@bastani/pi-ai 0.9.18 → 0.9.19-alpha.10

package/dist/types.js.map

@@ -1 +1 @@
-{"version":3,"file":"types.js","sourceRoot":"","sources":["../src/types.ts"],"names":[],"mappings":"","sourcesContent":["import type { TelemetryContext } from \"@earendil-works/pi-telemetry\";\nimport type { AnthropicOptions } from \"./api/anthropic-messages.ts\";\nimport type { AzureOpenAIResponsesOptions } from \"./api/azure-openai-responses.ts\";\nimport type { BedrockOptions } from \"./api/bedrock-converse-stream.ts\";\nimport type { GoogleOptions } from \"./api/google-generative-ai.ts\";\nimport type { GoogleVertexOptions } from \"./api/google-vertex.ts\";\nimport type { MistralOptions } from \"./api/mistral-conversations.ts\";\nimport type { OpenAICodexResponsesOptions } from \"./api/openai-codex-responses.ts\";\nimport type { OpenAICompletionsOptions } from \"./api/openai-completions.ts\";\nimport type { OpenAIResponsesOptions } from \"./api/openai-responses.ts\";\nimport type { PiMessagesOptions } from \"./api/pi-messages.ts\";\nimport type { AssistantMessageDiagnostic } from \"./utils/diagnostics.ts\";\nimport type { AssistantMessageEventStream } from \"./utils/event-stream.ts\";\n\nexport type { AssistantMessageEventStream } from \"./utils/event-stream.ts\";\n\nexport type KnownApi =\n\t| \"openai-completions\"\n\t| \"mistral-conversations\"\n\t| \"openai-responses\"\n\t| \"azure-openai-responses\"\n\t| \"openai-codex-responses\"\n\t| \"anthropic-messages\"\n\t| \"bedrock-converse-stream\"\n\t| \"google-generative-ai\"\n\t| \"google-vertex\"\n\t| \"pi-messages\";\n\nexport type Api = KnownApi | (string & {});\n\nexport type KnownImagesApi = \"openrouter-images\";\n\nexport type ImagesApi = KnownImagesApi | (string & {});\n\nexport type KnownProvider =\n\t| \"amazon-bedrock\"\n\t| \"ant-ling\"\n\t| \"anthropic\"\n\t| \"google\"\n\t| \"google-vertex\"\n\t| \"openai\"\n\t| \"azure-openai-responses\"\n\t| \"openai-codex\"\n\t| \"radius\"\n\t| \"nvidia\"\n\t| \"deepseek\"\n\t| \"github-copilot\"\n\t| \"xai\"\n\t| \"groq\"\n\t| \"cerebras\"\n\t| \"openrouter\"\n\t| \"vercel-ai-gateway\"\n\t| \"zai\"\n\t| \"zai-coding-cn\"\n\t| \"mistral\"\n\t| \"minimax\"\n\t| \"minimax-cn\"\n\t| \"moonshotai\"\n\t| \"moonshotai-cn\"\n\t| \"huggingface\"\n\t| \"fireworks\"\n\t| \"together\"\n\t| \"baseten\"\n\t| \"opencode\"\n\t| \"opencode-go\"\n\t| \"kimi-coding\"\n\t| \"cloudflare-workers-ai\"\n\t| \"cloudflare-ai-gateway\"\n\t| \"qwen-token-plan\"\n\t| \"qwen-token-plan-cn\"\n\t| \"qwen-token-plan-individual\"\n\t| \"xiaomi\"\n\t| \"xiaomi-token-plan-cn\"\n\t| \"xiaomi-token-plan-ams\"\n\t| \"xiaomi-token-plan-sgp\";\nexport type ProviderId = KnownProvider | string;\n\nexport type KnownImagesProvider = \"openrouter\";\n\nexport type ImagesProviderId = KnownImagesProvider | string;\n\nexport type ToolChoice = \"auto\" | \"none\";\nexport type ThinkingLevel = \"minimal\" | \"low\" | \"medium\" | \"high\" | \"xhigh\" | \"max\";\nexport type ModelThinkingLevel = \"off\" | ThinkingLevel;\nexport type ThinkingLevelMap = Partial<Record<ModelThinkingLevel, string | null>>;\nexport type ChatTemplateKwargValue =\n\t| string\n\t| number\n\t| boolean\n\t| null\n\t| {\n\t\t\t$var: \"thinking.enabled\" | \"thinking.effort\" | \"thinking.budget\";\n\t\t\tomitWhenOff?: boolean;\n\t  };\n\n/** Top-level request field used to cap reasoning tokens on OpenAI-compatible servers. */\nexport type ThinkingTokenBudgetField = \"thinking_token_budget\" | \"thinking_budget\" | \"thinking_budget_tokens\";\n\n/** Token budgets for each thinking level (token-based providers only) */\nexport interface ThinkingBudgets {\n\tminimal?: number;\n\tlow?: number;\n\tmedium?: number;\n\thigh?: number;\n}\n\n// Base options all providers share\nexport type CacheRetention = \"none\" | \"short\" | \"long\";\n\nexport type Transport = \"sse\" | \"websocket\" | \"websocket-cached\" | \"auto\";\n\n/** Provider-scoped environment overrides. Values take precedence over process.env. */\nexport type ProviderEnv = Record<string, string>;\nexport type ProviderHeaders = Record<string, string | null>;\nexport type FetchFunction = typeof globalThis.fetch;\nexport type SessionAffinityFormat = \"openai\" | \"openai-nosession\" | \"openrouter\";\n\nexport interface ProviderResponse {\n\tstatus: number;\n\theaders: Record<string, string>;\n}\n\n/** Authentication, HTTP transport, and lifecycle callbacks shared by provider requests. */\nexport interface ProviderRequestOptions<TModel = Model<Api>> {\n\tsignal?: AbortSignal;\n\t/** Explicit parent context for telemetry produced by this logical request. */\n\ttelemetryContext?: TelemetryContext;\n\tapiKey?: string;\n\t/**\n\t * Optional fetch implementation for provider HTTP requests.\n\t * Defaults to `globalThis.fetch`. Provider adapters that cannot inject a custom implementation may reject it.\n\t * This does not affect WebSocket transports.\n\t */\n\tfetch?: FetchFunction;\n\t/**\n\t * Provider-scoped environment values. These take precedence over process.env for\n\t * provider configuration such as regional settings, endpoint placeholders, and\n\t * proxy variables.\n\t */\n\tenv?: ProviderEnv;\n\t/**\n\t * Optional callback for inspecting or replacing provider payloads before sending.\n\t * Return undefined to keep the payload unchanged.\n\t */\n\tonPayload?: (payload: unknown, model: TModel) => unknown | undefined | Promise<unknown | undefined>;\n\t/**\n\t * Optional callback invoked after an HTTP response is received.\n\t */\n\tonResponse?: (response: ProviderResponse, model: TModel) => void | Promise<void>;\n\t/**\n\t * Optional custom HTTP headers to include in API requests.\n\t * Merged with provider defaults; caller values override default headers.\n\t * On AWS Bedrock these are injected via a Smithy `build`-step middleware so\n\t * they are covered by SigV4 signing; reserved headers (`x-amz-*`,\n\t * `authorization`, `host`) are silently ignored to preserve SigV4 / bearer auth.\n\t * A null value suppresses a provider/API default header with the same name.\n\t */\n\theaders?: ProviderHeaders;\n\t/**\n\t * HTTP request timeout in milliseconds for providers/SDKs that support it.\n\t * For example, OpenAI and Anthropic SDK clients default to 10 minutes.\n\t */\n\ttimeoutMs?: number;\n\t/**\n\t * Maximum retry attempts for providers/SDKs that support client-side retries.\n\t * For example, OpenAI and Anthropic SDK clients default to 2.\n\t */\n\tmaxRetries?: number;\n\t/**\n\t * Maximum delay in milliseconds to wait for a retry when the server requests a long wait.\n\t * If the server's requested delay exceeds this value, the request fails immediately\n\t * with an error containing the requested delay, allowing higher-level retry logic\n\t * to handle it with user visibility.\n\t * Default: 60000 (60 seconds). Set to 0 to disable the cap.\n\t */\n\tmaxRetryDelayMs?: number;\n}\n\nexport interface StreamOptions extends ProviderRequestOptions<Model<Api>> {\n\t/**\n\t * Optional callback invoked after an HTTP response is received and before\n\t * its body stream is consumed.\n\t */\n\tonResponse?: (response: ProviderResponse, model: Model<Api>) => void | Promise<void>;\n\ttemperature?: number;\n\t/**\n\t * Arbitrary sampling parameters merged into the request body as-is, after the named request\n\t * fields, so keys here override them. Lets custom OpenAI-compatible servers (llama.cpp, vLLM,\n\t * SGLang, ...) receive parameters pi does not model, e.g. `top_p`, `top_k`, `min_p`,\n\t * `repetition_penalty`. Merged over `Model.samplingParams` per key. Only applied by\n\t * OpenAI-compatible adapters (completions, responses, Azure responses); other APIs ignore it.\n\t */\n\tsamplingParams?: Record<string, unknown>;\n\tmaxTokens?: number;\n\t/**\n\t * Preferred transport for providers that support multiple transports.\n\t * Providers that do not support this option ignore it.\n\t */\n\ttransport?: Transport;\n\t/**\n\t * Prompt cache retention preference. Providers map this to their supported values.\n\t * Default: \"short\".\n\t */\n\tcacheRetention?: CacheRetention;\n\t/**\n\t * Optional session identifier for providers that support session-based caching.\n\t * Providers can use this to enable prompt caching, request routing, or other\n\t * session-aware features. Ignored by providers that don't support it.\n\t */\n\tsessionId?: string;\n\t/**\n\t * WebSocket connect timeout in milliseconds for providers that support\n\t * WebSocket transports. This covers the connection/open handshake only;\n\t * HTTP/SDK idleness uses `timeoutMs`, while gaps between decoded provider\n\t * stream events use `streamDeadlineMs`.\n\t */\n\twebsocketConnectTimeoutMs?: number;\n\t/**\n\t * Maximum idle gap in milliseconds between two provider stream events.\n\t *\n\t * Enforced below the HTTP layer, so a response body that stalls without ever\n\t * rejecting the adapter's async iterator (for example a decompression failure\n\t * that destroys the stream silently, #2553) still settles the attempt as a\n\t * retryable transport error instead of hanging forever. The timer restarts on\n\t * every event, so it never caps total stream duration.\n\t *\n\t * Defaults to 300000 ms. `0` (or any non-positive value) disables it.\n\t */\n\tstreamDeadlineMs?: number;\n\t/**\n\t * Optional metadata to include in API requests.\n\t * Providers extract the fields they understand and ignore the rest.\n\t * For example, Anthropic uses `user_id` for abuse tracking and rate limiting.\n\t */\n\tmetadata?: Record<string, unknown>;\n}\n\nexport type ProviderStreamOptions = StreamOptions & Record<string, unknown>;\n\nexport interface DeferredFetchOptions extends ProviderRequestOptions<Model<Api>> {\n\t/**\n\t * Maximum provider long-poll duration in milliseconds.\n\t * Defaults to 0, which performs one status check.\n\t */\n\twait?: number;\n}\n\n/** Request options for best-effort deferred-response cancellation. */\nexport type DeferredCancelOptions = ProviderRequestOptions<Model<Api>>;\n\n/**\n * Maps known APIs to their full provider-specific stream option types.\n * Type-only imports from API implementation modules are erased at emit, so\n * this is tree-shake safe.\n */\nexport interface ApiOptionsMap {\n\t\"anthropic-messages\": AnthropicOptions;\n\t\"openai-completions\": OpenAICompletionsOptions;\n\t\"openai-responses\": OpenAIResponsesOptions;\n\t\"openai-codex-responses\": OpenAICodexResponsesOptions;\n\t\"azure-openai-responses\": AzureOpenAIResponsesOptions;\n\t\"google-generative-ai\": GoogleOptions;\n\t\"google-vertex\": GoogleVertexOptions;\n\t\"mistral-conversations\": MistralOptions;\n\t\"bedrock-converse-stream\": BedrockOptions;\n\t\"pi-messages\": PiMessagesOptions;\n}\n\n/**\n * Full stream options for an API. Known APIs resolve to their concrete option\n * type; custom API strings fall back to the generic shape.\n */\nexport type ApiStreamOptions<TApi extends Api> = TApi extends keyof ApiOptionsMap\n\t? ApiOptionsMap[TApi]\n\t: StreamOptions & Record<string, unknown>;\n\n/**\n * The uniform stream contract of an API implementation module: every module\n * under `src/api/` exports `stream` and `streamSimple`; capable modules may also\n * export deferred-response methods. Lazy wrappers (`lazyApi()`) and provider\n * factories pass these around as values. This is the untyped dispatch shape;\n * per-API option typing lives on the implementation modules themselves and on\n * `Provider.stream()` via `ApiStreamOptions`.\n */\nexport interface ProviderStreams {\n\tstream(model: Model<Api>, context: Context, options?: StreamOptions): AssistantMessageEventStream;\n\tstreamSimple(model: Model<Api>, context: Context, options?: SimpleStreamOptions): AssistantMessageEventStream;\n\tfetchDeferred?(\n\t\tmodel: Model<Api>,\n\t\thandle: DeferredHandle,\n\t\toptions?: DeferredFetchOptions,\n\t): AssistantMessageEventStream;\n\tcancelDeferred?(model: Model<Api>, handle: DeferredHandle, options?: DeferredCancelOptions): Promise<void>;\n}\n\n/**\n * The uniform contract of an image-generation API implementation module:\n * every image API module under `src/api/` exports exactly `generateImages`,\n * so the module itself satisfies this interface. Lazy wrappers and image\n * provider factories pass these around as values.\n */\nexport interface ProviderImages {\n\tgenerateImages(\n\t\tmodel: ImagesModel<ImagesApi>,\n\t\tcontext: ImagesContext,\n\t\toptions?: ImagesOptions,\n\t): Promise<AssistantImages>;\n}\n\nexport interface ImagesOptions extends ProviderRequestOptions<ImagesModel<ImagesApi>> {\n\t/**\n\t * Optional metadata to include in API requests.\n\t * Providers extract the fields they understand and ignore the rest.\n\t */\n\tmetadata?: Record<string, unknown>;\n}\n\nexport type ProviderImagesOptions = ImagesOptions & Record<string, unknown>;\n\nexport interface AnthropicAllowedFallbackModel {\n\tprovider: ProviderId;\n\tmodel: string;\n\tcost: ModelCost;\n}\n\n// Unified options with reasoning passed to streamSimple() and completeSimple()\nexport interface SimpleStreamOptions extends StreamOptions {\n\t/** Provider-neutral tool selection for simple requests. When omitted, adapters use provider-specific behavior. */\n\ttoolChoice?: ToolChoice;\n\treasoning?: ThinkingLevel;\n\t/** Ask a capable provider to return a durable handle and continue the request asynchronously. */\n\tdeferred?: boolean | { window?: \"15m\" | \"1h\" | \"24h\" };\n\t/** Custom token budgets for thinking levels (token-based providers only) */\n\tthinkingBudgets?: ThinkingBudgets;\n}\n\n// Generic StreamFunction with typed options.\n//\n// Contract:\n// - Must return an AssistantMessageEventStream.\n// - Direct streamSimple() calls may throw synchronously when request auth is\n//   missing. Once a stream is returned, request/model/runtime failures should\n//   be encoded in the stream, not thrown.\n// - Error termination must produce an AssistantMessage with stopReason\n//   \"error\" or \"aborted\" and errorMessage, emitted via the stream protocol.\nexport type StreamFunction<TApi extends Api = Api, TOptions extends StreamOptions = StreamOptions> = (\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: TOptions,\n) => AssistantMessageEventStream;\n\nexport type ImagesFunction<TApi extends ImagesApi = ImagesApi, TOptions extends ImagesOptions = ImagesOptions> = (\n\tmodel: ImagesModel<TApi>,\n\tcontext: ImagesContext,\n\toptions?: TOptions,\n) => Promise<AssistantImages>;\n\nexport interface TextSignatureV1 {\n\tv: 1;\n\tid: string;\n\tphase?: \"commentary\" | \"final_answer\";\n}\n\nexport interface TextContent {\n\ttype: \"text\";\n\ttext: string;\n\ttextSignature?: string; // e.g., for OpenAI responses, message metadata (legacy id string or TextSignatureV1 JSON)\n}\n\nexport interface ThinkingContent {\n\ttype: \"thinking\";\n\tthinking: string;\n\tthinkingSignature?: string; // Provider-specific opaque or serialized reasoning replay data\n\t/** When true, the thinking content was redacted by safety filters. The opaque\n\t *  encrypted payload is stored in `thinkingSignature` so it can be passed back\n\t *  to the API for multi-turn continuity. */\n\tredacted?: boolean;\n}\n\n/**\n * Boundary marker emitted by Anthropic server-side fallback, where one model's output gives way\n * to the next after a classifier refusal.\n *\n * This is public content rather than a stream-only event because it must survive the round trip:\n * \"Keep it exactly where it appeared. The API uses its position to validate the thinking blocks\n * around it, so a request that echoes thinking blocks from both sides of the boundary is rejected\n * if the block is omitted or moved.\"\n * https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback\n */\nexport interface FallbackContent {\n\ttype: \"fallback\";\n\t/** The model that declined. Echoes the model string sent when the declining hop is the request's own model. */\n\tfromModel: string;\n\t/** Resolved id of the model that continues. Always present. */\n\ttoModel: string;\n}\n\nexport interface ImageContent {\n\ttype: \"image\";\n\tdata: string; // base64 encoded image data\n\tmimeType: string; // e.g., \"image/jpeg\", \"image/png\"\n}\n\n/**\n * A document sent as model input, currently PDF only.\n *\n * Anthropic documents PDF as a platform capability — \"All active models support PDF processing\" —\n * routed through the same vision path as images, so it is not a per-model modality. A model\n * advertises it through `Model.input` containing `\"pdf\"`, which generated metadata sets only for\n * the runtimes that can actually serialize a document block. Providers that cannot are sent a\n * visible placeholder instead, exactly as they are for images they cannot accept.\n * https://platform.claude.com/docs/en/build-with-claude/pdf-support\n */\nexport interface DocumentContent {\n\ttype: \"document\";\n\t/** Base64-encoded document bytes. Amazon Bedrock takes raw bytes and decodes this itself. */\n\tdata: string;\n\t/**\n\t * The only media type either serializer implements. It is a literal rather than `string`\n\t * because both paths hardcode PDF — the Anthropic block emits `BetaBase64PDFSource`, whose\n\t * `media_type` is itself a fixed literal, and Bedrock emits `DocumentFormat.PDF` — so a wider\n\t * type would let a caller label arbitrary bytes as a PDF. Widening this union later is\n\t * non-breaking; narrowing it after release would not be. Note that adding a format is not a\n\t * media-type swap on the Anthropic side: plain text needs a structurally different source\n\t * variant (`BetaPlainTextSource`, `type: \"text\"`), and Bedrock's `doc`/`docx`/`xls`/`xlsx`/\n\t * `csv`/`html`/`md` formats have no Anthropic source variant at all.\n\t */\n\tmimeType: \"application/pdf\";\n\t/**\n\t * Optional caller-supplied label. Amazon Bedrock requires a name and warns that it \"is\n\t * vulnerable to prompt injections\", so the Bedrock path sanitizes this to the characters AWS\n\t * permits and falls back to a neutral generated name.\n\t */\n\tname?: string;\n}\n\nexport interface ToolCall {\n\ttype: \"toolCall\";\n\tid: string;\n\tname: string;\n\targuments: Record<string, any>;\n\tthoughtSignature?: string; // Google-specific: opaque signature for reusing thought context\n\t/** OpenAI Responses namespace for calls to dynamically loaded or namespaced tools. */\n\tnamespace?: string;\n}\n\nexport interface Usage {\n\tinput: number;\n\toutput: number;\n\tcacheRead: number;\n\tcacheWrite: number;\n\t/** Subset of `cacheWrite` written with 1h retention. Only Anthropic reports this split. */\n\tcacheWrite1h?: number;\n\t/**\n\t * Reasoning/thinking tokens, when the provider reports them. This is a subset of\n\t * `output`: `output` already includes these tokens. Set to a number (possibly 0) by\n\t * providers that expose a reasoning breakdown; left undefined by providers that don't.\n\t */\n\treasoning?: number;\n\ttotalTokens: number;\n\tcost: {\n\t\tinput: number;\n\t\toutput: number;\n\t\tcacheRead: number;\n\t\tcacheWrite: number;\n\t\ttotal: number;\n\t};\n}\n\nexport type StopReason = \"pending\" | \"stop\" | \"length\" | \"toolUse\" | \"error\" | \"aborted\" | \"deferred\";\n\nexport type JsonValue = string | number | boolean | null | JsonValue[] | { [key: string]: JsonValue };\n\nexport interface DeferredHandle {\n\tprovider: string;\n\tmodelId: string;\n\tapi: string;\n\t/** Provider token, such as a response id or batch id plus row id. */\n\tid: string;\n\texpiresAt?: number;\n\tpollAfterMs?: number;\n\t/** Provider conversion data required to reconstruct the final assistant message. */\n\tdata?: JsonValue;\n}\n\nexport interface UserMessage {\n\trole: \"user\";\n\tcontent: string | (TextContent | ImageContent | DocumentContent)[];\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport interface AssistantMessage {\n\trole: \"assistant\";\n\tcontent: (TextContent | ThinkingContent | ToolCall | FallbackContent)[];\n\tapi: Api;\n\tprovider: ProviderId;\n\tmodel: string;\n\tresponseModel?: string; // Concrete `chunk.model` when different from the requested `model` (e.g. OpenRouter `auto` -> `anthropic/...`)\n\tresponseId?: string; // Provider-specific response/message identifier when the upstream API exposes one\n\t/** Exact provider-native effort level used for this response. Absent for legacy or unmanaged responses. */\n\tproviderThinkingLevel?: string;\n\tdiagnostics?: AssistantMessageDiagnostic[]; // Redacted provider/runtime diagnostics for failures and recoveries.\n\tusage: Usage;\n\tstopReason: StopReason;\n\tdeferred?: DeferredHandle;\n\terrorMessage?: string;\n\trawStopReason?: string;\n\t/**\n\t * Provider indication of whether the model explicitly ended its turn.\n\t * Preserved for debugging and does not currently affect agent control flow.\n\t */\n\tendTurn?: boolean;\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport interface ToolResultMessage<TDetails = any> {\n\trole: \"toolResult\";\n\ttoolCallId: string;\n\ttoolName: string;\n\tcontent: (TextContent | ImageContent)[]; // Supports text and images\n\tdetails?: TDetails;\n\t/** Usage from the tool execution itself, if available. Not part of main LLM context accounting. */\n\tusage?: Usage;\n\t/**\n\t * Names from `Context.tools` that became available after this result.\n\t * Providers with native deferred tool loading use this as the load point;\n\t * other providers ignore it and use `Context.tools` normally.\n\t */\n\taddedToolNames?: string[];\n\tisError: boolean;\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport type Message = UserMessage | AssistantMessage | ToolResultMessage;\n\nexport type ImagesInputContent = TextContent | ImageContent;\nexport type ImagesOutputContent = TextContent | ImageContent;\n\nexport interface ImagesContext {\n\tinput: ImagesInputContent[];\n}\n\nexport type ImagesStopReason = \"stop\" | \"error\" | \"aborted\";\n\nexport interface AssistantImages {\n\tapi: ImagesApi;\n\tprovider: ImagesProviderId;\n\tmodel: string;\n\toutput: ImagesOutputContent[];\n\tresponseId?: string;\n\tusage?: Usage;\n\tstopReason: ImagesStopReason;\n\terrorMessage?: string;\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nimport type { TSchema } from \"typebox\";\n\n/** OpenAI grammar variants for constrained sampling. */\nexport type GrammarFormat = \"openai_lark\" | \"openai_regex\";\n\nexport type GrammarVariants = Partial<Record<GrammarFormat, string>>;\n\n/**\n * Optional provider-side constrained sampling configs for a tool.\n *\n * The `json_schema` value roughly maps to the concept of `strict` in APIs which is\n * implemented as json-schema constrained sampling by APIs. Grammar variants let\n * callers provide provider-specific encodings of the same intended language.