1import type { TelemetryContext } from "@earendil-works/pi-telemetry";

2import type { AnthropicOptions } from "./api/anthropic-messages.ts";

3import type { AzureOpenAIResponsesOptions } from "./api/azure-openai-responses.ts";

4import type { BedrockOptions } from "./api/bedrock-converse-stream.ts";

5import type { GoogleOptions } from "./api/google-generative-ai.ts";

6import type { GoogleVertexOptions } from "./api/google-vertex.ts";

7import type { MistralOptions } from "./api/mistral-conversations.ts";

8import type { OpenAICodexResponsesOptions } from "./api/openai-codex-responses.ts";

9import type { OpenAICompletionsOptions } from "./api/openai-completions.ts";

10import type { OpenAIResponsesOptions } from "./api/openai-responses.ts";

11import type { PiMessagesOptions } from "./api/pi-messages.ts";

12import type { AssistantMessageDiagnostic } from "./utils/diagnostics.ts";

13import type { AssistantMessageEventStream } from "./utils/event-stream.ts";

14

15export type { AssistantMessageEventStream } from "./utils/event-stream.ts";

16

17export type KnownApi =

18 | "openai-completions"

19 | "mistral-conversations"

20 | "openai-responses"

21 | "azure-openai-responses"

22 | "openai-codex-responses"

23 | "anthropic-messages"

24 | "bedrock-converse-stream"

25 | "google-generative-ai"

26 | "google-vertex"

27 | "pi-messages";

28

29export type Api = KnownApi | (string & {});

30

31export type KnownImageApi = "openrouter-images";

32

33export type ImageApi = KnownImageApi | (string & {});

34

35export type KnownClassifierApi = "typesafe-system-one" | "cloudflare-workers-ai-system-one" | "llama-cpp-classify";

36

37export type ClassifierApi = KnownClassifierApi | (string & {});

38

39export type KnownProvider =

40 | "amazon-bedrock"

41 | "ant-ling"

42 | "anthropic"

43 | "google"

44 | "google-vertex"

45 | "openai"

46 | "azure-openai-responses"

47 | "openai-codex"

48 | "radius"

49 | "typesafe"

50 | "nvidia"

51 | "deepseek"

52 | "github-copilot"

53 | "xai"

54 | "groq"

55 | "cerebras"

56 | "openrouter"

57 | "vercel-ai-gateway"

58 | "zai"

59 | "zai-coding-cn"

60 | "mistral"

61 | "minimax"

62 | "minimax-cn"

63 | "moonshotai"

64 | "moonshotai-cn"

65 | "huggingface"

66 | "fireworks"

67 | "together"

68 | "baseten"

69 | "opencode"

70 | "opencode-go"

71 | "kimi-coding"

72 | "meta"

73 | "cloudflare-workers-ai"

74 | "cloudflare-ai-gateway"

75 | "qwen-token-plan"

76 | "qwen-token-plan-cn"

77 | "qwen-token-plan-individual"

78 | "xiaomi"

79 | "xiaomi-token-plan-cn"

80 | "xiaomi-token-plan-ams"

81 | "xiaomi-token-plan-sgp";

82export type ProviderId = KnownProvider | string;

83

84export type ToolChoice = "auto" | "none";

85export type ThinkingLevel = "minimal" | "low" | "medium" | "high" | "xhigh" | "max";

86export type ModelThinkingLevel = "off" | ThinkingLevel;

87export type ThinkingLevelMap = Partial<Record<ModelThinkingLevel, string | null>>;

88export type ChatTemplateKwargValue =

89 | string

90 | number

91 | boolean

92 | null

93 | {

94 $var: "thinking.enabled" | "thinking.effort" | "thinking.budget";

95 omitWhenOff?: boolean;

96 };

97

98/** Top-level request field used to cap reasoning tokens on OpenAI-compatible servers. */

99export type ThinkingTokenBudgetField = "thinking_token_budget" | "thinking_budget" | "thinking_budget_tokens";

100

101/** Token budgets for each thinking level (token-based providers only) */

102export interface ThinkingBudgets {

103 minimal?: number;

104 low?: number;

105 medium?: number;

106 high?: number;

107}

108

109// Base options all providers share

110export type CacheRetention = "none" | "short" | "long";

111

112/**

113 * Best-effort prompt cache lifetime in seconds for each retention tier a request can ask for.

114 * A missing tier means the lifetime is unknown; pi does not warm such caches.

115 */

116export type ModelPromptCache = Partial<Record<Exclude<CacheRetention, "none">, number>>;

117

118export type Transport = "sse" | "websocket" | "websocket-cached" | "auto";

119

120/** Provider-scoped environment overrides. Values take precedence over process.env. */

121export type ProviderEnv = Record<string, string>;

122export type ProviderHeaders = Record<string, string | null>;

123export type FetchFunction = typeof globalThis.fetch;

124export type SessionAffinityFormat = "openai" | "openai-nosession" | "openrouter";

125

126export interface ProviderResponse {

127 status: number;

128 headers: Record<string, string>;

129}

130

131/** Authentication, HTTP transport, and lifecycle callbacks shared by provider requests. */

132export interface ProviderRequestOptions<TModel = Model<Api>> {

133 signal?: AbortSignal;

134 /** Explicit parent context for telemetry produced by this logical request. */

135 telemetryContext?: TelemetryContext;

136 apiKey?: string;

137 /**

138 * Optional fetch implementation for provider HTTP requests.

139 * Defaults to `globalThis.fetch`. Provider adapters that cannot inject a custom implementation may reject it.

140 * This does not affect WebSocket transports.

141 */

142 fetch?: FetchFunction;

143 /**

144 * Provider-scoped environment values. These take precedence over process.env for

145 * provider configuration such as regional settings, endpoint placeholders, and

146 * proxy variables.

147 */

148 env?: ProviderEnv;

149 /**

150 * Optional callback for inspecting or replacing provider payloads before sending.

151 * Return undefined to keep the payload unchanged.

152 */

153 onPayload?: (payload: unknown, model: TModel) => unknown | undefined | Promise<unknown | undefined>;

154 /**

155 * Optional callback invoked after an HTTP response is received.

156 */

157 onResponse?: (response: ProviderResponse, model: TModel) => void | Promise<void>;

158 /**

159 * Optional custom HTTP headers to include in API requests.

160 * Merged with provider defaults; caller values override default headers.

161 * On AWS Bedrock these are injected via a Smithy `build`-step middleware so

162 * they are covered by SigV4 signing; reserved headers (`x-amz-*`,

163 * `authorization`, `host`) are silently ignored to preserve SigV4 / bearer auth.

