@averyyy/pi-ai 0.87.1-piclient.2 → 0.99.1-piclient.1

package/README.md

@@ -2,7 +2,7 @@
 
 Unified LLM API with provider collections, automatic auth resolution, token and cost tracking, and simple context persistence and hand-off to other models mid-session.
 
-**Note**: This library only includes models that support tool calling (function calling), as this is essential for agentic workflows.
+**Note**: The chat catalog only includes models that support tool calling (function calling), as this is essential for agentic workflows. Image and classifier catalogs use their operation-specific capabilities.
 
 ## Table of Contents
 
@@ -29,6 +29,7 @@ Unified LLM API with provider collections, automatic auth resolution, token and
   - [Compact Assistant Message Frames](#compact-assistant-message-frames)
 - [Image Input](#image-input)
 - [Image Generation](#image-generation)
+- [Classification](#classification)
 - [Thinking/Reasoning](#thinkingreasoning)
   - [Unified Interface](#unified-interface-streamsimplecompletesimple)
   - [Provider-Specific Options](#provider-specific-options-streamcomplete)
@@ -38,6 +39,7 @@ Unified LLM API with provider collections, automatic auth resolution, token and
   - [Aborting Requests](#aborting-requests)
   - [Continuing After Abort](#continuing-after-abort)
   - [Debugging Provider Payloads](#debugging-provider-payloads)
+  - [Observing Provider Stream Events](#observing-provider-stream-events)
 - [Custom Providers](#custom-providers)
   - [createProvider()](#createprovider)
   - [Calling API Implementations Directly](#calling-api-implementations-directly)
@@ -61,8 +63,9 @@ Unified LLM API with provider collections, automatic auth resolution, token and
 - **OpenAI**
 - **Ant Ling**
 - **Azure OpenAI (Responses)**
-- **OpenAI Codex** (ChatGPT Plus/Pro subscription, requires OAuth, see below)
+- **OpenAI Codex (legacy)** (ChatGPT Plus/Pro subscription, requires OAuth, see below)
 - **Radius** (API key or OAuth, with a dynamically refreshed gateway catalog)
+- **TypeSafe** (System One classifier API)
 - **DeepSeek**
 - **NVIDIA NIM**
 - **Anthropic**
@@ -275,7 +278,7 @@ Reads are synchronous and return the last-known lists:
 const providers = models.getProviders();           // registered Provider objects
 const provider = models.getProvider('anthropic');  // one provider
 
-const all = models.getModels();                    // every model across providers
+const all = models.getModels();                    // every chat model across providers
 const anthropicModels = models.getModels('anthropic');
 const model = models.getModel('anthropic', 'claude-sonnet-4-5');
 
@@ -288,7 +291,31 @@ for (const m of anthropicModels) {
 }
 ```
 
-Dynamically listed models are typed `Model<Api>`. Narrow with the `hasApi()` guard when you need API-specific option typing:
+The unqualified reads `getModels()`/`getModel()`/`getAvailable()` return chat models (`Model<Api>`) usable with `stream()`. The `*OfType` reads return one model type, and `getAllModels()`/`getAllAvailable()` return every type as `AnyModel`:
+
+```typescript
+const images = models.getModelsOfType('image', 'openrouter');           // ImageModel[]
+const flux = models.getModelOfType('image', 'openrouter', 'black-forest-labs/flux.2-pro');
+const jev = models.getModelOfType('classifier', 'typesafe', 'jev-latest');
+const availableImages = await models.getAvailableOfType('image');
+const everything = models.getAllModels();                               // AnyModel[]
+```
+
+The model's `type` decides which operation accepts it: chat models stream, `type: "image"` models generate images, and `type: "classifier"` models classify structured state. `type` is optional on chat models, so a model without `type` is a chat model. Do not compare `type` directly; narrow mixed lists with `isModelType()` or read the effective type with `getModelType()`:
+
+```typescript
+import { isModelType } from '@earendil-works/pi-ai';
+
+for (const model of models.getAllModels()) {
+  if (isModelType(model, 'image')) {
+    // model: ImageModel<ImageApi>
+  }
+}
+```
+
+IDs are unique within each provider and type; one upstream model may have separate entries for different operations. On a provider, `getModels()` returns chat models and the optional `getAllModels()` returns every type; providers with only chat models can omit it.
+
+Dynamically listed chat models are typed `Model<Api>`. Narrow with the `hasApi()` guard when you need API-specific option typing:
 
 ```typescript
 import { hasApi } from '@earendil-works/pi-ai';
@@ -305,12 +332,26 @@ if (m && hasApi(m, 'anthropic-messages')) {
 For tooling that wants the generated built-in catalog with full literal typing (provider and model IDs auto-complete), independent of any collection:
 
 ```typescript
-import { getBuiltinModel, getBuiltinModels, getBuiltinProviders } from '@earendil-works/pi-ai/providers/all';
+import {
+  getAllBuiltinModels,
+  getBuiltinClassifierModel,
+  getBuiltinClassifierModels,
+  getBuiltinImageModel,
+  getBuiltinImageModels,
+  getBuiltinModel,
+  getBuiltinModels,
+  getBuiltinProviders,
+} from '@earendil-works/pi-ai/providers/all';
 
 const model = getBuiltinModel('openai', 'gpt-4o-mini'); // typed Model<'openai-responses'>
-const radius = getBuiltinModel('radius', 'balanced');     // typed Model<'pi-messages'>
+const radius = getBuiltinModel('radius', 'balanced');   // typed Model<'pi-messages'>
+const flux = getBuiltinImageModel('openrouter', 'black-forest-labs/flux.2-pro');
+const jev = getBuiltinClassifierModel('typesafe', 'jev-latest');
 const providers = getBuiltinProviders();
-const anthropic = getBuiltinModels('anthropic');
+const openrouterChat = getBuiltinModels('openrouter');        // Model[]
+const openrouterImages = getBuiltinImageModels('openrouter'); // ImageModel[]
+const typesafeClassifiers = getBuiltinClassifierModels('typesafe'); // ClassifierModel[]
+const openrouterAll = getAllBuiltinModels('openrouter');      // AnyModel[]
 ```
 
 ### Dynamic Providers
@@ -422,6 +463,7 @@ Built-in providers resolve these env vars (Node.js; in browsers pass `apiKey` ex
 | Azure OpenAI | `AZURE_OPENAI_API_KEY` + `AZURE_OPENAI_BASE_URL` (e.g. `https://{resource}.ai.azure.com`) or `AZURE_OPENAI_RESOURCE_NAME`. Supports `*.openai.azure.com`, `*.cognitiveservices.azure.com` and `*.ai.azure.com`; root endpoints auto-normalize to `/openai/v1`. Optional: `AZURE_OPENAI_API_VERSION` (default `v1`), `AZURE_OPENAI_DEPLOYMENT_NAME_MAP`. |
 | Anthropic | `ANTHROPIC_API_KEY` or `ANTHROPIC_OAUTH_TOKEN` |
 | Radius | `RADIUS_API_KEY` |
+| TypeSafe | `TYPESAFE_API_KEY` |
 | DeepSeek | `DEEPSEEK_API_KEY` |
 | NVIDIA NIM | `NVIDIA_API_KEY` |
 | Google | `GEMINI_API_KEY` |
@@ -747,20 +789,19 @@ for (const block of response.content) {
 
 ## Image Generation
 
-Image generation uses a separate API surface from text/chat generation, mirroring the chat-side design: an `ImagesModels` collection holds `ImagesProvider`s, reads are sync, and auth resolves through the owning provider. Image generation is a one-shot API: `generateImages()` waits for the provider response and returns the final `AssistantImages` result — do not use the chat/stream APIs for it.
+Image models live in the same `Models` collection and on the same `Provider` as chat models, so one credential per provider covers both. They are typed `ImageModel` with `type: "image"` and are used through `generateImages()`, a one-shot API that waits for the provider response and returns the final `AssistantImages` result. Do not use the chat/stream APIs for them; `stream()` rejects image models.
 
 ### Basic Image Generation
 
 ```typescript
-import { builtinImagesModels } from '@earendil-works/pi-ai/providers/all';
+import { builtinModels } from '@earendil-works/pi-ai/providers/all';
 
-// Every built-in image-generation provider; accepts the same options as createModels()
-const imagesModels = builtinImagesModels();
+const models = builtinModels();
 
-const model = imagesModels.getModel('openrouter', 'google/gemini-2.5-flash-image')!;
+const model = models.getModelOfType('image', 'openrouter', 'google/gemini-2.5-flash-image')!;
 
 // Auth resolves through the provider (OPENROUTER_API_KEY here); explicit apiKey wins
-const result = await imagesModels.generateImages(model, {
+const result = await models.generateImages(model, {
   input: [{ type: 'text', text: 'Generate a red circle on a plain white background.' }]
 });
 
@@ -774,7 +815,33 @@ for (const block of result.output) {
 }
 ```
 
-Like the chat side, you can build the collection from parts: `createImagesModels({ credentials?, authContext? })`, the `openrouterImagesProvider()` factory from `@earendil-works/pi-ai/providers/openrouter-images`, and `createImagesProvider({ id, auth, models, refreshModels?, api })` for custom image providers (with `imagesModels.refresh(provider?)` for dynamic lists). Failures never reject — they return an `AssistantImages` with `stopReason: "error"`. The collection's provider-scoped `getAuth(providerId)` works exactly like the chat-side one.
+`generateImages()` accepts only `ImageModel` values. If an upstream model supports both chat and image generation, the catalog contains separate entries with the same provider and ID: `getModel()` returns its chat operation and `getModelOfType('image', ...)` returns its image operation. Failures never reject; they return an `AssistantImages` with `stopReason: "error"`, including unknown providers, unconfigured auth, and providers without an image implementation.
+
+A provider declares image support with the `images` option of [`createProvider()`](#createprovider): a map from `model.api` to an implementation with `generateImages()`. Image models go into the same `models` list as chat models. `api` becomes optional when `images` is present, so an image-only provider is just a provider without chat models:
+
+```typescript
+import { createProvider, envApiKeyAuth } from '@earendil-works/pi-ai';
+
+const pixels = createProvider({
+  id: 'pixels',
+  auth: { apiKey: envApiKeyAuth('Pixels API key', ['PIXELS_API_KEY']) },
+  models: [{
+    type: 'image',
+    id: 'flux-pro',
+    name: 'FLUX Pro',
+    api: 'pixels-images',
+    provider: 'pixels',
+    baseUrl: 'https://api.pixels.test/v1',
+    input: ['text'],
+    output: ['image'],
+    cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+  }],
+  images: {
+    'pixels-images': { generateImages: async (model, context, options) => { /* ... */ } },
+  },
+});
+models.setProvider(pixels);
+```
 
 The old global API (`getImageModel()` / `getImageModels()` / `getImageProviders()` / `generateImages()`) remains available on the [compat entrypoint](#migrating-from-the-old-global-api):
 
@@ -795,7 +862,7 @@ Some models also support image input:
 import { readFileSync } from 'fs';
 
 const imageBuffer = readFileSync('input.png');
-const result = await imagesModels.generateImages(model, {
+const result = await models.generateImages(model, {
   input: [
     { type: 'text', text: 'Create a variation of this image with a blue background.' },
     { type: 'image', data: imageBuffer.toString('base64'), mimeType: 'image/png' }
@@ -806,14 +873,13 @@ const result = await imagesModels.generateImages(model, {
 Check capabilities on the model metadata:
 
 ```typescript
-console.log(model.input);   // ['text', 'image']
-console.log(model.output);  // ['image'] or ['image', 'text']
+console.log(model.input);  // ['text'] or ['text', 'image']
+console.log(model.output); // ['image'] or ['image', 'text']
 ```
 
 ### Notes and Limitations
 
-- Image models live in `ImagesModels` collections, chat models in `Models` collections; the two are separate surfaces.
-- Use `generateImages()`, not the chat/stream APIs.
+- Image models and chat models share `Models` and `Provider`; list them with `getModelsOfType('image')` and run them with `generateImages()`, never the chat/stream APIs.
 - Image-generation models do not participate in tool calling.
 - Outputs are returned in `AssistantImages.output` and can include both base64-encoded `ImageContent` blocks and `TextContent` blocks.
 - Some models return only images, others return images plus text. Check `model.output`.
@@ -822,6 +888,97 @@ console.log(model.output);  // ['image'] or ['image', 'text']
 - If you want a model to analyze images in a conversation or call tools, use the regular chat APIs with a model that supports image input.
 - At the moment, image generation is available through only one provider, OpenRouter.
 