\n */\nexport type ConstrainedSamplingConfig =\n\t| {\n\t\t\ttype: \"json_schema\";\n\t\t\tstrict: \"prefer\" | \"require\";\n\t  }\n\t| {\n\t\t\ttype: \"grammar\";\n\t\t\tvariants: GrammarVariants;\n\t  };\n\nexport interface Tool<TParameters extends TSchema = TSchema> {\n\tname: string;\n\tdescription: string;\n\tparameters: TParameters;\n\tconstrainedSampling?: false | ConstrainedSamplingConfig;\n}\n\nexport interface Context {\n\tsystemPrompt?: string;\n\tmessages: Message[];\n\ttools?: Tool[];\n}\n\n/**\n * Event protocol for AssistantMessageEventStream.\n *\n * Successful streams emit `start` before partial updates and terminate with\n * `done`. A stream may terminate directly with `error` when request setup fails\n * before generation starts; after `start`, failures also terminate with `error`.\n * Direct `streamSimple()` calls throw synchronously when request auth is missing.\n * Updates and `done` must never appear before `start`.\n *\n * Tool-call arguments at `toolcall_start` are provider-specific;\n * `toolcall_delta` carries subsequent JSON updates.\n */\nexport type AssistantMessageEvent =\n\t| { type: \"start\"; partial: AssistantMessage }\n\t| { type: \"text_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"text_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| { type: \"text_end\"; contentIndex: number; content: string; partial: AssistantMessage }\n\t| { type: \"thinking_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"thinking_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| { type: \"thinking_end\"; contentIndex: number; content: string; partial: AssistantMessage }\n\t| { type: \"toolcall_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"toolcall_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| { type: \"toolcall_end\"; contentIndex: number; toolCall: ToolCall; partial: AssistantMessage }\n\t| {\n\t\t\ttype: \"done\";\n\t\t\treason: Extract<StopReason, \"stop\" | \"length\" | \"toolUse\" | \"deferred\">;\n\t\t\tmessage: AssistantMessage;\n\t  }\n\t| { type: \"error\"; reason: Extract<StopReason, \"aborted\" | \"error\">; error: AssistantMessage };\n\n/**\n * Compatibility settings for OpenAI-compatible completions APIs.\n * Use this to override URL-based auto-detection for custom providers.\n */\nexport interface OpenAICompletionsCompat {\n\t/** Whether the provider supports the `store` field. Default: auto-detected from URL. */\n\tsupportsStore?: boolean;\n\t/** Whether the provider supports the `developer` role (vs `system`). Default: auto-detected from URL. */\n\tsupportsDeveloperRole?: boolean;\n\t/** Whether the provider supports `reasoning_effort`. Default: auto-detected from URL. */\n\tsupportsReasoningEffort?: boolean;\n\t/**\n\t * Whether the model accepts the `temperature` request field. Claude Fable 5.1 rejects\n\t * non-default `temperature`, `top_p`, and `top_k` on every request, and OpenRouter's own\n\t * `supported_parameters` for that model omits `temperature`. When false, the provider\n\t * omits `temperature` and also strips `temperature`, `top_p`, and `top_k` from\n\t * `samplingParams`, which is otherwise merged last and would reopen them.\n\t * Default: true.\n\t */\n\tsupportsTemperature?: boolean;\n\t/**\n\t * Whether the model accepts forced tool use (`tool_choice` `\"required\"` or\n\t * `{ type: \"function\", ... }`). Claude Fable 5.1 rejects it on every request, whichever\n\t * platform serves the model. When false, the provider rejects a forced choice with an error\n\t * rather than sending a request the model cannot honor. `\"auto\"` and `\"none\"` are never\n\t * altered. https://platform.claude.com/docs/en/build-with-claude/thinking\n\t * Default: true.\n\t */\n\tsupportsForcedToolChoice?: boolean;\n\t/** Whether the provider supports `stream_options: { include_usage: true }` for token usage in streaming responses. Default: true. */\n\tsupportsUsageInStreaming?: boolean;\n\t/** Whether streamed responses include `finish_reason`. When false, pi infers `stop` or `toolUse` when the stream ends. Default: true. */\n\tsupportsFinishReason?: boolean;\n\t/** Which field to use for max tokens. Default: auto-detected from URL. */\n\tmaxTokensField?: \"max_completion_tokens\" | \"max_tokens\";\n\t/** Whether tool results require the `name` field. Default: auto-detected from URL. */\n\trequiresToolResultName?: boolean;\n\t/** Whether a user message after tool results requires an assistant message in between. Default: auto-detected from URL. */\n\trequiresAssistantAfterToolResult?: boolean;\n\t/** Whether thinking blocks must be converted to text blocks with <thinking> delimiters. Default: auto-detected from URL. */\n\trequiresThinkingAsText?: boolean;\n\t/** Whether all replayed assistant messages must include an empty reasoning_content field when reasoning is enabled. Default: auto-detected from URL. */\n\trequiresReasoningContentOnAssistantMessages?: boolean;\n\t/** 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\". */\n\tthinkingFormat?:\n\t\t| \"openai\"\n\t\t| \"openrouter\"\n\t\t| \"deepseek\"\n\t\t| \"together\"\n\t\t| \"baseten\"\n\t\t| \"zai\"\n\t\t| \"qwen\"\n\t\t| \"chat-template\"\n\t\t| \"qwen-chat-template\"\n\t\t| \"string-thinking\"\n\t\t| \"ant-ling\";\n\t/** 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. */\n\tchatTemplateKwargs?: Record<string, ChatTemplateKwargValue>;\n\t/** 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. */\n\tchatTemplateArgs?: Record<string, ChatTemplateKwargValue>;\n\t/** OpenRouter-compatible routing preferences sent as the `provider` request field. */\n\topenRouterRouting?: OpenRouterRouting;\n\t/** Vercel AI Gateway routing preferences. Only used when baseUrl points to Vercel AI Gateway. */\n\tvercelGatewayRouting?: VercelGatewayRouting;\n\t/** Whether z.ai supports top-level `tool_stream: true` for streaming tool call deltas. Default: false. */\n\tzaiToolStream?: boolean;\n\t/**\n\t * Top-level request field used to cap reasoning tokens from `thinkingBudgets`.\n\t * Reasoning and the answer share `max_tokens` on these endpoints, so without a budget a\n\t * reasoning-heavy turn can consume the whole response and emit no answer.\n\t * `\"thinking_token_budget\"` is vLLM, `\"thinking_budget\"` is Qwen/DashScope/SGLang,\n\t * `\"thinking_budget_tokens\"` is llama.cpp. Off by default; not set on the generated catalog.\n\t */\n\tthinkingTokenBudgetField?: ThinkingTokenBudgetField;\n\t/** Alias for `thinkingTokenBudgetField: \"thinking_token_budget\"` (vLLM). Prefer `thinkingTokenBudgetField`. Default: false. */\n\tsupportsThinkingTokenBudget?: boolean;\n\t/** 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. */\n\tsupportsOpenAIGrammarTools?: boolean;\n\t/** Whether the provider supports the `strict` field in tool definitions. Default: true. */\n\tsupportsStrictMode?: boolean;\n\t/** 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. */\n\tcacheControlFormat?: \"anthropic\";\n\t/** Whether to send session-affinity data from `options.sessionId`. Default: false. */\n\tsendSessionAffinityHeaders?: boolean;\n\t/** Provider-specific deferred tool serialization mode. */\n\tdeferredToolsMode?: \"kimi\";\n\t/** 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. */\n\tsessionAffinityFormat?: SessionAffinityFormat;\n\t/** Whether the provider supports long prompt cache retention (`prompt_cache_retention: \"24h\"` or Anthropic-style `cache_control.ttl: \"1h\"`, depending on format). Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n\t/**\n\t * vLLM scheduler priority sent as the top-level `priority` request field (lower values are\n\t * handled earlier; server default 0). Only meaningful when vLLM runs with\n\t * `--scheduling-policy priority`; useful for keeping background/batch work from stalling\n\t * interactive sessions. Off by default; not set on the generated catalog.\n\t */\n\tvllmPriority?: number;\n}\n\n/** Compatibility settings for OpenAI Responses APIs. */\nexport interface OpenAIResponsesCompat {\n\t/** Whether the provider supports the `developer` role (vs `system`). Default: true. */\n\tsupportsDeveloperRole?: boolean;\n\t/** 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. */\n\tsessionAffinityFormat?: SessionAffinityFormat;\n\t/** 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. */\n\tsupportsLongCacheRetention?: boolean;\n\t/** Whether the provider supports strict JSON-schema function tools. Defaults are API-specific; generated OpenAI models enable it explicitly. */\n\tsupportsStrictMode?: boolean;\n\t/** 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. */\n\tsupportsOpenAIGrammarTools?: boolean;\n\t/** Whether the model supports message-anchored `additional_tools` input items. Default: false. */\n\tsupportsAdditionalTools?: boolean;\n\t/** Whether the model supports client-executed tool search for deferred tools. Default: false. */\n\tsupportsToolSearch?: boolean;\n\t/** Whether the model accepts `prompt_cache_options` (OpenAI GPT-5.6+ prompt caching). Older OpenAI models reject the parameter. Default: false. */\n\tsupportsExplicitPromptCacheMode?: boolean;\n\t/** Whether the provider accepts the `max_output_tokens` parameter. Some Codex-protocol gateways reject it. Default: true. */\n\tsupportsMaxOutputTokens?: boolean;\n}\n\n/** Compatibility settings for Anthropic Messages-compatible APIs. */\nexport interface AnthropicMessagesCompat {\n\t/**\n\t * Whether the provider accepts per-tool `eager_input_streaming`.\n\t * When false, the Anthropic provider omits `tools[].eager_input_streaming`\n\t * and sends the legacy `fine-grained-tool-streaming-2025-05-14` beta header\n\t * for tool-enabled requests.\n\t * Default: true.\n\t */\n\tsupportsEagerToolInputStreaming?: boolean;\n\t/** Whether the provider supports Anthropic long cache retention (`cache_control.ttl: \"1h\"`). Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n\t/**\n\t * Whether to send the `x-session-affinity` header from `options.sessionId`\n\t * when caching is enabled. Required for providers like Fireworks that use\n\t * session affinity for prompt cache routing (requests to the same replica\n\t * maximize cache hits).\n\t * Default: false.\n\t */\n\tsendSessionAffinityHeaders?: boolean;\n\t/**\n\t * Whether the provider supports Anthropic-style `cache_control` markers on\n\t * tool definitions. When false, `cache_control` is omitted from tool params.\n\t * Some Anthropic-compatible providers (e.g., Fireworks) do not support this\n\t * field on tools and may reject or ignore it.\n\t * Default: true.\n\t */\n\tsupportsCacheControlOnTools?: boolean;\n\t/**\n\t * Whether the model accepts the Anthropic `temperature` request field.\n\t * Claude Opus 4.7+ rejects non-default temperature values.\n\t * Default: true.\n\t */\n\tsupportsTemperature?: boolean;\n\t/**\n\t * Whether to force adaptive thinking (`thinking.type: \"adaptive\"` plus\n\t * `output_config.effort`) regardless of the model id. Built-in models that\n\t * require adaptive thinking set this in generated metadata. Custom\n\t * Anthropic-compatible providers can set this to `true` for any model whose\n\t * upstream requires the adaptive format. Set to `false` to\n\t * opt out on overridden built-in models.\n\t * Default: false.\n\t */\n\tforceAdaptiveThinking?: boolean;\n\t/** Whether to replay empty thinking signatures as `signature: \"\"` instead of converting thinking to text. Default: false. */\n\tallowEmptySignature?: boolean;\n\t/** Whether the provider supports Anthropic strict tool schemas. Default: false; generated Anthropic models enable it explicitly. */\n\tsupportsStrictTools?: boolean;\n\t/** Whether the exact model transport supports effort-only system messages and thinking binding controls. Default: false. */\n\tsupportsMidConvoEffort?: boolean;\n\t/**\n\t * Models Anthropic accepts in `fallbacks` for server-side refusal fallback,\n\t * with local pricing metadata for returned fallback responses. When absent or\n\t * empty, callers must omit `fallbacks`; Anthropic rejects the field for models\n\t * with no permitted fallback targets.\n\t */\n\tallowedFallbackModels?: AnthropicAllowedFallbackModel[];\n\t/**\n\t * Whether the provider supports deferred tools loaded by `tool_reference`\n\t * blocks in tool results. Default: true for first-party Anthropic models\n\t * except Haiku and models older than Claude 4.5; false for other providers.\n\t */\n\tsupportsToolReferences?: boolean;\n\t/**\n\t * Whether the model runs Anthropic's preserved-thinking *conversation* check, which\n\t * rejects a request whose `system` prompt, `tools`, or earlier messages changed since a\n\t * replayed thinking block was produced. When true the provider sends the\n\t * `thinking-binding-controls-2026-08-01` beta header and\n\t * `thinking.block_binding.prefix_mismatch_behavior: \"drop_block\"`, so a changed prefix\n\t * drops the affected thinking blocks instead of failing the request with a 400.\n\t * Enforced by default for Anthropic accounts created on or after 2026-08-31.\n\t * Default: false.\n\t */\n\tenforcesPreservedThinkingBinding?: boolean;\n\t/**\n\t * Whether the API adjudicates thinking-block *model* binding itself, always dropping a\n\t * block the target model cannot read. When true, an assistant turn produced by a\n\t * different model on the same provider and API replays its signed `thinking` and\n\t * `redacted_thinking` blocks unchanged rather than being rewritten into plain text, so a\n\t * mid-conversation model switch keeps reasoning the target model is allowed to read.\n\t * Default: false.\n\t */\n\tdelegatesThinkingModelBinding?: boolean;\n\t/**\n\t * Whether the model accepts forced tool use (`tool_choice: {\"type\": \"any\"}` or\n\t * `{\"type\": \"tool\", ...}`). Claude Fable 5.1 and Claude Mythos 5.1 reject it on every\n\t * request with a 400; Anthropic's guidance for those models is to use\n\t * `tool_choice: {\"type\": \"auto\"}` with strict tool use or structured outputs instead.\n\t * When false, the provider rejects a forced choice with an error rather than sending a\n\t * request the model is guaranteed to refuse, or silently substituting a different one.\n\t * `auto` and `none` are never altered.\n\t * https://platform.claude.com/docs/en/build-with-claude/thinking\n\t * Default: true.\n\t */\n\tsupportsForcedToolChoice?: boolean;\n}\n\n/** Compatibility settings for Amazon Bedrock models. */\nexport interface BedrockCompat {\n\t/** Whether the model supports Bedrock strict tool schemas. Default: false. */\n\tsupportsStrictMode?: boolean;\n\t/**\n\t * Whether the model accepts forced tool use (`toolChoice` `\"any\"` or `{ type: \"tool\" }`).\n\t * Claude Fable 5.1 rejects it on every request with a 400, whichever platform serves the\n\t * model. When false, the provider rejects a forced choice with an error rather than sending\n\t * a request that is guaranteed to fail. `auto` and `none` are never altered.\n\t * https://platform.claude.com/docs/en/build-with-claude/thinking\n\t * Default: true.\n\t */\n\tsupportsForcedToolChoice?: boolean;\n\t/**\n\t * Whether the model accepts `inferenceConfig.temperature`. Claude Fable 5.1 rejects\n\t * non-default `temperature`, `top_p`, and `top_k` on every request. When false, the\n\t * provider omits the field.\n\t * Default: true.\n\t */\n\tsupportsTemperature?: boolean;\n}\n\n/**\n * OpenRouter provider routing preferences.\n * Controls which upstream providers OpenRouter routes requests to.\n * Sent as the `provider` field in the OpenRouter API request body.\n * @see https://openrouter.ai/docs/guides/routing/provider-selection\n */\nexport interface OpenRouterRouting {\n\t/** Whether to allow backup providers to serve requests. Default: true. */\n\tallow_fallbacks?: boolean;\n\t/** Whether to filter providers to only those that support all parameters in the request. Default: false. */\n\trequire_parameters?: boolean;\n\t/** Data collection setting. \"allow\" (default): allow providers that may store/train on data. \"deny\": only use providers that don't collect user data. */\n\tdata_collection?: \"deny\" | \"allow\";\n\t/** Whether to restrict routing to only ZDR (Zero Data Retention) endpoints. */\n\tzdr?: boolean;\n\t/** Whether to restrict routing to only models that allow text distillation. */\n\tenforce_distillable_text?: boolean;\n\t/** An ordered list of provider names/slugs to try in sequence, falling back to the next if unavailable. */\n\torder?: string[];\n\t/** List of provider names/slugs to exclusively allow for this request. */\n\tonly?: string[];\n\t/** List of provider names/slugs to skip for this request. */\n\tignore?: string[];\n\t/** A list of quantization levels to filter providers by (e.g., [\"fp16\", \"bf16\", \"fp8\", \"fp6\", \"int8\", \"int4\", \"fp4\", \"fp32\"]). */\n\tquantizations?: string[];\n\t/** Sorting strategy. Can be a string (e.g., \"price\", \"throughput\", \"latency\") or an object with `by` and `partition`. */\n\tsort?:\n\t\t| string\n\t\t| {\n\t\t\t\t/** The sorting metric: \"price\", \"throughput\", \"latency\". */\n\t\t\t\tby?: string;\n\t\t\t\t/** Partitioning strategy: \"model\" (default) or \"none\". */\n\t\t\t\tpartition?: string | null;\n\t\t  };\n\t/** Maximum price per million tokens (USD). */\n\tmax_price?: {\n\t\t/** Price per million prompt tokens. */\n\t\tprompt?: number | string;\n\t\t/** Price per million completion tokens. */\n\t\tcompletion?: number | string;\n\t\t/** Price per image. */\n\t\timage?: number | string;\n\t\t/** Price per audio unit. */\n\t\taudio?: number | string;\n\t\t/** Price per request. */\n\t\trequest?: number | string;\n\t};\n\t/** Preferred minimum throughput (tokens/second). Can be a number (applies to p50) or an object with percentile-specific cutoffs. */\n\tpreferred_min_throughput?:\n\t\t| number\n\t\t| {\n\t\t\t\t/** Minimum tokens/second at the 50th percentile. */\n\t\t\t\tp50?: number;\n\t\t\t\t/** Minimum tokens/second at the 75th percentile. */\n\t\t\t\tp75?: number;\n\t\t\t\t/** Minimum tokens/second at the 90th percentile. */\n\t\t\t\tp90?: number;\n\t\t\t\t/** Minimum tokens/second at the 99th percentile. */\n\t\t\t\tp99?: number;\n\t\t  };\n\t/** Preferred maximum latency (seconds). Can be a number (applies to p50) or an object with percentile-specific cutoffs. */\n\tpreferred_max_latency?:\n\t\t| number\n\t\t| {\n\t\t\t\t/** Maximum latency in seconds at the 50th percentile. */\n\t\t\t\tp50?: number;\n\t\t\t\t/** Maximum latency in seconds at the 75th percentile. */\n\t\t\t\tp75?: number;\n\t\t\t\t/** Maximum latency in seconds at the 90th percentile. */\n\t\t\t\tp90?: number;\n\t\t\t\t/** Maximum latency in seconds at the 99th percentile. */\n\t\t\t\tp99?: number;\n\t\t  };\n}\n\n/**\n * Vercel AI Gateway routing preferences.\n * Controls which upstream providers the gateway routes requests to.\n * @see https://vercel.com/docs/ai-gateway/models-and-providers/provider-options\n */\nexport interface VercelGatewayRouting {\n\t/** List of provider slugs to exclusively use for this request (e.g., [\"bedrock\", \"anthropic\"]). */\n\tonly?: string[];\n\t/** List of provider slugs to try in order (e.g., [\"anthropic\", \"openai\"]). */\n\torder?: string[];\n}\n\nexport interface ModelCostRates {\n\tinput: number; // $/million tokens\n\toutput: number; // $/million tokens\n\tcacheRead: number; // $/million tokens\n\tcacheWrite: number; // $/million tokens\n}\n\nexport interface ModelCostTier extends ModelCostRates {\n\t/** Use this tier for requests whose total input usage exceeds this token count. */\n\tinputTokensAbove: number;\n}\n\nexport interface ModelCost extends ModelCostRates {\n\t/** Request-wide pricing tiers. The highest matching input threshold applies to the full request. */\n\ttiers?: ModelCostTier[];\n}\n\n/**\n * Explicit routing metadata that marks a model as the fast-inference variant of another model.\n *\n * Presence of this field — never a `-fast` name suffix — is what gives a model fast semantics.\n * A provider, `models.json` custom model, or extension may own an exact `<base>-fast` ID without\n * this metadata, in which case it is an ordinary model and routes exactly as it is declared.\n */\nexport interface ModelFastRoute {\n\t/** Canonical ID of the normal-speed model this variant pairs with, in the same provider. */\n\tbaseModelId: string;\n\t/** Model ID to send upstream. OpenAI-style routing keeps the base ID; a provider with real fast siblings sends its own ID. */\n\tupstreamModelId: string;\n\t/** Service tier to send with the request. Set only for providers that route fast traffic through an OpenAI-style tier. */\n\tserviceTier?: \"priority\";\n}\n\n// Model interface for the unified model system\nexport interface Model<TApi extends Api> {\n\tid: string;\n\tname: string;\n\tapi: TApi;\n\tprovider: ProviderId;\n\tbaseUrl: string;\n\treasoning: boolean;\n\t/**\n\t * Maps pi thinking levels to provider/model-specific values.\n\t * Missing keys use provider defaults. null marks a level as unsupported.\n\t */\n\tthinkingLevelMap?: ThinkingLevelMap;\n\t/**\n\t * Modalities Atomic can send to this model. `\"pdf\"` is set only where a runtime can serialize\n\t * a document block — the Anthropic Messages and Amazon Bedrock Converse paths — so a provider\n\t * that publishes PDF support but has no such path here stays at `[\"text\", \"image\"]`.\n\t */\n\tinput: (\"text\" | \"image\" | \"pdf\")[];\n\tcost: ModelCost;\n\tcontextWindow: number;\n\tmaxTokens: number;\n\t/** Default sampling parameters for this model. See {@link StreamOptions.samplingParams}; per-request keys override these. */\n\tsamplingParams?: Record<string, unknown>;\n\theaders?: Record<string, string>;\n\t/**\n\t * Marks this model as the fast-inference variant of {@link ModelFastRoute.baseModelId} and carries the\n\t * upstream routing it needs. Absent on every normal model.\n\t */\n\tfastRoute?: ModelFastRoute;\n\t/** Compatibility overrides for OpenAI-compatible APIs. If not set, auto-detected from baseUrl. */\n\tcompat?: TApi extends \"openai-completions\"\n\t\t? OpenAICompletionsCompat\n\t\t: TApi extends \"openai-responses\" | \"azure-openai-responses\" | \"openai-codex-responses\"\n\t\t\t? OpenAIResponsesCompat\n\t\t\t: TApi extends \"anthropic-messages\"\n\t\t\t\t? AnthropicMessagesCompat\n\t\t\t\t: TApi extends \"bedrock-converse-stream\"\n\t\t\t\t\t? BedrockCompat\n\t\t\t\t\t: never;\n}\n\nexport interface ImagesModel<TApi extends ImagesApi>\n\textends Omit<Model<Api>, \"api\" | \"provider\" | \"reasoning\" | \"contextWindow\" | \"maxTokens\" | \"compat\"> {\n\tapi: TApi;\n\tprovider: ImagesProviderId;\n\toutput: (\"text\" | \"image\")[];\n}\n"]}