164 * A null value suppresses a provider/API default header with the same name.

165 */

166 headers?: ProviderHeaders;

167 /**

168 * HTTP request timeout in milliseconds for providers/SDKs that support it.

169 * For example, OpenAI and Anthropic SDK clients default to 10 minutes.

170 */

171 timeoutMs?: number;

172 /**

173 * Maximum retry attempts for providers/SDKs that support client-side retries.

174 * For example, OpenAI and Anthropic SDK clients default to 2.

175 */

176 maxRetries?: number;

177 /**

178 * Maximum delay in milliseconds to wait for a retry when the server requests a long wait.

179 * If the server's requested delay exceeds this value, the request fails immediately

180 * with an error containing the requested delay, allowing higher-level retry logic

181 * to handle it with user visibility.

182 * Default: 60000 (60 seconds). Set to 0 to disable the cap.

183 */

184 maxRetryDelayMs?: number;

185}

186

187export interface StreamOptions extends ProviderRequestOptions<Model<Api>> {

188 /**

189 * Optional callback invoked after an HTTP response is received and before

190 * its body stream is consumed.

191 */

192 onResponse?: (response: ProviderResponse, model: Model<Api>) => void | Promise<void>;

193 /**

194 * Optional observer for each parsed provider stream event before Pi normalization.

195 * Event data is adapter-owned and must be treated as read-only.

196 * Adapter support is explicit; unsupported adapters do not invoke it.

197 */

198 onProviderStreamEvent?: (data: unknown, model: Model<Api>) => void | Promise<void>;

199 temperature?: number;

200 /**

201 * Arbitrary sampling parameters merged into the request body as-is, after the named request

202 * fields, so keys here override them. Lets custom OpenAI-compatible servers (llama.cpp, vLLM,

203 * SGLang, ...) receive parameters pi does not model, e.g. `top_p`, `top_k`, `min_p`,

204 * `repetition_penalty`. Merged over `Model.samplingParams` per key. Only applied by

205 * OpenAI-compatible adapters (completions, responses, Azure responses); other APIs ignore it.

206 */

207 samplingParams?: Record<string, unknown>;

208 maxTokens?: number;

209 /**

210 * Preferred transport for providers that support multiple transports.

211 * Providers that do not support this option ignore it.

212 */

213 transport?: Transport;

214 /**

215 * Prompt cache retention preference. Providers map this to their supported values.

216 * Default: "short".

217 */

218 cacheRetention?: CacheRetention;

219 /**

220 * Optional session identifier for providers that support session-based caching.

221 * Providers can use this to enable prompt caching, request routing, or other

222 * session-aware features. Ignored by providers that don't support it.

223 */

224 sessionId?: string;

225 /**

226 * WebSocket connect timeout in milliseconds for providers that support

227 * WebSocket transports. This covers the connection/open handshake only;

228 * stream idleness after connection uses timeoutMs.

229 */

230 websocketConnectTimeoutMs?: number;

231 /**

232 * Optional metadata to include in API requests.

233 * Providers extract the fields they understand and ignore the rest.

234 * For example, Anthropic uses `user_id` for abuse tracking and rate limiting.

235 */

236 metadata?: Record<string, unknown>;

237}

238

239export type ProviderStreamOptions = StreamOptions & Record<string, unknown>;

240

241export interface DeferredFetchOptions extends ProviderRequestOptions<Model<Api>> {

242 /**

243 * Maximum provider long-poll duration in milliseconds.

244 * Defaults to 0, which performs one status check.

245 */

246 wait?: number;

247}

248

249/** Request options for best-effort deferred-response cancellation. */

250export type DeferredCancelOptions = ProviderRequestOptions<Model<Api>>;

251

252/**

253 * Maps known APIs to their full provider-specific stream option types.

254 * Type-only imports from API implementation modules are erased at emit, so

255 * this is tree-shake safe.

256 */

257export interface ApiOptionsMap {

258 "anthropic-messages": AnthropicOptions;

259 "openai-completions": OpenAICompletionsOptions;

260 "openai-responses": OpenAIResponsesOptions;

261 "openai-codex-responses": OpenAICodexResponsesOptions;

262 "azure-openai-responses": AzureOpenAIResponsesOptions;

263 "google-generative-ai": GoogleOptions;

264 "google-vertex": GoogleVertexOptions;

265 "mistral-conversations": MistralOptions;

266 "bedrock-converse-stream": BedrockOptions;

267 "pi-messages": PiMessagesOptions;

268}

269

270/**

271 * Full stream options for an API. Known APIs resolve to their concrete option

272 * type; custom API strings fall back to the generic shape.

273 */

274export type ApiStreamOptions<TApi extends Api> = TApi extends keyof ApiOptionsMap

275 ? ApiOptionsMap[TApi]

276 : StreamOptions & Record<string, unknown>;

277

278/**

279 * The uniform stream contract of an API implementation module: every module

280 * under `src/api/` exports `stream` and `streamSimple`; capable modules may also

281 * export deferred-response methods. Lazy wrappers (`lazyApi()`) and provider

282 * factories pass these around as values. This is the untyped dispatch shape;

283 * per-API option typing lives on the implementation modules themselves and on

284 * `Provider.stream()` via `ApiStreamOptions`.

285 */

286export interface ProviderStreams {

287 stream(model: Model<Api>, context: TranscriptContext, options?: StreamOptions): AssistantMessageEventStream;

288 streamSimple(

289 model: Model<Api>,

290 context: TranscriptContext,

291 options?: SimpleStreamOptions,

292 ): AssistantMessageEventStream;

293 fetchDeferred?(

294 model: Model<Api>,

295 handle: DeferredHandle,

296 options?: DeferredFetchOptions,

297 ): AssistantMessageEventStream;

298 cancelDeferred?(model: Model<Api>, handle: DeferredHandle, options?: DeferredCancelOptions): Promise<void>;

299}

300

301/**

302 * The uniform contract of an image-generation API implementation module:

303 * every image API module under `src/api/` exports exactly `generateImages`,

304 * so the module itself satisfies this interface. Lazy wrappers and

305 * `createProvider({ images })` pass these around as values.

306 */

307export interface ProviderImages {

308 generateImages(

309 model: ImageModel<ImageApi>,

310 context: ImagesContext,

311 options?: ImagesOptions,

312 ): Promise<AssistantImages>;

313}

314

315/** The uniform contract implemented by classifier API modules. */

316export interface ProviderClassifier {

317 classify(

318 model: ClassifierModel<ClassifierApi>,

319 context: ClassifierContext,

320 options?: ClassifierOptions,

321 ): Promise<ClassifierResult>;

322}

323

324export interface ClassifierOptions extends ProviderRequestOptions<ClassifierModel<ClassifierApi>> {

325 /**

326 * Divides the answer logits by this value before they are normalized into probabilities.