+## Classification
+
+Classifier models consume structured JSON state and answer one or more typed questions. They do not use chat or image-generation APIs. TypeSafe's Jev model is available from these built-in providers:
+
+| Provider | Model IDs | Auth |
+| --- | --- | --- |
+| `typesafe` | `jev-latest` | `TYPESAFE_API_KEY` |
+| `openrouter` | `typesafe/jev-1.13`, `~typesafe/jev-latest` | `OPENROUTER_API_KEY` or OpenRouter OAuth |
+| `cloudflare-workers-ai` | `typesafe/jev` | `CLOUDFLARE_API_KEY` and `CLOUDFLARE_ACCOUNT_ID` |
+| `vercel-ai-gateway` | `typesafe-ai/jev` | `AI_GATEWAY_API_KEY` |
+| `opencode` | `jev-1.13`, `jev-1.13-free` | `OPENCODE_API_KEY` |
+
+```typescript
+import { builtinModels } from '@earendil-works/pi-ai/providers/all';
+
+const models = builtinModels();
+const model = models.getModelOfType('classifier', 'typesafe', 'jev-latest')!;
+const result = await models.classify(model, {
+  state: { message: 'The change works perfectly, thanks.' },
+  questions: {
+    category: {
+      type: 'choice',
+      instructions: 'Classify the message.',
+      criteria: {
+        approval: 'The user approves of the result',
+        correction: 'The user requests a correction'
+      }
+    },
+    satisfaction: {
+      type: 'score',
+      instructions: 'Score user satisfaction.',
+      criteria: ['dissatisfied', 'neutral', 'satisfied']
+    },
+    approved: {
+      type: 'bool',
+      instructions: 'Does the user approve?',
+      criteria: { true: 'Approval', false: 'No approval' }
+    }
+  }
+});
+
+console.log(result.answers);
+```
+
+The public contract uses `bool` questions and `{ type: "bool", probability }` answers. The TypeSafe adapter translates those to and from its `noul` wire representation. Like image generation, `classify()` resolves to a result with `stopReason: "error"` instead of rejecting for provider, authentication, or response errors.
+
+When the service reports token counts, `result.usage` carries them with their cost at the model's catalog price, the same `Usage` shape as chat messages. All System One services report token counts; a request that was answered with malformed answers keeps its usage. Local classifiers such as `llama-cpp-classify` report no usage.
+
+`ClassifierOptions.temperature` divides the answer logits by the given value before they are normalized; values above 1 soften the distribution. APIs that cannot apply it, such as System One, ignore it.
+
+### Chat models on llama.cpp
+
+The `llama-cpp-classify` API turns a chat model served by llama.cpp's `llama-server` into a classifier. Each question becomes one chat prompt: the state, every question of the request, the state again, and the question with its answers under single-token labels (letters for a choice, `Yes`/`No` for a bool, digits for a score). The prompt up to the final question is shared by all questions of a request, so the server's prompt cache evaluates the state once per request. The server returns the log-probabilities of the next token, and the answer is the softmax over the label tokens. Choices support up to 62 options and scores up to 10 levels. The model's `baseUrl` is the server URL; a trailing `/v1` is ignored. In router mode, the model ID selects the model.
+
+```typescript
+import { createProvider } from '@earendil-works/pi-ai';
+import { llamaCppClassifyApi } from '@earendil-works/pi-ai/api/llama-cpp-classify.lazy';
+
+const provider = createProvider({
+  id: 'local-llama',
+  auth: { apiKey: { name: 'llama.cpp', resolve: async () => ({ auth: {} }) } },
+  models: [{
+    type: 'classifier',
+    id: 'qwen3-4b',
+    name: 'Qwen3 4B',
+    api: 'llama-cpp-classify',
+    provider: 'local-llama',
+    baseUrl: 'http://127.0.0.1:8080',
+    input: ['text'],
+    cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+    contextWindow: 32768
+  }],
+  classifiers: { 'llama-cpp-classify': llamaCppClassifyApi() }
+});
+```
+
+Raw label probabilities are usually overconfident; pass `temperature` above 1 to soften them.
+
+Custom providers register classifier models and implementations by API ID:
+
+```typescript
+createProvider({
+  id: 'classifier-service',
+  auth,
+  models: [model],
+  classifiers: {
+    'classifier-api': { classify: async (model, context, options) => result }
+  }
+});
+```
+
 ## Thinking/Reasoning
 
 Many models support thinking/reasoning capabilities where they can show their internal thought process. You can check if a model supports reasoning via the `reasoning` property. If you pass reasoning options to a non-reasoning model, they are silently ignored.
@@ -1030,11 +1187,30 @@ const response = await models.complete(model, context, {
 
 The callback is supported by `stream`, `complete`, `streamSimple`, and `completeSimple`.
 
+### Observing Provider Stream Events
+
+Use `onProviderStreamEvent` to inspect provider-specific fields that Pi does not include in `AssistantMessage`. The callback receives the parsed event available to the adapter before Pi normalizes it. Treat the event as read-only because mutations can affect normalization. This is not guaranteed to be the original HTTP bytes or SSE frame.
+
+```typescript
+const openRouterModel = models.getModel('openrouter', 'openrouter/auto')!;
+const response = await models.complete(openRouterModel, context, {
+  headers: { "X-OpenRouter-Metadata": "enabled" },
+  onProviderStreamEvent: (data) => {
+    const chunk = data as Record<string, unknown>;
+    if (chunk.openrouter_metadata) {
+      console.log(chunk.openrouter_metadata);
+    }
+  },
+});
+```
+
+Callbacks are awaited in stream order, so slow callbacks delay stream consumption and thrown errors fail the request. SDK-backed adapters can expose only fields retained by their SDK.
+
 ## Custom Providers
 
 ### createProvider()
 
-`createProvider()` builds a provider from parts: identity, auth, a model list, and an API implementation. Use it for local inference servers, proxies, or any OpenAI/Anthropic-compatible endpoint:
+`createProvider()` builds a provider from parts: identity, auth, a model list, and an API implementation (`api` for chat models, `images` for image generation, `classifiers` for classification; at least one is required, see [Image Generation](#image-generation)). Use it for local inference servers, proxies, or any OpenAI/Anthropic-compatible endpoint:
 
 ```typescript
 import { createModels, createProvider, envApiKeyAuth, type Model } from '@earendil-works/pi-ai';
@@ -1116,7 +1292,7 @@ const tenantGateway = createProvider({
 });
 ```
 