+{"version":3,"file":"types.js","sourceRoot":"","sources":["../src/types.ts"],"names":[],"mappings":"","sourcesContent":["import type { TelemetryContext } from \"@earendil-works/pi-telemetry\";\nimport type { AnthropicOptions } from \"./api/anthropic-messages.ts\";\nimport type { AzureOpenAIResponsesOptions } from \"./api/azure-openai-responses.ts\";\nimport type { BedrockOptions } from \"./api/bedrock-converse-stream.ts\";\nimport type { GoogleOptions } from \"./api/google-generative-ai.ts\";\nimport type { GoogleVertexOptions } from \"./api/google-vertex.ts\";\nimport type { MistralOptions } from \"./api/mistral-conversations.ts\";\nimport type { OpenAICodexResponsesOptions } from \"./api/openai-codex-responses.ts\";\nimport type { OpenAICompletionsOptions } from \"./api/openai-completions.ts\";\nimport type { OpenAIResponsesOptions } from \"./api/openai-responses.ts\";\nimport type { PiMessagesOptions } from \"./api/pi-messages.ts\";\nimport type { AssistantMessageDiagnostic } from \"./utils/diagnostics.ts\";\nimport type { AssistantMessageEventStream } from \"./utils/event-stream.ts\";\n\nexport type { AssistantMessageEventStream } from \"./utils/event-stream.ts\";\n\nexport type KnownApi =\n\t| \"openai-completions\"\n\t| \"mistral-conversations\"\n\t| \"openai-responses\"\n\t| \"azure-openai-responses\"\n\t| \"openai-codex-responses\"\n\t| \"anthropic-messages\"\n\t| \"bedrock-converse-stream\"\n\t| \"google-generative-ai\"\n\t| \"google-vertex\"\n\t| \"pi-messages\";\n\nexport type Api = KnownApi | (string & {});\n\nexport type KnownImagesApi = \"openrouter-images\";\n\nexport type ImagesApi = KnownImagesApi | (string & {});\n\nexport type KnownProvider =\n\t| \"amazon-bedrock\"\n\t| \"ant-ling\"\n\t| \"anthropic\"\n\t| \"google\"\n\t| \"google-vertex\"\n\t| \"openai\"\n\t| \"azure-openai-responses\"\n\t| \"openai-codex\"\n\t| \"radius\"\n\t| \"nvidia\"\n\t| \"deepseek\"\n\t| \"github-copilot\"\n\t| \"xai\"\n\t| \"groq\"\n\t| \"cerebras\"\n\t| \"openrouter\"\n\t| \"vercel-ai-gateway\"\n\t| \"zai\"\n\t| \"zai-coding-cn\"\n\t| \"mistral\"\n\t| \"minimax\"\n\t| \"minimax-cn\"\n\t| \"moonshotai\"\n\t| \"moonshotai-cn\"\n\t| \"huggingface\"\n\t| \"fireworks\"\n\t| \"together\"\n\t| \"baseten\"\n\t| \"opencode\"\n\t| \"opencode-go\"\n\t| \"kimi-coding\"\n\t| \"cloudflare-workers-ai\"\n\t| \"cloudflare-ai-gateway\"\n\t| \"qwen-token-plan\"\n\t| \"qwen-token-plan-cn\"\n\t| \"qwen-token-plan-individual\"\n\t| \"xiaomi\"\n\t| \"xiaomi-token-plan-cn\"\n\t| \"xiaomi-token-plan-ams\"\n\t| \"xiaomi-token-plan-sgp\";\nexport type ProviderId = KnownProvider | string;\n\nexport type KnownImagesProvider = \"openrouter\";\n\nexport type ImagesProviderId = KnownImagesProvider | string;\n\nexport type ToolChoice = \"auto\" | \"none\";\nexport type ThinkingLevel = \"minimal\" | \"low\" | \"medium\" | \"high\" | \"xhigh\" | \"max\";\nexport type ModelThinkingLevel = \"off\" | ThinkingLevel;\nexport type ThinkingLevelMap = Partial<Record<ModelThinkingLevel, string | null>>;\nexport type ChatTemplateKwargValue =\n\t| string\n\t| number\n\t| boolean\n\t| null\n\t| {\n\t\t\t$var: \"thinking.enabled\" | \"thinking.effort\" | \"thinking.budget\";\n\t\t\tomitWhenOff?: boolean;\n\t  };\n\n/** Top-level request field used to cap reasoning tokens on OpenAI-compatible servers. */\nexport type ThinkingTokenBudgetField = \"thinking_token_budget\" | \"thinking_budget\" | \"thinking_budget_tokens\";\n\n/** Token budgets for each thinking level (token-based providers only) */\nexport interface ThinkingBudgets {\n\tminimal?: number;\n\tlow?: number;\n\tmedium?: number;\n\thigh?: number;\n}\n\n// Base options all providers share\nexport type CacheRetention = \"none\" | \"short\" | \"long\";\n\nexport type Transport = \"sse\" | \"websocket\" | \"websocket-cached\" | \"auto\";\n\n/** Provider-scoped environment overrides. Values take precedence over process.env. */\nexport type ProviderEnv = Record<string, string>;\nexport type ProviderHeaders = Record<string, string | null>;\nexport type FetchFunction = typeof globalThis.fetch;\nexport type SessionAffinityFormat = \"openai\" | \"openai-nosession\" | \"openrouter\";\n\nexport interface ProviderResponse {\n\tstatus: number;\n\theaders: Record<string, string>;\n}\n\n/** Authentication, HTTP transport, and lifecycle callbacks shared by provider requests. */\nexport interface ProviderRequestOptions<TModel = Model<Api>> {\n\tsignal?: AbortSignal;\n\t/** Explicit parent context for telemetry produced by this logical request. */\n\ttelemetryContext?: TelemetryContext;\n\tapiKey?: string;\n\t/**\n\t * Optional fetch implementation for provider HTTP requests.\n\t * Defaults to `globalThis.fetch`. Provider adapters that cannot inject a custom implementation may reject it.\n\t * This does not affect WebSocket transports.\n\t */\n\tfetch?: FetchFunction;\n\t/**\n\t * Provider-scoped environment values. These take precedence over process.env for\n\t * provider configuration such as regional settings, endpoint placeholders, and\n\t * proxy variables.\n\t */\n\tenv?: ProviderEnv;\n\t/**\n\t * Optional callback for inspecting or replacing provider payloads before sending.\n\t * Return undefined to keep the payload unchanged.\n\t */\n\tonPayload?: (payload: unknown, model: TModel) => unknown | undefined | Promise<unknown | undefined>;\n\t/**\n\t * Optional callback invoked after an HTTP response is received.\n\t */\n\tonResponse?: (response: ProviderResponse, model: TModel) => void | Promise<void>;\n\t/**\n\t * Optional custom HTTP headers to include in API requests.\n\t * Merged with provider defaults; caller values override default headers.\n\t * On AWS Bedrock these are injected via a Smithy `build`-step middleware so\n\t * they are covered by SigV4 signing; reserved headers (`x-amz-*`,\n\t * `authorization`, `host`) are silently ignored to preserve SigV4 / bearer auth.\n\t * A null value suppresses a provider/API default header with the same name.\n\t */\n\theaders?: ProviderHeaders;\n\t/**\n\t * HTTP request timeout in milliseconds for providers/SDKs that support it.\n\t * For example, OpenAI and Anthropic SDK clients default to 10 minutes.\n\t */\n\ttimeoutMs?: number;\n\t/**\n\t * Maximum retry attempts for providers/SDKs that support client-side retries.\n\t * For example, OpenAI and Anthropic SDK clients default to 2.\n\t */\n\tmaxRetries?: number;\n\t/**\n\t * Maximum delay in milliseconds to wait for a retry when the server requests a long wait.\n\t * If the server's requested delay exceeds this value, the request fails immediately\n\t * with an error containing the requested delay, allowing higher-level retry logic\n\t * to handle it with user visibility.\n\t * Default: 60000 (60 seconds). Set to 0 to disable the cap.\n\t */\n\tmaxRetryDelayMs?: number;\n}\n\nexport interface StreamOptions extends ProviderRequestOptions<Model<Api>> {\n\t/**\n\t * Optional callback invoked after an HTTP response is received and before\n\t * its body stream is consumed.\n\t */\n\tonResponse?: (response: ProviderResponse, model: Model<Api>) => void | Promise<void>;\n\ttemperature?: number;\n\t/**\n\t * Arbitrary sampling parameters merged into the request body as-is, after the named request\n\t * fields, so keys here override them. Lets custom OpenAI-compatible servers (llama.cpp, vLLM,\n\t * SGLang, ...) receive parameters pi does not model, e.g. `top_p`, `top_k`, `min_p`,\n\t * `repetition_penalty`. Merged over `Model.samplingParams` per key. Only applied by\n\t * OpenAI-compatible adapters (completions, responses, Azure responses); other APIs ignore it.\n\t */\n\tsamplingParams?: Record<string, unknown>;\n\tmaxTokens?: number;\n\t/**\n\t * Preferred transport for providers that support multiple transports.\n\t * Providers that do not support this option ignore it.\n\t */\n\ttransport?: Transport;\n\t/**\n\t * Prompt cache retention preference. Providers map this to their supported values.\n\t * Default: \"short\".\n\t */\n\tcacheRetention?: CacheRetention;\n\t/**\n\t * Optional session identifier for providers that support session-based caching.\n\t * Providers can use this to enable prompt caching, request routing, or other\n\t * session-aware features. Ignored by providers that don't support it.\n\t */\n\tsessionId?: string;\n\t/**\n\t * WebSocket connect timeout in milliseconds for providers that support\n\t * WebSocket transports. This covers the connection/open handshake only;\n\t * HTTP/SDK idleness uses `timeoutMs`, while gaps between decoded provider\n\t * stream events use `streamDeadlineMs`.\n\t */\n\twebsocketConnectTimeoutMs?: number;\n\t/**\n\t * Maximum idle gap in milliseconds between two provider stream events.\n\t *\n\t * Enforced below the HTTP layer, so a response body that stalls without ever\n\t * rejecting the adapter's async iterator (for example a decompression failure\n\t * that destroys the stream silently, #2553) still settles the attempt as a\n\t * retryable transport error instead of hanging forever. The timer restarts on\n\t * every event, so it never caps total stream duration.\n\t *\n\t * Defaults to 300000 ms. `0` (or any non-positive value) disables it.\n\t */\n\tstreamDeadlineMs?: number;\n\t/**\n\t * Optional metadata to include in API requests.\n\t * Providers extract the fields they understand and ignore the rest.\n\t * For example, Anthropic uses `user_id` for abuse tracking and rate limiting.\n\t */\n\tmetadata?: Record<string, unknown>;\n}\n\nexport type ProviderStreamOptions = StreamOptions & Record<string, unknown>;\n\nexport interface DeferredFetchOptions extends ProviderRequestOptions<Model<Api>> {\n\t/**\n\t * Maximum provider long-poll duration in milliseconds.\n\t * Defaults to 0, which performs one status check.\n\t */\n\twait?: number;\n}\n\n/** Request options for best-effort deferred-response cancellation. */\nexport type DeferredCancelOptions = ProviderRequestOptions<Model<Api>>;\n\n/**\n * Maps known APIs to their full provider-specific stream option types.\n * Type-only imports from API implementation modules are erased at emit, so\n * this is tree-shake safe.\n */\nexport interface ApiOptionsMap {\n\t\"anthropic-messages\": AnthropicOptions;\n\t\"openai-completions\": OpenAICompletionsOptions;\n\t\"openai-responses\": OpenAIResponsesOptions;\n\t\"openai-codex-responses\": OpenAICodexResponsesOptions;\n\t\"azure-openai-responses\": AzureOpenAIResponsesOptions;\n\t\"google-generative-ai\": GoogleOptions;\n\t\"google-vertex\": GoogleVertexOptions;\n\t\"mistral-conversations\": MistralOptions;\n\t\"bedrock-converse-stream\": BedrockOptions;\n\t\"pi-messages\": PiMessagesOptions;\n}\n\n/**\n * Full stream options for an API. Known APIs resolve to their concrete option\n * type; custom API strings fall back to the generic shape.\n */\nexport type ApiStreamOptions<TApi extends Api> = TApi extends keyof ApiOptionsMap\n\t? ApiOptionsMap[TApi]\n\t: StreamOptions & Record<string, unknown>;\n\n/**\n * The uniform stream contract of an API implementation module: every module\n * under `src/api/` exports `stream` and `streamSimple`; capable modules may also\n * export deferred-response methods. Lazy wrappers (`lazyApi()`) and provider\n * factories pass these around as values. This is the untyped dispatch shape;\n * per-API option typing lives on the implementation modules themselves and on\n * `Provider.stream()` via `ApiStreamOptions`.\n */\nexport interface ProviderStreams {\n\tstream(model: Model<Api>, context: Context, options?: StreamOptions): AssistantMessageEventStream;\n\tstreamSimple(model: Model<Api>, context: Context, options?: SimpleStreamOptions): AssistantMessageEventStream;\n\tfetchDeferred?(\n\t\tmodel: Model<Api>,\n\t\thandle: DeferredHandle,\n\t\toptions?: DeferredFetchOptions,\n\t): AssistantMessageEventStream;\n\tcancelDeferred?(model: Model<Api>, handle: DeferredHandle, options?: DeferredCancelOptions): Promise<void>;\n}\n\n/**\n * The uniform contract of an image-generation API implementation module:\n * every image API module under `src/api/` exports exactly `generateImages`,\n * so the module itself satisfies this interface. Lazy wrappers and image\n * provider factories pass these around as values.\n */\nexport interface ProviderImages {\n\tgenerateImages(\n\t\tmodel: ImagesModel<ImagesApi>,\n\t\tcontext: ImagesContext,\n\t\toptions?: ImagesOptions,\n\t): Promise<AssistantImages>;\n}\n\nexport interface ImagesOptions extends ProviderRequestOptions<ImagesModel<ImagesApi>> {\n\t/**\n\t * Optional metadata to include in API requests.\n\t * Providers extract the fields they understand and ignore the rest.\n\t */\n\tmetadata?: Record<string, unknown>;\n}\n\nexport type ProviderImagesOptions = ImagesOptions & Record<string, unknown>;\n\nexport interface AnthropicAllowedFallbackModel {\n\tprovider: ProviderId;\n\tmodel: string;\n\tcost: ModelCost;\n}\n\n// Unified options with reasoning passed to streamSimple() and completeSimple()\nexport interface SimpleStreamOptions extends StreamOptions {\n\t/** Provider-neutral tool selection for simple requests. When omitted, adapters use provider-specific behavior. */\n\ttoolChoice?: ToolChoice;\n\treasoning?: ThinkingLevel;\n\t/** Ask a capable provider to return a durable handle and continue the request asynchronously. */\n\tdeferred?: boolean | { window?: \"15m\" | \"1h\" | \"24h\" };\n\t/** Custom token budgets for thinking levels (token-based providers only) */\n\tthinkingBudgets?: ThinkingBudgets;\n}\n\n// Generic StreamFunction with typed options.\n//\n// Contract:\n// - Must return an AssistantMessageEventStream.\n// - Direct streamSimple() calls may throw synchronously when request auth is\n//   missing. Once a stream is returned, request/model/runtime failures should\n//   be encoded in the stream, not thrown.\n// - Error termination must produce an AssistantMessage with stopReason\n//   \"error\" or \"aborted\" and errorMessage, emitted via the stream protocol.\nexport type StreamFunction<TApi extends Api = Api, TOptions extends StreamOptions = StreamOptions> = (\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: TOptions,\n) => AssistantMessageEventStream;\n\nexport type ImagesFunction<TApi extends ImagesApi = ImagesApi, TOptions extends ImagesOptions = ImagesOptions> = (\n\tmodel: ImagesModel<TApi>,\n\tcontext: ImagesContext,\n\toptions?: TOptions,\n) => Promise<AssistantImages>;\n\nexport interface TextSignatureV1 {\n\tv: 1;\n\tid: string;\n\tphase?: \"commentary\" | \"final_answer\";\n}\n\nexport interface TextContent {\n\ttype: \"text\";\n\ttext: string;\n\ttextSignature?: string; // e.g., for OpenAI responses, message metadata (legacy id string or TextSignatureV1 JSON)\n}\n\nexport interface ThinkingContent {\n\ttype: \"thinking\";\n\tthinking: string;\n\tthinkingSignature?: string; // Provider-specific opaque or serialized reasoning replay data\n\t/** When true, the thinking content was redacted by safety filters. The opaque\n\t *  encrypted payload is stored in `thinkingSignature` so it can be passed back\n\t *  to the API for multi-turn continuity. */\n\tredacted?: boolean;\n}\n\n/**\n * Boundary marker emitted by Anthropic server-side fallback, where one model's output gives way\n * to the next after a classifier refusal.\n *\n * This is public content rather than a stream-only event because it must survive the round trip:\n * \"Keep it exactly where it appeared. The API uses its position to validate the thinking blocks\n * around it, so a request that echoes thinking blocks from both sides of the boundary is rejected\n * if the block is omitted or moved.\"\n * https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback\n */\nexport interface FallbackContent {\n\ttype: \"fallback\";\n\t/** The model that declined. Echoes the model string sent when the declining hop is the request's own model. */\n\tfromModel: string;\n\t/** Resolved id of the model that continues. Always present. */\n\ttoModel: string;\n}\n\nexport interface ImageContent {\n\ttype: \"image\";\n\tdata: string; // base64 encoded image data\n\tmimeType: string; // e.g., \"image/jpeg\", \"image/png\"\n}\n\n/**\n * A document sent as model input, currently PDF only.\n *\n * Anthropic documents PDF as a platform capability — \"All active models support PDF processing\" —\n * routed through the same vision path as images, so it is not a per-model modality. A model\n * advertises it through `Model.input` containing `\"pdf\"`, which generated metadata sets only for\n * the runtimes that can actually serialize a document block. Providers that cannot are sent a\n * visible placeholder instead, exactly as they are for images they cannot accept.\n * https://platform.claude.com/docs/en/build-with-claude/pdf-support\n */\nexport interface DocumentContent {\n\ttype: \"document\";\n\t/** Base64-encoded document bytes. Amazon Bedrock takes raw bytes and decodes this itself. */\n\tdata: string;\n\t/**\n\t * The only media type either serializer implements. It is a literal rather than `string`\n\t * because both paths hardcode PDF — the Anthropic block emits `BetaBase64PDFSource`, whose\n\t * `media_type` is itself a fixed literal, and Bedrock emits `DocumentFormat.PDF` — so a wider\n\t * type would let a caller label arbitrary bytes as a PDF. Widening this union later is\n\t * non-breaking; narrowing it after release would not be. Note that adding a format is not a\n\t * media-type swap on the Anthropic side: plain text needs a structurally different source\n\t * variant (`BetaPlainTextSource`, `type: \"text\"`), and Bedrock's `doc`/`docx`/`xls`/`xlsx`/\n\t * `csv`/`html`/`md` formats have no Anthropic source variant at all.\n\t */\n\tmimeType: \"application/pdf\";\n\t/**\n\t * Optional caller-supplied label. Amazon Bedrock requires a name and warns that it \"is\n\t * vulnerable to prompt injections\", so the Bedrock path sanitizes this to the characters AWS\n\t * permits and falls back to a neutral generated name.\n\t */\n\tname?: string;\n}\n\nexport interface ToolCall {\n\ttype: \"toolCall\";\n\tid: string;\n\tname: string;\n\targuments: Record<string, any>;\n\tthoughtSignature?: string; // Google-specific: opaque signature for reusing thought context\n\t/** OpenAI Responses namespace for calls to dynamically loaded or namespaced tools. */\n\tnamespace?: string;\n}\n\nexport interface Usage {\n\tinput: number;\n\toutput: number;\n\tcacheRead: number;\n\tcacheWrite: number;\n\t/** Subset of `cacheWrite` written with 1h retention. Only Anthropic reports this split. */\n\tcacheWrite1h?: number;\n\t/**\n\t * Reasoning/thinking tokens, when the provider reports them. This is a subset of\n\t * `output`: `output` already includes these tokens. Set to a number (possibly 0) by\n\t * providers that expose a reasoning breakdown; left undefined by providers that don't.\n\t */\n\treasoning?: number;\n\ttotalTokens: number;\n\tcost: {\n\t\tinput: number;\n\t\toutput: number;\n\t\tcacheRead: number;\n\t\tcacheWrite: number;\n\t\ttotal: number;\n\t};\n}\n\nexport type StopReason = \"pending\" | \"stop\" | \"length\" | \"toolUse\" | \"error\" | \"aborted\" | \"deferred\";\n\nexport type JsonValue = string | number | boolean | null | JsonValue[] | { [key: string]: JsonValue };\n\nexport interface DeferredHandle {\n\tprovider: string;\n\tmodelId: string;\n\tapi: string;\n\t/** Provider token, such as a response id or batch id plus row id. */\n\tid: string;\n\texpiresAt?: number;\n\tpollAfterMs?: number;\n\t/** Provider conversion data required to reconstruct the final assistant message. */\n\tdata?: JsonValue;\n}\n\nexport interface UserMessage {\n\trole: \"user\";\n\tcontent: string | (TextContent | ImageContent | DocumentContent)[];\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport interface AssistantMessage {\n\trole: \"assistant\";\n\tcontent: (TextContent | ThinkingContent | ToolCall | FallbackContent)[];\n\tapi: Api;\n\tprovider: ProviderId;\n\tmodel: string;\n\tresponseModel?: string; // Concrete `chunk.model` when different from the requested `model` (e.g. OpenRouter `auto` -> `anthropic/...`)\n\tresponseId?: string; // Provider-specific response/message identifier when the upstream API exposes one\n\t/** Exact provider-native effort level used for this response. Absent for legacy or unmanaged responses. */\n\tproviderThinkingLevel?: string;\n\tdiagnostics?: AssistantMessageDiagnostic[]; // Redacted provider/runtime diagnostics for failures and recoveries.\n\tusage: Usage;\n\tstopReason: StopReason;\n\tdeferred?: DeferredHandle;\n\terrorMessage?: string;\n\trawStopReason?: string;\n\t/**\n\t * Provider indication of whether the model explicitly ended its turn.\n\t * Preserved for debugging and does not currently affect agent control flow.\n\t */\n\tendTurn?: boolean;\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport interface ToolResultMessage<TDetails = any> {\n\trole: \"toolResult\";\n\ttoolCallId: string;\n\ttoolName: string;\n\tcontent: (TextContent | ImageContent)[]; // Supports text and images\n\tdetails?: TDetails;\n\t/** Usage from the tool execution itself, if available. Not part of main LLM context accounting. */\n\tusage?: Usage;\n\t/**\n\t * Names from `Context.tools` that became available after this result.\n\t * Providers with native deferred tool loading use this as the load point;\n\t * other providers ignore it and use `Context.tools` normally.\n\t */\n\taddedToolNames?: string[];\n\tisError: boolean;\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport type Message = UserMessage | AssistantMessage | ToolResultMessage;\n\nexport type ImagesInputContent = TextContent | ImageContent;\nexport type ImagesOutputContent = TextContent | ImageContent;\n\nexport interface ImagesContext {\n\tinput: ImagesInputContent[];\n}\n\nexport type ImagesStopReason = \"stop\" | \"error\" | \"aborted\";\n\nexport interface AssistantImages {\n\tapi: ImagesApi;\n\tprovider: ImagesProviderId;\n\tmodel: string;\n\toutput: ImagesOutputContent[];\n\tresponseId?: string;\n\tusage?: Usage;\n\tstopReason: ImagesStopReason;\n\terrorMessage?: string;\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nimport type { TSchema } from \"typebox\";\n\n/** OpenAI grammar variants for constrained sampling. */\nexport type GrammarFormat = \"openai_lark\" | \"openai_regex\";\n\nexport type GrammarVariants = Partial<Record<GrammarFormat, string>>;\n\n/**\n * Optional provider-side constrained sampling configs for a tool.\n *\n * The `json_schema` value roughly maps to the concept of `strict` in APIs which is\n * implemented as json-schema constrained sampling by APIs. Grammar variants let\n * callers provide provider-specific encodings of the same intended language.