327 * Values above 1 soften the distribution; values below 1 sharpen it. Must be positive.

328 * APIs that cannot apply it ignore it.

329 */

330 temperature?: number;

331}

332

333export interface ImagesOptions extends ProviderRequestOptions<ImageModel<ImageApi>> {

334 /**

335 * Optional metadata to include in API requests.

336 * Providers extract the fields they understand and ignore the rest.

337 */

338 metadata?: Record<string, unknown>;

339}

340

341export type ProviderImagesOptions = ImagesOptions & Record<string, unknown>;

342

343export interface AnthropicAllowedFallbackModel {

344 provider: ProviderId;

345 model: string;

346 cost: ModelCost;

347}

348

349// Unified options with reasoning passed to streamSimple() and completeSimple()

350export interface SimpleStreamOptions extends StreamOptions {

351 /** Provider-neutral tool selection for simple requests. When omitted, adapters use provider-specific behavior. */

352 toolChoice?: ToolChoice;

353 reasoning?: ThinkingLevel;

354 /** Ask a capable provider to return a durable handle and continue the request asynchronously. */

355 deferred?: boolean | { window?: "15m" | "1h" | "24h" };

356 /** Custom token budgets for thinking levels (token-based providers only) */

357 thinkingBudgets?: ThinkingBudgets;

358}

359

360// Generic StreamFunction with typed options.

361//

362// Contract:

363// - Receives a normalized transcript: the system prompt and tools live in the

364// leading system message, never on the context itself.

365// - Must return an AssistantMessageEventStream.

366// - Direct streamSimple() calls may throw synchronously when request auth is

367// missing. Once a stream is returned, request/model/runtime failures should

368// be encoded in that stream.

369// - Error termination must produce an AssistantMessage with stopReason

370// "error" or "aborted" and errorMessage, emitted via the stream protocol.

371export type StreamFunction<TApi extends Api = Api, TOptions extends StreamOptions = StreamOptions> = (

372 model: Model<TApi>,

373 context: TranscriptContext,

374 options?: TOptions,

375) => AssistantMessageEventStream;

376

377export type ImagesFunction<TOptions extends ImagesOptions = ImagesOptions> = (

378 model: ImageModel<ImageApi>,

379 context: ImagesContext,

380 options?: TOptions,

381) => Promise<AssistantImages>;

382

383export type ClassifierFunction<TOptions extends ClassifierOptions = ClassifierOptions> = (

384 model: ClassifierModel<ClassifierApi>,

385 context: ClassifierContext,

386 options?: TOptions,

387) => Promise<ClassifierResult>;

388

389export interface TextSignatureV1 {

390 v: 1;

391 id: string;

392 phase?: "commentary" | "final_answer";

393}

394

395export interface TextContent {

396 type: "text";

397 text: string;

398 textSignature?: string; // e.g., for OpenAI responses, message metadata (legacy id string or TextSignatureV1 JSON)

399}

400

401export interface ThinkingContent {

402 type: "thinking";

403 thinking: string;

404 thinkingSignature?: string; // Provider-specific opaque or serialized reasoning replay data

405 /** When true, the thinking content was redacted by safety filters. The opaque

406 * encrypted payload is stored in `thinkingSignature` so it can be passed back

407 * to the API for multi-turn continuity. */

408 redacted?: boolean;

409}

410

411export interface ImageContent {

412 type: "image";

413 data: string; // base64 encoded image data

414 mimeType: string; // e.g., "image/jpeg", "image/png"

415}

416

417export interface ToolCall {

418 type: "toolCall";

419 id: string;

420 name: string;

421 arguments: JsonObject;

422 thoughtSignature?: string; // Google-specific: opaque signature for reusing thought context

423 /** OpenAI Responses namespace for calls to dynamically loaded or namespaced tools. */

424 namespace?: string;

425}

426

427export interface Usage {

428 input: number;

429 output: number;

430 cacheRead: number;

431 cacheWrite: number;

432 /** Subset of `cacheWrite` written with 1h retention. Only Anthropic reports this split. */

433 cacheWrite1h?: number;

434 /**

435 * Reasoning/thinking tokens, when the provider reports them. This is a subset of

436 * `output`: `output` already includes these tokens. Set to a number (possibly 0) by

437 * providers that expose a reasoning breakdown; left undefined by providers that don't.

438 */

439 reasoning?: number;

440 totalTokens: number;

441 cost: {

442 input: number;

443 output: number;

444 cacheRead: number;

445 cacheWrite: number;

446 total: number;

447 };

448}

449

450export type StopReason = "pending" | "stop" | "length" | "toolUse" | "error" | "aborted" | "deferred";

451

452export type JsonValue = null | boolean | number | string | readonly JsonValue[] | JsonObject;

453export type JsonObject = { [key: string]: JsonValue };

454

455type IsAny<T> = 0 extends 1 & T ? true : false;

456type IsExactlyJsonValue<T> = [T] extends [JsonValue] ? ([JsonValue] extends [T] ? true : false) : false;

457type IsJsonProperty<T> = IsAny<T> extends true

458 ? false

459 : unknown extends T

460 ? false

461 : [Exclude<T, undefined>] extends [never]

462 ? true

463 : IsJsonCompatible<Exclude<T, undefined>>;

464type InvalidJsonKeys<T extends object> = {

465 [TKey in keyof T]-?: TKey extends string | number ? (IsJsonProperty<T[TKey]> extends true ? never : TKey) : TKey;

466}[keyof T];

467type IsJsonCompatible<T> = IsAny<T> extends true

468 ? false

469 : unknown extends T

470 ? false

471 : IsExactlyJsonValue<T> extends true

472 ? true

473 : T extends null | boolean | number | string

474 ? true

475 : T extends undefined

476 ? false

477 : T extends readonly (infer TItem)[]

478 ? IsJsonCompatible<TItem>

479 : T extends (...args: never[]) => unknown

480 ? false

481 : T extends object

482 ? [InvalidJsonKeys<T>] extends [never]

483 ? true

484 : false

485 : false;

486

487/** The JSON representation of a typed in-memory value. Optional object properties remain optional. */

488export type JsonRepresentation<T> = IsAny<T> extends true

489 ? JsonValue

490 : unknown extends T

491 ? JsonValue

492 : [T] extends [JsonValue]

493 ? T

494 : T extends readonly unknown[]

495 ? { [TKey in keyof T]: JsonRepresentation<Exclude<T[TKey], undefined>> }

496 : T extends object

497 ? { [TKey in keyof T]: JsonRepresentation<Exclude<T[TKey], undefined>> }

498 : never;

499

500export interface DeferredHandle {

501 provider: string;

502 modelId: string;

503 api: string;

504 /** Provider token, such as a response id or batch id plus row id. */

505 id: string;

506 expiresAt?: number;

507 pollAfterMs?: number;

508 /** Provider conversion data required to reconstruct the final assistant message. */

509 data?: JsonValue;

510}

511

512/**

513 * System instructions and tool declarations at one point in the transcript.