-Dynamic model lists use `fetchModels`. `Models.refresh()` refreshes every configured dynamic provider, passing its effective API-key or refreshed OAuth credential. A `ModelsStore` persists dynamic catalogs; both stores default to in-memory implementations. Its `read`, `write`, and `delete` operations accept optional cancellation, and `Models` binds those waits to the provider refresh signal.
+Dynamic model lists use `fetchModels`, which can return models of every type. `Models.refresh()` refreshes every configured dynamic provider, passing its effective API-key or refreshed OAuth credential. A `ModelsStore` persists dynamic catalogs; both stores default to in-memory implementations. Its `read`, `write`, and `delete` operations accept optional cancellation, and `Models` binds those waits to the provider refresh signal.
 
 ```typescript
 const models = createModels({ credentials, modelsStore });
@@ -1138,7 +1314,7 @@ for (const [provider, error] of result.errors) console.error(provider, error);
 
 Use `models.refresh({ providers: ['openrouter'] })` to restrict work to selected providers, `models.refresh({ allowNetwork: false })` to restore persisted catalogs without network access, or `models.refresh({ force: true })` to bypass provider freshness checks. Model reads stay synchronous and return the last restored or refreshed list.
 
-`createProvider()` handles dynamic publication and persistence automatically. Handwritten `Provider.refreshModels()` implementations receive the read-only `context.stored` snapshot and publish through `context.publish({ persist?, update? })`. Omit `persist` to leave storage unchanged, pass a `ModelsStoreEntry` to write it, or pass `persist: null` to delete it. Publication is generation-checked; put synchronous in-memory catalog changes in `update` rather than mutating state before publication.
+`createProvider()` handles dynamic publication and persistence automatically. Handwritten `Provider.refreshModels()` implementations receive the read-only `context.stored` snapshot and publish through `context.publish({ persist?, update? })`. Omit `persist` to leave storage unchanged, pass a `ModelsStoreEntry` to write it, or pass `persist: null` to delete it. `ModelsStoreEntry.models` contains models of every type. Publication is generation-checked; put synchronous in-memory catalog changes in `update` rather than mutating state before publication.
 
 Custom models can carry `headers` (e.g. proxies behind bot detection) and `compat` flags. `Models.getAuth(model)` includes those model headers, and stream methods merge them before explicit request headers and `transformHeaders`. See [OpenAI Compatibility Settings](#openai-compatibility-settings).
 
@@ -1553,7 +1729,8 @@ Use this when one process needs different provider settings per request, or when
 Several providers support OAuth authentication instead of static API keys:
 
 - **Anthropic** (Claude Pro/Max subscription)
-- **OpenAI Codex** (ChatGPT Plus/Pro subscription, access to GPT-5.x Codex models)
+- **OpenAI** (Sign in with ChatGPT: uses the ChatGPT subscription with the OpenAI API)
+- **OpenAI Codex (legacy)** (ChatGPT Plus/Pro subscription, access to GPT-5.x Codex models)
 - **GitHub Copilot** (Copilot subscription)
 - **OpenRouter** (OAuth PKCE that mints a user-controlled API key)
 
@@ -1633,7 +1810,7 @@ Built-in login and refresh flows are private provider implementations. Use provi
 
 Provider notes:
 
-**OpenAI Codex**: Requires a ChatGPT Plus or Pro subscription. Provides access to GPT-5.x Codex models with extended context windows and reasoning capabilities. The library automatically handles session-based prompt caching when `sessionId` is provided in stream options unless `cacheRetention` is `"none"`. You can set `transport` in stream options to `"sse"`, `"websocket"`, or `"auto"` for Codex Responses transport selection. When using WebSocket with a `sessionId` and cache retention enabled, connections are reused per session and expire after 5 minutes of inactivity. Call `cleanupSessionResources(sessionId)` when finished so the pooled connection does not keep the process alive.
+**OpenAI Codex (legacy)**: Superseded by Sign in with ChatGPT on the OpenAI provider. Requires a ChatGPT Plus or Pro subscription. Provides access to GPT-5.x Codex models with extended context windows and reasoning capabilities. The library automatically handles session-based prompt caching when `sessionId` is provided in stream options unless `cacheRetention` is `"none"`. You can set `transport` in stream options to `"sse"`, `"websocket"`, or `"auto"` for Codex Responses transport selection. When using WebSocket with a `sessionId` and cache retention enabled, connections are reused per session and expire after 5 minutes of inactivity. Call `cleanupSessionResources(sessionId)` when finished so the pooled connection does not keep the process alive.
 
 **Azure OpenAI (Responses)**: Uses the Responses API only. Set `AZURE_OPENAI_API_KEY` and either `AZURE_OPENAI_BASE_URL` or `AZURE_OPENAI_RESOURCE_NAME`. `AZURE_OPENAI_BASE_URL` supports both `https://<resource>.openai.azure.com` and `https://<resource>.cognitiveservices.azure.com`; root endpoints are normalized to `.../openai/v1` automatically. Use `AZURE_OPENAI_API_VERSION` (defaults to `v1`) to override the API version if needed. Deployment names are treated as model IDs by default, override with `azureDeploymentName` or `AZURE_OPENAI_DEPLOYMENT_NAME_MAP` using comma-separated `model-id=deployment` pairs (for example `gpt-4o-mini=my-deployment,gpt-4o=prod`). Legacy deployment-based URLs are intentionally unsupported.
 
@@ -1662,6 +1839,9 @@ Compat is a strict superset of the root entrypoint, so a file can switch its imp
 | `getEnvApiKey('openai')` | `await models.getAuth(model.provider)` |
 | `streamAnthropic(model, ctx, opts)` | `stream` from `@earendil-works/pi-ai/api/anthropic-messages`, or a provider in a collection |
 | `registerFauxProvider()` | `fauxProvider()` + `models.setProvider()` |
+| `getImageModel('openrouter', id)` / `generateImages(model, ctx, { apiKey })` | `models.getModelOfType('image', 'openrouter', id)` / `models.generateImages(model, ctx)` |
+
+The separate `ImagesModels`/`ImagesProvider` collection that existed briefly (`createImagesModels()`, `createImagesProvider()`, `openrouterImagesProvider()`, `builtinImagesModels()`) is gone: image models now live on the regular provider. Replace `builtinImagesModels()` with `builtinModels()`, `imagesModels.getModel()` with `models.getModelOfType('image', ...)`, and `createImagesProvider({ models, api })` with `createProvider({ models, images })`. The old plural image type names are removed; use `ImageModel` and `ImageApi`, and add `type: "image"` to image model literals.
 
 ## Development
 
@@ -1686,11 +1866,11 @@ Create a new API implementation file (for example `bedrock-converse-stream.ts`)
 
 Add a lazy wrapper `src/api/<api-id>.lazy.ts` (`<name>Api()` via `lazyApi()`) so providers can reference the implementation without importing its SDK. Add any root-level `export type` re-exports in `src/index.ts` that should remain available from `@earendil-works/pi-ai`.
 
-#### 3. Model Generation (`scripts/generate-models.ts`, `scripts/generate-image-models.ts`)
+#### 3. Model Generation (`scripts/generate-models.ts`)
 
 - Add logic to fetch and parse models from the provider's source (e.g., models.dev API)
-- Map chat/tool-capable provider model data to the standardized `Model` interface via `scripts/generate-models.ts`; hydration groups the ignored `src/providers/data/<id>.json` values by API, while stable `src/providers/<id>.models.ts` wrappers derive exact model/API types directly from those JSON keys
-- Map image-generation provider model data to the standardized `ImagesModel` interface via `scripts/generate-image-models.ts`
+- Map chat/tool-capable provider data to `Model`, image-generation data to `ImageModel`, and models.dev `type: "decision"` entries to `ClassifierModel`; hydration groups the ignored `src/providers/data/<id>.json` values by API while stable `src/providers/<id>.models.ts` wrappers derive exact model/API types directly from those JSON keys
+- Keep model ids unique within each provider and model type; emit separate entries when an upstream model supports multiple operations
 - Handle provider-specific quirks (pricing format, capability flags, model ID transformations)
 
 #### 4. Provider Factory (`src/providers/<id>.ts`)

package/dist/api/anthropic-messages.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"anthropic-messages.d.ts","sourceRoot":"","sources":["../../src/api/anthropic-messages.ts"],"names":[],"mappings":"AAAA,OAAO,SAAS,MAAM,mBAAmB,CAAC;AAa1C,OAAO,KAAK,EASX,mBAAmB,EAEnB,cAAc,EACd,aAAa,EAMb,MAAM,aAAa,CAAC;AAiJrB,MAAM,MAAM,eAAe,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,GAAG,OAAO,GAAG,KAAK,CAAC;AAE1E,MAAM,MAAM,wBAAwB,GAAG,YAAY,GAAG,SAAS,CAAC;AA2ChE,MAAM,WAAW,gBAAiB,SAAQ,aAAa;IACtD;;;;;;OAMG;IACH,eAAe,CAAC,EAAE,OAAO,CAAC;IAC1B;;;;OAIG;IACH,oBAAoB,CAAC,EAAE,MAAM,CAAC;IAC9B;;;;;;;;;;;OAWG;IACH,MAAM,CAAC,EAAE,eAAe,CAAC;IACzB;;;;;;;;;;;OAWG;IACH,eAAe,CAAC,EAAE,wBAAwB,CAAC;IAC3C;;;;;OAKG;IACH,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAC9B;;;;OAIG;IACH,UAAU,CAAC,EAAE,MAAM,GAAG,KAAK,GAAG,MAAM,GAAG;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAA;KAAE,CAAC;IACtE;;;;OAIG;IACH,MAAM,CAAC,EAAE,SAAS,CAAC;CACnB;AAqOD,eAAO,MAAM,MAAM,EAAE,cAAc,CAAC,oBAAoB,EAAE,gBAAgB,CAgUzE,CAAC;AA2BF,eAAO,MAAM,YAAY,EAAE,cAAc,CAAC,oBAAoB,EAAE,mBAAmB,CA8ClF,CAAC","sourcesContent":["import Anthropic from \"@anthropic-ai/sdk\";\nimport type {\n\tBetaStopReason,\n\tBetaThinkingDroppedInputTransformation,\n\tBetaTool,\n\tBetaCacheControlEphemeral as CacheControlEphemeral,\n\tBetaContentBlockParam as ContentBlockParam,\n\tMessageCreateParamsStreaming,\n\tBetaMessageParam as MessageParam,\n\tBetaRawMessageStreamEvent as RawMessageStreamEvent,\n\tBetaRefusalStopDetails as RefusalStopDetails,\n} from \"@anthropic-ai/sdk/resources/beta/messages/messages.js\";\nimport { calculateCost } from \"../models.ts\";\nimport type {\n\tApi,\n\tAssistantMessage,\n\tCacheRetention,\n\tImageContent,\n\tMessage,\n\tModel,\n\tProviderEnv,\n\tProviderHeaders,\n\tSimpleStreamOptions,\n\tStopReason,\n\tStreamFunction,\n\tStreamOptions,\n\tTextContent,\n\tThinkingContent,\n\tTool,\n\tToolCall,\n\tToolResultMessage,\n} from \"../types.ts\";\nimport { appendAssistantMessageDiagnostic } from \"../utils/diagnostics.ts\";\nimport { AssistantMessageEventStream } from \"../utils/event-stream.ts\";\nimport { headersToRecord } from \"../utils/headers.ts\";\nimport { parseJsonWithRepair, parseStreamingJson } from \"../utils/json-parse.ts\";\nimport { getPiUserAgent } from \"../utils/pi-user-agent.ts\";\nimport { getProviderEnvValue } from \"../utils/provider-env.ts\";\nimport { retryProviderRequest } from \"../utils/provider-retry.ts\";\nimport { sanitizeSurrogates } from \"../utils/sanitize-unicode.ts\";\nimport { getSystemMessageText, renderSystemMessageUpdate } from \"../utils/text.ts\";\nimport {\n\tgetCurrentTools,\n\tgetDeclaredTools,\n\tgetInitialSystemMessage,\n\thasToolRedefinitions,\n\tresolveTranscript,\n\ttype TranscriptContext,\n} from \"../utils/transcript.ts\";\n\nimport { getJsonSchemaToolParameters, resolveJsonSchemaStrictSampling } from \"./constrained-sampling.ts\";\nimport { buildCopilotDynamicHeaders, hasCopilotVisionInput } from \"./github-copilot-headers.ts\";\nimport { adjustMaxTokensForThinking, buildBaseOptions, clampMaxTokensToContext } from \"./simple-options.ts\";\nimport { transformMessages } from \"./transform-messages.ts\";\n\n/**\n * Resolve cache retention preference.