\n */\nexport type ConstrainedSamplingConfig =\n\t| {\n\t\t\ttype: \"json_schema\";\n\t\t\tstrict: \"prefer\" | \"require\";\n\t  }\n\t| {\n\t\t\ttype: \"grammar\";\n\t\t\tvariants: GrammarVariants;\n\t  };\n\nexport interface Tool<TParameters extends TSchema = TSchema> {\n\tname: string;\n\tdescription: string;\n\tparameters: TParameters;\n\tconstrainedSampling?: false | ConstrainedSamplingConfig;\n}\n\nexport interface Context {\n\tsystemPrompt?: string;\n\tmessages: Message[];\n\ttools?: Tool[];\n}\n\n/**\n * Event protocol for AssistantMessageEventStream.\n *\n * Successful streams emit `start` before partial updates and terminate with\n * `done`. A stream may terminate directly with `error` when request setup fails\n * before generation starts; after `start`, failures also terminate with `error`.\n * Direct `streamSimple()` calls throw synchronously when request auth is missing.\n * Updates and `done` must never appear before `start`.\n *\n * Tool-call arguments at `toolcall_start` are provider-specific;\n * `toolcall_delta` carries subsequent JSON updates.\n */\nexport type AssistantMessageEvent =\n\t| { type: \"start\"; partial: AssistantMessage }\n\t| { type: \"text_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"text_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| { type: \"text_end\"; contentIndex: number; content: string; partial: AssistantMessage }\n\t| { type: \"thinking_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"thinking_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| { type: \"thinking_end\"; contentIndex: number; content: string; partial: AssistantMessage }\n\t| { type: \"toolcall_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"toolcall_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| { type: \"toolcall_end\"; contentIndex: number; toolCall: ToolCall; partial: AssistantMessage }\n\t| {\n\t\t\ttype: \"done\";\n\t\t\treason: Extract<StopReason, \"stop\" | \"length\" | \"toolUse\" | \"deferred\">;\n\t\t\tmessage: AssistantMessage;\n\t  }\n\t| { type: \"error\"; reason: Extract<StopReason, \"aborted\" | \"error\">; error: AssistantMessage };\n\n/**\n * Compatibility settings for OpenAI-compatible completions APIs.\n * Use this to override URL-based auto-detection for custom providers.\n */\nexport interface OpenAICompletionsCompat {\n\t/** Whether the provider supports the `store` field. Default: auto-detected from URL. */\n\tsupportsStore?: boolean;\n\t/** Whether the provider supports the `developer` role (vs `system`). Default: auto-detected from URL. */\n\tsupportsDeveloperRole?: boolean;\n\t/** Whether the provider supports `reasoning_effort`. Default: auto-detected from URL. */\n\tsupportsReasoningEffort?: boolean;\n\t/**\n\t * Whether the model accepts the `temperature` request field. Claude Fable 5.1 rejects\n\t * non-default `temperature`, `top_p`, and `top_k` on every request, and OpenRouter's own\n\t * `supported_parameters` for that model omits `temperature`. When false, the provider\n\t * omits `temperature` and also strips `temperature`, `top_p`, and `top_k` from\n\t * `samplingParams`, which is otherwise merged last and would reopen them.\n\t * Default: true.\n\t */\n\tsupportsTemperature?: boolean;\n\t/**\n\t * Whether the model accepts forced tool use (`tool_choice` `\"required\"` or\n\t * `{ type: \"function\", ... }`). Claude Fable 5.1 rejects it on every request, whichever\n\t * platform serves the model. When false, the provider rejects a forced choice with an error\n\t * rather than sending a request the model cannot honor. `\"auto\"` and `\"none\"` are never\n\t * altered. https://platform.claude.com/docs/en/build-with-claude/thinking\n\t * Default: true.\n\t */\n\tsupportsForcedToolChoice?: boolean;\n\t/** Whether the provider supports `stream_options: { include_usage: true }` for token usage in streaming responses. Default: true. */\n\tsupportsUsageInStreaming?: boolean;\n\t/** Whether streamed responses include `finish_reason`. When false, pi infers `stop` or `toolUse` when the stream ends. Default: true. */\n\tsupportsFinishReason?: boolean;\n\t/** Which field to use for max tokens. Default: auto-detected from URL. */\n\tmaxTokensField?: \"max_completion_tokens\" | \"max_tokens\";\n\t/** Whether tool results require the `name` field. Default: auto-detected from URL. */\n\trequiresToolResultName?: boolean;\n\t/** Whether a user message after tool results requires an assistant message in between. Default: auto-detected from URL. */\n\trequiresAssistantAfterToolResult?: boolean;\n\t/** Whether thinking blocks must be converted to text blocks with <thinking> delimiters. Default: auto-detected from URL. */\n\trequiresThinkingAsText?: boolean;\n\t/** Whether all replayed assistant messages must include an empty reasoning_content field when reasoning is enabled. Default: auto-detected from URL. */\n\trequiresReasoningContentOnAssistantMessages?: boolean;\n\t/** 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\". */\n\tthinkingFormat?:\n\t\t| \"openai\"\n\t\t| \"openrouter\"\n\t\t| \"deepseek\"\n\t\t| \"together\"\n\t\t| \"baseten\"\n\t\t| \"zai\"\n\t\t| \"qwen\"\n\t\t| \"chat-template\"\n\t\t| \"qwen-chat-template\"\n\t\t| \"string-thinking\"\n\t\t| \"ant-ling\";\n\t/** Kwargs to send as `chat_template_kwargs` when `thinkingFormat` is `chat-template`. Use `{ \"$var\": \"thinking.enabled\" }`, `{ \"$var\": \"thinking.effort\" }`, or `{ \"$var\": \"thinking.budget\" }` for pi-controlled thinking values. */\n\tchatTemplateKwargs?: Record<string, ChatTemplateKwargValue>;\n\t/** Arguments to send as `chat_template_args` when `thinkingFormat` is `baseten`. Use `{ \"$var\": \"thinking.enabled\" }`, `{ \"$var\": \"thinking.effort\" }`, or `{ \"$var\": \"thinking.budget\" }` for pi-controlled thinking values. */\n\tchatTemplateArgs?: Record<string, ChatTemplateKwargValue>;\n\t/** OpenRouter-compatible routing preferences sent as the `provider` request field. */\n\topenRouterRouting?: OpenRouterRouting;\n\t/** Vercel AI Gateway routing preferences. Only used when baseUrl points to Vercel AI Gateway. */\n\tvercelGatewayRouting?: VercelGatewayRouting;\n\t/** Whether z.ai supports top-level `tool_stream: true` for streaming tool call deltas. Default: false. */\n\tzaiToolStream?: boolean;\n\t/**\n\t * Top-level request field used to cap reasoning tokens from `thinkingBudgets`.\n\t * Reasoning and the answer share `max_tokens` on these endpoints, so without a budget a\n\t * reasoning-heavy turn can consume the whole response and emit no answer.\n\t * `\"thinking_token_budget\"` is vLLM, `\"thinking_budget\"` is Qwen/DashScope/SGLang,\n\t * `\"thinking_budget_tokens\"` is llama.cpp. Off by default; not set on the generated catalog.\n\t */\n\tthinkingTokenBudgetField?: ThinkingTokenBudgetField;\n\t/** Alias for `thinkingTokenBudgetField: \"thinking_token_budget\"` (vLLM). Prefer `thinkingTokenBudgetField`. Default: false. */\n\tsupportsThinkingTokenBudget?: boolean;\n\t/** 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. */\n\tsupportsOpenAIGrammarTools?: boolean;\n\t/** Whether the provider supports the `strict` field in tool definitions. Default: true. */\n\tsupportsStrictMode?: boolean;\n\t/** 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. */\n\tcacheControlFormat?: \"anthropic\";\n\t/** Whether to send session-affinity data from `options.sessionId`. Default: true for OpenRouter endpoints, false otherwise. */\n\tsendSessionAffinityHeaders?: boolean;\n\t/** Provider-specific deferred tool serialization mode. */\n\tdeferredToolsMode?: \"kimi\";\n\t/** 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. */\n\tsessionAffinityFormat?: SessionAffinityFormat;\n\t/** Whether the provider supports long prompt cache retention (`prompt_cache_retention: \"24h\"` or Anthropic-style `cache_control.ttl: \"1h\"`, depending on format). Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n\t/**\n\t * vLLM scheduler priority sent as the top-level `priority` request field (lower values are\n\t * handled earlier; server default 0). Only meaningful when vLLM runs with\n\t * `--scheduling-policy priority`; useful for keeping background/batch work from stalling\n\t * interactive sessions. Off by default; not set on the generated catalog.\n\t */\n\tvllmPriority?: number;\n}\n\n/** Compatibility settings for OpenAI Responses APIs. */\nexport interface OpenAIResponsesCompat {\n\t/** Whether the provider supports the `developer` role (vs `system`). Default: true. */\n\tsupportsDeveloperRole?: boolean;\n\t/** 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. */\n\tsessionAffinityFormat?: SessionAffinityFormat;\n\t/** 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. */\n\tsupportsLongCacheRetention?: boolean;\n\t/** Whether the provider supports strict JSON-schema function tools. Defaults are API-specific; generated OpenAI models enable it explicitly. */\n\tsupportsStrictMode?: boolean;\n\t/** 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. */\n\tsupportsOpenAIGrammarTools?: boolean;\n\t/** Whether the model supports message-anchored `additional_tools` input items. Default: false. */\n\tsupportsAdditionalTools?: boolean;\n\t/** Whether the model supports client-executed tool search for deferred tools. Default: false. */\n\tsupportsToolSearch?: boolean;\n\t/** Whether the model accepts `prompt_cache_options` (OpenAI GPT-5.6+ prompt caching). Older OpenAI models reject the parameter. Default: false. */\n\tsupportsExplicitPromptCacheMode?: boolean;\n\t/** Whether the provider accepts the `max_output_tokens` parameter. Some Codex-protocol gateways reject it. Default: true. */\n\tsupportsMaxOutputTokens?: boolean;\n}\n\n/** Compatibility settings for Anthropic Messages-compatible APIs. */\nexport interface AnthropicMessagesCompat {\n\t/**\n\t * Whether the provider accepts per-tool `eager_input_streaming`.\n\t * When false, the Anthropic provider omits `tools[].eager_input_streaming`\n\t * and sends the legacy `fine-grained-tool-streaming-2025-05-14` beta header\n\t * for tool-enabled requests.\n\t * Default: true.\n\t */\n\tsupportsEagerToolInputStreaming?: boolean;\n\t/** Whether the provider supports Anthropic long cache retention (`cache_control.ttl: \"1h\"`). Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n\t/**\n\t * Whether to send a session-affinity header from `options.sessionId`\n\t * when caching is enabled. Required for providers like Fireworks that use\n\t * session affinity for prompt cache routing (requests to the same replica\n\t * maximize cache hits).\n\t * Default: true for OpenRouter endpoints, false otherwise.\n\t */\n\tsendSessionAffinityHeaders?: boolean;\n\t/** Session-affinity format. `\"openrouter\"` sends `x-session-id`; defaults to that format for OpenRouter endpoints and `x-session-affinity` elsewhere. */\n\tsessionAffinityFormat?: \"openrouter\";\n\t/**\n\t * Whether the provider supports Anthropic-style `cache_control` markers on\n\t * tool definitions. When false, `cache_control` is omitted from tool params.\n\t * Some Anthropic-compatible providers (e.g., Fireworks) do not support this\n\t * field on tools and may reject or ignore it.\n\t * Default: true.\n\t */\n\tsupportsCacheControlOnTools?: boolean;\n\t/**\n\t * Whether the model accepts the Anthropic `temperature` request field.\n\t * Claude Opus 4.7+ rejects non-default temperature values.\n\t * Default: true.\n\t */\n\tsupportsTemperature?: boolean;\n\t/**\n\t * Whether to force adaptive thinking (`thinking.type: \"adaptive\"` plus\n\t * `output_config.effort`) regardless of the model id. Built-in models that\n\t * require adaptive thinking set this in generated metadata. Custom\n\t * Anthropic-compatible providers can set this to `true` for any model whose\n\t * upstream requires the adaptive format. Set to `false` to\n\t * opt out on overridden built-in models.\n\t * Default: false.\n\t */\n\tforceAdaptiveThinking?: boolean;\n\t/** Whether to replay empty thinking signatures as `signature: \"\"` instead of converting thinking to text. Default: false. */\n\tallowEmptySignature?: boolean;\n\t/** Whether the provider supports Anthropic strict tool schemas. Default: false; generated Anthropic models enable it explicitly. */\n\tsupportsStrictTools?: boolean;\n\t/** Whether the exact model transport supports effort-only system messages and thinking binding controls. Default: false. */\n\tsupportsMidConvoEffort?: boolean;\n\t/**\n\t * Models Anthropic accepts in `fallbacks` for server-side refusal fallback,\n\t * with local pricing metadata for returned fallback responses. When absent or\n\t * empty, callers must omit `fallbacks`; Anthropic rejects the field for models\n\t * with no permitted fallback targets.\n\t */\n\tallowedFallbackModels?: AnthropicAllowedFallbackModel[];\n\t/**\n\t * Whether the provider supports deferred tools loaded by `tool_reference`\n\t * blocks in tool results. Default: true for first-party Anthropic models\n\t * except Haiku and models older than Claude 4.5; false for other providers.\n\t */\n\tsupportsToolReferences?: boolean;\n\t/**\n\t * Whether the model runs Anthropic's preserved-thinking *conversation* check, which\n\t * rejects a request whose `system` prompt, `tools`, or earlier messages changed since a\n\t * replayed thinking block was produced. When true the provider sends the\n\t * `thinking-binding-controls-2026-08-01` beta header and\n\t * `thinking.block_binding.prefix_mismatch_behavior: \"drop_block\"`, so a changed prefix\n\t * drops the affected thinking blocks instead of failing the request with a 400.\n\t * Enforced by default for Anthropic accounts created on or after 2026-08-31.\n\t * Default: false.\n\t */\n\tenforcesPreservedThinkingBinding?: boolean;\n\t/**\n\t * Whether the API adjudicates thinking-block *model* binding itself, always dropping a\n\t * block the target model cannot read. When true, an assistant turn produced by a\n\t * different model on the same provider and API replays its signed `thinking` and\n\t * `redacted_thinking` blocks unchanged rather than being rewritten into plain text, so a\n\t * mid-conversation model switch keeps reasoning the target model is allowed to read.\n\t * Default: false.\n\t */\n\tdelegatesThinkingModelBinding?: boolean;\n\t/**\n\t * Whether the model accepts forced tool use (`tool_choice: {\"type\": \"any\"}` or\n\t * `{\"type\": \"tool\", ...}`). Claude Fable 5.1 and Claude Mythos 5.1 reject it on every\n\t * request with a 400; Anthropic's guidance for those models is to use\n\t * `tool_choice: {\"type\": \"auto\"}` with strict tool use or structured outputs instead.\n\t * When false, the provider rejects a forced choice with an error rather than sending a\n\t * request the model is guaranteed to refuse, or silently substituting a different one.\n\t * `auto` and `none` are never altered.\n\t * https://platform.claude.com/docs/en/build-with-claude/thinking\n\t * Default: true.\n\t */\n\tsupportsForcedToolChoice?: boolean;\n}\n\n/** Compatibility settings for Amazon Bedrock models. */\nexport interface BedrockCompat {\n\t/** Whether the model supports Bedrock strict tool schemas. Default: false. */\n\tsupportsStrictMode?: boolean;\n\t/**\n\t * Whether the model accepts forced tool use (`toolChoice` `\"any\"` or `{ type: \"tool\" }`).\n\t * Claude Fable 5.1 rejects it on every request with a 400, whichever platform serves the\n\t * model. When false, the provider rejects a forced choice with an error rather than sending\n\t * a request that is guaranteed to fail. `auto` and `none` are never altered.\n\t * https://platform.claude.com/docs/en/build-with-claude/thinking\n\t * Default: true.\n\t */\n\tsupportsForcedToolChoice?: boolean;\n\t/**\n\t * Whether the model accepts `inferenceConfig.temperature`. Claude Fable 5.1 rejects\n\t * non-default `temperature`, `top_p`, and `top_k` on every request. When false, the\n\t * provider omits the field.\n\t * Default: true.\n\t */\n\tsupportsTemperature?: boolean;\n}\n\n/**\n * OpenRouter provider routing preferences.\n * Controls which upstream providers OpenRouter routes requests to.\n * Sent as the `provider` field in the OpenRouter API request body.\n * @see https://openrouter.ai/docs/guides/routing/provider-selection\n */\nexport interface OpenRouterRouting {\n\t/** Whether to allow backup providers to serve requests. Default: true. */\n\tallow_fallbacks?: boolean;\n\t/** Whether to filter providers to only those that support all parameters in the request. Default: false. */\n\trequire_parameters?: boolean;\n\t/** Data collection setting. \"allow\" (default): allow providers that may store/train on data. \"deny\": only use providers that don't collect user data. */\n\tdata_collection?: \"deny\" | \"allow\";\n\t/** Whether to restrict routing to only ZDR (Zero Data Retention) endpoints. */\n\tzdr?: boolean;\n\t/** Whether to restrict routing to only models that allow text distillation. */\n\tenforce_distillable_text?: boolean;\n\t/** An ordered list of provider names/slugs to try in sequence, falling back to the next if unavailable. */\n\torder?: string[];\n\t/** List of provider names/slugs to exclusively allow for this request. */\n\tonly?: string[];\n\t/** List of provider names/slugs to skip for this request. */\n\tignore?: string[];\n\t/** A list of quantization levels to filter providers by (e.g., [\"fp16\", \"bf16\", \"fp8\", \"fp6\", \"int8\", \"int4\", \"fp4\", \"fp32\"]). */\n\tquantizations?: string[];\n\t/** Sorting strategy. Can be a string (e.g., \"price\", \"throughput\", \"latency\") or an object with `by` and `partition`. */\n\tsort?:\n\t\t| string\n\t\t| {\n\t\t\t\t/** The sorting metric: \"price\", \"throughput\", \"latency\". */\n\t\t\t\tby?: string;\n\t\t\t\t/** Partitioning strategy: \"model\" (default) or \"none\". */\n\t\t\t\tpartition?: string | null;\n\t\t  };\n\t/** Maximum price per million tokens (USD). */\n\tmax_price?: {\n\t\t/** Price per million prompt tokens. */\n\t\tprompt?: number | string;\n\t\t/** Price per million completion tokens. */\n\t\tcompletion?: number | string;\n\t\t/** Price per image. */\n\t\timage?: number | string;\n\t\t/** Price per audio unit. */\n\t\taudio?: number | string;\n\t\t/** Price per request. */\n\t\trequest?: number | string;\n\t};\n\t/** Preferred minimum throughput (tokens/second). Can be a number (applies to p50) or an object with percentile-specific cutoffs. */\n\tpreferred_min_throughput?:\n\t\t| number\n\t\t| {\n\t\t\t\t/** Minimum tokens/second at the 50th percentile. */\n\t\t\t\tp50?: number;\n\t\t\t\t/** Minimum tokens/second at the 75th percentile. */\n\t\t\t\tp75?: number;\n\t\t\t\t/** Minimum tokens/second at the 90th percentile. */\n\t\t\t\tp90?: number;\n\t\t\t\t/** Minimum tokens/second at the 99th percentile. */\n\t\t\t\tp99?: number;\n\t\t  };\n\t/** Preferred maximum latency (seconds). Can be a number (applies to p50) or an object with percentile-specific cutoffs. */\n\tpreferred_max_latency?:\n\t\t| number\n\t\t| {\n\t\t\t\t/** Maximum latency in seconds at the 50th percentile. */\n\t\t\t\tp50?: number;\n\t\t\t\t/** Maximum latency in seconds at the 75th percentile. */\n\t\t\t\tp75?: number;\n\t\t\t\t/** Maximum latency in seconds at the 90th percentile. */\n\t\t\t\tp90?: number;\n\t\t\t\t/** Maximum latency in seconds at the 99th percentile. */\n\t\t\t\tp99?: number;\n\t\t  };\n}\n\n/**\n * Vercel AI Gateway routing preferences.\n * Controls which upstream providers the gateway routes requests to.\n * @see https://vercel.com/docs/ai-gateway/models-and-providers/provider-options\n */\nexport interface VercelGatewayRouting {\n\t/** List of provider slugs to exclusively use for this request (e.g., [\"bedrock\", \"anthropic\"]). */\n\tonly?: string[];\n\t/** List of provider slugs to try in order (e.g., [\"anthropic\", \"openai\"]). */\n\torder?: string[];\n}\n\nexport interface ModelCostRates {\n\tinput: number; // $/million tokens\n\toutput: number; // $/million tokens\n\tcacheRead: number; // $/million tokens\n\tcacheWrite: number; // $/million tokens\n}\n\nexport interface ModelCostTier extends ModelCostRates {\n\t/** Use this tier for requests whose total input usage exceeds this token count. */\n\tinputTokensAbove: number;\n}\n\nexport interface ModelCost extends ModelCostRates {\n\t/** Request-wide pricing tiers. The highest matching input threshold applies to the full request. */\n\ttiers?: ModelCostTier[];\n}\n\n/**\n * Explicit routing metadata that marks a model as the fast-inference variant of another model.\n *\n * Presence of this field — never a `-fast` name suffix — is what gives a model fast semantics.\n * A provider, `models.json` custom model, or extension may own an exact `<base>-fast` ID without\n * this metadata, in which case it is an ordinary model and routes exactly as it is declared.\n */\nexport interface ModelFastRoute {\n\t/** Canonical ID of the normal-speed model this variant pairs with, in the same provider. */\n\tbaseModelId: string;\n\t/** Model ID to send upstream. OpenAI-style routing keeps the base ID; a provider with real fast siblings sends its own ID. */\n\tupstreamModelId: string;\n\t/** Service tier to send with the request. Set only for providers that route fast traffic through an OpenAI-style tier. */\n\tserviceTier?: \"priority\";\n}\n\n// Model interface for the unified model system\nexport interface Model<TApi extends Api> {\n\tid: string;\n\tname: string;\n\tapi: TApi;\n\tprovider: ProviderId;\n\tbaseUrl: string;\n\treasoning: boolean;\n\t/**\n\t * Maps pi thinking levels to provider/model-specific values.\n\t * Missing keys use provider defaults. null marks a level as unsupported.\n\t */\n\tthinkingLevelMap?: ThinkingLevelMap;\n\t/**\n\t * Modalities Atomic can send to this model. `\"pdf\"` is set only where a runtime can serialize\n\t * a document block — the Anthropic Messages and Amazon Bedrock Converse paths — so a provider\n\t * that publishes PDF support but has no such path here stays at `[\"text\", \"image\"]`.\n\t */\n\tinput: (\"text\" | \"image\" | \"pdf\")[];\n\tcost: ModelCost;\n\tcontextWindow: number;\n\tmaxTokens: number;\n\t/** Default sampling parameters for this model. See {@link StreamOptions.samplingParams}; per-request keys override these. */\n\tsamplingParams?: Record<string, unknown>;\n\theaders?: Record<string, string>;\n\t/**\n\t * Marks this model as the fast-inference variant of {@link ModelFastRoute.baseModelId} and carries the\n\t * upstream routing it needs. Absent on every normal model.\n\t */\n\tfastRoute?: ModelFastRoute;\n\t/** Compatibility overrides for OpenAI-compatible APIs. If not set, auto-detected from baseUrl. */\n\tcompat?: TApi extends \"openai-completions\"\n\t\t? OpenAICompletionsCompat\n\t\t: TApi extends \"openai-responses\" | \"azure-openai-responses\" | \"openai-codex-responses\"\n\t\t\t? OpenAIResponsesCompat\n\t\t\t: TApi extends \"anthropic-messages\"\n\t\t\t\t? AnthropicMessagesCompat\n\t\t\t\t: TApi extends \"bedrock-converse-stream\"\n\t\t\t\t\t? BedrockCompat\n\t\t\t\t\t: never;\n}\n\nexport interface ImagesModel<TApi extends ImagesApi>\n\textends Omit<Model<Api>, \"api\" | \"provider\" | \"reasoning\" | \"contextWindow\" | \"maxTokens\" | \"compat\"> {\n\tapi: TApi;\n\tprovider: ImagesProviderId;\n\toutput: (\"text\" | \"image\")[];\n}\n"]}