514 *

515 * The leading system message is the system prompt. Later system messages change it:

516 * `content` adds instructions from that point on, `sections` replace or remove named

517 * prompt sections, and `toolsAdded`/`toolsRemoved` change the tool set. Replaying

518 * every system message in order yields the current prompt and tools. Providers that

519 * accept system messages mid-conversation send each one in place; other providers

520 * rebuild the leading system message from the replayed state.

521 */

522export interface SystemMessage {

523 role: "system";

524 /** Instruction text. On the leading message this is the base prompt; later, additional instructions. */

525 content: string | TextContent[];

526 /**

527 * Named, ordered prompt sections rendered verbatim after `content`. The leading message

528 * declares them; later messages replace sections by name, and `null` removes one. Keep

529 * each section self-delimiting (a tag, a heading) so the model can relate an update to

530 * the original. Avoid integer-like names; JSON objects reorder those.

531 */

532 sections?: Record<string, string | null>;

533 /** Complete definitions of tools that become available at this point. */

534 toolsAdded?: Tool[];

535 /** Tools that stop being available at this point. */

536 toolsRemoved?: ToolReference[];

537 timestamp: number; // Unix timestamp in milliseconds

538}

539

540export interface UserMessage {

541 role: "user";

542 content: string | (TextContent | ImageContent)[];

543 timestamp: number; // Unix timestamp in milliseconds

544}

545

546export interface AssistantMessage {

547 role: "assistant";

548 content: (TextContent | ThinkingContent | ToolCall)[];

549 api: Api;

550 provider: ProviderId;

551 model: string;

552 responseModel?: string; // Concrete model reported by the provider when different from the requested `model`

553 responseId?: string; // Provider-specific response/message identifier when the upstream API exposes one

554 /** Exact provider-native effort level used for this response. Absent for legacy or unmanaged responses. */

555 providerThinkingLevel?: string;

556 /** Pi thinking level the agent loop requested for this response. Absent outside the agent loop and for legacy responses. */

557 thinkingLevel?: ModelThinkingLevel;

558 diagnostics?: AssistantMessageDiagnostic[]; // Redacted provider/runtime diagnostics for failures and recoveries.

559 usage: Usage;

560 stopReason: StopReason;

561 deferred?: DeferredHandle;

562 errorMessage?: string;

563 rawStopReason?: string;

564 /**

565 * Provider indication of whether the model explicitly ended its turn.

566 * Preserved for debugging and does not currently affect agent control flow.

567 */

568 endTurn?: boolean;

569 timestamp: number; // Unix timestamp in milliseconds

570}

571

572/** A tool call that another tool made while it ran, for example from a codemode script. */

573export interface NestedToolCallRecord {

574 id: string;

575 name: string;

576 /** Omitted when over the size limits; `argumentsBytes` then gives their size. */

577 arguments?: JsonObject;

578 /** UTF-8 size of the arguments as JSON, set when `arguments` is omitted. */

579 argumentsBytes?: number;

580 /** `unfinished`: the call was still running when the calling tool finished. */

581 status: "ok" | "error" | "unfinished";

582 durationMs?: number;

583 /** Error text, truncated. */

584 error?: string;

585}

586

587/** Bounded record of the nested calls a tool made. Results are not recorded. */

588export interface NestedToolCalls {

589 calls: NestedToolCallRecord[];

590 /** False when calls were dropped, arguments omitted, or calls had not finished. */

591 complete: boolean;

592}

593

594export type ToolResultMessage<TDetails = JsonValue> = IsJsonCompatible<TDetails> extends true

595 ? {

596 role: "toolResult";

597 toolCallId: string;

598 toolName: string;

599 content: (TextContent | ImageContent)[]; // Supports text and images

600 details?: JsonRepresentation<TDetails>;

601 /** Usage from the tool execution itself, if available. Not part of main LLM context accounting. */

602 usage?: Usage;

603 /** Calls this tool made to other tools. Kept for the session record; not sent to the model. */

604 nestedCalls?: NestedToolCalls;

605 isError: boolean;

606 timestamp: number; // Unix timestamp in milliseconds

607 }

608 : never;

609

610export type Message = SystemMessage | UserMessage | AssistantMessage | ToolResultMessage;

611

612export type ImagesInputContent = TextContent | ImageContent;

613export type ImagesOutputContent = TextContent | ImageContent;

614

615export interface ImagesContext {

616 input: ImagesInputContent[];

617}

618

619export type ImagesStopReason = "stop" | "error" | "aborted";

620

621export interface AssistantImages {

622 api: ImageApi;

623 provider: ProviderId;

624 model: string;

625 output: ImagesOutputContent[];

626 responseId?: string;

627 usage?: Usage;

628 stopReason: ImagesStopReason;

629 errorMessage?: string;

630 timestamp: number; // Unix timestamp in milliseconds

631}

632

633export interface ClassifierChoiceQuestion {

634 type: "choice";

635 instructions: string;

636 criteria: Record<string, string>;

637}

638

639export interface ClassifierScoreQuestion {

640 type: "score";

641 instructions: string;

642 criteria: string[];

643}

644

645export interface ClassifierBoolQuestion {

646 type: "bool";

647 instructions: string;

648 criteria: { true: string; false: string };

649}

650

651export type ClassifierQuestion = ClassifierChoiceQuestion | ClassifierScoreQuestion | ClassifierBoolQuestion;

652

653export interface ClassifierContext {

654 state: JsonObject;

655 questions: Record<string, ClassifierQuestion>;

656}

657

658export interface ClassifierChoiceAnswer {

659 type: "choice";

660 choice: string;

661 probabilities: Record<string, number>;

662 confidence: number;

663}

664

665export interface ClassifierScoreAnswer {

666 type: "score";

667 score: number;

668 confidence: number;