\n * Defaults to \"short\" and uses PI_CACHE_RETENTION for backward compatibility.\n */\nfunction resolveCacheRetention(cacheRetention?: CacheRetention, env?: ProviderEnv): CacheRetention {\n\tif (cacheRetention) {\n\t\treturn cacheRetention;\n\t}\n\tif (getProviderEnvValue(\"PI_CACHE_RETENTION\", env) === \"long\") {\n\t\treturn \"long\";\n\t}\n\treturn \"short\";\n}\n\nfunction getCacheControl(\n\tmodel: Model<\"anthropic-messages\">,\n\tcacheRetention?: CacheRetention,\n\tenv?: ProviderEnv,\n): { retention: CacheRetention; cacheControl?: CacheControlEphemeral } {\n\tconst retention = resolveCacheRetention(cacheRetention, env);\n\tif (retention === \"none\") {\n\t\treturn { retention };\n\t}\n\tconst ttl = retention === \"long\" && getAnthropicCompat(model).supportsLongCacheRetention ? \"1h\" : undefined;\n\treturn {\n\t\tretention,\n\t\tcacheControl: { type: \"ephemeral\", ...(ttl && { ttl }) },\n\t};\n}\n\n// Stealth mode: Mimic Claude Code's tool naming exactly\nconst claudeCodeVersion = \"2.1.280\";\n\n// Claude Code 2.x tool names (canonical casing)\n// Source: https://cchistory.mariozechner.at/data/prompts-2.1.11.md\n// To update: https://github.com/badlogic/cchistory\nconst claudeCodeTools = [\n\t\"Read\",\n\t\"Write\",\n\t\"Edit\",\n\t\"Bash\",\n\t\"Grep\",\n\t\"Glob\",\n\t\"AskUserQuestion\",\n\t\"EnterPlanMode\",\n\t\"ExitPlanMode\",\n\t\"KillShell\",\n\t\"NotebookEdit\",\n\t\"Skill\",\n\t\"Task\",\n\t\"TaskOutput\",\n\t\"TodoWrite\",\n\t\"WebFetch\",\n\t\"WebSearch\",\n];\n\nconst ccToolLookup = new Map(claudeCodeTools.map((t) => [t.toLowerCase(), t]));\n\n// Convert tool name to CC canonical casing if it matches (case-insensitive)\nconst toClaudeCodeName = (name: string) => ccToolLookup.get(name.toLowerCase()) ?? name;\nconst fromClaudeCodeName = (name: string, tools?: Tool[]) => {\n\tif (tools && tools.length > 0) {\n\t\tconst lowerName = name.toLowerCase();\n\t\tconst matchedTool = tools.find((tool) => tool.name.toLowerCase() === lowerName);\n\t\tif (matchedTool) return matchedTool.name;\n\t}\n\treturn name;\n};\n\n/**\n * Convert content blocks to Anthropic API format\n */\nfunction convertContentBlocks(content: (TextContent | ImageContent)[]):\n\t| string\n\t| Array<\n\t\t\t| { type: \"text\"; text: string }\n\t\t\t| {\n\t\t\t\t\ttype: \"image\";\n\t\t\t\t\tsource: {\n\t\t\t\t\t\ttype: \"base64\";\n\t\t\t\t\t\tmedia_type: \"image/jpeg\" | \"image/png\" | \"image/gif\" | \"image/webp\";\n\t\t\t\t\t\tdata: string;\n\t\t\t\t\t};\n\t\t\t  }\n\t  > {\n\t// If only text blocks, return as concatenated string for simplicity\n\tconst hasImages = content.some((c) => c.type === \"image\");\n\tif (!hasImages) {\n\t\treturn sanitizeSurrogates(content.map((c) => (c as TextContent).text).join(\"\\n\"));\n\t}\n\n\t// If we have images, convert to content block array\n\tconst blocks = content.map((block) => {\n\t\tif (block.type === \"text\") {\n\t\t\treturn {\n\t\t\t\ttype: \"text\" as const,\n\t\t\t\ttext: sanitizeSurrogates(block.text),\n\t\t\t};\n\t\t}\n\t\treturn {\n\t\t\ttype: \"image\" as const,\n\t\t\tsource: {\n\t\t\t\ttype: \"base64\" as const,\n\t\t\t\tmedia_type: block.mimeType as \"image/jpeg\" | \"image/png\" | \"image/gif\" | \"image/webp\",\n\t\t\t\tdata: block.data,\n\t\t\t},\n\t\t};\n\t});\n\n\t// If only images (no text), add placeholder text block\n\tconst hasText = blocks.some((b) => b.type === \"text\");\n\tif (!hasText) {\n\t\tblocks.unshift({\n\t\t\ttype: \"text\" as const,\n\t\t\ttext: \"(see attached image)\",\n\t\t});\n\t}\n\n\treturn blocks;\n}\n\nexport type AnthropicEffort = \"low\" | \"medium\" | \"high\" | \"xhigh\" | \"max\";\n\nexport type AnthropicThinkingDisplay = \"summarized\" | \"omitted\";\n\nconst FINE_GRAINED_TOOL_STREAMING_BETA = \"fine-grained-tool-streaming-2025-05-14\";\nconst INTERLEAVED_THINKING_BETA = \"interleaved-thinking-2025-05-14\";\nconst SERVER_SIDE_FALLBACK_BETA = \"server-side-fallback-2026-07-01\";\nconst MID_CONVERSATION_OUTPUT_CONFIG_BETA = \"mid-conversation-output-config-2026-07-01\";\nconst THINKING_BINDING_CONTROLS_BETA = \"thinking-binding-controls-2026-08-01\";\nconst MID_CONVERSATION_TOOL_CHANGES_BETA = \"mid-conversation-tool-changes-2026-07-01\";\n\n/**\n * Stable deferred tool declared whenever native tool changes are in use. Anthropic adds\n * hidden prompt scaffolding as soon as any tool has `defer_loading`; declaring this\n * placeholder from the first request keeps that scaffolding in the cached prefix, so the\n * first real late tool does not invalidate the cache (measured: full miss without it).\n * It is never activated and the model cannot see it.\n */\nconst DEFERRED_TOOL_PLACEHOLDER: BetaTool = {\n\tname: \"__pi_deferred_placeholder__\",\n\tdescription: \"Reserved placeholder. Never available. Never call this.\",\n\tinput_schema: { type: \"object\", properties: {}, required: [] },\n\tdefer_loading: true,\n};\n\nfunction shouldUseServerSideFallbackBeta(model: Model<\"anthropic-messages\">): boolean {\n\treturn (model.compat?.allowedFallbackModels?.length ?? 0) > 0;\n}\n\nfunction getAnthropicCompat(model: Model<\"anthropic-messages\">) {\n\tconst isOpenRouter = model.provider === \"openrouter\" || model.baseUrl.includes(\"openrouter.ai\");\n\treturn {\n\t\tsupportsEagerToolInputStreaming: model.compat?.supportsEagerToolInputStreaming ?? true,\n\t\tsupportsLongCacheRetention: model.compat?.supportsLongCacheRetention ?? true,\n\t\tsendSessionAffinityHeaders: model.compat?.sendSessionAffinityHeaders ?? isOpenRouter,\n\t\tsessionAffinityFormat: model.compat?.sessionAffinityFormat ?? (isOpenRouter ? \"openrouter\" : undefined),\n\t\tsupportsCacheControlOnTools: model.compat?.supportsCacheControlOnTools ?? true,\n\t\tsupportsTemperature: model.compat?.supportsTemperature ?? true,\n\t\tallowEmptySignature: model.compat?.allowEmptySignature ?? false,\n\t\tsupportsStrictTools: model.compat?.supportsStrictTools ?? false,\n\t\tsupportsMidConvoSystemMessages: model.compat?.supportsMidConvoSystemMessages ?? false,\n\t\tsupportsMidConvoToolChanges: model.compat?.supportsMidConvoToolChanges ?? false,\n\t};\n}\n\nexport interface AnthropicOptions extends StreamOptions {\n\t/**\n\t * Enable extended thinking.\n\t * For adaptive thinking models: the model decides when/how much to think.\n\t * For older models: uses budget-based thinking with thinkingBudgetTokens.\n\t * Default: undefined (thinking is omitted unless `streamSimple()` maps\n\t * a simple reasoning level to this option, or callers set it explicitly).\n\t */\n\tthinkingEnabled?: boolean;\n\t/**\n\t * Token budget for extended thinking (older models only).\n\t * Ignored for adaptive thinking models.\n\t * Default: 1024 when `thinkingEnabled` is true and no budget is provided.\n\t */\n\tthinkingBudgetTokens?: number;\n\t/**\n\t * Effort level for adaptive thinking models.\n\t * Controls how much thinking Claude allocates:\n\t * - \"max\": Always thinks with no constraints (Opus 4.6 only)\n\t * - \"xhigh\": Highest reasoning level (Opus 4.7+, Fable 5)\n\t * - \"high\": Always thinks, deep reasoning\n\t * - \"medium\": Moderate thinking, may skip for simple queries\n\t * - \"low\": Minimal thinking, skips for simple tasks\n\t * Ignored for older models.\n\t * Default: omitted unless `streamSimple()` maps a simple reasoning\n\t * level to this option.\n\t */\n\teffort?: AnthropicEffort;\n\t/**\n\t * Controls how thinking content is returned in API responses.\n\t * - \"summarized\": Thinking blocks contain summarized thinking text.\n\t * - \"omitted\": Thinking blocks return an empty thinking field; the encrypted\n\t *   signature still travels back for multi-turn continuity. Use for faster\n\t *   time-to-first-text-token when your UI does not surface thinking.\n\t *\n\t * Note: Anthropic's API default for Claude Opus 4.7 and Claude Mythos Preview\n\t * is \"omitted\". We default to \"summarized\" here to keep behavior consistent\n\t * with older Claude 4 models. Set this explicitly to \"omitted\" to opt in.\n\t * Default: \"summarized\" when thinking is enabled.\n\t */\n\tthinkingDisplay?: AnthropicThinkingDisplay;\n\t/**\n\t * Whether to request the interleaved thinking beta header for non-adaptive\n\t * thinking models. Adaptive thinking models have interleaved thinking built in,\n\t * so the header is skipped for them regardless of this setting.\n\t * Default: true.\n\t */\n\tinterleavedThinking?: boolean;\n\t/**\n\t * Anthropic tool choice behavior. String values map to Anthropic's built-in\n\t * choices; `{ type: \"tool\", name }` forces a specific tool.\n\t * Default: omitted (Anthropic default behavior, currently equivalent to auto).\n\t */\n\ttoolChoice?: \"auto\" | \"any\" | \"none\" | { type: \"tool\"; name: string };\n\t/**\n\t * Pre-built Anthropic client instance. When provided, skips internal client\n\t * construction entirely. Use this to inject alternative SDK clients such as\n\t * `AnthropicVertex` that shares the same messaging API.\n\t */\n\tclient?: Anthropic;\n}\n\nfunction mergeHeaders(...headerSources: (ProviderHeaders | undefined)[]): ProviderHeaders {\n\tconst merged: ProviderHeaders = {};\n\tfor (const headers of headerSources) {\n\t\tif (headers) {\n\t\t\tObject.assign(merged, headers);\n\t\t}\n\t}\n\treturn merged;\n}\n\nfunction mergeClientHeaders(...headerSources: (ProviderHeaders | undefined)[]): ProviderHeaders {\n\treturn mergeHeaders({ \"User-Agent\": getPiUserAgent() }, ...headerSources);\n}\n\nfunction hasHeader(headers: ProviderHeaders | undefined, name: string): boolean {\n\tif (!headers) return false;\n\tconst expected = name.toLowerCase();\n\tfor (const [key, value] of Object.entries(headers)) {\n\t\tif (key.toLowerCase() === expected && value !