package/dist/utils/event-stream.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"event-stream.d.ts","sourceRoot":"","sources":["../../src/utils/event-stream.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,qBAAqB,EAAE,MAAM,aAAa,CAAC;AAG3E,qBAAa,WAAW,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,CAAE,YAAW,aAAa,CAAC,CAAC,CAAC;IAC7D,OAAO,CAAC,KAAK,CAAW;IACxB,OAAO,CAAC,OAAO,CAA8C;IAC7D,OAAO,CAAC,IAAI,CAAS;IACrB,OAAO,CAAC,kBAAkB,CAAa;IACvC,OAAO,CAAC,kBAAkB,CAAuB;IACjD,OAAO,CAAC,UAAU,CAAwB;IAC1C,OAAO,CAAC,aAAa,CAAkB;IAEvC,YAAY,UAAU,EAAE,CAAC,KAAK,EAAE,CAAC,KAAK,OAAO,EAAE,aAAa,EAAE,CAAC,KAAK,EAAE,CAAC,KAAK,CAAC,EAM5E;IAED,IAAI,CAAC,KAAK,EAAE,CAAC,GAAG,IAAI,CAenB;IAED,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC,GAAG,IAAI,CAUpB;IAEM,CAAC,MAAM,CAAC,aAAa,CAAC,IAAI,aAAa,CAAC,CAAC,CAAC,CAYhD;IAED,MAAM,IAAI,OAAO,CAAC,CAAC,CAAC,CAEnB;CACD;AAED,qBAAa,2BAA4B,SAAQ,WAAW,CAAC,qBAAqB,EAAE,gBAAgB,CAAC;IACpG,cAYC;CACD;AAED,+EAA+E;AAC/E,wBAAgB,iCAAiC,IAAI,2BAA2B,CAE/E","sourcesContent":["import type { AssistantMessage, AssistantMessageEvent } from \"../types.ts\";\n\n// Generic event stream class for async iteration\nexport class EventStream<T, R = T> implements AsyncIterable<T> {\n\tprivate queue: T[] = [];\n\tprivate waiting: ((value: IteratorResult<T>) => void)[] = [];\n\tprivate done = false;\n\tprivate finalResultPromise: Promise<R>;\n\tprivate resolveFinalResult!: (result: R) => void;\n\tprivate isComplete: (event: T) => boolean;\n\tprivate extractResult: (event: T) => R;\n\n\tconstructor(isComplete: (event: T) => boolean, extractResult: (event: T) => R) {\n\t\tthis.isComplete = isComplete;\n\t\tthis.extractResult = extractResult;\n\t\tthis.finalResultPromise = new Promise((resolve) => {\n\t\t\tthis.resolveFinalResult = resolve;\n\t\t});\n\t}\n\n\tpush(event: T): void {\n\t\tif (this.done) return;\n\n\t\tif (this.isComplete(event)) {\n\t\t\tthis.done = true;\n\t\t\tthis.resolveFinalResult(this.extractResult(event));\n\t\t}\n\n\t\t// Deliver to waiting consumer or queue it\n\t\tconst waiter = this.waiting.shift();\n\t\tif (waiter) {\n\t\t\twaiter({ value: event, done: false });\n\t\t} else {\n\t\t\tthis.queue.push(event);\n\t\t}\n\t}\n\n\tend(result?: R): void {\n\t\tthis.done = true;\n\t\tif (result !== undefined) {\n\t\t\tthis.resolveFinalResult(result);\n\t\t}\n\t\t// Notify all waiting consumers that we're done\n\t\twhile (this.waiting.length > 0) {\n\t\t\tconst waiter = this.waiting.shift()!;\n\t\t\twaiter({ value: undefined as any, done: true });\n\t\t}\n\t}\n\n\tasync *[Symbol.asyncIterator](): AsyncIterator<T> {\n\t\twhile (true) {\n\t\t\tif (this.queue.length > 0) {\n\t\t\t\tyield this.queue.shift()!;\n\t\t\t} else if (this.done) {\n\t\t\t\treturn;\n\t\t\t} else {\n\t\t\t\tconst result = await new Promise<IteratorResult<T>>((resolve) => this.waiting.push(resolve));\n\t\t\t\tif (result.done) return;\n\t\t\t\tyield result.value;\n\t\t\t}\n\t\t}\n\t}\n\n\tresult(): Promise<R> {\n\t\treturn this.finalResultPromise;\n\t}\n}\n\nexport class AssistantMessageEventStream extends EventStream<AssistantMessageEvent, AssistantMessage> {\n\tconstructor() {\n\t\tsuper(\n\t\t\t(event) => event.type === \"done\" || event.type === \"error\",\n\t\t\t(event) => {\n\t\t\t\tif (event.type === \"done\") {\n\t\t\t\t\treturn event.message;\n\t\t\t\t} else if (event.type === \"error\") {\n\t\t\t\t\treturn event.error;\n\t\t\t\t}\n\t\t\t\tthrow new Error(\"Unexpected event type for final result\");\n\t\t\t},\n\t\t);\n\t}\n}\n\n/** Factory function for AssistantMessageEventStream (for use in extensions) */\nexport function createAssistantMessageEventStream(): AssistantMessageEventStream {\n\treturn new AssistantMessageEventStream();\n}\n"]}
+{"version":3,"file":"event-stream.d.ts","sourceRoot":"","sources":["../../src/utils/event-stream.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,qBAAqB,EAAE,MAAM,aAAa,CAAC;AAyB3E,qBAAa,WAAW,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,CAAE,YAAW,aAAa,CAAC,CAAC,CAAC;IAC7D,OAAO,CAAC,KAAK,CAAsB;IACnC,OAAO,CAAC,OAAO,CAAuD;IACtE,OAAO,CAAC,IAAI,CAAS;IACrB,OAAO,CAAC,kBAAkB,CAAa;IACvC,OAAO,CAAC,kBAAkB,CAAuB;IACjD,OAAO,CAAC,UAAU,CAAwB;IAC1C,OAAO,CAAC,aAAa,CAAkB;IAEvC,YAAY,UAAU,EAAE,CAAC,KAAK,EAAE,CAAC,KAAK,OAAO,EAAE,aAAa,EAAE,CAAC,KAAK,EAAE,CAAC,KAAK,CAAC,EAM5E;IAED,IAAI,CAAC,KAAK,EAAE,CAAC,GAAG,IAAI,CAenB;IAED,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC,GAAG,IAAI,CAUpB;IAEM,CAAC,MAAM,CAAC,aAAa,CAAC,IAAI,aAAa,CAAC,CAAC,CAAC,CAYhD;IAED,MAAM,IAAI,OAAO,CAAC,CAAC,CAAC,CAEnB;CACD;AAED,qBAAa,2BAA4B,SAAQ,WAAW,CAAC,qBAAqB,EAAE,gBAAgB,CAAC;IACpG,cAYC;CACD;AAED,+EAA+E;AAC/E,wBAAgB,iCAAiC,IAAI,2BAA2B,CAE/E","sourcesContent":["import type { AssistantMessage, AssistantMessageEvent } from \"../types.ts\";\n\nclass FifoQueue<T> {\n\tprivate incoming: T[] = [];\n\tprivate outgoing: T[] = [];\n\n\tget length(): number {\n\t\treturn this.incoming.length + this.outgoing.length;\n\t}\n\n\tenqueue(value: T): void {\n\t\tthis.incoming.push(value);\n\t}\n\n\tdequeue(): T | undefined {\n\t\tif (this.outgoing.length === 0) {\n\t\t\twhile (this.incoming.length > 0) {\n\t\t\t\tthis.outgoing.push(this.incoming.pop()!);\n\t\t\t}\n\t\t}\n\t\treturn this.outgoing.pop();\n\t}\n}\n\n// Generic event stream class for async iteration\nexport class EventStream<T, R = T> implements AsyncIterable<T> {\n\tprivate queue = new FifoQueue<T>();\n\tprivate waiting = new FifoQueue<(value: IteratorResult<T>) => void>();\n\tprivate done = false;\n\tprivate finalResultPromise: Promise<R>;\n\tprivate resolveFinalResult!: (result: R) => void;\n\tprivate isComplete: (event: T) => boolean;\n\tprivate extractResult: (event: T) => R;\n\n\tconstructor(isComplete: (event: T) => boolean, extractResult: (event: T) => R) {\n\t\tthis.isComplete = isComplete;\n\t\tthis.extractResult = extractResult;\n\t\tthis.finalResultPromise = new Promise((resolve) => {\n\t\t\tthis.resolveFinalResult = resolve;\n\t\t});\n\t}\n\n\tpush(event: T): void {\n\t\tif (this.done) return;\n\n\t\tif (this.isComplete(event)) {\n\t\t\tthis.done = true;\n\t\t\tthis.resolveFinalResult(this.extractResult(event));\n\t\t}\n\n\t\t// Deliver to waiting consumer or queue it\n\t\tconst waiter = this.waiting.dequeue();\n\t\tif (waiter) {\n\t\t\twaiter({ value: event, done: false });\n\t\t} else {\n\t\t\tthis.queue.enqueue(event);\n\t\t}\n\t}\n\n\tend(result?: R): void {\n\t\tthis.done = true;\n\t\tif (result !== undefined) {\n\t\t\tthis.resolveFinalResult(result);\n\t\t}\n\t\t// Notify all waiting consumers that we're done\n\t\twhile (this.waiting.length > 0) {\n\t\t\tconst waiter = this.waiting.dequeue()!;\n\t\t\twaiter({ value: undefined as any, done: true });\n\t\t}\n\t}\n\n\tasync *[Symbol.asyncIterator](): AsyncIterator<T> {\n\t\twhile (true) {\n\t\t\tif (this.queue.length > 0) {\n\t\t\t\tyield this.queue.dequeue()!;\n\t\t\t} else if (this.done) {\n\t\t\t\treturn;\n\t\t\t} else {\n\t\t\t\tconst result = await new Promise<IteratorResult<T>>((resolve) => this.waiting.enqueue(resolve));\n\t\t\t\tif (result.done) return;\n\t\t\t\tyield result.value;\n\t\t\t}\n\t\t}\n\t}\n\n\tresult(): Promise<R> {\n\t\treturn this.finalResultPromise;\n\t}\n}\n\nexport class AssistantMessageEventStream extends EventStream<AssistantMessageEvent, AssistantMessage> {\n\tconstructor() {\n\t\tsuper(\n\t\t\t(event) => event.type === \"done\" || event.type === \"error\",\n\t\t\t(event) => {\n\t\t\t\tif (event.type === \"done\") {\n\t\t\t\t\treturn event.message;\n\t\t\t\t} else if (event.type === \"error\") {\n\t\t\t\t\treturn event.error;\n\t\t\t\t}\n\t\t\t\tthrow new Error(\"Unexpected event type for final result\");\n\t\t\t},\n\t\t);\n\t}\n}\n\n/** Factory function for AssistantMessageEventStream (for use in extensions) */\nexport function createAssistantMessageEventStream(): AssistantMessageEventStream {\n\treturn new AssistantMessageEventStream();\n}\n"]}

package/dist/utils/event-stream.js

@@ -1,7 +1,25 @@
+class FifoQueue {
+    incoming = [];
+    outgoing = [];
+    get length() {
+        return this.incoming.length + this.outgoing.length;
+    }
+    enqueue(value) {
+        this.incoming.push(value);
+    }
+    dequeue() {
+        if (this.outgoing.length === 0) {
+            while (this.incoming.length > 0) {
+                this.outgoing.push(this.incoming.pop());
+            }
+        }
+        return this.outgoing.pop();
+    }
+}
 // Generic event stream class for async iteration
 export class EventStream {
-    queue = [];
-    waiting = [];
+    queue = new FifoQueue();
+    waiting = new FifoQueue();
     done = false;
     finalResultPromise;
     resolveFinalResult;
@@ -22,12 +40,12 @@ export class EventStream {
             this.resolveFinalResult(this.extractResult(event));
         }
         // Deliver to waiting consumer or queue it
-        const waiter = this.waiting.shift();
+        const waiter = this.waiting.dequeue();
         if (waiter) {
             waiter({ value: event, done: false });
         }
         else {
-            this.queue.push(event);
+            this.queue.enqueue(event);
         }
     }
     end(result) {
@@ -37,20 +55,20 @@ export class EventStream {
         }
         // Notify all waiting consumers that we're done
         while (this.waiting.length > 0) {
-            const waiter = this.waiting.shift();
+            const waiter = this.waiting.dequeue();
             waiter({ value: undefined, done: true });
         }
     }
     async *[Symbol.asyncIterator]() {
         while (true) {
             if (this.queue.length > 0) {
-                yield this.queue.shift();
+                yield this.queue.dequeue();
             }
             else if (this.done) {
                 return;
             }
             else {
-                const result = await new Promise((resolve) => this.waiting.push(resolve));
+                const result = await new Promise((resolve) => this.waiting.enqueue(resolve));
                 if (result.done)
                     return;
                 yield result.value;

package/dist/utils/event-stream.js.map

@@ -1 +1 @@
-{"version":3,"file":"event-stream.js","sourceRoot":"","sources":["../../src/utils/event-stream.ts"],"names":[],"mappings":"AAEA,iDAAiD;AACjD,MAAM,OAAO,WAAW;IACf,KAAK,GAAQ,EAAE,CAAC;IAChB,OAAO,GAA2C,EAAE,CAAC;IACrD,IAAI,GAAG,KAAK,CAAC;IACb,kBAAkB,CAAa;IAC/B,kBAAkB,CAAuB;IACzC,UAAU,CAAwB;IAClC,aAAa,CAAkB;IAEvC,YAAY,UAAiC,EAAE,aAA8B,EAAE;QAC9E,IAAI,CAAC,UAAU,GAAG,UAAU,CAAC;QAC7B,IAAI,CAAC,aAAa,GAAG,aAAa,CAAC;QACnC,IAAI,CAAC,kBAAkB,GAAG,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC;YAClD,IAAI,CAAC,kBAAkB,GAAG,OAAO,CAAC;QAAA,CAClC,CAAC,CAAC;IAAA,CACH;IAED,IAAI,CAAC,KAAQ,EAAQ;QACpB,IAAI,IAAI,CAAC,IAAI;YAAE,OAAO;QAEtB,IAAI,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC,EAAE,CAAC;YAC5B,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC;YACjB,IAAI,CAAC,kBAAkB,CAAC,IAAI,CAAC,aAAa,CAAC,KAAK,CAAC,CAAC,CAAC;QACpD,CAAC;QAED,0CAA0C;QAC1C,MAAM,MAAM,GAAG,IAAI,CAAC,OAAO,CAAC,KAAK,EAAE,CAAC;QACpC,IAAI,MAAM,EAAE,CAAC;YACZ,MAAM,CAAC,EAAE,KAAK,EAAE,KAAK,EAAE,IAAI,EAAE,KAAK,EAAE,CAAC,CAAC;QACvC,CAAC;aAAM,CAAC;YACP,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;QACxB,CAAC;IAAA,CACD;IAED,GAAG,CAAC,MAAU,EAAQ;QACrB,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC;QACjB,IAAI,MAAM,KAAK,SAAS,EAAE,CAAC;YAC1B,IAAI,CAAC,kBAAkB,CAAC,MAAM,CAAC,CAAC;QACjC,CAAC;QACD,+CAA+C;QAC/C,OAAO,IAAI,CAAC,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YAChC,MAAM,MAAM,GAAG,IAAI,CAAC,OAAO,CAAC,KAAK,EAAG,CAAC;YACrC,MAAM,CAAC,EAAE,KAAK,EAAE,SAAgB,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC,CAAC;QACjD,CAAC;IAAA,CACD;IAED,KAAK,CAAC,CAAC,CAAC,MAAM,CAAC,aAAa,CAAC,GAAqB;QACjD,OAAO,IAAI,EAAE,CAAC;YACb,IAAI,IAAI,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;gBAC3B,MAAM,IAAI,CAAC,KAAK,CAAC,KAAK,EAAG,CAAC;YAC3B,CAAC;iBAAM,IAAI,IAAI,CAAC,IAAI,EAAE,CAAC;gBACtB,OAAO;YACR,CAAC;iBAAM,CAAC;gBACP,MAAM,MAAM,GAAG,MAAM,IAAI,OAAO,CAAoB,CAAC,OAAO,EAAE,EAAE,CAAC,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC,CAAC;gBAC7F,IAAI,MAAM,CAAC,IAAI;oBAAE,OAAO;gBACxB,MAAM,MAAM,CAAC,KAAK,CAAC;YACpB,CAAC;QACF,CAAC;IAAA,CACD;IAED,MAAM,GAAe;QACpB,OAAO,IAAI,CAAC,kBAAkB,CAAC;IAAA,CAC/B;CACD;AAED,MAAM,OAAO,2BAA4B,SAAQ,WAAoD;IACpG,cAAc;QACb,KAAK,CACJ,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,IAAI,KAAK,MAAM,IAAI,KAAK,CAAC,IAAI,KAAK,OAAO,EAC1D,CAAC,KAAK,EAAE,EAAE,CAAC;YACV,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;gBAC3B,OAAO,KAAK,CAAC,OAAO,CAAC;YACtB,CAAC;iBAAM,IAAI,KAAK,CAAC,IAAI,KAAK,OAAO,EAAE,CAAC;gBACnC,OAAO,KAAK,CAAC,KAAK,CAAC;YACpB,CAAC;YACD,MAAM,IAAI,KAAK,CAAC,wCAAwC,CAAC,CAAC;QAAA,CAC1D,CACD,CAAC;IAAA,CACF;CACD;AAED,+EAA+E;AAC/E,MAAM,UAAU,iCAAiC,GAAgC;IAChF,OAAO,IAAI,2BAA2B,EAAE,CAAC;AAAA,CACzC","sourcesContent":["import type { AssistantMessage, AssistantMessageEvent } from \"../types.ts\";\n\n// Generic event stream class for async iteration\nexport class EventStream<T, R = T> implements AsyncIterable<T> {\n\tprivate queue: T[] = [];\n\tprivate waiting: ((value: IteratorResult<T>) => void)[] = [];\n\tprivate done = false;\n\tprivate finalResultPromise: Promise<R>;\n\tprivate resolveFinalResult!: (result: R) => void;\n\tprivate isComplete: (event: T) => boolean;\n\tprivate extractResult: (event: T) => R;\n\n\tconstructor(isComplete: (event: T) => boolean, extractResult: (event: T) => R) {\n\t\tthis.isComplete = isComplete;\n\t\tthis.extractResult = extractResult;\n\t\tthis.finalResultPromise = new Promise((resolve) => {\n\t\t\tthis.resolveFinalResult = resolve;\n\t\t});\n\t}\n\n\tpush(event: T): void {\n\t\tif (this.done) return;\n\n\t\tif (this.isComplete(event)) {\n\t\t\tthis.done = true;\n\t\t\tthis.resolveFinalResult(this.extractResult(event));\n\t\t}\n\n\t\t// Deliver to waiting consumer or queue it\n\t\tconst waiter = this.waiting.shift();\n\t\tif (waiter) {\n\t\t\twaiter({ value: event, done: false });\n\t\t} else {\n\t\t\tthis.queue.push(event);\n\t\t}\n\t}\n\n\tend(result?: R): void {\n\t\tthis.done = true;\n\t\tif (result !== undefined) {\n\t\t\tthis.resolveFinalResult(result);\n\t\t}\n\t\t// Notify all waiting consumers that we're done\n\t\twhile (this.waiting.length > 0) {\n\t\t\tconst waiter = this.waiting.shift()!;\n\t\t\twaiter({ value: undefined as any, done: true });\n\t\t}\n\t}\n\n\tasync *[Symbol.asyncIterator](): AsyncIterator<T> {\n\t\twhile (true) {\n\t\t\tif (this.queue.length > 0) {\n\t\t\t\tyield this.queue.shift()!;\n\t\t\t} else if (this.done) {\n\t\t\t\treturn;\n\t\t\t} else {\n\t\t\t\tconst result = await new Promise<IteratorResult<T>>((resolve) => this.waiting.push(resolve));\n\t\t\t\tif (result.done) return;\n\t\t\t\tyield result.value;\n\t\t\t}\n\t\t}\n\t}\n\n\tresult(): Promise<R> {\n\t\treturn this.finalResultPromise;\n\t}\n}\n\nexport class AssistantMessageEventStream extends EventStream<AssistantMessageEvent, AssistantMessage> {\n\tconstructor() {\n\t\tsuper(\n\t\t\t(event) => event.type === \"done\" || event.type === \"error\",\n\t\t\t(event) => {\n\t\t\t\tif (event.type === \"done\") {\n\t\t\t\t\treturn event.message;\n\t\t\t\t} else if (event.type === \"error\") {\n\t\t\t\t\treturn event.error;\n\t\t\t\t}\n\t\t\t\tthrow new Error(\"Unexpected event type for final result\");\n\t\t\t},\n\t\t);\n\t}\n}\n\n/** Factory function for AssistantMessageEventStream (for use in extensions) */\nexport function createAssistantMessageEventStream(): AssistantMessageEventStream {\n\treturn new AssistantMessageEventStream();\n}\n"]}
+{"version":3,"file":"event-stream.js","sourceRoot":"","sources":["../../src/utils/event-stream.ts"],"names":[],"mappings":"AAEA,MAAM,SAAS;IACN,QAAQ,GAAQ,EAAE,CAAC;IACnB,QAAQ,GAAQ,EAAE,CAAC;IAE3B,IAAI,MAAM,GAAW;QACpB,OAAO,IAAI,CAAC,QAAQ,CAAC,MAAM,GAAG,IAAI,CAAC,QAAQ,CAAC,MAAM,CAAC;IAAA,CACnD;IAED,OAAO,CAAC,KAAQ,EAAQ;QACvB,IAAI,CAAC,QAAQ,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;IAAA,CAC1B;IAED,OAAO,GAAkB;QACxB,IAAI,IAAI,CAAC,QAAQ,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YAChC,OAAO,IAAI,CAAC,QAAQ,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;gBACjC,IAAI,CAAC,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,GAAG,EAAG,CAAC,CAAC;YAC1C,CAAC;QACF,CAAC;QACD,OAAO,IAAI,CAAC,QAAQ,CAAC,GAAG,EAAE,CAAC;IAAA,CAC3B;CACD;AAED,iDAAiD;AACjD,MAAM,OAAO,WAAW;IACf,KAAK,GAAG,IAAI,SAAS,EAAK,CAAC;IAC3B,OAAO,GAAG,IAAI,SAAS,EAAsC,CAAC;IAC9D,IAAI,GAAG,KAAK,CAAC;IACb,kBAAkB,CAAa;IAC/B,kBAAkB,CAAuB;IACzC,UAAU,CAAwB;IAClC,aAAa,CAAkB;IAEvC,YAAY,UAAiC,EAAE,aAA8B,EAAE;QAC9E,IAAI,CAAC,UAAU,GAAG,UAAU,CAAC;QAC7B,IAAI,CAAC,aAAa,GAAG,aAAa,CAAC;QACnC,IAAI,CAAC,kBAAkB,GAAG,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC;YAClD,IAAI,CAAC,kBAAkB,GAAG,OAAO,CAAC;QAAA,CAClC,CAAC,CAAC;IAAA,CACH;IAED,IAAI,CAAC,KAAQ,EAAQ;QACpB,IAAI,IAAI,CAAC,IAAI;YAAE,OAAO;QAEtB,IAAI,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC,EAAE,CAAC;YAC5B,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC;YACjB,IAAI,CAAC,kBAAkB,CAAC,IAAI,CAAC,aAAa,CAAC,KAAK,CAAC,CAAC,CAAC;QACpD,CAAC;QAED,0CAA0C;QAC1C,MAAM,MAAM,GAAG,IAAI,CAAC,OAAO,CAAC,OAAO,EAAE,CAAC;QACtC,IAAI,MAAM,EAAE,CAAC;YACZ,MAAM,CAAC,EAAE,KAAK,EAAE,KAAK,EAAE,IAAI,EAAE,KAAK,EAAE,CAAC,CAAC;QACvC,CAAC;aAAM,CAAC;YACP,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;QAC3B,CAAC;IAAA,CACD;IAED,GAAG,CAAC,MAAU,EAAQ;QACrB,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC;QACjB,IAAI,MAAM,KAAK,SAAS,EAAE,CAAC;YAC1B,IAAI,CAAC,kBAAkB,CAAC,MAAM,CAAC,CAAC;QACjC,CAAC;QACD,+CAA+C;QAC/C,OAAO,IAAI,CAAC,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YAChC,MAAM,MAAM,GAAG,IAAI,CAAC,OAAO,CAAC,OAAO,EAAG,CAAC;YACvC,MAAM,CAAC,EAAE,KAAK,EAAE,SAAgB,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC,CAAC;QACjD,CAAC;IAAA,CACD;IAED,KAAK,CAAC,CAAC,CAAC,MAAM,CAAC,aAAa,CAAC,GAAqB;QACjD,OAAO,IAAI,EAAE,CAAC;YACb,IAAI,IAAI,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;gBAC3B,MAAM,IAAI,CAAC,KAAK,CAAC,OAAO,EAAG,CAAC;YAC7B,CAAC;iBAAM,IAAI,IAAI,CAAC,IAAI,EAAE,CAAC;gBACtB,OAAO;YACR,CAAC;iBAAM,CAAC;gBACP,MAAM,MAAM,GAAG,MAAM,IAAI,OAAO,CAAoB,CAAC,OAAO,EAAE,EAAE,CAAC,IAAI,CAAC,OAAO,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC,CAAC;gBAChG,IAAI,MAAM,CAAC,IAAI;oBAAE,OAAO;gBACxB,MAAM,MAAM,CAAC,KAAK,CAAC;YACpB,CAAC;QACF,CAAC;IAAA,CACD;IAED,MAAM,GAAe;QACpB,OAAO,IAAI,CAAC,kBAAkB,CAAC;IAAA,CAC/B;CACD;AAED,MAAM,OAAO,2BAA4B,SAAQ,WAAoD;IACpG,cAAc;QACb,KAAK,CACJ,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,IAAI,KAAK,MAAM,IAAI,KAAK,CAAC,IAAI,KAAK,OAAO,EAC1D,CAAC,KAAK,EAAE,EAAE,CAAC;YACV,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;gBAC3B,OAAO,KAAK,CAAC,OAAO,CAAC;YACtB,CAAC;iBAAM,IAAI,KAAK,CAAC,IAAI,KAAK,OAAO,EAAE,CAAC;gBACnC,OAAO,KAAK,CAAC,KAAK,CAAC;YACpB,CAAC;YACD,MAAM,IAAI,KAAK,CAAC,wCAAwC,CAAC,CAAC;QAAA,CAC1D,CACD,CAAC;IAAA,CACF;CACD;AAED,+EAA+E;AAC/E,MAAM,UAAU,iCAAiC,GAAgC;IAChF,OAAO,IAAI,2BAA2B,EAAE,CAAC;AAAA,CACzC","sourcesContent":["import type { AssistantMessage, AssistantMessageEvent } from \"../types.ts\";\n\nclass FifoQueue<T> {\n\tprivate incoming: T[] = [];\n\tprivate outgoing: T[] = [];\n\n\tget length(): number {\n\t\treturn this.incoming.length + this.outgoing.length;\n\t}\n\n\tenqueue(value: T): void {\n\t\tthis.incoming.push(value);\n\t}\n\n\tdequeue(): T | undefined {\n\t\tif (this.outgoing.length === 0) {\n\t\t\twhile (this.incoming.length > 0) {\n\t\t\t\tthis.outgoing.push(this.incoming.pop()!);\n\t\t\t}\n\t\t}\n\t\treturn this.outgoing.pop();\n\t}\n}\n\n// Generic event stream class for async iteration\nexport class EventStream<T, R = T> implements AsyncIterable<T> {\n\tprivate queue = new FifoQueue<T>();\n\tprivate waiting = new FifoQueue<(value: IteratorResult<T>) => void>();\n\tprivate done = false;\n\tprivate finalResultPromise: Promise<R>;\n\tprivate resolveFinalResult!: (result: R) => void;\n\tprivate isComplete: (event: T) => boolean;\n\tprivate extractResult: (event: T) => R;\n\n\tconstructor(isComplete: (event: T) => boolean, extractResult: (event: T) => R) {\n\t\tthis.isComplete = isComplete;\n\t\tthis.extractResult = extractResult;\n\t\tthis.finalResultPromise = new Promise((resolve) => {\n\t\t\tthis.resolveFinalResult = resolve;\n\t\t});\n\t}\n\n\tpush(event: T): void {\n\t\tif (this.done) return;\n\n\t\tif (this.isComplete(event)) {\n\t\t\tthis.done = true;\n\t\t\tthis.resolveFinalResult(this.extractResult(event));\n\t\t}\n\n\t\t// Deliver to waiting consumer or queue it\n\t\tconst waiter = this.waiting.dequeue();\n\t\tif (waiter) {\n\t\t\twaiter({ value: event, done: false });\n\t\t} else {\n\t\t\tthis.queue.enqueue(event);\n\t\t}\n\t}\n\n\tend(result?: R): void {\n\t\tthis.done = true;\n\t\tif (result !== undefined) {\n\t\t\tthis.resolveFinalResult(result);\n\t\t}\n\t\t// Notify all waiting consumers that we're done\n\t\twhile (this.waiting.length > 0) {\n\t\t\tconst waiter = this.waiting.dequeue()!;\n\t\t\twaiter({ value: undefined as any, done: true });\n\t\t}\n\t}\n\n\tasync *[Symbol.asyncIterator](): AsyncIterator<T> {\n\t\twhile (true) {\n\t\t\tif (this.queue.length > 0) {\n\t\t\t\tyield this.queue.dequeue()!;\n\t\t\t} else if (this.done) {\n\t\t\t\treturn;\n\t\t\t} else {\n\t\t\t\tconst result = await new Promise<IteratorResult<T>>((resolve) => this.waiting.enqueue(resolve));\n\t\t\t\tif (result.done) return;\n\t\t\t\tyield result.value;\n\t\t\t}\n\t\t}\n\t}\n\n\tresult(): Promise<R> {\n\t\treturn this.finalResultPromise;\n\t}\n}\n\nexport class AssistantMessageEventStream extends EventStream<AssistantMessageEvent, AssistantMessage> {\n\tconstructor() {\n\t\tsuper(\n\t\t\t(event) => event.type === \"done\" || event.type === \"error\",\n\t\t\t(event) => {\n\t\t\t\tif (event.type === \"done\") {\n\t\t\t\t\treturn event.message;\n\t\t\t\t} else if (event.type === \"error\") {\n\t\t\t\t\treturn event.error;\n\t\t\t\t}\n\t\t\t\tthrow new Error(\"Unexpected event type for final result\");\n\t\t\t},\n\t\t);\n\t}\n}\n\n/** Factory function for AssistantMessageEventStream (for use in extensions) */\nexport function createAssistantMessageEventStream(): AssistantMessageEventStream {\n\treturn new AssistantMessageEventStream();\n}\n"]}

package/dist/utils/retry.d.ts

@@ -1,6 +1,7 @@
 import type { AssistantMessage } from "../types.ts";
 /**
  * Retry policy: bounded attempts with exponential backoff (`baseDelayMs * 2^(attempt-1)`).
