@bastani/pi-ai 0.9.20-alpha.2 → 0.9.20-alpha.4

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: 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"]}
+{"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\n/**\n * Best-effort prompt cache lifetime in seconds for each retention tier a request can ask for.\n * A missing tier means the lifetime is unknown; pi does not warm such caches.\n */\nexport type ModelPromptCache = Partial<Record<Exclude<CacheRetention, \"none\">, number>>;\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: TranscriptContext, options?: StreamOptions): AssistantMessageEventStream;\n\tstreamSimple(\n\t\tmodel: Model<Api>,\n\t\tcontext: TranscriptContext,\n\t\toptions?: SimpleStreamOptions,\n\t): 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// - Receives a normalized transcript: the system prompt and tools live in the\n//   leading system message, never on the context itself.\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: TranscriptContext,\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: JsonObject;\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. Anthropic and Bedrock report 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 = null | boolean | number | string | readonly JsonValue[] | JsonObject;\nexport type JsonObject = { [key: string]: JsonValue };\n\ntype IsAny<T> = 0 extends 1 & T ? true : false;\ntype IsExactlyJsonValue<T> = [T] extends [JsonValue] ? ([JsonValue] extends [T] ? true : false) : false;\ntype IsJsonProperty<T> =\n\tIsAny<T> extends true\n\t\t? false\n\t\t: unknown extends T\n\t\t\t? false\n\t\t\t: [Exclude<T, undefined>] extends [never]\n\t\t\t\t? true\n\t\t\t\t: IsJsonCompatible<Exclude<T, undefined>>;\ntype InvalidJsonKeys<T extends object> = {\n\t[TKey in keyof T]-?: TKey extends string | number ? (IsJsonProperty<T[TKey]> extends true ? never : TKey) : TKey;\n}[keyof T];\ntype IsJsonCompatible<T> =\n\tIsAny<T> extends true\n\t\t? false\n\t\t: unknown extends T\n\t\t\t? false\n\t\t\t: IsExactlyJsonValue<T> extends true\n\t\t\t\t? true\n\t\t\t\t: T extends null | boolean | number | string\n\t\t\t\t\t? true\n\t\t\t\t\t: T extends undefined\n\t\t\t\t\t\t? false\n\t\t\t\t\t\t: T extends readonly (infer TItem)[]\n\t\t\t\t\t\t\t? IsJsonCompatible<TItem>\n\t\t\t\t\t\t\t: T extends (...args: never[]) => unknown\n\t\t\t\t\t\t\t\t? false\n\t\t\t\t\t\t\t\t: T extends object\n\t\t\t\t\t\t\t\t\t? [InvalidJsonKeys<T>] extends [never]\n\t\t\t\t\t\t\t\t\t\t? true\n\t\t\t\t\t\t\t\t\t\t: false\n\t\t\t\t\t\t\t\t\t: false;\n\n/** The JSON representation of a typed in-memory value. Optional object properties remain optional. */\nexport type JsonRepresentation<T> =\n\tIsAny<T> extends true\n\t\t? JsonValue\n\t\t: unknown extends T\n\t\t\t? JsonValue\n\t\t\t: [T] extends [JsonValue]\n\t\t\t\t? T\n\t\t\t\t: T extends readonly unknown[]\n\t\t\t\t\t? { [TKey in keyof T]: JsonRepresentation<Exclude<T[TKey], undefined>> }\n\t\t\t\t\t: T extends object\n\t\t\t\t\t\t? { [TKey in keyof T]: JsonRepresentation<Exclude<T[TKey], undefined>> }\n\t\t\t\t\t\t: never;\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\n/**\n * System instructions and tool declarations at one point in the transcript.\n *\n * The leading system message is the system prompt. Later system messages change it:\n * `content` adds instructions from that point on, `sections` replace or remove named\n * prompt sections, and `toolsAdded`/`toolsRemoved` change the tool set. Replaying\n * every system message in order yields the current prompt and tools. Providers that\n * accept system messages mid-conversation send each one in place; other providers\n * rebuild the leading system message from the replayed state.\n */\nexport interface SystemMessage {\n\trole: \"system\";\n\t/** Instruction text. On the leading message this is the base prompt; later, additional instructions. */\n\tcontent: string | TextContent[];\n\t/**\n\t * Named, ordered prompt sections rendered verbatim after `content`. The leading message\n\t * declares them; later messages replace sections by name, and `null` removes one. Keep\n\t * each section self-delimiting (a tag, a heading) so the model can relate an update to\n\t * the original. Avoid integer-like names; JSON objects reorder those.