== null && value.trim().length > 0) return true;\n\t}\n\treturn false;\n}\n\nfunction assertRequestAuth(provider: string, apiKey: string | undefined, headers: ProviderHeaders | undefined): void {\n\tif (apiKey) return;\n\tif (\n\t\thasHeader(headers, \"authorization\") ||\n\t\thasHeader(headers, \"x-api-key\") ||\n\t\thasHeader(headers, \"cf-aig-authorization\")\n\t) {\n\t\treturn;\n\t}\n\tthrow new Error(`No API key for provider: ${provider}`);\n}\n\ninterface ServerSentEvent {\n\tevent: string | null;\n\tdata: string;\n\traw: string[];\n}\n\ninterface SseDecoderState {\n\tevent: string | null;\n\tdata: string[];\n\traw: string[];\n}\n\nconst ANTHROPIC_MESSAGE_EVENTS: ReadonlySet<string> = new Set([\n\t\"message_start\",\n\t\"message_delta\",\n\t\"message_stop\",\n\t\"content_block_start\",\n\t\"content_block_delta\",\n\t\"content_block_stop\",\n]);\n\nfunction flushSseEvent(state: SseDecoderState): ServerSentEvent | null {\n\tif (!state.event && state.data.length === 0) {\n\t\treturn null;\n\t}\n\n\tconst event: ServerSentEvent = {\n\t\tevent: state.event,\n\t\tdata: state.data.join(\"\\n\"),\n\t\traw: [...state.raw],\n\t};\n\tstate.event = null;\n\tstate.data = [];\n\tstate.raw = [];\n\treturn event;\n}\n\nfunction decodeSseLine(line: string, state: SseDecoderState): ServerSentEvent | null {\n\tif (line === \"\") {\n\t\treturn flushSseEvent(state);\n\t}\n\n\tstate.raw.push(line);\n\tif (line.startsWith(\":\")) {\n\t\treturn null;\n\t}\n\n\tconst delimiterIndex = line.indexOf(\":\");\n\tconst fieldName = delimiterIndex === -1 ? line : line.slice(0, delimiterIndex);\n\tlet value = delimiterIndex === -1 ? \"\" : line.slice(delimiterIndex + 1);\n\tif (value.startsWith(\" \")) {\n\t\tvalue = value.slice(1);\n\t}\n\n\tif (fieldName === \"event\") {\n\t\tstate.event = value;\n\t} else if (fieldName === \"data\") {\n\t\tstate.data.push(value);\n\t}\n\n\treturn null;\n}\n\nfunction nextLineBreakIndex(text: string): number {\n\tconst carriageReturnIndex = text.indexOf(\"\\r\");\n\tconst newlineIndex = text.indexOf(\"\\n\");\n\tif (carriageReturnIndex === -1) {\n\t\treturn newlineIndex;\n\t}\n\tif (newlineIndex === -1) {\n\t\treturn carriageReturnIndex;\n\t}\n\treturn Math.min(carriageReturnIndex, newlineIndex);\n}\n\nfunction consumeLine(text: string): { line: string; rest: string } | null {\n\tconst lineBreakIndex = nextLineBreakIndex(text);\n\tif (lineBreakIndex === -1) {\n\t\treturn null;\n\t}\n\n\tlet nextIndex = lineBreakIndex + 1;\n\tif (text[lineBreakIndex] === \"\\r\" && text[nextIndex] === \"\\n\") {\n\t\tnextIndex += 1;\n\t}\n\n\treturn {\n\t\tline: text.slice(0, lineBreakIndex),\n\t\trest: text.slice(nextIndex),\n\t};\n}\n\nasync function* iterateSseMessages(\n\tbody: ReadableStream<Uint8Array>,\n\tsignal?: AbortSignal,\n): AsyncGenerator<ServerSentEvent> {\n\tconst reader = body.getReader();\n\tconst decoder = new TextDecoder();\n\tconst state: SseDecoderState = { event: null, data: [], raw: [] };\n\tlet buffer = \"\";\n\n\ttry {\n\t\twhile (true) {\n\t\t\tif (signal?.aborted) {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\n\t\t\tconst { value, done } = await reader.read();\n\t\t\tif (done) {\n\t\t\t\tbreak;\n\t\t\t}\n\n\t\t\tbuffer += decoder.decode(value, { stream: true });\n\t\t\tlet consumed = consumeLine(buffer);\n\t\t\twhile (consumed) {\n\t\t\t\tbuffer = consumed.rest;\n\t\t\t\tconst event = decodeSseLine(consumed.line, state);\n\t\t\t\tif (event) {\n\t\t\t\t\tyield event;\n\t\t\t\t}\n\t\t\t\tconsumed = consumeLine(buffer);\n\t\t\t}\n\t\t}\n\n\t\tbuffer += decoder.decode();\n\t\tlet consumed = consumeLine(buffer);\n\t\twhile (consumed) {\n\t\t\tbuffer = consumed.rest;\n\t\t\tconst event = decodeSseLine(consumed.line, state);\n\t\t\tif (event) {\n\t\t\t\tyield event;\n\t\t\t}\n\t\t\tconsumed = consumeLine(buffer);\n\t\t}\n\n\t\tif (buffer.length > 0) {\n\t\t\tconst event = decodeSseLine(buffer, state);\n\t\t\tif (event) {\n\t\t\t\tyield event;\n\t\t\t}\n\t\t}\n\n\t\tconst trailingEvent = flushSseEvent(state);\n\t\tif (trailingEvent) {\n\t\t\tyield trailingEvent;\n\t\t}\n\t} finally {\n\t\treader.releaseLock();\n\t}\n}\n\nasync function* iterateAnthropicEvents(\n\tresponse: Response,\n\tsignal?: AbortSignal,\n): AsyncGenerator<RawMessageStreamEvent> {\n\tif (!response.body) {\n\t\tthrow new Error(\"Attempted to iterate over an Anthropic response with no body\");\n\t}\n\n\tlet sawMessageStart = false;\n\tlet sawMessageEnd = false;\n\n\tfor await (const sse of iterateSseMessages(response.body, signal)) {\n\t\tif (sse.event === \"error\") {\n\t\t\tthrow new Error(sse.data);\n\t\t}\n\n\t\tif (!ANTHROPIC_MESSAGE_EVENTS.has(sse.event ?? \"\")) {\n\t\t\tcontinue;\n\t\t}\n\n\t\ttry {\n\t\t\tconst event = parseJsonWithRepair<RawMessageStreamEvent>(sse.data);\n\t\t\tif (event.type === \"message_start\") {\n\t\t\t\tsawMessageStart = true;\n\t\t\t} else if (event.type === \"message_stop\") {\n\t\t\t\tsawMessageEnd = true;\n\t\t\t}\n\t\t\tyield event;\n\t\t} catch (error) {\n\t\t\tconst message = error instanceof Error ? error.message : String(error);\n\t\t\tthrow new Error(\n\t\t\t\t`Could not parse Anthropic SSE event ${sse.event}: ${message}; data=${sse.data}; raw=${sse.raw.join(\"\\\\n\")}`,\n\t\t\t);\n\t\t}\n\t}\n\n\tif (sawMessageStart && !sawMessageEnd) {\n\t\tthrow new Error(\"Anthropic stream ended before message_stop\");\n\t}\n}\n\nexport const stream: StreamFunction<\"anthropic-messages\", AnthropicOptions> = (\n\tmodel: Model<\"anthropic-messages\">,\n\tcontext: TranscriptContext,\n\toptions?: AnthropicOptions,\n): AssistantMessageEventStream => {\n\tconst stream = new AssistantMessageEventStream();\n\tconst normalizedContext = resolveTranscript(context, getAnthropicCompat(model).supportsMidConvoSystemMessages);\n\tconst currentTools = getCurrentTools(normalizedContext.messages);\n\n\t(async () => {\n\t\tconst providerThinkingLevel = model.compat?.supportsMidConvoEffort ? (options?.effort ?? \"high\") : undefined;\n\t\tconst output: AssistantMessage = {\n\t\t\trole: \"assistant\",\n\t\t\tcontent: [],\n\t\t\tapi: model.api as Api,\n\t\t\tprovider: model.provider,\n\t\t\tmodel: model.id,\n\t\t\t...(providerThinkingLevel === undefined ? {} : { providerThinkingLevel }),\n\t\t\tusage: {\n\t\t\t\tinput: 0,\n\t\t\t\toutput: 0,\n\t\t\t\tcacheRead: 0,\n\t\t\t\tcacheWrite: 0,\n\t\t\t\ttotalTokens: 0,\n\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t},\n\t\t\tstopReason: \"pending\",\n\t\t\ttimestamp: Date.now(),\n\t\t};\n\n\t\ttry {\n\t\t\tlet client: Anthropic;\n\t\t\tlet isOAuth: boolean;\n\t\t\tlet usageModel = model;\n\t\t\tlet inputTransformations: BetaThinkingDroppedInputTransformation[] | undefined;\n\n\t\t\tif (options?.client) {\n\t\t\t\tclient = options.client;\n\t\t\t\tisOAuth = false;\n\t\t\t} else {\n\t\t\t\tconst apiKey = options?.apiKey;\n\t\t\t\tassertRequestAuth(model.provider, apiKey, options?.headers);\n\n\t\t\t\tlet copilotDynamicHeaders: Record<string, string> | undefined;\n\t\t\t\tif (model.provider === \"github-copilot\") {\n\t\t\t\t\tconst hasImages = hasCopilotVisionInput(normalizedContext.messages);\n\t\t\t\t\tcopilotDynamicHeaders = buildCopilotDynamicHeaders({\n\t\t\t\t\t\tmessages: normalizedContext.messages,\n\t\t\t\t\t\thasImages,\n\t\t\t\t\t});\n\t\t\t\t}\n\n\t\t\t\tconst cacheRetention = resolveCacheRetention(options?.cacheRetention, options?.env);\n\t\t\t\tconst cacheSessionId = cacheRetention === \"none\" ? undefined : options?.sessionId;\n\n\t\t\t\tconst created = createClient(\n\t\t\t\t\tmodel,\n\t\t\t\t\tapiKey,\n\t\t\t\t\toptions?.headers,\n\t\t\t\t\toptions?.fetch,\n\t\t\t\t\tcopilotDynamicHeaders,\n\t\t\t\t\tcacheSessionId,\n\t\t\t\t);\n\t\t\t\tclient = created.client;\n\t\t\t\tisOAuth = created.isOAuthToken;\n\t\t\t}\n\t\t\tlet params = buildParams(model, normalizedContext, isOAuth, options);\n\t\t\tconst nextParams = await options?.onPayload?.(params, model);\n\t\t\tif (nextParams !== undefined) {\n\t\t\t\tparams = { ...(nextParams as MessageCreateParamsStreaming), stream: true };\n\t\t\t}\n\t\t\tconst requestOptions = {\n\t\t\t\t...(options?.signal ? { signal: options.signal } : {}),\n\t\t\t\t...(options?.timeoutMs !== undefined ? { timeout: options.timeoutMs } : {}),\n\t\t\t\tmaxRetries: 0,\n\t\t\t};\n\t\t\tconst response = await retryProviderRequest(\n\t\t\t\t() => client.beta.messages.create(params, requestOptions).asResponse(),\n\t\t\t\t{\n\t\t\t\t\tmaxRetries: options?.maxRetries,\n\t\t\t\t\tmaxRetryDelayMs: options?.maxRetryDelayMs,\n\t\t\t\t\tsignal: options?.signal,\n\t\t\t\t},\n\t\t\t);\n\t\t\tawait options?.onResponse?.({ status: response.status, headers: headersToRecord(response.headers) }, model);\n\t\t\tstream.push({ type: \"start\", partial: output });\n\n\t\t\ttype Block = (ThinkingContent | TextContent | (ToolCall & { partialJson: string })) & { index: number };\n\t\t\tconst blocks = output.content as Block[];\n\n\t\t\tfor await (const event of iterateAnthropicEvents(response, options?.signal)) {\n\t\t\t\tif (event.type === \"message_start\") {\n\t\t\t\t\toutput.responseId = event.message.id;\n\t\t\t\t\tconst transformations = event.message.input_transformations;\n\t\t\t\t\tif (Array.isArray(transformations)) inputTransformations = transformations;\n\t\t\t\t\tconst responseModel = event.message.model;\n\t\t\t\t\tif (responseModel !== model.id) output.responseModel = responseModel;\n\t\t\t\t\tconst fallbackCost =\n\t\t\t\t\t\tresponseModel === model.id\n\t\t\t\t\t\t\t? undefined\n\t\t\t\t\t\t\t: model.compat?.allowedFallbackModels?.find(\n\t\t\t\t\t\t\t\t\t(fallback) => fallback.provider === model.provider && fallback.model === responseModel,\n\t\t\t\t\t\t\t\t)?.cost;\n\t\t\t\t\tusageModel = fallbackCost ? { ...model, id: responseModel, cost: fallbackCost } : model;\n\t\t\t\t\t// Capture initial token usage from message_start event\n\t\t\t\t\t// This ensures we have input token counts even if the stream is aborted early\n\t\t\t\t\toutput.usage.input = event.message.usage.input_tokens || 0;\n\t\t\t\t\toutput.usage.output = event.message.usage.output_tokens || 0;\n\t\t\t\t\toutput.usage.cacheRead = event.message.usage.cache_read_input_tokens || 0;\n\t\t\t\t\toutput.usage.cacheWrite = event.message.usage.cache_creation_input_tokens || 0;\n\t\t\t\t\toutput.usage.cacheWrite1h = event.message.usage.cache_creation?.ephemeral_1h_input_tokens || 0;\n\t\t\t\t\t// Anthropic doesn't provide total_tokens, compute from components\n\t\t\t\t\toutput.usage.totalTokens =\n\t\t\t\t\t\toutput.usage.input + output.usage.output + output.usage.cacheRead + output.usage.cacheWrite;\n\t\t\t\t\tcalculateCost(usageModel, output.usage);\n\t\t\t\t} else if (event.type === \"content_block_start\") {\n\t\t\t\t\tif (event.content_block.type === \"fallback\") {\n\t\t\t\t\t\tif (output.content.length > 0) {\n\t\t\t\t\t\t\tthrow new Error(\"Anthropic performed an unsupported mid-output model fallback\");\n\t\t\t\t\t\t}\n\t\t\t\t\t\tcontinue;\n\t\t\t\t\t}\n\t\t\t\t\tif (event.content_block.type === \"text\") {\n\t\t\t\t\t\tconst block: Block = {\n\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\ttext: event.content_block.text ?? \"\",\n\t\t\t\t\t\t\tindex: event.index,\n\t\t\t\t\t\t};\n\t\t\t\t\t\toutput.content.push(block);\n\t\t\t\t\t\tstream.push({ type: \"text_start\", contentIndex: output.content.length - 1, partial: output });\n\t\t\t\t\t} else if (event.content_block.type === \"thinking\") {\n\t\t\t\t\t\tconst block: Block = {\n\t\t\t\t\t\t\ttype: \"thinking\",\n\t\t\t\t\t\t\tthinking: event.content_block.thinking ?? \"\",\n\t\t\t\t\t\t\tthinkingSignature: event.content_block.signature ?? \"\",\n\t\t\t\t\t\t\tindex: event.index,\n\t\t\t\t\t\t};\n\t\t\t\t\t\toutput.content.push(block);\n\t\t\t\t\t\tstream.push({ type: \"thinking_start\", contentIndex: output.content.length - 1, partial: output });\n\t\t\t\t\t} else if (event.content_block.type === \"redacted_thinking\") {\n\t\t\t\t\t\tconst block: Block = {\n\t\t\t\t\t\t\ttype: \"thinking\",\n\t\t\t\t\t\t\tthinking: \"[Reasoning redacted]\",\n\t\t\t\t\t\t\tthinkingSignature: event.content_block.data,\n\t\t\t\t\t\t\tredacted: true,\n\t\t\t\t\t\t\tindex: event.index,\n\t\t\t\t\t\t};\n\t\t\t\t\t\toutput.content.push(block);\n\t\t\t\t\t\tstream.push({ type: \"thinking_start\", contentIndex: output.content.length - 1, partial: output });\n\t\t\t\t\t} else if (event.content_block.type === \"tool_use\") {\n\t\t\t\t\t\tconst block: Block = {\n\t\t\t\t\t\t\ttype: \"toolCall\",\n\t\t\t\t\t\t\tid: event.content_block.id,\n\t\t\t\t\t\t\tname: isOAuth\n\t\t\t\t\t\t\t\t? fromClaudeCodeName(event.content_block.name, currentTools)\n\t\t\t\t\t\t\t\t: event.content_block.name,\n\t\t\t\t\t\t\targuments: (event.content_block.input as Record<string, any>) ?? {},\n\t\t\t\t\t\t\tpartialJson: \"\",\n\t\t\t\t\t\t\tindex: event.index,\n\t\t\t\t\t\t};\n\t\t\t\t\t\toutput.content.push(block);\n\t\t\t\t\t\tstream.push({ type: \"toolcall_start\", contentIndex: output.content.length - 1, partial: output });\n\t\t\t\t\t}\n\t\t\t\t} else if (event.type === \"content_block_delta\") {\n\t\t\t\t\tif (event.delta.type === \"text_delta\") {\n\t\t\t\t\t\tconst index = blocks.findIndex((b) => b.index === event.index);\n\t\t\t\t\t\tconst block = blocks[index];\n\t\t\t\t\t\tif (block && block.type === \"text\") {\n\t\t\t\t\t\t\tblock.text += event.delta.text;\n\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\ttype: \"text_delta\",\n\t\t\t\t\t\t\t\tcontentIndex: index,\n\t\t\t\t\t\t\t\tdelta: event.delta.text,\n\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t} else if (event.delta.type === \"thinking_delta\") {\n\t\t\t\t\t\tconst index = blocks.findIndex((b) => b.index === event.index);\n\t\t\t\t\t\tconst block = blocks[index];\n\t\t\t\t\t\tif (block && block.type === \"thinking\") {\n\t\t\t\t\t\t\tblock.thinking += event.delta.thinking;\n\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\ttype: \"thinking_delta\",\n\t\t\t\t\t\t\t\tcontentIndex: index,\n\t\t\t\t\t\t\t\tdelta: event.delta.thinking,\n\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t} else if (event.delta.type === \"input_json_delta\") {\n\t\t\t\t\t\tconst index = blocks.findIndex((b) => b.index === event.index);\n\t\t\t\t\t\tconst block = blocks[index];\n\t\t\t\t\t\tif (block && block.type === \"toolCall\") {\n\t\t\t\t\t\t\tblock.partialJson += event.delta.partial_json;\n\t\t\t\t\t\t\tblock.arguments = parseStreamingJson(block.partialJson);\n\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\ttype: \"toolcall_delta\",\n\t\t\t\t\t\t\t\tcontentIndex: index,\n\t\t\t\t\t\t\t\tdelta: event.delta.partial_json,\n\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t} else if (event.delta.type === \"signature_delta\") {\n\t\t\t\t\t\tconst index = blocks.findIndex((b) => b.index === event.index);\n\t\t\t\t\t\tconst block = blocks[index];\n\t\t\t\t\t\tif (block && block.type === \"thinking\") {\n\t\t\t\t\t\t\tblock.thinkingSignature = block.thinkingSignature || \"\";\n\t\t\t\t\t\t\tblock.thinkingSignature += event.delta.signature;\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t} else if (event.type === \"content_block_stop\") {\n\t\t\t\t\tconst index = blocks.findIndex((b) => b.index === event.index);\n\t\t\t\t\tconst block = blocks[index];\n\t\t\t\t\tif (block) {\n\t\t\t\t\t\tdelete (block as any).index;\n\t\t\t\t\t\tif (block.type === \"text\") {\n\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\ttype: \"text_end\",\n\t\t\t\t\t\t\t\tcontentIndex: index,\n\t\t\t\t\t\t\t\tcontent: block.text,\n\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t} else if (block.type === \"thinking\") {\n\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\ttype: \"thinking_end\",\n\t\t\t\t\t\t\t\tcontentIndex: index,\n\t\t\t\t\t\t\t\tcontent: block.thinking,\n\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t} else if (block.type === \"toolCall\") {\n\t\t\t\t\t\t\tblock.arguments = parseStreamingJson(block.partialJson);\n\t\t\t\t\t\t\t// Finalize in-place and strip the scratch buffer so replay only\n\t\t\t\t\t\t\t// carries parsed arguments.\n\t\t\t\t\t\t\tdelete (block as { partialJson?: string }).partialJson;\n\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\ttype: \"toolcall_end\",\n\t\t\t\t\t\t\t\tcontentIndex: index,\n\t\t\t\t\t\t\t\ttoolCall: block,\n\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t} else if (event.type === \"message_delta\") {\n\t\t\t\t\tconst transformations = event.input_transformations;\n\t\t\t\t\tif (Array.isArray(transformations)) inputTransformations = transformations;\n\t\t\t\t\tif (event.delta.stop_reason) {\n\t\t\t\t\t\toutput.rawStopReason = event.delta.stop_reason;\n\t\t\t\t\t\tconst stopReasonResult = mapStopReason(event.delta.stop_reason, event.delta.stop_details);\n\t\t\t\t\t\toutput.stopReason = stopReasonResult.stopReason;\n\t\t\t\t\t\tif (stopReasonResult.errorMessage) {\n\t\t\t\t\t\t\toutput.errorMessage = stopReasonResult.errorMessage;\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t\t// Only update usage fields if present (not null).\n\t\t\t\t\t// Preserves input_tokens from message_start when proxies omit it in message_delta.\n\t\t\t\t\tif (event.usage) {\n\t\t\t\t\t\tif (event.usage.input_tokens != null) {\n\t\t\t\t\t\t\toutput.usage.input = event.usage.input_tokens;\n\t\t\t\t\t\t}\n\t\t\t\t\t\tif (event.usage.output_tokens != null) {\n\t\t\t\t\t\t\toutput.usage.output = event.usage.output_tokens;\n\t\t\t\t\t\t}\n\t\t\t\t\t\tif (event.usage.cache_read_input_tokens != null) {\n\t\t\t\t\t\t\toutput.usage.cacheRead = event.usage.cache_read_input_tokens;\n\t\t\t\t\t\t}\n\t\t\t\t\t\tif (event.usage.cache_creation_input_tokens != null) {\n\t\t\t\t\t\t\toutput.usage.cacheWrite = event.usage.cache_creation_input_tokens;\n\t\t\t\t\t\t}\n\t\t\t\t\t\t// Anthropic reports reasoning tokens as a subset of output tokens.\n\t\t\t\t\t\tconst thinkingTokens = event.usage.output_tokens_details?.thinking_tokens;\n\t\t\t\t\t\tif (thinkingTokens != null) {\n\t\t\t\t\t\t\toutput.usage.reasoning = thinkingTokens;\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t\t// Anthropic doesn't provide total_tokens, compute from components\n\t\t\t\t\toutput.usage.totalTokens =\n\t\t\t\t\t\toutput.usage.input + output.usage.output + output.usage.cacheRead + output.usage.cacheWrite;\n\t\t\t\t\tcalculateCost(usageModel, output.usage);\n\t\t\t\t}\n\t\t\t}\n\n\t\t\tif (options?.signal?.aborted) {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\n\t\t\tif (output.stopReason === \"pending\") {\n\t\t\t\tthrow new Error(\"Anthropic stream ended without a stop reason\");\n\t\t\t}\n\t\t\tif (output.stopReason === \"aborted\" || output.stopReason === \"error\") {\n\t\t\t\tthrow new Error(output.errorMessage || \"An unknown error occurred\");\n\t\t\t}\n\t\t\tif (inputTransformations && inputTransformations.length > 0) {\n\t\t\t\tappendAssistantMessageDiagnostic(output, {\n\t\t\t\t\ttype: \"anthropic_input_transformations\",\n\t\t\t\t\ttimestamp: Date.now(),\n\t\t\t\t\tdetails: {\n\t\t\t\t\t\ttransformations: inputTransformations.map((transformation) => ({\n\t\t\t\t\t\t\ttype: transformation.type ?? undefined,\n\t\t\t\t\t\t\tpath: transformation.path ?? undefined,\n\t\t\t\t\t\t\treason: transformation.reason ?? undefined,\n\t\t\t\t\t\t})),\n\t\t\t\t\t},\n\t\t\t\t});\n\t\t\t}\n\n\t\t\tstream.push({ type: \"done\", reason: output.stopReason, message: output });\n\t\t\tstream.end();\n\t\t} catch (error) {\n\t\t\tfor (const block of output.content) {\n\t\t\t\tdelete (block as { index?: number }).index;\n\t\t\t\t// partialJson is only a streaming scratch buffer; never persist it.\n\t\t\t\tdelete (block as { partialJson?: string }).partialJson;\n\t\t\t}\n\t\t\toutput.stopReason = options?.signal?.aborted ? \"aborted\" : \"error\";\n\t\t\toutput.errorMessage = error instanceof Error ? error.message : JSON.stringify(error);\n\t\t\tstream.push({ type: \"error\", reason: output.stopReason, error: output });\n\t\t\tstream.end();\n\t\t}\n\t})();\n\n\treturn stream;\n};\n\n/**\n * Map ThinkingLevel to Anthropic effort levels for adaptive thinking.\n * Note: effort \"max\" is available on all adaptive-thinking Claude models, while native\n * \"xhigh\" is only available on Opus 4.7/4.8, Sonnet 5, and Fable 5.\n */\nfunction mapThinkingLevelToEffort(\n\tmodel: Model<\"anthropic-messages\">,\n\tlevel: SimpleStreamOptions[\"reasoning\"],\n): AnthropicEffort {\n\tconst mapped = level ? model.thinkingLevelMap?.[level] : undefined;\n\tif (typeof mapped === \"string\") return mapped as AnthropicEffort;\n\n\tswitch (level) {\n\t\tcase \"minimal\":\n\t\tcase \"low\":\n\t\t\treturn \"low\";\n\t\tcase \"medium\":\n\t\t\treturn \"medium\";\n\t\tcase \"high\":\n\t\t\treturn \"high\";\n\t\tdefault:\n\t\t\treturn \"high\";\n\t}\n}\n\nexport const streamSimple: StreamFunction<\"anthropic-messages\", SimpleStreamOptions> = (\n\tmodel: Model<\"anthropic-messages\">,\n\tcontext: TranscriptContext,\n\toptions?: SimpleStreamOptions,\n): AssistantMessageEventStream => {\n\tassertRequestAuth(model.provider, options?.apiKey, options?.headers);\n\n\tconst base = {\n\t\t...buildBaseOptions(model, context, options, options?.apiKey),\n\t\ttoolChoice: options?.toolChoice,\n\t} satisfies AnthropicOptions;\n\tif (!options?.reasoning) {\n\t\treturn stream(model, context, {\n\t\t\t...base,\n\t\t\tthinkingEnabled: false,\n\t\t} satisfies AnthropicOptions);\n\t}\n\n\t// For models with adaptive thinking: use an effort level.\n\t// For older models: use budget-based thinking.\n\tif (model.compat?.forceAdaptiveThinking === true) {\n\t\tconst effort = mapThinkingLevelToEffort(model, options.reasoning);\n\t\treturn stream(model, context, {\n\t\t\t...base,\n\t\t\tthinkingEnabled: true,\n\t\t\teffort,\n\t\t} satisfies AnthropicOptions);\n\t}\n\n\t// Undefined means the caller did not request an output cap; let the helper use the model cap.\n\t// Do not coerce to 0 here, or the thinking budget would become the entire max_tokens value.\n\tconst adjusted = adjustMaxTokensForThinking(\n\t\tbase.maxTokens,\n\t\tmodel.maxTokens,\n\t\toptions.reasoning,\n\t\toptions.thinkingBudgets,\n\t);\n\n\tconst maxTokens = clampMaxTokensToContext(model, context, adjusted.maxTokens);\n\n\treturn stream(model, context, {\n\t\t...base,\n\t\tmaxTokens,\n\t\tthinkingEnabled: true,\n\t\tthinkingBudgetTokens: Math.min(adjusted.thinkingBudget, Math.max(0, maxTokens - 1024)),\n\t} satisfies AnthropicOptions);\n};\n\nfunction isOAuthToken(apiKey: string): boolean {\n\treturn apiKey.includes(\"sk-ant-oat\");\n}\n\nfunction createClient(\n\tmodel: Model<\"anthropic-messages\">,\n\tapiKey: string | undefined,\n\toptionsHeaders?: ProviderHeaders,\n\tfetch?: typeof globalThis.fetch,\n\tdynamicHeaders?: Record<string, string>,\n\tsessionId?: string,\n): { client: Anthropic; isOAuthToken: boolean } {\n\t// Copilot: Bearer auth.\n\tif (model.provider === \"github-copilot\") {\n\t\tconst client = new Anthropic({\n\t\t\tapiKey: null,\n\t\t\tauthToken: apiKey ?? null,\n\t\t\tbaseURL: model.baseUrl,\n\t\t\tdangerouslyAllowBrowser: true,\n\t\t\tfetch,\n\t\t\tdefaultHeaders: mergeClientHeaders(\n\t\t\t\t{\n\t\t\t\t\taccept: \"application/json\",\n\t\t\t\t\t\"anthropic-dangerous-direct-browser-access\": \"true\",\n\t\t\t\t},\n\t\t\t\tmodel.headers,\n\t\t\t\tdynamicHeaders,\n\t\t\t\toptionsHeaders,\n\t\t\t),\n\t\t});\n\n\t\treturn { client, isOAuthToken: false };\n\t}\n\n\t// OAuth: Bearer auth, Claude Code identity headers\n\tif (apiKey && isOAuthToken(apiKey)) {\n\t\tconst client = new Anthropic({\n\t\t\tapiKey: null,\n\t\t\tauthToken: apiKey,\n\t\t\tbaseURL: model.baseUrl,\n\t\t\tdangerouslyAllowBrowser: true,\n\t\t\tfetch,\n\t\t\tdefaultHeaders: mergeClientHeaders(\n\t\t\t\t{\n\t\t\t\t\taccept: \"application/json\",\n\t\t\t\t\t\"anthropic-dangerous-direct-browser-access\": \"true\",\n\t\t\t\t\t\"user-agent\": `claude-cli/${claudeCodeVersion}`,\n\t\t\t\t\t\"x-app\": \"cli\",\n\t\t\t\t},\n\t\t\t\tmodel.headers,\n\t\t\t\toptionsHeaders,\n\t\t\t),\n\t\t});\n\n\t\treturn { client, isOAuthToken: true };\n\t}\n\n\t// API key or header-owned auth.