669}

670

671export interface ClassifierBoolAnswer {

672 type: "bool";

673 probability: number;

674}

675

676export type ClassifierAnswer = ClassifierChoiceAnswer | ClassifierScoreAnswer | ClassifierBoolAnswer;

677export type ClassifierStopReason = "stop" | "error" | "aborted";

678

679export interface ClassifierResult {

680 api: ClassifierApi;

681 provider: ProviderId;

682 model: string;

683 answers: Record<string, ClassifierAnswer>;

684 /** Token usage and its cost at the model's catalog price, when the service reports token counts. */

685 usage?: Usage;

686 stopReason: ClassifierStopReason;

687 errorMessage?: string;

688 timestamp: number; // Unix timestamp in milliseconds

689}

690

691import type { TSchema } from "typebox";

692

693/** OpenAI grammar variants for constrained sampling. */

694export type GrammarFormat = "openai_lark" | "openai_regex";

695

696export type GrammarVariants = Partial<Record<GrammarFormat, string>>;

697

698/**

699 * Optional provider-side constrained sampling configs for a tool.

700 *

701 * The `json_schema` value roughly maps to the concept of `strict` in APIs which is

702 * implemented as json-schema constrained sampling by APIs. Grammar variants let

703 * callers provide provider-specific encodings of the same intended language.

704 */

705export type ConstrainedSamplingConfig =

706 | {

707 type: "json_schema";

708 strict: "prefer" | "require";

709 }

710 | {

711 type: "grammar";

712 variants: GrammarVariants;

713 };

714

715export interface Tool<TParameters extends TSchema = TSchema> {

716 name: string;

717 description: string;

718 parameters: TParameters;

719 constrainedSampling?: false | ConstrainedSamplingConfig;

720}

721

722export interface ToolReference {

723 name: string;

724}

725

726/**

727 * Request input accepted by the public stream entry points (`Models.stream()`,

728 * `streamSimple()`, ...). `systemPrompt` and `tools` are shorthand for a leading

729 * system message; `normalizeContext()` folds them into one before the request

730 * reaches a provider.

731 */

732export interface Context {

733 systemPrompt?: string;

734 messages: Message[];

735 tools?: Tool[];

736}

737

738declare const transcriptContextBrand: unique symbol;

739

740/**

741 * Normalized request context passed to providers and API implementations. The

742 * prompt and tool declarations are carried by the transcript's system messages.

743 * Only `normalizeContext()` produces this type, so a raw `Context` cannot reach

744 * provider code by accident.

745 */

746export type TranscriptContext = {

747 messages: Message[];

748 readonly [transcriptContextBrand]: true;

749};

750

751/**

752 * Event protocol for AssistantMessageEventStream.

753 *

754 * Successful streams emit `start` before partial updates and terminate with

755 * `done`. A stream may terminate directly with `error` when request setup fails

756 * before generation starts; after `start`, failures also terminate with `error`.

757 * Direct `streamSimple()` calls throw synchronously when request auth is missing.

758 * Updates and `done` must never appear before `start`.

759 *

760 * `partial` is the shared live response-so-far helper, not an event-time

761 * snapshot. Text and thinking blocks are empty when their `*_start` event is

762 * emitted and grow only through their corresponding `*_delta` events until the

763 * authoritative `*_end`. Redacted thinking may be complete at start and emit no

764 * deltas. Tool-call arguments at `toolcall_start` are provider-specific;

765 * `toolcall_delta` carries subsequent JSON updates.

766 */

767export type AssistantMessageEvent =

768 | { type: "start"; partial: AssistantMessage }

769 | { type: "text_start"; contentIndex: number; partial: AssistantMessage }

770 | { type: "text_delta"; contentIndex: number; delta: string; partial: AssistantMessage }

771 | { type: "text_end"; contentIndex: number; content: string; partial: AssistantMessage }

772 | { type: "thinking_start"; contentIndex: number; partial: AssistantMessage }

773 | { type: "thinking_delta"; contentIndex: number; delta: string; partial: AssistantMessage }

774 | { type: "thinking_end"; contentIndex: number; content: string; partial: AssistantMessage }

775 | { type: "toolcall_start"; contentIndex: number; partial: AssistantMessage }

776 | { type: "toolcall_delta"; contentIndex: number; delta: string; partial: AssistantMessage }

777 | { type: "toolcall_end"; contentIndex: number; toolCall: ToolCall; partial: AssistantMessage }

778 | {

779 type: "done";

780 reason: Extract<StopReason, "stop" | "length" | "toolUse" | "deferred">;

781 message: AssistantMessage;

782 }

783 | { type: "error"; reason: Extract<StopReason, "aborted" | "error">; error: AssistantMessage };

784

785/**

786 * Compatibility settings for OpenAI-compatible completions APIs.

787 * Use this to override URL-based auto-detection for custom providers.

788 */

789export interface OpenAICompletionsCompat {

790 /** Whether the provider supports the `store` field. Default: auto-detected from URL. */

791 supportsStore?: boolean;

792 /** Whether the provider supports the `developer` role (vs `system`). Default: auto-detected from URL. */

793 supportsDeveloperRole?: boolean;

794 /** Whether the provider supports `reasoning_effort`. Default: auto-detected from URL. */

795 supportsReasoningEffort?: boolean;

796 /** Whether the provider supports `stream_options: { include_usage: true }` for token usage in streaming responses. Default: true. */

797 supportsUsageInStreaming?: boolean;

798 /** Whether streamed responses include `finish_reason`. When false, pi infers `stop` or `toolUse` when the stream ends. Default: true. */

799 supportsFinishReason?: boolean;

800 /** Which field to use for max tokens. Default: auto-detected from URL. */

801 maxTokensField?: "max_completion_tokens" | "max_tokens";

802 /** Whether tool results require the `name` field. Default: auto-detected from URL. */

803 requiresToolResultName?: boolean;

804 /** Whether a user message after tool results requires an assistant message in between. Default: auto-detected from URL. */

805 requiresAssistantAfterToolResult?: boolean;

806 /** Whether thinking blocks must be converted to text blocks with <thinking> delimiters. Default: auto-detected from URL. */

807 requiresThinkingAsText?: boolean;

808 /** Whether all replayed assistant messages must include an empty reasoning_content field when reasoning is enabled. Default: auto-detected from URL. */

809 requiresReasoningContentOnAssistantMessages?: boolean;