+ * Each delay is capped by `maxAgentDelayMs` (60 seconds by default).
  * Matches `settings.retry` (`enabled`, `maxRetries`, `baseDelayMs`) in coding-agent; kept
  * here so the classifier and the policy-driven retry loop live together and stay reusable
  * by the SDK and other callers.
@@ -11,7 +12,11 @@ export interface RetryPolicy {
     maxRetries: number;
     /** Base delay in ms. Per-attempt delay is `baseDelayMs * 2^(attempt-1)` before jitter. */
     baseDelayMs: number;
+    /** Cap for agent-level retry delays in ms. Defaults to 60 seconds. */
+    maxAgentDelayMs?: number;
 }
+export declare const DEFAULT_MAX_AGENT_RETRY_DELAY_MS = 60000;
+export declare function retryDelayMs(policy: Pick<RetryPolicy, "baseDelayMs" | "maxAgentDelayMs">, attempt: number): number;
 /** Optional callbacks emitted by {@link retryAssistantCall} around each retry. */
 export interface RetryCallbacks {
     /** Emitted before the backoff sleep of each retry attempt (1-indexed). */

package/dist/utils/retry.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"retry.d.ts","sourceRoot":"","sources":["../../src/utils/retry.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AA2GpD;;;;;GAKG;AACH,MAAM,WAAW,WAAW;IAC3B,OAAO,EAAE,OAAO,CAAC;IACjB,qFAAqF;IACrF,UAAU,EAAE,MAAM,CAAC;IACnB,0FAA0F;IAC1F,WAAW,EAAE,MAAM,CAAC;CACpB;AAED,kFAAkF;AAClF,MAAM,WAAW,cAAc;IAC9B,0EAA0E;IAC1E,gBAAgB,CAAC,EAAE,CAClB,OAAO,EAAE,MAAM,EACf,WAAW,EAAE,MAAM,EACnB,OAAO,EAAE,MAAM,EACf,YAAY,EAAE,MAAM,KAChB,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IAC1B,mFAAmF;IACnF,mBAAmB,CAAC,EAAE,MAAM,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IACjD,mFAAmF;IACnF,eAAe,CAAC,EAAE,CAAC,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE,UAAU,CAAC,EAAE,MAAM,KAAK,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;CACnG;AA0BD;;;;;;;;;;;;;;;;;GAiBG;AACH,wBAAsB,kBAAkB,CACvC,OAAO,EAAE,MAAM,OAAO,CAAC,gBAAgB,CAAC,EACxC,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC,gBAAgB,CAAC,CA4C3B;AAED;;;;;;;;GAQG;AACH,wBAAgB,yBAAyB,CAAC,OAAO,EAAE,gBAAgB,GAAG,OAAO,CAK5E","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\n\nfunction buildProviderErrorPattern(patterns: readonly string[]): RegExp {\n\treturn new RegExp(patterns.join(\"|\"), \"i\");\n}\n\nconst NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN = buildProviderErrorPattern([\n\t// OpenCode Go/free-tier limits returned as 429 JSON error types by OpenCode's\n\t// Zen API. These are subscription/account limits, not transient throttles.\n\t\"GoUsageLimitError\",\n\t\"FreeUsageLimitError\",\n\n\t// OpenCode Go subscription-limit text asks users to enable available-balance\n\t// usage after rolling/weekly/monthly limits are reached.\n\t\"Monthly usage limit reached\",\n\t\"available balance\",\n\n\t// Generic quota/budget/billing exhaustion. `insufficient_quota` is OpenAI's\n\t// quota/billing error code; the other strings cover common gateway wording.\n\t\"insufficient_quota\",\n\t\"out of budget\",\n\t\"quota exceeded\",\n\t\"billing\",\n]);\n\nconst RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([\n\t// Generic provider load, HTTP status, and server-side transient failures.\n\t\"overloaded\",\n\t\"rate.?limit\",\n\t\"too many requests\",\n\t\"429\",\n\t\"500\",\n\t\"502\",\n\t\"503\",\n\t\"504\",\n\t\"524\",\n\t\"service.?unavailable\",\n\t\"server.?error\",\n\t\"internal.?error\",\n\n\t// Wrapper/provider text for transient upstream failures, including OpenRouter\n\t// \"Provider returned error\" responses (#2264).\n\t\"provider.?returned.?error\",\n\t\"exceeded request buffer limit while retrying upstream\",\n\n\t// Network, proxy, and fetch transport failures. This includes OpenAI Codex\n\t// raw-fetch failures such as \"upstream connect\", \"connection refused\", and\n\t// \"reset before headers\" (#733), plus OpenRouter connection drops (#3317).\n\t\"network.?error\",\n\t\"connection.?error\",\n\t\"connection.?refused\",\n\t\"connection.?lost\",\n\t\"other side closed\",\n\t\"fetch failed\",\n\t\"getaddrinfo\",\n\t\"ENOTFOUND\",\n\t\"EAI_AGAIN\",\n\t\"upstream.?connect\",\n\t\"reset before headers\",\n\t\"socket hang up\",\n\t\"socket connection was closed\",\n\t\"timed? out\",\n\t\"timeout\",\n\t\"terminated\",\n\n\t// WebSocket transports can report close/error text instead of HTTP/fetch text.\n\t\"websocket.?closed\",\n\t\"websocket.?error\",\n\n\t// Response-body decompression failures. Node/Bun zlib rejects a body whose\n\t// Content-Encoding lied or whose compressed bytes were corrupted, and the\n\t// fetch/WebSocket layer wraps it with a \"Library error:\" prefix. GitHub\n\t// Copilot on the default `transport: \"auto\"` reports this as\n\t// \"Library error: zlib error: incorrect header check\" (#2553). The transfer\n\t// is already unusable, and a fresh attempt normally succeeds, so treat every\n\t// variant as a transient transport failure rather than a deterministic error.\n\t\"zlib\",\n\t\"incorrect header check\",\n\t\"decompress\",\n\t\"library error\",\n\n\t// Idle-stream deadline enforced below the HTTP layer by `withStreamDeadline`\n\t// so a stream that never settles cannot hang an attempt forever (#2553).\n\t\"stream deadline exceeded\",\n\n\t// Premature stream endings from SDKs and transports. Anthropic can throw\n\t// \"stream ended without ...\" and \"Anthropic stream ended before message_stop\"\n\t// (#4433); Bedrock/Smithy can throw an HTTP/2 no-response error (#3594).\n\t\"ended without\",\n\t\"stream ended before message_stop\",\n\t\"stream ended before a terminal response event\",\n\t\"http2 request did not get a response\",\n\n\t// Provider-requested retry delay cap failures should flow through the outer\n\t// retry policy so callers can surface/abort the backoff (#1123).\n\t\"retry delay\",\n\n\t// Explicit retry guidance emitted mid-stream by OpenAI Responses and Bedrock\n\t// stream exceptions (#6019).\n\t\"you can retry your request\",\n\t\"try your request again\",\n\t\"please retry your request\",\n\n\t// gRPC based providers (e.g. NVIDIA NIM)\n\t\"ResourceExhausted\",\n]);\n\n/**\n * Retry policy: bounded attempts with exponential backoff (`baseDelayMs * 2^(attempt-1)`).\n * Matches `settings.retry` (`enabled`, `maxRetries`, `baseDelayMs`) in coding-agent; kept\n * here so the classifier and the policy-driven retry loop live together and stay reusable\n * by the SDK and other callers.\n */\nexport interface RetryPolicy {\n\tenabled: boolean;\n\t/** Max retry attempts (0 = no retries). The initial call never counts as a retry. */\n\tmaxRetries: number;\n\t/** Base delay in ms. Per-attempt delay is `baseDelayMs * 2^(attempt-1)` before jitter. */\n\tbaseDelayMs: number;\n}\n\n/** Optional callbacks emitted by {@link retryAssistantCall} around each retry. */\nexport interface RetryCallbacks {\n\t/** Emitted before the backoff sleep of each retry attempt (1-indexed). */\n\tonRetryScheduled?: (\n\t\tattempt: number,\n\t\tmaxAttempts: number,\n\t\tdelayMs: number,\n\t\terrorMessage: string,\n\t) => void | Promise<void>;\n\t/** Emitted after the backoff sleep, immediately before the retried call starts. */\n\tonRetryAttemptStart?: () => void | Promise<void>;\n\t/** Emitted once when the loop ends: success if a later call completed normally. */\n\tonRetryFinished?: (success: boolean, attempt: number, finalError?: string) => void | Promise<void>;\n}\n\nclass RetrySleepAbortError extends Error {\n\tconstructor() {\n\t\tsuper(\"Aborted\");\n\t}\n}\n\nfunction sleep(ms: number, signal?: AbortSignal): Promise<void> {\n\treturn new Promise((resolve, reject) => {\n\t\tif (signal?.aborted) {\n\t\t\treject(new RetrySleepAbortError());\n\t\t\treturn;\n\t\t}\n\t\tconst timeout = setTimeout(resolve, ms);\n\t\tsignal?.addEventListener(\n\t\t\t\"abort\",\n\t\t\t() => {\n\t\t\t\tclearTimeout(timeout);\n\t\t\t\treject(new RetrySleepAbortError());\n\t\t\t},\n\t\t\t{ once: true },\n\t\t);\n\t});\n}\n\n/**\n * Run a single assistant-producing call with bounded retry on transient errors.\n *\n * Behavior:\n * - A successful response is returned immediately. Aborts are terminal and never\n *   retried, but reported as unsuccessful if they happen after a retry was scheduled.\n *   Aborts during the backoff sleep are normalized to an aborted `AssistantMessage`\n *   too, so callers do not need to care when cancellation happened.\n * - A non-retryable error (per {@link isRetryableAssistantError}, including quota/\n *   billing exhaustion) is returned immediately so deterministic errors fail fast.\n * - Otherwise retries up to `maxRetries` times with exponential backoff, emitting\n *   `onRetryScheduled` before each sleep, `onRetryAttemptStart` after each sleep before\n *   the retried call starts, and `onRetryFinished` once at the end (whether the loop\n *   ends in success, exhausted retries, or an aborted backoff).\n *\n * When `policy` is undefined or disabled, the first response is returned unchanged\n * (equivalent to calling `produce()` directly).\n */\nexport async function retryAssistantCall(\n\tproduce: () => Promise<AssistantMessage>,\n\tpolicy: RetryPolicy | undefined,\n\tsignal: AbortSignal | undefined,\n\tcallbacks?: RetryCallbacks,\n): Promise<AssistantMessage> {\n\tconst maxAttempts = policy?.enabled ? policy.maxRetries : 0;\n\n\tlet attempt = 0;\n\tlet lastRetry: { attempt: number; errorMessage: string } | undefined;\n\tfor (;;) {\n\t\tconst response = await produce();\n\n\t\t// Abort: terminal but not successful. Never retry an aborted message.\n\t\tif (response.stopReason === \"aborted\") {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(false, lastRetry.attempt);\n\t\t\treturn response;\n\t\t}\n\n\t\t// Success: non-error, non-abort responses return as-is.\n\t\tif (response.stopReason !== \"error\") {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(true, lastRetry.attempt);\n\t\t\treturn response;\n\t\t}\n\n\t\t// Non-retryable, or budget exhausted: return the final error message.\n\t\tif (attempt >= maxAttempts || !isRetryableAssistantError(response)) {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(false, lastRetry.attempt, response.errorMessage);\n\t\t\treturn response;\n\t\t}\n\n\t\tattempt++;\n\t\tlastRetry = { attempt, errorMessage: response.errorMessage || \"Unknown error\" };\n\t\tconst delayMs = policy!.baseDelayMs * 2 ** (attempt - 1);\n\t\tawait callbacks?.onRetryScheduled?.(attempt, maxAttempts, delayMs, lastRetry.errorMessage);\n\n\t\t// Normalize aborts during retry backoff to the same AssistantMessage shape as\n\t\t// provider stream aborts, so callers do not need to care when cancellation happened.\n\t\ttry {\n\t\t\tawait sleep(delayMs, signal);\n\t\t} catch (error) {\n\t\t\tawait callbacks?.onRetryFinished?.(false, attempt, lastRetry.errorMessage);\n\t\t\tif (error instanceof RetrySleepAbortError) {\n\t\t\t\treturn { ...response, stopReason: \"aborted\", errorMessage: undefined };\n\t\t\t}\n\t\t\tthrow error;\n\t\t}\n\t\tawait callbacks?.onRetryAttemptStart?.();\n\t}\n}\n\n/**\n * Classifies whether a failed assistant message looks like a transient provider\n * or transport error, so callers can decide if the last assistant turn should be\n * restarted.\n *\n * This does not implement retry policy. Callers should first handle context\n * overflow separately, then apply their own retry budget, backoff, and reporting\n * before restarting the assistant turn.\n */\nexport function isRetryableAssistantError(message: AssistantMessage): boolean {\n\tif (message.stopReason !== \"error\" || !message.errorMessage) return false;\n\tconst errorMessage = message.errorMessage;\n\tif (NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN.test(errorMessage)) return false;\n\treturn RETRYABLE_PROVIDER_ERROR_PATTERN.test(errorMessage);\n}\n"]}
+{"version":3,"file":"retry.d.ts","sourceRoot":"","sources":["../../src/utils/retry.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AA2GpD;;;;;;GAMG;AACH,MAAM,WAAW,WAAW;IAC3B,OAAO,EAAE,OAAO,CAAC;IACjB,qFAAqF;IACrF,UAAU,EAAE,MAAM,CAAC;IACnB,0FAA0F;IAC1F,WAAW,EAAE,MAAM,CAAC;IACpB,sEAAsE;IACtE,eAAe,CAAC,EAAE,MAAM,CAAC;CACzB;AAED,eAAO,MAAM,gCAAgC,QAAS,CAAC;AAEvD,wBAAgB,YAAY,CAAC,MAAM,EAAE,IAAI,CAAC,WAAW,EAAE,aAAa,GAAG,iBAAiB,CAAC,EAAE,OAAO,EAAE,MAAM,GAAG,MAAM,CAIlH;AAED,kFAAkF;AAClF,MAAM,WAAW,cAAc;IAC9B,0EAA0E;IAC1E,gBAAgB,CAAC,EAAE,CAClB,OAAO,EAAE,MAAM,EACf,WAAW,EAAE,MAAM,EACnB,OAAO,EAAE,MAAM,EACf,YAAY,EAAE,MAAM,KAChB,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IAC1B,mFAAmF;IACnF,mBAAmB,CAAC,EAAE,MAAM,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IACjD,mFAAmF;IACnF,eAAe,CAAC,EAAE,CAAC,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE,UAAU,CAAC,EAAE,MAAM,KAAK,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;CACnG;AA0BD;;;;;;;;;;;;;;;;;GAiBG;AACH,wBAAsB,kBAAkB,CACvC,OAAO,EAAE,MAAM,OAAO,CAAC,gBAAgB,CAAC,EACxC,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC,gBAAgB,CAAC,CA4C3B;AAED;;;;;;;;GAQG;AACH,wBAAgB,yBAAyB,CAAC,OAAO,EAAE,gBAAgB,GAAG,OAAO,CAK5E","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\n\nfunction buildProviderErrorPattern(patterns: readonly string[]): RegExp {\n\treturn new RegExp(patterns.join(\"|\"), \"i\");\n}\n\nconst NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN = buildProviderErrorPattern([\n\t// OpenCode Go/free-tier limits returned as 429 JSON error types by OpenCode's\n\t// Zen API. These are subscription/account limits, not transient throttles.\n\t\"GoUsageLimitError\",\n\t\"FreeUsageLimitError\",\n\n\t// OpenCode Go subscription-limit text asks users to enable available-balance\n\t// usage after rolling/weekly/monthly limits are reached.\n\t\"Monthly usage limit reached\",\n\t\"available balance\",\n\n\t// Generic quota/budget/billing exhaustion. `insufficient_quota` is OpenAI's\n\t// quota/billing error code; the other strings cover common gateway wording.\n\t\"insufficient_quota\",\n\t\"out of budget\",\n\t\"quota exceeded\",\n\t\"billing\",\n]);\n\nconst RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([\n\t// Generic provider load, HTTP status, and server-side transient failures.\n\t\"overloaded\",\n\t\"rate.?limit\",\n\t\"too many requests\",\n\t\"429\",\n\t\"500\",\n\t\"502\",\n\t\"503\",\n\t\"504\",\n\t\"524\",\n\t\"service.?unavailable\",\n\t\"server.?error\",\n\t\"internal.?error\",\n\n\t// Wrapper/provider text for transient upstream failures, including OpenRouter\n\t// \"Provider returned error\" responses (#2264).\n\t\"provider.?returned.?error\",\n\t\"exceeded request buffer limit while retrying upstream\",\n\n\t// Network, proxy, and fetch transport failures. This includes OpenAI Codex\n\t// raw-fetch failures such as \"upstream connect\", \"connection refused\", and\n\t// \"reset before headers\" (#733), plus OpenRouter connection drops (#3317).\n\t\"network.?error\",\n\t\"connection.?error\",\n\t\"connection.?refused\",\n\t\"connection.?lost\",\n\t\"other side closed\",\n\t\"fetch failed\",\n\t\"getaddrinfo\",\n\t\"ENOTFOUND\",\n\t\"EAI_AGAIN\",\n\t\"upstream.?connect\",\n\t\"reset before headers\",\n\t\"socket hang up\",\n\t\"socket connection was closed\",\n\t\"timed? out\",\n\t\"timeout\",\n\t\"terminated\",\n\n\t// WebSocket transports can report close/error text instead of HTTP/fetch text.\n\t\"websocket.?closed\",\n\t\"websocket.?error\",\n\n\t// Response-body decompression failures. Node/Bun zlib rejects a body whose\n\t// Content-Encoding lied or whose compressed bytes were corrupted, and the\n\t// fetch/WebSocket layer wraps it with a \"Library error:\" prefix. GitHub\n\t// Copilot on the default `transport: \"auto\"` reports this as\n\t// \"Library error: zlib error: incorrect header check\" (#2553). The transfer\n\t// is already unusable, and a fresh attempt normally succeeds, so treat every\n\t// variant as a transient transport failure rather than a deterministic error.\n\t\"zlib\",\n\t\"incorrect header check\",\n\t\"decompress\",\n\t\"library error\",\n\n\t// Idle-stream deadline enforced below the HTTP layer by `withStreamDeadline`\n\t// so a stream that never settles cannot hang an attempt forever (#2553).\n\t\"stream deadline exceeded\",\n\n\t// Premature stream endings from SDKs and transports. Anthropic can throw\n\t// \"stream ended without ...\" and \"Anthropic stream ended before message_stop\"\n\t// (#4433); Bedrock/Smithy can throw an HTTP/2 no-response error (#3594).\n\t\"ended without\",\n\t\"stream ended before message_stop\",\n\t\"stream ended before a terminal response event\",\n\t\"http2 request did not get a response\",\n\n\t// Provider-requested retry delay cap failures should flow through the outer\n\t// retry policy so callers can surface/abort the backoff (#1123).\n\t\"retry delay\",\n\n\t// Explicit retry guidance emitted mid-stream by OpenAI Responses and Bedrock\n\t// stream exceptions (#6019).\n\t\"you can retry your request\",\n\t\"try your request again\",\n\t\"please retry your request\",\n\n\t// gRPC based providers (e.g. NVIDIA NIM)\n\t\"ResourceExhausted\",\n]);\n\n/**\n * Retry policy: bounded attempts with exponential backoff (`baseDelayMs * 2^(attempt-1)`).\n * Each delay is capped by `maxAgentDelayMs` (60 seconds by default).\n * Matches `settings.retry` (`enabled`, `maxRetries`, `baseDelayMs`) in coding-agent; kept\n * here so the classifier and the policy-driven retry loop live together and stay reusable\n * by the SDK and other callers.