\n\tconst compat = getAnthropicCompat(model);\n\tconst sessionAffinityHeaders: ProviderHeaders = {};\n\tif (sessionId && compat.sendSessionAffinityHeaders) {\n\t\tconst header = compat.sessionAffinityFormat === \"openrouter\" ? \"x-session-id\" : \"x-session-affinity\";\n\t\tsessionAffinityHeaders[header] = sessionId;\n\t}\n\tconst defaultHeaders = mergeClientHeaders(\n\t\t{\n\t\t\taccept: \"application/json\",\n\t\t\t\"anthropic-dangerous-direct-browser-access\": \"true\",\n\t\t},\n\t\tsessionAffinityHeaders,\n\t\tmodel.headers,\n\t\toptionsHeaders,\n\t);\n\tconst client = new Anthropic({\n\t\tapiKey: apiKey ?? null,\n\t\tauthToken: null,\n\t\tbaseURL: model.baseUrl,\n\t\tdangerouslyAllowBrowser: true,\n\t\tfetch,\n\t\tdefaultHeaders,\n\t});\n\n\treturn { client, isOAuthToken: false };\n}\n\nfunction getBetaFeatures(\n\tmodel: Model<\"anthropic-messages\">,\n\tcontext: TranscriptContext,\n\tisOAuthToken: boolean,\n\tnativeToolChanges: boolean,\n\toptions?: AnthropicOptions,\n): NonNullable<MessageCreateParamsStreaming[\"betas\"]> {\n\tlet configuredFeatures: string | null | undefined;\n\tfor (const headers of [model.headers, options?.headers]) {\n\t\tfor (const [name, value] of Object.entries(headers ?? {})) {\n\t\t\tif (name.toLowerCase() === \"anthropic-beta\") configuredFeatures = value;\n\t\t}\n\t}\n\tif (configuredFeatures === null) return [];\n\tif (configuredFeatures !== undefined) {\n\t\treturn [\n\t\t\t...new Set(\n\t\t\t\tconfiguredFeatures\n\t\t\t\t\t.split(\",\")\n\t\t\t\t\t.map((feature) => feature.trim())\n\t\t\t\t\t.filter((feature) => feature.length > 0),\n\t\t\t),\n\t\t];\n\t}\n\n\tconst features: NonNullable<MessageCreateParamsStreaming[\"betas\"]> = [];\n\tif (isOAuthToken) features.push(\"claude-code-20250219\", \"oauth-2025-04-20\");\n\tif (shouldUseFineGrainedToolStreamingBeta(model, context)) features.push(FINE_GRAINED_TOOL_STREAMING_BETA);\n\tif (\n\t\tmodel.reasoning &&\n\t\toptions?.thinkingEnabled === true &&\n\t\t(options.interleavedThinking ?? true) &&\n\t\tmodel.compat?.forceAdaptiveThinking !== true\n\t) {\n\t\tfeatures.push(INTERLEAVED_THINKING_BETA);\n\t}\n\tif (shouldUseServerSideFallbackBeta(model)) features.push(SERVER_SIDE_FALLBACK_BETA);\n\tif (model.compat?.supportsMidConvoEffort === true) {\n\t\tfeatures.push(MID_CONVERSATION_OUTPUT_CONFIG_BETA, THINKING_BINDING_CONTROLS_BETA);\n\t}\n\tif (nativeToolChanges) features.push(MID_CONVERSATION_TOOL_CHANGES_BETA);\n\treturn [...new Set(features)];\n}\n\nfunction buildParams(\n\tmodel: Model<\"anthropic-messages\">,\n\tcontext: TranscriptContext,\n\tisOAuthToken: boolean,\n\toptions?: AnthropicOptions,\n): MessageCreateParamsStreaming {\n\tconst { cacheControl } = getCacheControl(model, options?.cacheRetention, options?.env);\n\tconst compat = getAnthropicCompat(model);\n\tconst initialSystemMessage = getInitialSystemMessage(context.messages);\n\tconst initialSystemText = initialSystemMessage ? getSystemMessageText(initialSystemMessage) : \"\";\n\tconst transformedMessages = transformMessages(context.messages, model, normalizeToolCallId);\n\tconst conversationMessages = initialSystemMessage ? transformedMessages.slice(1) : transformedMessages;\n\t// Native tool changes reference tools by name, so a redefined name cannot be expressed,\n\t// and Anthropic rejects a tool list where every tool is deferred, so there must be an\n\t// initial active tool to anchor the deferred ones. Otherwise the current tool list is sent.\n\tconst initialTools = initialSystemMessage?.toolsAdded ?? [];\n\tconst nativeToolChanges =\n\t\tcompat.supportsMidConvoSystemMessages &&\n\t\tcompat.supportsMidConvoToolChanges &&\n\t\tinitialTools.length > 0 &&\n\t\t!hasToolRedefinitions(context.messages);\n\tconst converted = convertMessages(\n\t\tconversationMessages,\n\t\tisOAuthToken,\n\t\tcacheControl,\n\t\tcompat.allowEmptySignature,\n\t\tmodel.compat?.supportsMidConvoEffort === true ? model.provider : undefined,\n\t\tnativeToolChanges,\n\t);\n\tconst activeEffort = options?.effort ?? \"high\";\n\tconst betaFeatures = getBetaFeatures(model, context, isOAuthToken, nativeToolChanges, options);\n\tconst params: MessageCreateParamsStreaming = {\n\t\tmodel: model.id,\n\t\tmessages:\n\t\t\tmodel.compat?.supportsMidConvoEffort === true\n\t\t\t\t? insertThinkingLevelMessages(converted, activeEffort)\n\t\t\t\t: converted.messages,\n\t\tmax_tokens: options?.maxTokens ?? model.maxTokens,\n\t\tstream: true,\n\t\t...(betaFeatures.length > 0 ? { betas: betaFeatures } : {}),\n\t};\n\n\t// For OAuth tokens, we MUST include Claude Code identity\n\tif (isOAuthToken) {\n\t\tparams.system = [\n\t\t\t{\n\t\t\t\ttype: \"text\",\n\t\t\t\ttext: \"You are Claude Code, Anthropic's official CLI for Claude.\",\n\t\t\t\t...(cacheControl ? { cache_control: cacheControl } : {}),\n\t\t\t},\n\t\t];\n\t\tif (initialSystemText) {\n\t\t\tparams.system.push({\n\t\t\t\ttype: \"text\",\n\t\t\t\ttext: sanitizeSurrogates(initialSystemText),\n\t\t\t\t...(cacheControl ? { cache_control: cacheControl } : {}),\n\t\t\t});\n\t\t}\n\t} else if (initialSystemText) {\n\t\t// Add cache control to system prompt for non-OAuth tokens\n\t\tparams.system = [\n\t\t\t{\n\t\t\t\ttype: \"text\",\n\t\t\t\ttext: sanitizeSurrogates(initialSystemText),\n\t\t\t\t...(cacheControl ? { cache_control: cacheControl } : {}),\n\t\t\t},\n\t\t];\n\t}\n\n\t// Temperature is incompatible with extended thinking and unsupported on Claude Opus 4.7+.\n\tif (\n\t\toptions?.temperature !== undefined &&\n\t\t!options?.thinkingEnabled &&\n\t\tmodel.compat?.supportsMidConvoEffort !== true &&\n\t\tcompat.supportsTemperature\n\t) {\n\t\tparams.temperature = options.temperature;\n\t}\n\n\tconst toolCacheControl = compat.supportsCacheControlOnTools ? cacheControl : undefined;\n\tif (nativeToolChanges) {\n\t\t// Initial tools stay active with the cache breakpoint on the last one. Every later\n\t\t// declaration is deferred and only surfaced by its `tool_addition` block; removed\n\t\t// tools stay declared and are withdrawn by `tool_removal`. The request-level list\n\t\t// therefore only grows, keeping the cached prefix intact across tool changes.\n\t\tconst initialNames = new Set(initialTools.map((tool) => tool.name));\n\t\tconst laterTools = getDeclaredTools(context.messages).filter((tool) => !initialNames.has(tool.name));\n\t\tparams.tools = [\n\t\t\t...convertTools(\n\t\t\t\tinitialTools,\n\t\t\t\tisOAuthToken,\n\t\t\t\tcompat.supportsEagerToolInputStreaming,\n\t\t\t\tcompat.supportsStrictTools,\n\t\t\t\ttoolCacheControl,\n\t\t\t),\n\t\t\tDEFERRED_TOOL_PLACEHOLDER,\n\t\t\t...convertTools(\n\t\t\t\tlaterTools,\n\t\t\t\tisOAuthToken,\n\t\t\t\tcompat.supportsEagerToolInputStreaming,\n\t\t\t\tcompat.supportsStrictTools,\n\t\t\t).map((tool) => ({ ...tool, defer_loading: true })),\n\t\t];\n\t} else {\n\t\tconst tools = getCurrentTools(context.messages);\n\t\tif (tools.length > 0) {\n\t\t\tparams.tools = convertTools(\n\t\t\t\ttools,\n\t\t\t\tisOAuthToken,\n\t\t\t\tcompat.supportsEagerToolInputStreaming,\n\t\t\t\tcompat.supportsStrictTools,\n\t\t\t\ttoolCacheControl,\n\t\t\t);\n\t\t}\n\t}\n\n\t// Managed effort models always use adaptive thinking so prefix mismatches can\n\t// be dropped instead of surfacing as persistent 400 responses.\n\tif (model.compat?.supportsMidConvoEffort === true) {\n\t\tparams.thinking = {\n\t\t\ttype: \"adaptive\",\n\t\t\tdisplay: options?.thinkingDisplay ?? \"summarized\",\n\t\t\tblock_binding: { prefix_mismatch_behavior: \"drop_block\" },\n\t\t};\n\t\tparams.output_config = { effort: \"high\" };\n\t} else if (model.reasoning) {\n\t\tif (options?.thinkingEnabled) {\n\t\t\t// Default to \"summarized\" so Opus 4.7 and Mythos Preview behave like\n\t\t\t// older Claude 4 models (whose API default is also \"summarized\").\n\t\t\tconst display: AnthropicThinkingDisplay = options.thinkingDisplay ?? \"summarized\";\n\t\t\tif (model.compat?.forceAdaptiveThinking === true) {\n\t\t\t\t// Adaptive thinking: Claude decides when and how much to think.\n\t\t\t\tparams.thinking = { type: \"adaptive\", display };\n\t\t\t\tif (options.effort) {\n\t\t\t\t\tparams.output_config = { effort: options.effort };\n\t\t\t\t}\n\t\t\t} else {\n\t\t\t\t// Budget-based thinking for older models\n\t\t\t\tparams.thinking = {\n\t\t\t\t\ttype: \"enabled\",\n\t\t\t\t\tbudget_tokens: options.thinkingBudgetTokens || 1024,\n\t\t\t\t\tdisplay,\n\t\t\t\t};\n\t\t\t}\n\t\t} else if (options?.thinkingEnabled === false && model.thinkingLevelMap?.off !== null) {\n\t\t\tparams.thinking = { type: \"disabled\" };\n\t\t}\n\t}\n\n\tif (options?.metadata) {\n\t\tconst userId = options.metadata.user_id;\n\t\tif (typeof userId === \"string\") {\n\t\t\tparams.metadata = { user_id: userId };\n\t\t}\n\t}\n\n\tif (options?.toolChoice) {\n\t\tif (typeof options.toolChoice === \"string\") {\n\t\t\tparams.tool_choice = { type: options.toolChoice };\n\t\t} else {\n\t\t\tparams.tool_choice = options.toolChoice;\n\t\t}\n\t}\n\n\tconst allowedFallbackModels = model.compat?.allowedFallbackModels;\n\tif (allowedFallbackModels && allowedFallbackModels.length > 0) {\n\t\tparams.fallbacks = allowedFallbackModels.map((fallback) => ({ model: fallback.model }));\n\t}\n\n\treturn params;\n}\n\n// Normalize tool call IDs to match Anthropic's required pattern and length\nfunction normalizeToolCallId(id: string): string {\n\treturn id.replace(/[^a-zA-Z0-9_-]/g, \"_\").slice(0, 64);\n}\n\nfunction convertToolResult(msg: ToolResultMessage): ContentBlockParam {\n\treturn {\n\t\ttype: \"tool_result\",\n\t\ttool_use_id: msg.toolCallId,\n\t\tcontent: convertContentBlocks(msg.content),\n\t\tis_error: msg.isError,\n\t};\n}\n\ninterface ConvertedAnthropicMessages {\n\tmessages: MessageParam[];\n\tassistantLevels: Map<number, AnthropicEffort>;\n}\n\nfunction convertMessages(\n\ttransformedMessages: Message[],\n\tisOAuthToken: boolean,\n\tcacheControl?: CacheControlEphemeral,\n\tallowEmptySignature = false,\n\tmanagedProvider?: string,\n\tnativeToolChanges = false,\n): ConvertedAnthropicMessages {\n\tconst params: MessageParam[] = [];\n\tconst assistantLevels = new Map<number, AnthropicEffort>();\n\t// Later system messages are held back and emitted directly before the next assistant\n\t// message (or at the end of the transcript). Anthropic requires `tool_result` blocks to\n\t// immediately follow their `tool_use`, so a system message between them is rejected; this\n\t// also mirrors where the managed-effort system messages are inserted. As a result an\n\t// update placed before a user message in the transcript lands after it on the wire.\n\tconst pendingSystemMessages: MessageParam[] = [];\n\tconst flushPendingSystemMessages = (): void => {\n\t\tparams.push(...pendingSystemMessages);\n\t\tpendingSystemMessages.length = 0;\n\t};\n\n\tfor (let i = 0; i < transformedMessages.length; i++) {\n\t\tconst msg = transformedMessages[i];\n\n\t\tif (msg.role === \"system\") {\n\t\t\t// Later system messages only reach this point when the model accepts them natively;\n\t\t\t// otherwise the transcript was collapsed into the leading message before conversion.\n\t\t\tconst text = renderSystemMessageUpdate(msg);\n\t\t\tconst blocks: ContentBlockParam[] = [];\n\t\t\tif (text.length > 0) blocks.push({ type: \"text\", text: sanitizeSurrogates(text) });\n\t\t\tif (nativeToolChanges) {\n\t\t\t\tfor (const tool of msg.toolsRemoved ?? []) {\n\t\t\t\t\tblocks.push({\n\t\t\t\t\t\ttype: \"tool_removal\",\n\t\t\t\t\t\ttool: { type: \"tool_reference\", name: isOAuthToken ? toClaudeCodeName(tool.name) : tool.name },\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t\tfor (const tool of msg.toolsAdded ?? []) {\n\t\t\t\t\tblocks.push({\n\t\t\t\t\t\ttype: \"tool_addition\",\n\t\t\t\t\t\ttool: { type: \"tool_reference\", name: isOAuthToken ? toClaudeCodeName(tool.name) : tool.name },\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t}\n\t\t\tif (blocks.length > 0) pendingSystemMessages.push({ role: \"system\", content: blocks });\n\t\t} else if (msg.role === \"user\") {\n\t\t\tif (typeof msg.content === \"string\") {\n\t\t\t\tif (msg.content.trim().length > 0) {\n\t\t\t\t\tparams.push({\n\t\t\t\t\t\trole: \"user\",\n\t\t\t\t\t\tcontent: sanitizeSurrogates(msg.content),\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t} else {\n\t\t\t\tconst blocks: ContentBlockParam[] = msg.content.map((item) => {\n\t\t\t\t\tif (item.type === \"text\") {\n\t\t\t\t\t\treturn {\n\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\ttext: sanitizeSurrogates(item.text),\n\t\t\t\t\t\t};\n\t\t\t\t\t} else {\n\t\t\t\t\t\treturn {\n\t\t\t\t\t\t\ttype: \"image\",\n\t\t\t\t\t\t\tsource: {\n\t\t\t\t\t\t\t\ttype: \"base64\",\n\t\t\t\t\t\t\t\tmedia_type: item.mimeType as \"image/jpeg\" | \"image/png\" | \"image/gif\" | \"image/webp\",\n\t\t\t\t\t\t\t\tdata: item.data,\n\t\t\t\t\t\t\t},\n\t\t\t\t\t\t};\n\t\t\t\t\t}\n\t\t\t\t});\n\t\t\t\tconst filteredBlocks = blocks.filter((b) => {\n\t\t\t\t\tif (b.type === \"text\") {\n\t\t\t\t\t\treturn b.text.trim().length > 0;\n\t\t\t\t\t}\n\t\t\t\t\treturn true;\n\t\t\t\t});\n\t\t\t\tif (filteredBlocks.length === 0) continue;\n\t\t\t\tparams.