\n\t */\n\tsections?: Record<string, string | null>;\n\t/** Complete definitions of tools that become available at this point. */\n\ttoolsAdded?: Tool[];\n\t/** Tools that stop being available at this point. */\n\ttoolsRemoved?: ToolReference[];\n\ttimestamp: number; // Unix timestamp in milliseconds\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 model reported by the provider when different from the requested `model`\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 type ToolResultMessage<TDetails = JsonValue> =\n\tIsJsonCompatible<TDetails> extends true\n\t\t? {\n\t\t\t\trole: \"toolResult\";\n\t\t\t\ttoolCallId: string;\n\t\t\t\ttoolName: string;\n\t\t\t\tcontent: (TextContent | ImageContent)[]; // Supports text and images\n\t\t\t\tdetails?: JsonRepresentation<TDetails>;\n\t\t\t\t/** Usage from the tool execution itself, if available. Not part of main LLM context accounting. */\n\t\t\t\tusage?: Usage;\n\t\t\t\tisError: boolean;\n\t\t\t\ttimestamp: number; // Unix timestamp in milliseconds\n\t\t\t}\n\t\t: never;\n\nexport type Message = SystemMessage | 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 ToolReference {\n\tname: string;\n}\n\n/**\n * Request input accepted by the public stream entry points (`Models.stream()`,\n * `streamSimple()`, ...). `systemPrompt` and `tools` are shorthand for a leading\n * system message; `normalizeContext()` folds them into one before the request\n * reaches a provider.\n */\nexport interface Context {\n\tsystemPrompt?: string;\n\tmessages: Message[];\n\ttools?: Tool[];\n}\n\ndeclare const transcriptContextBrand: unique symbol;\n\n/**\n * Normalized request context passed to providers and API implementations. The\n * prompt and tool declarations are carried by the transcript's system messages.\n * Only `normalizeContext()` produces this type, so a raw `Context` cannot reach\n * provider code by accident.\n */\nexport type TranscriptContext = {\n\tmessages: Message[];\n\treadonly [transcriptContextBrand]: true;\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 exact model accepts system or developer messages after the conversation has started. When false, later system messages are folded into the leading system message. Default: false; the generated model catalog enables it for verified models. */\n\tsupportsMidConvoSystemMessages?: boolean;\n\t/** Whether system messages can introduce additional tools mid-conversation. Requires `supportsMidConvoSystemMessages`. Default: false; the generated model catalog enables it for capable models. */\n\tsupportsMidConvoToolAdditions?: 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/** 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/** Whether the exact model accepts developer or system messages after the conversation has started. When false, later system messages are folded into the leading system message. Default: false; the generated model catalog enables it for verified models. */\n\tsupportsMidConvoSystemMessages?: 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 transcript-anchored additions. 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/** Whether the exact model accepts system-role messages inside the conversation. When false, later system messages are folded into the top-level system prompt. Default: false. */\n\tsupportsMidConvoSystemMessages?: boolean;\n\t/** Whether the exact model accepts mid-conversation `tool_addition` and `tool_removal` blocks. Requires `supportsMidConvoSystemMessages`. Default: false. */\n\tsupportsMidConvoToolChanges?: 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 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/** Compatibility settings for the Mistral chat API. */\nexport