\n */\nexport interface RetryPolicy {\n\tenabled: boolean;\n\t/** Max retry attempts (0 = no retries). The initial call never counts as a retry. */\n\tmaxRetries: number;\n\t/** Base delay in ms. Per-attempt delay is `baseDelayMs * 2^(attempt-1)` before jitter. */\n\tbaseDelayMs: number;\n\t/** Cap for agent-level retry delays in ms. Defaults to 60 seconds. */\n\tmaxAgentDelayMs?: number;\n}\n\nexport const DEFAULT_MAX_AGENT_RETRY_DELAY_MS = 60_000;\n\nexport function retryDelayMs(policy: Pick<RetryPolicy, \"baseDelayMs\" | \"maxAgentDelayMs\">, attempt: number): number {\n\tconst delay = policy.baseDelayMs * 2 ** Math.max(0, attempt - 1);\n\tconst safeDelay = Number.isSafeInteger(delay) ? delay : Number.MAX_SAFE_INTEGER;\n\treturn Math.min(safeDelay, policy.maxAgentDelayMs ?? DEFAULT_MAX_AGENT_RETRY_DELAY_MS);\n}\n\n/** Optional callbacks emitted by {@link retryAssistantCall} around each retry. */\nexport interface RetryCallbacks {\n\t/** Emitted before the backoff sleep of each retry attempt (1-indexed). */\n\tonRetryScheduled?: (\n\t\tattempt: number,\n\t\tmaxAttempts: number,\n\t\tdelayMs: number,\n\t\terrorMessage: string,\n\t) => void | Promise<void>;\n\t/** Emitted after the backoff sleep, immediately before the retried call starts. */\n\tonRetryAttemptStart?: () => void | Promise<void>;\n\t/** Emitted once when the loop ends: success if a later call completed normally. */\n\tonRetryFinished?: (success: boolean, attempt: number, finalError?: string) => void | Promise<void>;\n}\n\nclass RetrySleepAbortError extends Error {\n\tconstructor() {\n\t\tsuper(\"Aborted\");\n\t}\n}\n\nfunction sleep(ms: number, signal?: AbortSignal): Promise<void> {\n\treturn new Promise((resolve, reject) => {\n\t\tif (signal?.aborted) {\n\t\t\treject(new RetrySleepAbortError());\n\t\t\treturn;\n\t\t}\n\t\tconst timeout = setTimeout(resolve, ms);\n\t\tsignal?.addEventListener(\n\t\t\t\"abort\",\n\t\t\t() => {\n\t\t\t\tclearTimeout(timeout);\n\t\t\t\treject(new RetrySleepAbortError());\n\t\t\t},\n\t\t\t{ once: true },\n\t\t);\n\t});\n}\n\n/**\n * Run a single assistant-producing call with bounded retry on transient errors.\n *\n * Behavior:\n * - A successful response is returned immediately. Aborts are terminal and never\n *   retried, but reported as unsuccessful if they happen after a retry was scheduled.\n *   Aborts during the backoff sleep are normalized to an aborted `AssistantMessage`\n *   too, so callers do not need to care when cancellation happened.\n * - A non-retryable error (per {@link isRetryableAssistantError}, including quota/\n *   billing exhaustion) is returned immediately so deterministic errors fail fast.\n * - Otherwise retries up to `maxRetries` times with exponential backoff, emitting\n *   `onRetryScheduled` before each sleep, `onRetryAttemptStart` after each sleep before\n *   the retried call starts, and `onRetryFinished` once at the end (whether the loop\n *   ends in success, exhausted retries, or an aborted backoff).\n *\n * When `policy` is undefined or disabled, the first response is returned unchanged\n * (equivalent to calling `produce()` directly).\n */\nexport async function retryAssistantCall(\n\tproduce: () => Promise<AssistantMessage>,\n\tpolicy: RetryPolicy | undefined,\n\tsignal: AbortSignal | undefined,\n\tcallbacks?: RetryCallbacks,\n): Promise<AssistantMessage> {\n\tconst maxAttempts = policy?.enabled ? policy.maxRetries : 0;\n\n\tlet attempt = 0;\n\tlet lastRetry: { attempt: number; errorMessage: string } | undefined;\n\tfor (;;) {\n\t\tconst response = await produce();\n\n\t\t// Abort: terminal but not successful. Never retry an aborted message.\n\t\tif (response.stopReason === \"aborted\") {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(false, lastRetry.attempt);\n\t\t\treturn response;\n\t\t}\n\n\t\t// Success: non-error, non-abort responses return as-is.\n\t\tif (response.stopReason !== \"error\") {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(true, lastRetry.attempt);\n\t\t\treturn response;\n\t\t}\n\n\t\t// Non-retryable, or budget exhausted: return the final error message.\n\t\tif (attempt >= maxAttempts || !isRetryableAssistantError(response)) {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(false, lastRetry.attempt, response.errorMessage);\n\t\t\treturn response;\n\t\t}\n\n\t\tattempt++;\n\t\tlastRetry = { attempt, errorMessage: response.errorMessage || \"Unknown error\" };\n\t\tconst delayMs = retryDelayMs(policy!, attempt);\n\t\tawait callbacks?.onRetryScheduled?.(attempt, maxAttempts, delayMs, lastRetry.errorMessage);\n\n\t\t// Normalize aborts during retry backoff to the same AssistantMessage shape as\n\t\t// provider stream aborts, so callers do not need to care when cancellation happened.\n\t\ttry {\n\t\t\tawait sleep(delayMs, signal);\n\t\t} catch (error) {\n\t\t\tawait callbacks?.onRetryFinished?.(false, attempt, lastRetry.errorMessage);\n\t\t\tif (error instanceof RetrySleepAbortError) {\n\t\t\t\treturn { ...response, stopReason: \"aborted\", errorMessage: undefined };\n\t\t\t}\n\t\t\tthrow error;\n\t\t}\n\t\tawait callbacks?.onRetryAttemptStart?.();\n\t}\n}\n\n/**\n * Classifies whether a failed assistant message looks like a transient provider\n * or transport error, so callers can decide if the last assistant turn should be\n * restarted.\n *\n * This does not implement retry policy. Callers should first handle context\n * overflow separately, then apply their own retry budget, backoff, and reporting\n * before restarting the assistant turn.\n */\nexport function isRetryableAssistantError(message: AssistantMessage): boolean {\n\tif (message.stopReason !== \"error\" || !message.errorMessage) return false;\n\tconst errorMessage = message.errorMessage;\n\tif (NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN.test(errorMessage)) return false;\n\treturn RETRYABLE_PROVIDER_ERROR_PATTERN.test(errorMessage);\n}\n"]}

package/dist/utils/retry.js

@@ -89,6 +89,12 @@ const RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
     // gRPC based providers (e.g. NVIDIA NIM)
     "ResourceExhausted",
 ]);
+export const DEFAULT_MAX_AGENT_RETRY_DELAY_MS = 60_000;
+export function retryDelayMs(policy, attempt) {
+    const delay = policy.baseDelayMs * 2 ** Math.max(0, attempt - 1);
+    const safeDelay = Number.isSafeInteger(delay) ? delay : Number.MAX_SAFE_INTEGER;
+    return Math.min(safeDelay, policy.maxAgentDelayMs ?? DEFAULT_MAX_AGENT_RETRY_DELAY_MS);
+}
 class RetrySleepAbortError extends Error {
     constructor() {
         super("Aborted");
@@ -151,7 +157,7 @@ export async function retryAssistantCall(produce, policy, signal, callbacks) {
         }
         attempt++;
         lastRetry = { attempt, errorMessage: response.errorMessage || "Unknown error" };
-        const delayMs = policy.baseDelayMs * 2 ** (attempt - 1);
+        const delayMs = retryDelayMs(policy, attempt);
         await callbacks?.onRetryScheduled?.(attempt, maxAttempts, delayMs, lastRetry.errorMessage);
         // Normalize aborts during retry backoff to the same AssistantMessage shape as
         // provider stream aborts, so callers do not need to care when cancellation happened.

package/dist/utils/retry.js.map

@@ -1 +1 @@
-{"version":3,"file":"retry.js","sourceRoot":"","sources":["../../src/utils/retry.ts"],"names":[],"mappings":"AAEA,SAAS,yBAAyB,CAAC,QAA2B,EAAU;IACvE,OAAO,IAAI,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,GAAG,CAAC,CAAC;AAAA,CAC3C;AAED,MAAM,0CAA0C,GAAG,yBAAyB,CAAC;IAC5E,8EAA8E;IAC9E,2EAA2E;IAC3E,mBAAmB;IACnB,qBAAqB;IAErB,6EAA6E;IAC7E,yDAAyD;IACzD,6BAA6B;IAC7B,mBAAmB;IAEnB,4EAA4E;IAC5E,4EAA4E;IAC5E,oBAAoB;IACpB,eAAe;IACf,gBAAgB;IAChB,SAAS;CACT,CAAC,CAAC;AAEH,MAAM,gCAAgC,GAAG,yBAAyB,CAAC;IAClE,0EAA0E;IAC1E,YAAY;IACZ,aAAa;IACb,mBAAmB;IACnB,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,sBAAsB;IACtB,eAAe;IACf,iBAAiB;IAEjB,8EAA8E;IAC9E,+CAA+C;IAC/C,2BAA2B;IAC3B,uDAAuD;IAEvD,2EAA2E;IAC3E,2EAA2E;IAC3E,2EAA2E;IAC3E,gBAAgB;IAChB,mBAAmB;IACnB,qBAAqB;IACrB,kBAAkB;IAClB,mBAAmB;IACnB,cAAc;IACd,aAAa;IACb,WAAW;IACX,WAAW;IACX,mBAAmB;IACnB,sBAAsB;IACtB,gBAAgB;IAChB,8BAA8B;IAC9B,YAAY;IACZ,SAAS;IACT,YAAY;IAEZ,+EAA+E;IAC/E,mBAAmB;IACnB,kBAAkB;IAElB,2EAA2E;IAC3E,0EAA0E;IAC1E,wEAAwE;IACxE,6DAA6D;IAC7D,4EAA4E;IAC5E,6EAA6E;IAC7E,8EAA8E;IAC9E,MAAM;IACN,wBAAwB;IACxB,YAAY;IACZ,eAAe;IAEf,6EAA6E;IAC7E,yEAAyE;IACzE,0BAA0B;IAE1B,yEAAyE;IACzE,8EAA8E;IAC9E,yEAAyE;IACzE,eAAe;IACf,kCAAkC;IAClC,+CAA+C;IAC/C,sCAAsC;IAEtC,4EAA4E;IAC5E,iEAAiE;IACjE,aAAa;IAEb,6EAA6E;IAC7E,6BAA6B;IAC7B,4BAA4B;IAC5B,wBAAwB;IACxB,2BAA2B;IAE3B,yCAAyC;IACzC,mBAAmB;CACnB,CAAC,CAAC;AA+BH,MAAM,oBAAqB,SAAQ,KAAK;IACvC,cAAc;QACb,KAAK,CAAC,SAAS,CAAC,CAAC;IAAA,CACjB;CACD;AAED,SAAS,KAAK,CAAC,EAAU,EAAE,MAAoB,EAAiB;IAC/D,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE,CAAC;QACvC,IAAI,MAAM,EAAE,OAAO,EAAE,CAAC;YACrB,MAAM,CAAC,IAAI,oBAAoB,EAAE,CAAC,CAAC;YACnC,OAAO;QACR,CAAC;QACD,MAAM,OAAO,GAAG,UAAU,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC;QACxC,MAAM,EAAE,gBAAgB,CACvB,OAAO,EACP,GAAG,EAAE,CAAC;YACL,YAAY,CAAC,OAAO,CAAC,CAAC;YACtB,MAAM,CAAC,IAAI,oBAAoB,EAAE,CAAC,CAAC;QAAA,CACnC,EACD,EAAE,IAAI,EAAE,IAAI,EAAE,CACd,CAAC;IAAA,CACF,CAAC,CAAC;AAAA,CACH;AAED;;;;;;;;;;;;;;;;;GAiBG;AACH,MAAM,CAAC,KAAK,UAAU,kBAAkB,CACvC,OAAwC,EACxC,MAA+B,EAC/B,MAA+B,EAC/B,SAA0B,EACE;IAC5B,MAAM,WAAW,GAAG,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC;IAE5D,IAAI,OAAO,GAAG,CAAC,CAAC;IAChB,IAAI,SAAgE,CAAC;IACrE,SAAS,CAAC;QACT,MAAM,QAAQ,GAAG,MAAM,OAAO,EAAE,CAAC;QAEjC,sEAAsE;QACtE,IAAI,QAAQ,CAAC,UAAU,KAAK,SAAS,EAAE,CAAC;YACvC,IAAI,SAAS;gBAAE,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC,OAAO,CAAC,CAAC;YAC5E,OAAO,QAAQ,CAAC;QACjB,CAAC;QAED,wDAAwD;QACxD,IAAI,QAAQ,CAAC,UAAU,KAAK,OAAO,EAAE,CAAC;YACrC,IAAI,SAAS;gBAAE,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,IAAI,EAAE,SAAS,CAAC,OAAO,CAAC,CAAC;YAC3E,OAAO,QAAQ,CAAC;QACjB,CAAC;QAED,sEAAsE;QACtE,IAAI,OAAO,IAAI,WAAW,IAAI,CAAC,yBAAyB,CAAC,QAAQ,CAAC,EAAE,CAAC;YACpE,IAAI,SAAS;gBAAE,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC,OAAO,EAAE,QAAQ,CAAC,YAAY,CAAC,CAAC;YACnG,OAAO,QAAQ,CAAC;QACjB,CAAC;QAED,OAAO,EAAE,CAAC;QACV,SAAS,GAAG,EAAE,OAAO,EAAE,YAAY,EAAE,QAAQ,CAAC,YAAY,IAAI,eAAe,EAAE,CAAC;QAChF,MAAM,OAAO,GAAG,MAAO,CAAC,WAAW,GAAG,CAAC,IAAI,CAAC,OAAO,GAAG,CAAC,CAAC,CAAC;QACzD,MAAM,SAAS,EAAE,gBAAgB,EAAE,CAAC,OAAO,EAAE,WAAW,EAAE,OAAO,EAAE,SAAS,CAAC,YAAY,CAAC,CAAC;QAE3F,8EAA8E;QAC9E,qFAAqF;QACrF,IAAI,CAAC;YACJ,MAAM,KAAK,CAAC,OAAO,EAAE,MAAM,CAAC,CAAC;QAC9B,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,KAAK,EAAE,OAAO,EAAE,SAAS,CAAC,YAAY,CAAC,CAAC;YAC3E,IAAI,KAAK,YAAY,oBAAoB,EAAE,CAAC;gBAC3C,OAAO,EAAE,GAAG,QAAQ,EAAE,UAAU,EAAE,SAAS,EAAE,YAAY,EAAE,SAAS,EAAE,CAAC;YACxE,CAAC;YACD,MAAM,KAAK,CAAC;QACb,CAAC;QACD,MAAM,SAAS,EAAE,mBAAmB,EAAE,EAAE,CAAC;IAC1C,CAAC;AAAA,CACD;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,yBAAyB,CAAC,OAAyB,EAAW;IAC7E,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,CAAC,OAAO,CAAC,YAAY;QAAE,OAAO,KAAK,CAAC;IAC1E,MAAM,YAAY,GAAG,OAAO,CAAC,YAAY,CAAC;IAC1C,IAAI,0CAA0C,CAAC,IAAI,CAAC,YAAY,CAAC;QAAE,OAAO,KAAK,CAAC;IAChF,OAAO,gCAAgC,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC;AAAA,CAC3D","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\n\nfunction buildProviderErrorPattern(patterns: readonly string[]): RegExp {\n\treturn new RegExp(patterns.join(\"|\"), \"i\");\n}\n\nconst NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN = buildProviderErrorPattern([\n\t// OpenCode Go/free-tier limits returned as 429 JSON error types by OpenCode's\n\t// Zen API. These are subscription/account limits, not transient throttles.\n\t\"GoUsageLimitError\",\n\t\"FreeUsageLimitError\",\n\n\t// OpenCode Go subscription-limit text asks users to enable available-balance\n\t// usage after rolling/weekly/monthly limits are reached.\n\t\"Monthly usage limit reached\",\n\t\"available balance\",\n\n\t// Generic quota/budget/billing exhaustion. `insufficient_quota` is OpenAI's\n\t// quota/billing error code; the other strings cover common gateway wording.\n\t\"insufficient_quota\",\n\t\"out of budget\",\n\t\"quota exceeded\",\n\t\"billing\",\n]);\n\nconst RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([\n\t// Generic provider load, HTTP status, and server-side transient failures.\n\t\"overloaded\",\n\t\"rate.?limit\",\n\t\"too many requests\",\n\t\"429\",\n\t\"500\",\n\t\"502\",\n\t\"503\",\n\t\"504\",\n\t\"524\",\n\t\"service.?unavailable\",\n\t\"server.?error\",\n\t\"internal.?error\",\n\n\t// Wrapper/provider text for transient upstream failures, including OpenRouter\n\t// \"Provider returned error\" responses (#2264).\n\t\"provider.?returned.?error\",\n\t\"exceeded request buffer limit while retrying upstream\",\n\n\t// Network, proxy, and fetch transport failures. This includes OpenAI Codex\n\t// raw-fetch failures such as \"upstream connect\", \"connection refused\", and\n\t// \"reset before headers\" (#733), plus OpenRouter connection drops (#3317).\n\t\"network.?error\",\n\t\"connection.?error\",\n\t\"connection.?refused\",\n\t\"connection.?lost\",\n\t\"other side closed\",\n\t\"fetch failed\",\n\t\"getaddrinfo\",\n\t\"ENOTFOUND\",\n\t\"EAI_AGAIN\",\n\t\"upstream.?connect\",\n\t\"reset before headers\",\n\t\"socket hang up\",\n\t\"socket connection was closed\",\n\t\"timed? out\",\n\t\"timeout\",\n\t\"terminated\",\n\n\t// WebSocket transports can report close/error text instead of HTTP/fetch text.\n\t\"websocket.?closed\",\n\t\"websocket.?error\",\n\n\t// Response-body decompression failures. Node/Bun zlib rejects a body whose\n\t// Content-Encoding lied or whose compressed bytes were corrupted, and the\n\t// fetch/WebSocket layer wraps it with a \"Library error:\" prefix. GitHub\n\t// Copilot on the default `transport: \"auto\"` reports this as\n\t// \"Library error: zlib error: incorrect header check\" (#2553). The transfer\n\t// is already unusable, and a fresh attempt normally succeeds, so treat every\n\t// variant as a transient transport failure rather than a deterministic error.\n\t\"zlib\",\n\t\"incorrect header check\",\n\t\"decompress\",\n\t\"library error\",\n\n\t// Idle-stream deadline enforced below the HTTP layer by `withStreamDeadline`\n\t// so a stream that never settles cannot hang an attempt forever (#2553).\n\t\"stream deadline exceeded\",\n\n\t// Premature stream endings from SDKs and transports. Anthropic can throw\n\t// \"stream ended without ...\" and \"Anthropic stream ended before message_stop\"\n\t// (#4433); Bedrock/Smithy can throw an HTTP/2 no-response error (#3594).\n\t\"ended without\",\n\t\"stream ended before message_stop\",\n\t\"stream ended before a terminal response event\",\n\t\"http2 request did not get a response\",\n\n\t// Provider-requested retry delay cap failures should flow through the outer\n\t// retry policy so callers can surface/abort the backoff (#1123).\n\t\"retry delay\",\n\n\t// Explicit retry guidance emitted mid-stream by OpenAI Responses and Bedrock\n\t// stream exceptions (#6019).\n\t\"you can retry your request\",\n\t\"try your request again\",\n\t\"please retry your request\",\n\n\t// gRPC based providers (e.g. NVIDIA NIM)\n\t\"ResourceExhausted\",\n]);\n\n/**\n * Retry policy: bounded attempts with exponential backoff (`baseDelayMs * 2^(attempt-1)`).\n * Matches `settings.retry` (`enabled`, `maxRetries`, `baseDelayMs`) in coding-agent; kept\n * here so the classifier and the policy-driven retry loop live together and stay reusable\n * by the SDK and other callers.\n */\nexport interface RetryPolicy {\n\tenabled: boolean;\n\t/** Max retry attempts (0 = no retries). The initial call never counts as a retry. */\n\tmaxRetries: number;\n\t/** Base delay in ms. Per-attempt delay is `baseDelayMs * 2^(attempt-1)` before jitter. */\n\tbaseDelayMs: number;\n}\n\n/** Optional callbacks emitted by {@link retryAssistantCall} around each retry. */\nexport interface RetryCallbacks {\n\t/** Emitted before the backoff sleep of each retry attempt (1-indexed). */\n\tonRetryScheduled?: (\n\t\tattempt: number,\n\t\tmaxAttempts: number,\n\t\tdelayMs: number,\n\t\terrorMessage: string,\n\t) => void | Promise<void>;\n\t/** Emitted after the backoff sleep, immediately before the retried call starts. */\n\tonRetryAttemptStart?: () => void | Promise<void>;\n\t/** Emitted once when the loop ends: success if a later call completed normally. */\n\tonRetryFinished?: (success: boolean, attempt: number, finalError?: string) => void | Promise<void>;\n}\n\nclass RetrySleepAbortError extends Error {\n\tconstructor() {\n\t\tsuper(\"Aborted\");\n\t}\n}\n\nfunction sleep(ms: number, signal?: AbortSignal): Promise<void> {\n\treturn new Promise((resolve, reject) => {\n\t\tif (signal?.aborted) {\n\t\t\treject(new RetrySleepAbortError());\n\t\t\treturn;\n\t\t}\n\t\tconst timeout = setTimeout(resolve, ms);\n\t\tsignal?.addEventListener(\n\t\t\t\"abort\",\n\t\t\t() => {\n\t\t\t\tclearTimeout(timeout);\n\t\t\t\treject(new RetrySleepAbortError());\n\t\t\t},\n\t\t\t{ once: true },\n\t\t);\n\t});\n}\n\n/**\n * Run a single assistant-producing call with bounded retry on transient errors.\n *\n * Behavior:\n * - A successful response is returned immediately. Aborts are terminal and never\n *   retried, but reported as unsuccessful if they happen after a retry was scheduled.\n *   Aborts during the backoff sleep are normalized to an aborted `AssistantMessage`\n *   too, so callers do not need to care when cancellation happened.\n * - A non-retryable error (per {@link isRetryableAssistantError}, including quota/\n *   billing exhaustion) is returned immediately so deterministic errors fail fast.\n * - Otherwise retries up to `maxRetries` times with exponential backoff, emitting\n *   `onRetryScheduled` before each sleep, `onRetryAttemptStart` after each sleep before\n *   the retried call starts, and `onRetryFinished` once at the end (whether the loop\n *   ends in success, exhausted retries, or an aborted backoff).\n *\n * When `policy` is undefined or disabled, the first response is returned unchanged\n * (equivalent to calling `produce()` directly).\n */\nexport async function retryAssistantCall(\n\tproduce: () => Promise<AssistantMessage>,\n\tpolicy: RetryPolicy | undefined,\n\tsignal: AbortSignal | undefined,\n\tcallbacks?: RetryCallbacks,\n): Promise<AssistantMessage> {\n\tconst maxAttempts = policy?.enabled ? policy.maxRetries : 0;\n\n\tlet attempt = 0;\n\tlet lastRetry: { attempt: number; errorMessage: string } | undefined;\n\tfor (;;) {\n\t\tconst response = await produce();\n\n\t\t// Abort: terminal but not successful. Never retry an aborted message.\n\t\tif (response.stopReason === \"aborted\") {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(false, lastRetry.attempt);\n\t\t\treturn response;\n\t\t}\n\n\t\t// Success: non-error, non-abort responses return as-is.\n\t\tif (response.stopReason !== \"error\") {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(true, lastRetry.attempt);\n\t\t\treturn response;\n\t\t}\n\n\t\t// Non-retryable, or budget exhausted: return the final error message.\n\t\tif (attempt >= maxAttempts || !isRetryableAssistantError(response)) {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(false, lastRetry.attempt, response.errorMessage);\n\t\t\treturn response;\n\t\t}\n\n\t\tattempt++;\n\t\tlastRetry = { attempt, errorMessage: response.errorMessage || \"Unknown error\" };\n\t\tconst delayMs = policy!.baseDelayMs * 2 ** (attempt - 1);\n\t\tawait callbacks?.onRetryScheduled?.(attempt, maxAttempts, delayMs, lastRetry.errorMessage);\n\n\t\t// Normalize aborts during retry backoff to the same AssistantMessage shape as\n\t\t// provider stream aborts, so callers do not need to care when cancellation happened.\n\t\ttry {\n\t\t\tawait sleep(delayMs, signal);\n\t\t} catch (error) {\n\t\t\tawait callbacks?.onRetryFinished?.(false, attempt, lastRetry.errorMessage);\n\t\t\tif (error instanceof RetrySleepAbortError) {\n\t\t\t\treturn { ...response, stopReason: \"aborted\", errorMessage: undefined };\n\t\t\t}\n\t\t\tthrow error;\n\t\t}\n\t\tawait callbacks?.onRetryAttemptStart?.();\n\t}\n}\n\n/**\n * Classifies whether a failed assistant message looks like a transient provider\n * or transport error, so callers can decide if the last assistant turn should be\n * restarted.\n *\n * This does not implement retry policy. Callers should first handle context\n * overflow separately, then apply their own retry budget, backoff, and reporting\n * before restarting the assistant turn.\n */\nexport function isRetryableAssistantError(message: AssistantMessage): boolean {\n\tif (message.stopReason !== \"error\" || !message.errorMessage) return false;\n\tconst errorMessage = message.errorMessage;\n\tif (NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN.test(errorMessage)) return false;\n\treturn RETRYABLE_PROVIDER_ERROR_PATTERN.test(errorMessage);\n}\n"]}
