks the provider and bills through Unified Billing or a key stored on the
gateway.

Any other model object — one from pi's Anthropic, OpenAI or Groq registry, with the vendor's
own id — goes to AI Gateway's **provider-native path** through the binding's `fetch`, with the
provider slug taken from the model's `baseUrl` host. The gateway forwards it to the vendor with
the credential it holds and streams the vendor's own SSE back untouched.

Either way no vendor credential leaves your Worker, and the model's `api`, `provider` and
`id` are never rewritten, so thinking signatures and encrypted reasoning items replay on the
next turn.

The `api` markers in that table are the only ones this package routes, and the table is the
whole list: `cloudflare-ai`, `anthropic-messages`, `openai-responses`, `openai-completions`.
Check your model's `api` before you pass it — several pi registries declare something else and
are not routable today, among them `mistral` (`mistral-conversations`), `google`
(`google-generative-ai`), `google-vertex`, `amazon-bedrock`, `azure-openai-responses` and
`openai-codex`.

Workers AI stays on `env.AI.run`. Its departures from strict OpenAI chat completions (native
events, heartbeats, per-delta usage, per-family request quirks) are absorbed by the same
compat layer the [AI SDK provider](./models.md) uses, so the parser only ever sees strict
OpenAI chunks. Anything the layer had to drop is recorded on the message as a
`cloudflare-compat` diagnostic.

An `api` this package cannot route, or a `baseUrl` whose host AI Gateway does not serve, ends
the stream with an `error` event naming it.

### Thinking shapes

On the `openai-completions` wire this package builds one thinking shape: OpenAI's
`reasoning_effort`, with the level remapped through the model's own `thinkingLevelMap`, and
only when the model reasons and its `compat` does not say `supportsReasoningEffort: false`.

A model whose `compat.thinkingFormat` is anything else — `deepseek`, `zai`, `openrouter`,
`together`, `ant-ling`, `qwen`, `qwen-chat-template`, `chat-template`, `string-thinking` — has
its reasoning level **dropped**, and the drop recorded on the message as a `cloudflare-compat`
diagnostic. Those shapes are the vendor's, and building them here would mean keeping vendor
wire formats in `agents`, which is the one thing this package does not do: reach such a model
through pi-ai's own provider if you need its thinking shape. The `anthropic-messages` and
`openai-responses` wires build the thinking shape their own model metadata declares.

## Unified billing and BYOK

Most vendors work with no key at all: the gateway bills them to your Cloudflare account. A
vendor that is not on unified billing answers with its own error until you store a key on the
gateway — Google AI Studio, reached through a routable model, returns `403 unregistered
callers`. That error is the vendor's, surfaced as-is; the gateway's own failures (no such
gateway, an unpaid account) become a `CloudflareAIError`, so fallback and retry treat them the
way they treat Workers AI failures. A model whose `api` is not one of the four routed markers
never gets that far: the stream ends with an `error` event naming the `api`.

`byokAlias` picks a stored key by alias when the gateway holds several for one provider. It is
sent as `cf-aig-byok-alias` on a catalog model and on any other model object. `zdr` turns Zero
Data Retention on or off for a Unified Billing request, and applies to model objects outside
pi's gateway catalog only: `/run` does not read it.

## Observability

Every assistant message carries a `cloudflare` diagnostic with the gateway correlation for its
response:

```ts
const message = await ai.complete(model, context);
const cloudflare = message.diagnostics?.find(
  (d) => d.type === "cloudflare"
)?.details;
// { model, gateway, provider?, specifier?, logId?, eventId?, cacheStatus?,
//   step?, traceId?, runId? }
```

A failed dispatch ends the stream with an `error` event whose message carries a
`cloudflare-error` diagnostic (`status`, `code`, `logId`), and `onResponse` fires with the raw
status and headers for every request.

## Stream resume

Through the binding, a model from pi's AI Gateway catalog streams from `/run`, which names the run in
`cf-aig-run-id` and keeps its events. If the connection drops, the provider reconnects through
the binding and replays from the last complete event, so the stream continues with nothing lost
or repeated. It gives up after five drops or once the gateway no longer holds the run, and an
abort never resumes. Workers AI streams and models on the provider-native path have no run id
and are not resumed. The run id is on the `cloudflare` diagnostic as `runId`.

## Fallback

```ts
ai(claude, { fallback: ["@cf/zai-org/glm-5.3-flash"] });
```

Legs may be Workers AI ids or models, in any mix. They are tried in order: a leg is abandoned
when its stream ends with an `error` before it produced any output, and once a leg has
produced output it is committed to. The message that answers carries a `cloudflare-fallback`
diagnostic listing the abandoned attempts, and `message.model` names the model that actually
answered.

A leg travels with the chain's gateway options — the chain's `id`, `cacheTtl` and the rest win
over a leg's own, and `metadata` merges — as they do in the [AI SDK provider](./models.md). A
leg built as `ai(model, options)` keeps everything else it was built with. A leg named by a
`@cf/` id has no options of its own, so it also inherits the chain's `headers`, session
affinity and Workers AI knobs; it never inherits the chain's model metadata (`name`, `cost`,
`contextWindow`, `maxTokens`, `reasoning`, `input`), which describes the model it was written
on — the leg's usage and cost are its own registry entry's.

## Images

```ts
const images = ai.images("@cf/black-forest-labs/flux-1-schnell", { steps: 8 });
const result = await ai.generateImages(images, {
  input: [{ type: "text", text: "A red bicycle" }]
});
```

Image generation is Workers AI only. The knobs a model takes are the model's:
`flux-1-schnell` takes a prompt and `steps` and nothing else, so `width`,
`height`, `seed`, `guidance` and `negativePrompt` are dropped for it and listed
in a `cloudflare-compat` diagnostic on the result rather than sent and ignored.
Every other model in the catalog takes the full set.

## Requirements

This subpath needs `@earendil-works/pi-ai@~1.5.0`, an optional peer of `agents`. `agents`
takes no vendor package as a dependency, optional or otherwise. An unlisted `@cf/` id resolves with default metadata (a
128k context window, 8192 output tokens, zero cost); pass overrides when you know better.

## See also

- [Models for the AI SDK](./models.md) — the same factory for the Vercel AI SDK
- [pi-ai](https://github.com/earendil-works/pi/tree/main/packages/ai) and
  [pi-durable](https://github.com/earendil-works/pi/tree/main/packages/durable)