810 /** Format for reasoning/thinking parameter. "openai" uses reasoning_effort, "openrouter" uses reasoning: { effort }, "deepseek" uses thinking: { type } plus reasoning_effort when supported, "together" uses reasoning: { enabled } plus reasoning_effort when supported, "baseten" uses configurable chat_template_args plus reasoning_effort when supported, "zai" uses thinking: { type }, "qwen" uses top-level enable_thinking: boolean, "qwen-chat-template" uses chat_template_kwargs.enable_thinking and preserve_thinking, "chat-template" uses configurable chat_template_kwargs, "string-thinking" uses top-level thinking: string, and "ant-ling" uses reasoning: { effort } only when the mapped effort is non-null. Default: "openai". */

811 thinkingFormat?:

812 | "openai"

813 | "openrouter"

814 | "deepseek"

815 | "together"

816 | "baseten"

817 | "zai"

818 | "qwen"

819 | "chat-template"

820 | "qwen-chat-template"

821 | "string-thinking"

822 | "ant-ling";

823 /** Kwargs to send as `chat_template_kwargs` when `thinkingFormat` is `chat-template`. Use `{ "$var": "thinking.enabled" }`, `{ "$var": "thinking.effort" }`, or `{ "$var": "thinking.budget" }` for pi-controlled thinking values. */

824 chatTemplateKwargs?: Record<string, ChatTemplateKwargValue>;

825 /** Arguments to send as `chat_template_args` when `thinkingFormat` is `baseten`. Use `{ "$var": "thinking.enabled" }`, `{ "$var": "thinking.effort" }`, or `{ "$var": "thinking.budget" }` for pi-controlled thinking values. */

826 chatTemplateArgs?: Record<string, ChatTemplateKwargValue>;

827 /** OpenRouter-compatible routing preferences sent as the `provider` request field. */

828 openRouterRouting?: OpenRouterRouting;

829 /** Vercel AI Gateway routing preferences. Only used when baseUrl points to Vercel AI Gateway. */

830 vercelGatewayRouting?: VercelGatewayRouting;

831 /** Whether z.ai supports top-level `tool_stream: true` for streaming tool call deltas. Default: false. */

832 zaiToolStream?: boolean;

833 /**

834 * Top-level request field used to cap reasoning tokens from `thinkingBudgets`.

835 * Reasoning and the answer share `max_tokens` on these endpoints, so without a budget a

836 * reasoning-heavy turn can consume the whole response and emit no answer.

837 * `"thinking_token_budget"` is vLLM, `"thinking_budget"` is Qwen/DashScope/SGLang,

838 * `"thinking_budget_tokens"` is llama.cpp. Off by default; not set on the generated catalog.

839 */

840 thinkingTokenBudgetField?: ThinkingTokenBudgetField;

841 /** Alias for `thinkingTokenBudgetField: "thinking_token_budget"` (vLLM). Prefer `thinkingTokenBudgetField`. Default: false. */

842 supportsThinkingTokenBudget?: boolean;

843 /** Whether the provider supports OpenAI custom tools with Lark/regex grammar formats. When false, grammar-constrained tools fall back to normal function tools. Default: false; the generated model catalog enables it for capable models. */

844 supportsOpenAIGrammarTools?: boolean;

845 /** Whether the exact model accepts system or developer messages after the conversation has started. When false, later system messages are folded into the leading system message. Default: false; the generated model catalog enables it for verified models. */

846 supportsMidConvoSystemMessages?: boolean;

847 /** Whether system messages can introduce additional tools mid-conversation. Requires `supportsMidConvoSystemMessages`. Default: false; the generated model catalog enables it for capable models. */

848 supportsMidConvoToolAdditions?: boolean;

849 /** Whether the provider supports the `strict` field in tool definitions. Default: false; generated capable models enable it explicitly. */

850 supportsStrictMode?: boolean;

851 /** Cache control convention for prompt caching. "anthropic" applies Anthropic-style `cache_control` markers to the system prompt, last tool definition, and last user, assistant, or tool-result text content. */

852 cacheControlFormat?: "anthropic";

853 /** Whether to send session-affinity data from `options.sessionId`. Default: true for OpenRouter endpoints, false otherwise. */

854 sendSessionAffinityHeaders?: boolean;

855 /** Session-affinity header format: `openai` sends `session_id`, `x-client-request-id`, and `x-session-affinity`; `openai-nosession` sends `x-client-request-id` and `x-session-affinity`; `openrouter` sends `x-session-id`. Does not affect the `prompt_cache_key` body param, which is governed by cache retention. Default: auto-detected. */

856 sessionAffinityFormat?: SessionAffinityFormat;

857 /** Whether the provider supports long prompt cache retention (`prompt_cache_retention: "24h"` or Anthropic-style `cache_control.ttl: "1h"`, depending on format). Default: true. */

858 supportsLongCacheRetention?: boolean;

859 /**

860 * vLLM scheduler priority sent as the top-level `priority` request field (lower values are

861 * handled earlier; server default 0). Only meaningful when vLLM runs with

862 * `--scheduling-policy priority`; useful for keeping background/batch work from stalling

863 * interactive sessions. Off by default; not set on the generated catalog.

864 */

865 vllmPriority?: number;

866}

867

868/** Compatibility settings for OpenAI Responses APIs. */

869export interface OpenAIResponsesCompat {

870 /** Whether the provider supports the `developer` role (vs `system`). Default: true. */

871 supportsDeveloperRole?: boolean;

872 /** Whether the exact model accepts developer or system messages after the conversation has started. When false, later system messages are folded into the leading system message. Default: false; the generated model catalog enables it for verified models. */

873 supportsMidConvoSystemMessages?: boolean;

874 /** Session-affinity header format: `openai` sends `session_id` and `x-client-request-id`; `openai-nosession` sends `x-client-request-id`; `openrouter` sends `x-session-id`. Does not affect the `prompt_cache_key` body param, which is governed by cache retention. Default: auto-detected. */

875 sessionAffinityFormat?: SessionAffinityFormat;

876 /** Whether the provider supports long prompt cache retention. This uses `prompt_cache_options.ttl: "30m"` on GPT-5.6+ and `prompt_cache_retention: "24h"` on earlier models. Default: true. */

877 supportsLongCacheRetention?: boolean;

878 /** Whether the provider supports strict JSON-schema function tools. Defaults are API-specific; generated OpenAI models enable it explicitly. */

879 supportsStrictMode?: boolean;

880 /** Whether to emit OpenAI custom tools with Lark/regex grammar formats. When false, grammar-constrained tools fall back to normal function tools. Default: false; the generated model catalog enables it for capable models. */

881 supportsOpenAIGrammarTools?: boolean;

882 /** Whether the model supports message-anchored `additional_tools` input items. Default: false. */

883 supportsAdditionalTools?: boolean;

884 /** Whether the model supports client-executed tool search for transcript-anchored additions. Default: false. */

885 supportsToolSearch?: boolean;

886 /** Whether the model accepts `prompt_cache_options` (OpenAI GPT-5.6+ prompt caching). Older OpenAI models reject the parameter. Default: false. */

887 supportsExplicitPromptCacheMode?: boolean;

888 /** Whether the provider accepts the `max_output_tokens` parameter. Some Codex-protocol gateways reject it. Default: true. */

889 supportsMaxOutputTokens?: boolean;

890}

891

892/** Compatibility settings for Anthropic Messages-compatible APIs. */

893export interface AnthropicMessagesCompat {

894 /**

895 * Whether the provider accepts per-tool `eager_input_streaming`.