interface MistralConversationsCompat {\n\t/** Whether the exact model accepts system messages after the conversation has started. When false, later system messages are folded into the leading system message. Default: false. */\n\tsupportsMidConvoSystemMessages?: 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\t/** Prompt cache lifetimes per retention tier. Unset when the provider's cache behavior is unknown. */\n\tpromptCache?: ModelPromptCache;\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: TApi extends \"mistral-conversations\"\n\t\t\t\t\t\t? MistralConversationsCompat\n\t\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/diagnostics.d.ts

@@ -1,3 +1,4 @@
+import type { JsonObject } from "../types.ts";
 export interface DiagnosticErrorInfo {
     name?: string;
     message: string;
@@ -8,11 +9,11 @@ export interface AssistantMessageDiagnostic {
     type: string;
     timestamp: number;
     error?: DiagnosticErrorInfo;
-    details?: Record<string, unknown>;
+    details?: JsonObject;
 }
 export declare function formatThrownValue(value: unknown): string;
 export declare function extractDiagnosticError(error: unknown): DiagnosticErrorInfo;
-export declare function createAssistantMessageDiagnostic(type: string, error: unknown, details?: Record<string, unknown>): AssistantMessageDiagnostic;
+export declare function createAssistantMessageDiagnostic(type: string, error: unknown, details?: JsonObject): AssistantMessageDiagnostic;
 export declare function appendAssistantMessageDiagnostic<T extends {
     diagnostics?: AssistantMessageDiagnostic[];
 }>(message: T, diagnostic: AssistantMessageDiagnostic): void;

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

@@ -1 +1 @@
-{"version":3,"file":"diagnostics.d.ts","sourceRoot":"","sources":["../../src/utils/diagnostics.ts"],"names":[],"mappings":"AAAA,MAAM,WAAW,mBAAmB;IACnC,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,OAAO,EAAE,MAAM,CAAC;IAChB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,IAAI,CAAC,EAAE,MAAM,GAAG,MAAM,CAAC;CACvB;AAED,MAAM,WAAW,0BAA0B;IAC1C,IAAI,EAAE,MAAM,CAAC;IACb,SAAS,EAAE,MAAM,CAAC;IAClB,KAAK,CAAC,EAAE,mBAAmB,CAAC;IAC5B,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;CAClC;AAED,wBAAgB,iBAAiB,CAAC,KAAK,EAAE,OAAO,GAAG,MAAM,CAIxD;AAED,wBAAgB,sBAAsB,CAAC,KAAK,EAAE,OAAO,GAAG,mBAAmB,CAS1E;AAED,wBAAgB,gCAAgC,CAC/C,IAAI,EAAE,MAAM,EACZ,KAAK,EAAE,OAAO,EACd,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAC/B,0BAA0B,CAE5B;AAED,wBAAgB,gCAAgC,CAAC,CAAC,SAAS;IAAE,WAAW,CAAC,EAAE,0BAA0B,EAAE,CAAA;CAAE,EACxG,OAAO,EAAE,CAAC,EACV,UAAU,EAAE,0BAA0B,GACpC,IAAI,CAEN","sourcesContent":["export interface DiagnosticErrorInfo {\n\tname?: string;\n\tmessage: string;\n\tstack?: string;\n\tcode?: string | number;\n}\n\nexport interface AssistantMessageDiagnostic {\n\ttype: string;\n\ttimestamp: number;\n\terror?: DiagnosticErrorInfo;\n\tdetails?: Record<string, unknown>;\n}\n\nexport function formatThrownValue(value: unknown): string {\n\tif (value instanceof Error) return value.message || value.name;\n\tif (typeof value === \"string\") return value;\n\treturn String(value);\n}\n\nexport function extractDiagnosticError(error: unknown): DiagnosticErrorInfo {\n\tif (!(error instanceof Error)) return { name: \"ThrownValue\", message: formatThrownValue(error) };\n\tconst code = (error as Error & { code?: unknown }).code;\n\treturn {\n\t\tname: error.name || undefined,\n\t\tmessage: error.message || error.name,\n\t\tstack: error.stack,\n\t\tcode: typeof code === \"string\" || typeof code === \"number\" ? code : undefined,\n\t};\n}\n\nexport function createAssistantMessageDiagnostic(\n\ttype: string,\n\terror: unknown,\n\tdetails?: Record<string, unknown>,\n): AssistantMessageDiagnostic {\n\treturn { type, timestamp: Date.now(), error: extractDiagnosticError(error), details };\n}\n\nexport function appendAssistantMessageDiagnostic<T extends { diagnostics?: AssistantMessageDiagnostic[] }>(\n\tmessage: T,\n\tdiagnostic: AssistantMessageDiagnostic,\n): void {\n\tmessage.diagnostics = [...(message.diagnostics ?? []), diagnostic];\n}\n"]}