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tcontent: filteredBlocks,\n\t\t\t\t});\n\t\t\t}\n\t\t} else if (msg.role === \"assistant\") {\n\t\t\tflushPendingSystemMessages();\n\t\t\tconst blocks: ContentBlockParam[] = [];\n\n\t\t\tfor (const block of msg.content) {\n\t\t\t\tif (block.type === \"text\") {\n\t\t\t\t\tif (block.text.trim().length === 0) continue;\n\t\t\t\t\tblocks.push({\n\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\ttext: sanitizeSurrogates(block.text),\n\t\t\t\t\t});\n\t\t\t\t} else if (block.type === \"thinking\") {\n\t\t\t\t\t// Redacted thinking: pass the opaque payload back as redacted_thinking\n\t\t\t\t\tif (block.redacted) {\n\t\t\t\t\t\tblocks.push({\n\t\t\t\t\t\t\ttype: \"redacted_thinking\",\n\t\t\t\t\t\t\tdata: block.thinkingSignature!,\n\t\t\t\t\t\t});\n\t\t\t\t\t\tcontinue;\n\t\t\t\t\t}\n\t\t\t\t\tconst thinkingSignature = block.thinkingSignature;\n\t\t\t\t\tconst hasThinkingSignature = !!thinkingSignature && thinkingSignature.trim().length > 0;\n\t\t\t\t\tif (block.thinking.trim().length === 0 && !hasThinkingSignature) continue;\n\t\t\t\t\t// If thinking signature is missing/empty (e.g., from aborted stream),\n\t\t\t\t\t// convert to plain text for Anthropic. Some compatible providers emit\n\t\t\t\t\t// and accept empty signatures, so let marked models preserve the block.\n\t\t\t\t\tif (!hasThinkingSignature) {\n\t\t\t\t\t\tblocks.push(\n\t\t\t\t\t\t\tallowEmptySignature\n\t\t\t\t\t\t\t\t? {\n\t\t\t\t\t\t\t\t\t\ttype: \"thinking\",\n\t\t\t\t\t\t\t\t\t\tthinking: sanitizeSurrogates(block.thinking),\n\t\t\t\t\t\t\t\t\t\tsignature: \"\",\n\t\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\t: {\n\t\t\t\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\t\t\t\ttext: sanitizeSurrogates(block.thinking),\n\t\t\t\t\t\t\t\t\t},\n\t\t\t\t\t\t);\n\t\t\t\t\t} else {\n\t\t\t\t\t\tblocks.push({\n\t\t\t\t\t\t\ttype: \"thinking\",\n\t\t\t\t\t\t\tthinking: sanitizeSurrogates(block.thinking),\n\t\t\t\t\t\t\tsignature: thinkingSignature,\n\t\t\t\t\t\t});\n\t\t\t\t\t}\n\t\t\t\t} else if (block.type === \"toolCall\") {\n\t\t\t\t\tblocks.push({\n\t\t\t\t\t\ttype: \"tool_use\",\n\t\t\t\t\t\tid: block.id,\n\t\t\t\t\t\tname: isOAuthToken ? toClaudeCodeName(block.name) : block.name,\n\t\t\t\t\t\tinput: block.arguments ?? {},\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t}\n\t\t\tif (blocks.length === 0) continue;\n\t\t\tconst messageIndex = params.length;\n\t\t\tparams.push({\n\t\t\t\trole: \"assistant\",\n\t\t\t\tcontent: blocks,\n\t\t\t});\n\t\t\tif (\n\t\t\t\tmanagedProvider !== undefined &&\n\t\t\t\tmsg.api === \"anthropic-messages\" &&\n\t\t\t\tmsg.provider === managedProvider &&\n\t\t\t\tisAnthropicEffort(msg.providerThinkingLevel)\n\t\t\t) {\n\t\t\t\tassistantLevels.set(messageIndex, msg.providerThinkingLevel);\n\t\t\t}\n\t\t} else if (msg.role === \"toolResult\") {\n\t\t\t// Collect all consecutive toolResult messages, needed for z.ai Anthropic endpoint.\n\t\t\tconst toolResults: ContentBlockParam[] = [];\n\t\t\tlet j = i;\n\t\t\twhile (j < transformedMessages.length && transformedMessages[j].role === \"toolResult\") {\n\t\t\t\ttoolResults.push(convertToolResult(transformedMessages[j] as ToolResultMessage));\n\t\t\t\tj++;\n\t\t\t}\n\n\t\t\t// Skip the messages we've already processed.\n\t\t\ti = j - 1;\n\n\t\t\tparams.push({\n\t\t\t\trole: \"user\",\n\t\t\t\tcontent: toolResults,\n\t\t\t});\n\t\t}\n\t}\n\n\tflushPendingSystemMessages();\n\n\t// Add cache_control to the last user or system message to cache conversation history\n\tif (cacheControl && params.length > 0) {\n\t\tconst lastMessage = params[params.length - 1];\n\t\tif (lastMessage.role === \"user\" || lastMessage.role === \"system\") {\n\t\t\tif (Array.isArray(lastMessage.content)) {\n\t\t\t\tconst lastBlock = lastMessage.content[lastMessage.content.length - 1];\n\t\t\t\tif (\n\t\t\t\t\tlastBlock &&\n\t\t\t\t\t(lastBlock.type === \"text\" ||\n\t\t\t\t\t\tlastBlock.type === \"image\" ||\n\t\t\t\t\t\tlastBlock.type === \"tool_result\" ||\n\t\t\t\t\t\tlastBlock.type === \"tool_addition\" ||\n\t\t\t\t\t\tlastBlock.type === \"tool_removal\")\n\t\t\t\t) {\n\t\t\t\t\t(lastBlock as any).cache_control = cacheControl;\n\t\t\t\t}\n\t\t\t} else if (typeof lastMessage.content === \"string\") {\n\t\t\t\tlastMessage.content = [\n\t\t\t\t\t{\n\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\ttext: lastMessage.content,\n\t\t\t\t\t\tcache_control: cacheControl,\n\t\t\t\t\t},\n\t\t\t\t] as any;\n\t\t\t}\n\t\t}\n\t}\n\n\treturn { messages: params, assistantLevels };\n}\n\nfunction isAnthropicEffort(value: unknown): value is AnthropicEffort {\n\treturn value === \"low\" || value === \"medium\" || value === \"high\" || value === \"xhigh\" || value === \"max\";\n}\n\nfunction insertThinkingLevelMessages(\n\tconverted: ConvertedAnthropicMessages,\n\tactiveEffort: AnthropicEffort,\n): MessageParam[] {\n\tconst messages: MessageParam[] = [];\n\tfor (let index = 0; index < converted.messages.length; index++) {\n\t\tconst historicalEffort = converted.assistantLevels.get(index);\n\t\tif (historicalEffort !== undefined) {\n\t\t\tmessages.push({ role: \"system\", content: [], output_config: { effort: historicalEffort } });\n\t\t}\n\t\tmessages.push(converted.messages[index]);\n\t}\n\tmessages.push({ role: \"system\", content: [], output_config: { effort: activeEffort } });\n\treturn messages;\n}\n\nfunction shouldUseFineGrainedToolStreamingBeta(\n\tmodel: Model<\"anthropic-messages\">,\n\tcontext: TranscriptContext,\n): boolean {\n\treturn getCurrentTools(context.messages).length > 0 && !getAnthropicCompat(model).supportsEagerToolInputStreaming;\n}\n\nfunction convertTools(\n\ttools: Tool[],\n\tisOAuthToken: boolean,\n\tsupportsEagerToolInputStreaming: boolean,\n\tsupportsStrictTools: boolean,\n\tcacheControl?: CacheControlEphemeral,\n): BetaTool[] {\n\tif (!tools) return [];\n\n\treturn tools.map((tool, index) => {\n\t\tconst strict = resolveJsonSchemaStrictSampling(tool, supportsStrictTools);\n\t\tconst parameters = getJsonSchemaToolParameters(tool, strict);\n\t\tconst schema = parameters as { properties?: unknown; required?: string[] };\n\t\tconst legacyInputSchema = {\n\t\t\ttype: \"object\" as const,\n\t\t\tproperties: schema.properties ?? {},\n\t\t\trequired: schema.required ?? [],\n\t\t};\n\t\tconst inputSchema =\n\t\t\tstrict === true\n\t\t\t\t? {\n\t\t\t\t\t\t...(parameters as Record<string, unknown>),\n\t\t\t\t\t\t...legacyInputSchema,\n\t\t\t\t\t}\n\t\t\t\t: legacyInputSchema;\n\n\t\treturn {\n\t\t\tname: isOAuthToken ? toClaudeCodeName(tool.name) : tool.name,\n\t\t\tdescription: tool.description,\n\t\t\t...(supportsEagerToolInputStreaming ? { eager_input_streaming: true } : {}),\n\t\t\t...(strict === true ? { strict: true } : {}),\n\t\t\tinput_schema: inputSchema,\n\t\t\t...(cacheControl && index === tools.length - 1 ? { cache_control: cacheControl } : {}),\n\t\t};\n\t});\n}\n\nfunction mapStopReason(\n\treason: BetaStopReason | string,\n\tstopDetails?: RefusalStopDetails | null,\n): { stopReason: StopReason; errorMessage?: string } {\n\tswitch (reason) {\n\t\tcase \"end_turn\":\n\t\t\treturn { stopReason: \"stop\" };\n\t\tcase \"max_tokens\":\n\t\t\treturn { stopReason: \"length\" };\n\t\tcase \"tool_use\":\n\t\t\treturn { stopReason: \"toolUse\" };\n\t\tcase \"refusal\":\n\t\t\treturn {\n\t\t\t\tstopReason: \"error\",\n\t\t\t\terrorMessage: stopDetails?.explanation || `The model refused to complete the request`,\n\t\t\t};\n\t\tcase \"pause_turn\": // Stop is good enough -> resubmit\n\t\t\treturn { stopReason: \"stop\" };\n\t\tcase \"stop_sequence\":\n\t\t\treturn { stopReason: \"stop\" }; // We don't supply stop sequences, so this should never happen\n\t\tcase \"sensitive\": // Content flagged by safety filters (not yet in SDK types)\n\t\t\treturn { stopReason: \"error\", errorMessage: \"Provider stopped with: sensitive\" };\n\t\tdefault:\n\t\t\t// Handle unknown stop reasons gracefully (API may add new values)\n\t\t\tthrow new Error(`Unhandled stop reason: ${reason}`);\n\t}\n}\n"]}
+{"version":3,"file":"anthropic-messages.d.ts","sourceRoot":"","sources":["../../src/api/anthropic-messages.ts"],"names":[],"mappings":"AAAA,OAAO,SAAS,MAAM,mBAAmB,CAAC;AAa1C,OAAO,KAAK,EASX,mBAAmB,EAEnB,cAAc,EACd,aAAa,EAMb,MAAM,aAAa,CAAC;AAiJrB,MAAM,MAAM,eAAe,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,GAAG,OAAO,GAAG,KAAK,CAAC;AAE1E,MAAM,MAAM,wBAAwB,GAAG,YAAY,GAAG,SAAS,CAAC;AA2ChE,MAAM,WAAW,gBAAiB,SAAQ,aAAa;IACtD;;;;;;OAMG;IACH,eAAe,CAAC,EAAE,OAAO,CAAC;IAC1B;;;;OAIG;IACH,oBAAoB,CAAC,EAAE,MAAM,CAAC;IAC9B;;;;;;;;;;;OAWG;IACH,MAAM,CAAC,EAAE,eAAe,CAAC;IACzB;;;;;;;;;;;OAWG;IACH,eAAe,CAAC,EAAE,wBAAwB,CAAC;IAC3C;;;;;OAKG;IACH,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAC9B;;;;OAIG;IACH,UAAU,CAAC,EAAE,MAAM,GAAG,KAAK,GAAG,MAAM,GAAG;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAA;KAAE,CAAC;IACtE;;;;OAIG;IACH,MAAM,CAAC,EAAE,SAAS,CAAC;CACnB;AAqOD,eAAO,MAAM,MAAM,EAAE,cAAc,CAAC,oBAAoB,EAAE,gBAAgB,CAwUzE,CAAC;AA2BF,eAAO,MAAM,YAAY,EAAE,cAAc,CAAC,oBAAoB,EAAE,mBAAmB,CA8ClF,CAAC"}

package/dist/api/anthropic-messages.js

@@ -399,6 +399,7 @@ export const stream = (model, context, options) => {
             stream.push({ type: "start", partial: output });
             const blocks = output.content;
             for await (const event of iterateAnthropicEvents(response, options?.signal)) {
+                await options?.onProviderStreamEvent?.(event, model);
                 if (event.type === "message_start") {
                     output.responseId = event.message.id;
                     const transformations = event.message.input_transformations;
@@ -587,6 +588,11 @@ export const stream = (model, context, options) => {
                         if (event.usage.cache_creation_input_tokens != null) {
                             output.usage.cacheWrite = event.usage.cache_creation_input_tokens;
                         }
+                        // Vercel AI Gateway includes the TTL breakdown in deltas, though the SDK only types it on message_start.
+                        const cacheCreation = event.usage.cache_creation;
+                        if (cacheCreation?.ephemeral_1h_input_tokens != null) {
+                            output.usage.cacheWrite1h = cacheCreation.ephemeral_1h_input_tokens;
+                        }
                         // Anthropic reports reasoning tokens as a subset of output tokens.
                         const thinkingTokens = event.usage.output_tokens_details?.thinking_tokens;
                         if (thinkingTokens != null) {