896 * When false, the Anthropic provider omits `tools[].eager_input_streaming`

897 * and sends the legacy `fine-grained-tool-streaming-2025-05-14` beta header

898 * for tool-enabled requests.

899 * Default: true.

900 */

901 supportsEagerToolInputStreaming?: boolean;

902 /** Whether the provider supports Anthropic long cache retention (`cache_control.ttl: "1h"`). Default: true. */

903 supportsLongCacheRetention?: boolean;

904 /**

905 * Whether to send the `x-session-affinity` header from `options.sessionId`

906 * when caching is enabled. Required for providers like Fireworks that use

907 * session affinity for prompt cache routing (requests to the same replica

908 * maximize cache hits).

909 * Default: false.

910 */

911 sendSessionAffinityHeaders?: boolean;

912 /** Session-affinity format. `"openrouter"` sends `x-session-id`; when unset, sends `x-session-affinity`. */

913 sessionAffinityFormat?: "openrouter";

914 /**

915 * Whether the provider supports Anthropic-style `cache_control` markers on

916 * tool definitions. When false, `cache_control` is omitted from tool params.

917 * Some Anthropic-compatible providers (e.g., Fireworks) do not support this

918 * field on tools and may reject or ignore it.

919 * Default: true.

920 */

921 supportsCacheControlOnTools?: boolean;

922 /**

923 * Whether the model accepts the Anthropic `temperature` request field.

924 * Claude Opus 4.7+ rejects non-default temperature values.

925 * Default: true.

926 */

927 supportsTemperature?: boolean;

928 /**

929 * Whether to force adaptive thinking (`thinking.type: "adaptive"` plus

930 * `output_config.effort`) regardless of the model id. Built-in models that

931 * require adaptive thinking set this in generated metadata. Custom

932 * Anthropic-compatible providers can set this to `true` for any model whose

933 * upstream requires the adaptive format. Set to `false` to

934 * opt out on overridden built-in models.

935 * Default: false.

936 */

937 forceAdaptiveThinking?: boolean;

938 /** Whether to replay empty thinking signatures as `signature: ""` instead of converting thinking to text. Default: false. */

939 allowEmptySignature?: boolean;

940 /** Whether the provider supports Anthropic strict tool schemas. Default: false; generated Anthropic models enable it explicitly. */

941 supportsStrictTools?: boolean;

942 /** Whether the exact model transport supports effort-only system messages and thinking binding controls. Default: false. */

943 supportsMidConvoEffort?: boolean;

944 /** Whether the exact model accepts system-role messages inside the conversation. When false, later system messages are folded into the top-level system prompt. Default: false. */

945 supportsMidConvoSystemMessages?: boolean;

946 /** Whether the exact model accepts mid-conversation `tool_addition` and `tool_removal` blocks. Requires `supportsMidConvoSystemMessages`. Default: false. */

947 supportsMidConvoToolChanges?: boolean;

948 /**

949 * Models Anthropic accepts in `fallbacks` for server-side refusal fallback,

950 * with local pricing metadata for returned fallback responses. When absent or

951 * empty, callers must omit `fallbacks`; Anthropic rejects the field for models

952 * with no permitted fallback targets.

953 */

954 allowedFallbackModels?: AnthropicAllowedFallbackModel[];

955}

956

957/** Compatibility settings for Amazon Bedrock models. */

958export interface BedrockCompat {

959 /** Whether the model supports Bedrock strict tool schemas. Default: false. */

960 supportsStrictMode?: boolean;

961}

962

963/** Compatibility settings for the Mistral chat API. */

964export interface MistralConversationsCompat {

965 /** Whether the exact model accepts system messages after the conversation has started. When false, later system messages are folded into the leading system message. Default: false. */

966 supportsMidConvoSystemMessages?: boolean;

967}

968

969/**

970 * OpenRouter provider routing preferences.

971 * Controls which upstream providers OpenRouter routes requests to.

972 * Sent as the `provider` field in the OpenRouter API request body.

973 * @see https://openrouter.ai/docs/guides/routing/provider-selection

974 */

975export interface OpenRouterRouting {

976 /** Whether to allow backup providers to serve requests. Default: true. */

977 allow_fallbacks?: boolean;

978 /** Whether to filter providers to only those that support all parameters in the request. Default: false. */

979 require_parameters?: boolean;

980 /** Data collection setting. "allow" (default): allow providers that may store/train on data. "deny": only use providers that don't collect user data. */

981 data_collection?: "deny" | "allow";

982 /** Whether to restrict routing to only ZDR (Zero Data Retention) endpoints. */

983 zdr?: boolean;

984 /** Whether to restrict routing to only models that allow text distillation. */

985 enforce_distillable_text?: boolean;

986 /** An ordered list of provider names/slugs to try in sequence, falling back to the next if unavailable. */

987 order?: string[];

988 /** List of provider names/slugs to exclusively allow for this request. */

989 only?: string[];

990 /** List of provider names/slugs to skip for this request. */

991 ignore?: string[];

992 /** A list of quantization levels to filter providers by (e.g., ["fp16", "bf16", "fp8", "fp6", "int8", "int4", "fp4", "fp32"]). */

993 quantizations?: string[];

994 /** Sorting strategy. Can be a string (e.g., "price", "throughput", "latency") or an object with `by` and `partition`. */

995 sort?:

996 | string

997 | {

998 /** The sorting metric: "price", "throughput", "latency". */

999 by?: string;

1000 /** Partitioning strategy: "model" (default) or "none". */

1001 partition?: string | null;

1002 };

1003 /** Maximum price per million tokens (USD). */

1004 max_price?: {

1005 /** Price per million prompt tokens. */

1006 prompt?: number | string;