+{"version":3,"file":"retry.js","sourceRoot":"","sources":["../../src/utils/retry.ts"],"names":[],"mappings":"AAEA,SAAS,yBAAyB,CAAC,QAA2B,EAAU;IACvE,OAAO,IAAI,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,GAAG,CAAC,CAAC;AAAA,CAC3C;AAED,MAAM,0CAA0C,GAAG,yBAAyB,CAAC;IAC5E,8EAA8E;IAC9E,2EAA2E;IAC3E,mBAAmB;IACnB,qBAAqB;IAErB,6EAA6E;IAC7E,yDAAyD;IACzD,6BAA6B;IAC7B,mBAAmB;IAEnB,4EAA4E;IAC5E,4EAA4E;IAC5E,oBAAoB;IACpB,eAAe;IACf,gBAAgB;IAChB,SAAS;CACT,CAAC,CAAC;AAEH,MAAM,gCAAgC,GAAG,yBAAyB,CAAC;IAClE,0EAA0E;IAC1E,YAAY;IACZ,aAAa;IACb,mBAAmB;IACnB,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,sBAAsB;IACtB,eAAe;IACf,iBAAiB;IAEjB,8EAA8E;IAC9E,+CAA+C;IAC/C,2BAA2B;IAC3B,uDAAuD;IAEvD,2EAA2E;IAC3E,2EAA2E;IAC3E,2EAA2E;IAC3E,gBAAgB;IAChB,mBAAmB;IACnB,qBAAqB;IACrB,kBAAkB;IAClB,mBAAmB;IACnB,cAAc;IACd,aAAa;IACb,WAAW;IACX,WAAW;IACX,mBAAmB;IACnB,sBAAsB;IACtB,gBAAgB;IAChB,8BAA8B;IAC9B,YAAY;IACZ,SAAS;IACT,YAAY;IAEZ,+EAA+E;IAC/E,mBAAmB;IACnB,kBAAkB;IAElB,2EAA2E;IAC3E,0EAA0E;IAC1E,wEAAwE;IACxE,6DAA6D;IAC7D,4EAA4E;IAC5E,6EAA6E;IAC7E,8EAA8E;IAC9E,MAAM;IACN,wBAAwB;IACxB,YAAY;IACZ,eAAe;IAEf,6EAA6E;IAC7E,yEAAyE;IACzE,0BAA0B;IAE1B,yEAAyE;IACzE,8EAA8E;IAC9E,yEAAyE;IACzE,eAAe;IACf,kCAAkC;IAClC,+CAA+C;IAC/C,sCAAsC;IAEtC,4EAA4E;IAC5E,iEAAiE;IACjE,aAAa;IAEb,6EAA6E;IAC7E,6BAA6B;IAC7B,4BAA4B;IAC5B,wBAAwB;IACxB,2BAA2B;IAE3B,yCAAyC;IACzC,mBAAmB;CACnB,CAAC,CAAC;AAmBH,MAAM,CAAC,MAAM,gCAAgC,GAAG,MAAM,CAAC;AAEvD,MAAM,UAAU,YAAY,CAAC,MAA4D,EAAE,OAAe,EAAU;IACnH,MAAM,KAAK,GAAG,MAAM,CAAC,WAAW,GAAG,CAAC,IAAI,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,OAAO,GAAG,CAAC,CAAC,CAAC;IACjE,MAAM,SAAS,GAAG,MAAM,CAAC,aAAa,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,MAAM,CAAC,gBAAgB,CAAC;IAChF,OAAO,IAAI,CAAC,GAAG,CAAC,SAAS,EAAE,MAAM,CAAC,eAAe,IAAI,gCAAgC,CAAC,CAAC;AAAA,CACvF;AAiBD,MAAM,oBAAqB,SAAQ,KAAK;IACvC,cAAc;QACb,KAAK,CAAC,SAAS,CAAC,CAAC;IAAA,CACjB;CACD;AAED,SAAS,KAAK,CAAC,EAAU,EAAE,MAAoB,EAAiB;IAC/D,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE,CAAC;QACvC,IAAI,MAAM,EAAE,OAAO,EAAE,CAAC;YACrB,MAAM,CAAC,IAAI,oBAAoB,EAAE,CAAC,CAAC;YACnC,OAAO;QACR,CAAC;QACD,MAAM,OAAO,GAAG,UAAU,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC;QACxC,MAAM,EAAE,gBAAgB,CACvB,OAAO,EACP,GAAG,EAAE,CAAC;YACL,YAAY,CAAC,OAAO,CAAC,CAAC;YACtB,MAAM,CAAC,IAAI,oBAAoB,EAAE,CAAC,CAAC;QAAA,CACnC,EACD,EAAE,IAAI,EAAE,IAAI,EAAE,CACd,CAAC;IAAA,CACF,CAAC,CAAC;AAAA,CACH;AAED;;;;;;;;;;;;;;;;;GAiBG;AACH,MAAM,CAAC,KAAK,UAAU,kBAAkB,CACvC,OAAwC,EACxC,MAA+B,EAC/B,MAA+B,EAC/B,SAA0B,EACE;IAC5B,MAAM,WAAW,GAAG,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC;IAE5D,IAAI,OAAO,GAAG,CAAC,CAAC;IAChB,IAAI,SAAgE,CAAC;IACrE,SAAS,CAAC;QACT,MAAM,QAAQ,GAAG,MAAM,OAAO,EAAE,CAAC;QAEjC,sEAAsE;QACtE,IAAI,QAAQ,CAAC,UAAU,KAAK,SAAS,EAAE,CAAC;YACvC,IAAI,SAAS;gBAAE,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC,OAAO,CAAC,CAAC;YAC5E,OAAO,QAAQ,CAAC;QACjB,CAAC;QAED,wDAAwD;QACxD,IAAI,QAAQ,CAAC,UAAU,KAAK,OAAO,EAAE,CAAC;YACrC,IAAI,SAAS;gBAAE,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,IAAI,EAAE,SAAS,CAAC,OAAO,CAAC,CAAC;YAC3E,OAAO,QAAQ,CAAC;QACjB,CAAC;QAED,sEAAsE;QACtE,IAAI,OAAO,IAAI,WAAW,IAAI,CAAC,yBAAyB,CAAC,QAAQ,CAAC,EAAE,CAAC;YACpE,IAAI,SAAS;gBAAE,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC,OAAO,EAAE,QAAQ,CAAC,YAAY,CAAC,CAAC;YACnG,OAAO,QAAQ,CAAC;QACjB,CAAC;QAED,OAAO,EAAE,CAAC;QACV,SAAS,GAAG,EAAE,OAAO,EAAE,YAAY,EAAE,QAAQ,CAAC,YAAY,IAAI,eAAe,EAAE,CAAC;QAChF,MAAM,OAAO,GAAG,YAAY,CAAC,MAAO,EAAE,OAAO,CAAC,CAAC;QAC/C,MAAM,SAAS,EAAE,gBAAgB,EAAE,CAAC,OAAO,EAAE,WAAW,EAAE,OAAO,EAAE,SAAS,CAAC,YAAY,CAAC,CAAC;QAE3F,8EAA8E;QAC9E,qFAAqF;QACrF,IAAI,CAAC;YACJ,MAAM,KAAK,CAAC,OAAO,EAAE,MAAM,CAAC,CAAC;QAC9B,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,KAAK,EAAE,OAAO,EAAE,SAAS,CAAC,YAAY,CAAC,CAAC;YAC3E,IAAI,KAAK,YAAY,oBAAoB,EAAE,CAAC;gBAC3C,OAAO,EAAE,GAAG,QAAQ,EAAE,UAAU,EAAE,SAAS,EAAE,YAAY,EAAE,SAAS,EAAE,CAAC;YACxE,CAAC;YACD,MAAM,KAAK,CAAC;QACb,CAAC;QACD,MAAM,SAAS,EAAE,mBAAmB,EAAE,EAAE,CAAC;IAC1C,CAAC;AAAA,CACD;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,yBAAyB,CAAC,OAAyB,EAAW;IAC7E,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,CAAC,OAAO,CAAC,YAAY;QAAE,OAAO,KAAK,CAAC;IAC1E,MAAM,YAAY,GAAG,OAAO,CAAC,YAAY,CAAC;IAC1C,IAAI,0CAA0C,CAAC,IAAI,CAAC,YAAY,CAAC;QAAE,OAAO,KAAK,CAAC;IAChF,OAAO,gCAAgC,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC;AAAA,CAC3D","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\n\nfunction buildProviderErrorPattern(patterns: readonly string[]): RegExp {\n\treturn new RegExp(patterns.join(\"|\"), \"i\");\n}\n\nconst NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN = buildProviderErrorPattern([\n\t// OpenCode Go/free-tier limits returned as 429 JSON error types by OpenCode's\n\t// Zen API. These are subscription/account limits, not transient throttles.\n\t\"GoUsageLimitError\",\n\t\"FreeUsageLimitError\",\n\n\t// OpenCode Go subscription-limit text asks users to enable available-balance\n\t// usage after rolling/weekly/monthly limits are reached.\n\t\"Monthly usage limit reached\",\n\t\"available balance\",\n\n\t// Generic quota/budget/billing exhaustion. `insufficient_quota` is OpenAI's\n\t// quota/billing error code; the other strings cover common gateway wording.\n\t\"insufficient_quota\",\n\t\"out of budget\",\n\t\"quota exceeded\",\n\t\"billing\",\n]);\n\nconst RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([\n\t// Generic provider load, HTTP status, and server-side transient failures.\n\t\"overloaded\",\n\t\"rate.?limit\",\n\t\"too many requests\",\n\t\"429\",\n\t\"500\",\n\t\"502\",\n\t\"503\",\n\t\"504\",\n\t\"524\",\n\t\"service.?unavailable\",\n\t\"server.?error\",\n\t\"internal.?error\",\n\n\t// Wrapper/provider text for transient upstream failures, including OpenRouter\n\t// \"Provider returned error\" responses (#2264).\n\t\"provider.?returned.?error\",\n\t\"exceeded request buffer limit while retrying upstream\",\n\n\t// Network, proxy, and fetch transport failures. This includes OpenAI Codex\n\t// raw-fetch failures such as \"upstream connect\", \"connection refused\", and\n\t// \"reset before headers\" (#733), plus OpenRouter connection drops (#3317).\n\t\"network.?error\",\n\t\"connection.?error\",\n\t\"connection.?refused\",\n\t\"connection.?lost\",\n\t\"other side closed\",\n\t\"fetch failed\",\n\t\"getaddrinfo\",\n\t\"ENOTFOUND\",\n\t\"EAI_AGAIN\",\n\t\"upstream.?connect\",\n\t\"reset before headers\",\n\t\"socket hang up\",\n\t\"socket connection was closed\",\n\t\"timed? out\",\n\t\"timeout\",\n\t\"terminated\",\n\n\t// WebSocket transports can report close/error text instead of HTTP/fetch text.\n\t\"websocket.?closed\",\n\t\"websocket.?error\",\n\n\t// Response-body decompression failures. Node/Bun zlib rejects a body whose\n\t// Content-Encoding lied or whose compressed bytes were corrupted, and the\n\t// fetch/WebSocket layer wraps it with a \"Library error:\" prefix. GitHub\n\t// Copilot on the default `transport: \"auto\"` reports this as\n\t// \"Library error: zlib error: incorrect header check\" (#2553). The transfer\n\t// is already unusable, and a fresh attempt normally succeeds, so treat every\n\t// variant as a transient transport failure rather than a deterministic error.\n\t\"zlib\",\n\t\"incorrect header check\",\n\t\"decompress\",\n\t\"library error\",\n\n\t// Idle-stream deadline enforced below the HTTP layer by `withStreamDeadline`\n\t// so a stream that never settles cannot hang an attempt forever (#2553).\n\t\"stream deadline exceeded\",\n\n\t// Premature stream endings from SDKs and transports. Anthropic can throw\n\t// \"stream ended without ...\" and \"Anthropic stream ended before message_stop\"\n\t// (#4433); Bedrock/Smithy can throw an HTTP/2 no-response error (#3594).\n\t\"ended without\",\n\t\"stream ended before message_stop\",\n\t\"stream ended before a terminal response event\",\n\t\"http2 request did not get a response\",\n\n\t// Provider-requested retry delay cap failures should flow through the outer\n\t// retry policy so callers can surface/abort the backoff (#1123).\n\t\"retry delay\",\n\n\t// Explicit retry guidance emitted mid-stream by OpenAI Responses and Bedrock\n\t// stream exceptions (#6019).\n\t\"you can retry your request\",\n\t\"try your request again\",\n\t\"please retry your request\",\n\n\t// gRPC based providers (e.g. NVIDIA NIM)\n\t\"ResourceExhausted\",\n]);\n\n/**\n * Retry policy: bounded attempts with exponential backoff (`baseDelayMs * 2^(attempt-1)`).\n * Each delay is capped by `maxAgentDelayMs` (60 seconds by default).\n * Matches `settings.retry` (`enabled`, `maxRetries`, `baseDelayMs`) in coding-agent; kept\n * here so the classifier and the policy-driven retry loop live together and stay reusable\n * by the SDK and other callers.\n */\nexport interface RetryPolicy {\n\tenabled: boolean;\n\t/** Max retry attempts (0 = no retries). The initial call never counts as a retry. */\n\tmaxRetries: number;\n\t/** Base delay in ms. Per-attempt delay is `baseDelayMs * 2^(attempt-1)` before jitter. */\n\tbaseDelayMs: number;\n\t/** Cap for agent-level retry delays in ms. Defaults to 60 seconds. */\n\tmaxAgentDelayMs?: number;\n}\n\nexport const DEFAULT_MAX_AGENT_RETRY_DELAY_MS = 60_000;\n\nexport function retryDelayMs(policy: Pick<RetryPolicy, \"baseDelayMs\" | \"maxAgentDelayMs\">, attempt: number): number {\n\tconst delay = policy.baseDelayMs * 2 ** Math.max(0, attempt - 1);\n\tconst safeDelay = Number.isSafeInteger(delay) ? delay : Number.MAX_SAFE_INTEGER;\n\treturn Math.min(safeDelay, policy.maxAgentDelayMs ?? DEFAULT_MAX_AGENT_RETRY_DELAY_MS);\n}\n\n/** Optional callbacks emitted by {@link retryAssistantCall} around each retry. */\nexport interface RetryCallbacks {\n\t/** Emitted before the backoff sleep of each retry attempt (1-indexed). */\n\tonRetryScheduled?: (\n\t\tattempt: number,\n\t\tmaxAttempts: number,\n\t\tdelayMs: number,\n\t\terrorMessage: string,\n\t) => void | Promise<void>;\n\t/** Emitted after the backoff sleep, immediately before the retried call starts. */\n\tonRetryAttemptStart?: () => void | Promise<void>;\n\t/** Emitted once when the loop ends: success if a later call completed normally. */\n\tonRetryFinished?: (success: boolean, attempt: number, finalError?: string) => void | Promise<void>;\n}\n\nclass RetrySleepAbortError extends Error {\n\tconstructor() {\n\t\tsuper(\"Aborted\");\n\t}\n}\n\nfunction sleep(ms: number, signal?: AbortSignal): Promise<void> {\n\treturn new Promise((resolve, reject) => {\n\t\tif (signal?.aborted) {\n\t\t\treject(new RetrySleepAbortError());\n\t\t\treturn;\n\t\t}\n\t\tconst timeout = setTimeout(resolve, ms);\n\t\tsignal?.addEventListener(\n\t\t\t\"abort\",\n\t\t\t() => {\n\t\t\t\tclearTimeout(timeout);\n\t\t\t\treject(new RetrySleepAbortError());\n\t\t\t},\n\t\t\t{ once: true },\n\t\t);\n\t});\n}\n\n/**\n * Run a single assistant-producing call with bounded retry on transient errors.\n *\n * Behavior:\n * - A successful response is returned immediately. Aborts are terminal and never\n *   retried, but reported as unsuccessful if they happen after a retry was scheduled.\n *   Aborts during the backoff sleep are normalized to an aborted `AssistantMessage`\n *   too, so callers do not need to care when cancellation happened.\n * - A non-retryable error (per {@link isRetryableAssistantError}, including quota/\n *   billing exhaustion) is returned immediately so deterministic errors fail fast.\n * - Otherwise retries up to `maxRetries` times with exponential backoff, emitting\n *   `onRetryScheduled` before each sleep, `onRetryAttemptStart` after each sleep before\n *   the retried call starts, and `onRetryFinished` once at the end (whether the loop\n *   ends in success, exhausted retries, or an aborted backoff).\n *\n * When `policy` is undefined or disabled, the first response is returned unchanged\n * (equivalent to calling `produce()` directly).\n */\nexport async function retryAssistantCall(\n\tproduce: () => Promise<AssistantMessage>,\n\tpolicy: RetryPolicy | undefined,\n\tsignal: AbortSignal | undefined,\n\tcallbacks?: RetryCallbacks,\n): Promise<AssistantMessage> {\n\tconst maxAttempts = policy?.enabled ? policy.maxRetries : 0;\n\n\tlet attempt = 0;\n\tlet lastRetry: { attempt: number; errorMessage: string } | undefined;\n\tfor (;;) {\n\t\tconst response = await produce();\n\n\t\t// Abort: terminal but not successful. Never retry an aborted message.\n\t\tif (response.stopReason === \"aborted\") {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(false, lastRetry.attempt);\n\t\t\treturn response;\n\t\t}\n\n\t\t// Success: non-error, non-abort responses return as-is.\n\t\tif (response.stopReason !== \"error\") {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(true, lastRetry.attempt);\n\t\t\treturn response;\n\t\t}\n\n\t\t// Non-retryable, or budget exhausted: return the final error message.\n\t\tif (attempt >= maxAttempts || !isRetryableAssistantError(response)) {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(false, lastRetry.attempt, response.errorMessage);\n\t\t\treturn response;\n\t\t}\n\n\t\tattempt++;\n\t\tlastRetry = { attempt, errorMessage: response.errorMessage || \"Unknown error\" };\n\t\tconst delayMs = retryDelayMs(policy!, attempt);\n\t\tawait callbacks?.onRetryScheduled?.(attempt, maxAttempts, delayMs, lastRetry.errorMessage);\n\n\t\t// Normalize aborts during retry backoff to the same AssistantMessage shape as\n\t\t// provider stream aborts, so callers do not need to care when cancellation happened.\n\t\ttry {\n\t\t\tawait sleep(delayMs, signal);\n\t\t} catch (error) {\n\t\t\tawait callbacks?.onRetryFinished?.(false, attempt, lastRetry.errorMessage);\n\t\t\tif (error instanceof RetrySleepAbortError) {\n\t\t\t\treturn { ...response, stopReason: \"aborted\", errorMessage: undefined };\n\t\t\t}\n\t\t\tthrow error;\n\t\t}\n\t\tawait callbacks?.onRetryAttemptStart?.();\n\t}\n}\n\n/**\n * Classifies whether a failed assistant message looks like a transient provider\n * or transport error, so callers can decide if the last assistant turn should be\n * restarted.\n *\n * This does not implement retry policy. Callers should first handle context\n * overflow separately, then apply their own retry budget, backoff, and reporting\n * before restarting the assistant turn.\n */\nexport function isRetryableAssistantError(message: AssistantMessage): boolean {\n\tif (message.stopReason !== \"error\" || !message.errorMessage) return false;\n\tconst errorMessage = message.errorMessage;\n\tif (NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN.test(errorMessage)) return false;\n\treturn RETRYABLE_PROVIDER_ERROR_PATTERN.test(errorMessage);\n}\n"]}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bastani/pi-ai",
-  "version": "0.9.18",
+  "version": "0.9.19-alpha.10",
   "description": "Bastani-branded unified LLM API with automatic model discovery and provider configuration",
   "type": "module",
   "main": "./dist/index.js",
@@ -71,13 +71,13 @@
     "build:catalog": "npm run generate-models && npm run build:offline"
   },
   "dependencies": {
-    "@anthropic-ai/sdk": "0.91.1",
-    "@aws-sdk/client-bedrock-runtime": "3.1048.0",
+    "@anthropic-ai/sdk": "0.124.0",
+    "@aws-sdk/client-bedrock-runtime": "3.1127.0",
     "@earendil-works/pi-telemetry": "0.85.1",
-    "@google/genai": "1.52.0",
-    "@smithy/node-http-handler": "4.7.3",
-    "http-proxy-agent": "7.0.2",
-    "https-proxy-agent": "7.0.6",
+    "@google/genai": "2.21.0",
+    "@smithy/node-http-handler": "4.12.1",
+    "http-proxy-agent": "9.1.0",
+    "https-proxy-agent": "9.1.0",
     "openai": "6.40.0",
     "partial-json": "0.1.7",
     "typebox": "1.3.7"
@@ -88,7 +88,7 @@
     "canvas": "3.2.3",
     "shx": "0.4.0",
     "typescript": "5.9.3",
-    "vitest": "4.1.9"
+    "vitest": "4.1.11"
   },
   "keywords": [
     "ai",