1007 /** Price per million completion tokens. */

1008 completion?: number | string;

1009 /** Price per image. */

1010 image?: number | string;

1011 /** Price per audio unit. */

1012 audio?: number | string;

1013 /** Price per request. */

1014 request?: number | string;

1015 };

1016 /** Preferred minimum throughput (tokens/second). Can be a number (applies to p50) or an object with percentile-specific cutoffs. */

1017 preferred_min_throughput?:

1018 | number

1019 | {

1020 /** Minimum tokens/second at the 50th percentile. */

1021 p50?: number;

1022 /** Minimum tokens/second at the 75th percentile. */

1023 p75?: number;

1024 /** Minimum tokens/second at the 90th percentile. */

1025 p90?: number;

1026 /** Minimum tokens/second at the 99th percentile. */

1027 p99?: number;

1028 };

1029 /** Preferred maximum latency (seconds). Can be a number (applies to p50) or an object with percentile-specific cutoffs. */

1030 preferred_max_latency?:

1031 | number

1032 | {

1033 /** Maximum latency in seconds at the 50th percentile. */

1034 p50?: number;

1035 /** Maximum latency in seconds at the 75th percentile. */

1036 p75?: number;

1037 /** Maximum latency in seconds at the 90th percentile. */

1038 p90?: number;

1039 /** Maximum latency in seconds at the 99th percentile. */

1040 p99?: number;

1041 };

1042}

1043

1044/**

1045 * Vercel AI Gateway routing preferences.

1046 * Controls which upstream providers the gateway routes requests to.

1047 * @see https://vercel.com/docs/ai-gateway/models-and-providers/provider-options

1048 */

1049export interface VercelGatewayRouting {

1050 /** List of provider slugs to exclusively use for this request (e.g., ["bedrock", "anthropic"]). */

1051 only?: string[];

1052 /** List of provider slugs to try in order (e.g., ["anthropic", "openai"]). */

1053 order?: string[];

1054}

1055

1056export interface ModelCostRates {

1057 input: number; // $/million tokens

1058 output: number; // $/million tokens

1059 cacheRead: number; // $/million tokens

1060 cacheWrite: number; // $/million tokens

1061}

1062

1063export interface ModelCostTier extends ModelCostRates {

1064 /** Use this tier for requests whose total input usage exceeds this token count. */

1065 inputTokensAbove: number;

1066}

1067

1068export interface ModelCost extends ModelCostRates {

1069 /** Request-wide pricing tiers. The highest matching input threshold applies to the full request. */

1070 tiers?: ModelCostTier[];

1071}

1072

1073export interface ModelImageResizeOptions {

1074 maxWidth?: number;

1075 maxHeight?: number;

1076 /** Maximum base64-encoded payload size in bytes. */

1077 maxBytes?: number;

1078 jpegQuality?: number;

1079}

1080

1081export interface ModelImageInputLimits {

1082 /** Cache-safe resize profile applied before a new image enters conversation history. */

1083 resize?: ModelImageResizeOptions;

1084 /** Maximum images accepted in one provider message. */

1085 maxPerMessage?: number;

1086 /** Maximum images accepted across one provider request. */

1087 maxPerRequest?: number;

1088}

1089

1090export interface ModelInputLimits {

1091 /** Maximum serialized provider request size in bytes. */

1092 maxRequestBytes?: number;

1093 images?: ModelImageInputLimits;

1094}

1095

1096/** Fields shared by every catalog entry, regardless of what you can do with it. */

1097export interface BaseModel<TApi extends string> {

1098 id: string;

1099 name: string;

1100 api: TApi;

1101 provider: ProviderId;

1102 baseUrl: string;

1103 input: ("text" | "image")[];

1104 /** Provider input limits and cache-safe preprocessing metadata. */

1105 inputLimits?: ModelInputLimits;

1106 cost: ModelCost;

1107 headers?: Record<string, string>;

1108}

1109

1110/** Chat model: usable with `stream()` and friends. */

1111export interface Model<TApi extends Api> extends BaseModel<TApi> {

1112 /**

1113 * Optional: chat is the default model type, so models without `type` are chat

1114 * models. Narrow mixed model lists with `isModelType()` instead of comparing

1115 * `type` directly.

1116 */

1117 type?: "chat";

1118 reasoning: boolean;

1119 /**

1120 * Maps pi thinking levels to provider/model-specific values.

1121 * Missing keys use provider defaults. null marks a level as unsupported.

1122 */

1123 thinkingLevelMap?: ThinkingLevelMap;

1124 /** Prompt cache lifetimes per retention tier. Unset when the provider's cache behavior is unknown. */

1125 promptCache?: ModelPromptCache;

1126 contextWindow: number;

1127 maxTokens: number;

1128 /** Default sampling parameters for this model. See {@link StreamOptions.samplingParams}; per-request keys override these. */

1129 samplingParams?: Record<string, unknown>;

1130 /** Compatibility overrides for OpenAI-compatible APIs. If not set, auto-detected from baseUrl. */

1131 compat?: TApi extends "openai-completions"

1132 ? OpenAICompletionsCompat

1133 : TApi extends "openai-responses" | "azure-openai-responses" | "openai-codex-responses"

1134 ? OpenAIResponsesCompat

1135 : TApi extends "anthropic-messages"

1136 ? AnthropicMessagesCompat

1137 : TApi extends "bedrock-converse-stream"

1138 ? BedrockCompat

1139 : TApi extends "mistral-conversations"

1140 ? MistralConversationsCompat

1141 : never;

1142}

1143

1144/** Image-generation model: usable with `generateImages()` only. */

1145export interface ImageModel<TApi extends ImageApi> extends BaseModel<TApi> {

1146 type: "image";

1147 /** Output modalities. Always includes `"image"`; `"text"` means the model can also return text blocks. */

1148 output: ("text" | "image")[];

1149}

1150

1151/** Structured classifier model: usable with `classify()` only. */

1152export interface ClassifierModel<TApi extends ClassifierApi> extends BaseModel<TApi> {

1153 type: "classifier";

1154 contextWindow: number;

1155}

1156

1157/** Model shape for each model type. */

1158export interface ModelTypeMap {

1159 chat: Model<Api>;

1160 image: ImageModel<ImageApi>;

1161 classifier: ClassifierModel<ClassifierApi>;

1162}

1163

1164/** What a catalog entry is for. Decides which `Models` operation accepts it. */

1165export type ModelType = keyof ModelTypeMap;

1166

1167/** Anything a provider can list. Narrow with `isModelType()`. */

1168export type AnyModel = ModelTypeMap[ModelType];