+{"version":3,"file":"diagnostics.d.ts","sourceRoot":"","sources":["../../src/utils/diagnostics.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AAE9C,MAAM,WAAW,mBAAmB;IACnC,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,OAAO,EAAE,MAAM,CAAC;IAChB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,IAAI,CAAC,EAAE,MAAM,GAAG,MAAM,CAAC;CACvB;AAED,MAAM,WAAW,0BAA0B;IAC1C,IAAI,EAAE,MAAM,CAAC;IACb,SAAS,EAAE,MAAM,CAAC;IAClB,KAAK,CAAC,EAAE,mBAAmB,CAAC;IAC5B,OAAO,CAAC,EAAE,UAAU,CAAC;CACrB;AAED,wBAAgB,iBAAiB,CAAC,KAAK,EAAE,OAAO,GAAG,MAAM,CAIxD;AAED,wBAAgB,sBAAsB,CAAC,KAAK,EAAE,OAAO,GAAG,mBAAmB,CAS1E;AAED,wBAAgB,gCAAgC,CAC/C,IAAI,EAAE,MAAM,EACZ,KAAK,EAAE,OAAO,EACd,OAAO,CAAC,EAAE,UAAU,GAClB,0BAA0B,CAE5B;AAED,wBAAgB,gCAAgC,CAAC,CAAC,SAAS;IAAE,WAAW,CAAC,EAAE,0BAA0B,EAAE,CAAA;CAAE,EACxG,OAAO,EAAE,CAAC,EACV,UAAU,EAAE,0BAA0B,GACpC,IAAI,CAEN","sourcesContent":["import type { JsonObject } from \"../types.ts\";\n\nexport interface DiagnosticErrorInfo {\n\tname?: string;\n\tmessage: string;\n\tstack?: string;\n\tcode?: string | number;\n}\n\nexport interface AssistantMessageDiagnostic {\n\ttype: string;\n\ttimestamp: number;\n\terror?: DiagnosticErrorInfo;\n\tdetails?: JsonObject;\n}\n\nexport function formatThrownValue(value: unknown): string {\n\tif (value instanceof Error) return value.message || value.name;\n\tif (typeof value === \"string\") return value;\n\treturn String(value);\n}\n\nexport function extractDiagnosticError(error: unknown): DiagnosticErrorInfo {\n\tif (!(error instanceof Error)) return { name: \"ThrownValue\", message: formatThrownValue(error) };\n\tconst code = (error as Error & { code?: unknown }).code;\n\treturn {\n\t\tname: error.name || undefined,\n\t\tmessage: error.message || error.name,\n\t\tstack: error.stack,\n\t\tcode: typeof code === \"string\" || typeof code === \"number\" ? code : undefined,\n\t};\n}\n\nexport function createAssistantMessageDiagnostic(\n\ttype: string,\n\terror: unknown,\n\tdetails?: JsonObject,\n): AssistantMessageDiagnostic {\n\treturn { type, timestamp: Date.now(), error: extractDiagnosticError(error), details };\n}\n\nexport function appendAssistantMessageDiagnostic<T extends { diagnostics?: AssistantMessageDiagnostic[] }>(\n\tmessage: T,\n\tdiagnostic: AssistantMessageDiagnostic,\n): void {\n\tmessage.diagnostics = [...(message.diagnostics ?? []), diagnostic];\n}\n"]}

package/dist/utils/diagnostics.js.map

@@ -1 +1 @@
-{"version":3,"file":"diagnostics.js","sourceRoot":"","sources":["../../src/utils/diagnostics.ts"],"names":[],"mappings":"AAcA,MAAM,UAAU,iBAAiB,CAAC,KAAc,EAAU;IACzD,IAAI,KAAK,YAAY,KAAK;QAAE,OAAO,KAAK,CAAC,OAAO,IAAI,KAAK,CAAC,IAAI,CAAC;IAC/D,IAAI,OAAO,KAAK,KAAK,QAAQ;QAAE,OAAO,KAAK,CAAC;IAC5C,OAAO,MAAM,CAAC,KAAK,CAAC,CAAC;AAAA,CACrB;AAED,MAAM,UAAU,sBAAsB,CAAC,KAAc,EAAuB;IAC3E,IAAI,CAAC,CAAC,KAAK,YAAY,KAAK,CAAC;QAAE,OAAO,EAAE,IAAI,EAAE,aAAa,EAAE,OAAO,EAAE,iBAAiB,CAAC,KAAK,CAAC,EAAE,CAAC;IACjG,MAAM,IAAI,GAAI,KAAoC,CAAC,IAAI,CAAC;IACxD,OAAO;QACN,IAAI,EAAE,KAAK,CAAC,IAAI,IAAI,SAAS;QAC7B,OAAO,EAAE,KAAK,CAAC,OAAO,IAAI,KAAK,CAAC,IAAI;QACpC,KAAK,EAAE,KAAK,CAAC,KAAK;QAClB,IAAI,EAAE,OAAO,IAAI,KAAK,QAAQ,IAAI,OAAO,IAAI,KAAK,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,SAAS;KAC7E,CAAC;AAAA,CACF;AAED,MAAM,UAAU,gCAAgC,CAC/C,IAAY,EACZ,KAAc,EACd,OAAiC,EACJ;IAC7B,OAAO,EAAE,IAAI,EAAE,SAAS,EAAE,IAAI,CAAC,GAAG,EAAE,EAAE,KAAK,EAAE,sBAAsB,CAAC,KAAK,CAAC,EAAE,OAAO,EAAE,CAAC;AAAA,CACtF;AAED,MAAM,UAAU,gCAAgC,CAC/C,OAAU,EACV,UAAsC,EAC/B;IACP,OAAO,CAAC,WAAW,GAAG,CAAC,GAAG,CAAC,OAAO,CAAC,WAAW,IAAI,EAAE,CAAC,EAAE,UAAU,CAAC,CAAC;AAAA,CACnE","sourcesContent":["export interface DiagnosticErrorInfo {\n\tname?: string;\n\tmessage: string;\n\tstack?: string;\n\tcode?: string | number;\n}\n\nexport interface AssistantMessageDiagnostic {\n\ttype: string;\n\ttimestamp: number;\n\terror?: DiagnosticErrorInfo;\n\tdetails?: Record<string, unknown>;\n}\n\nexport function formatThrownValue(value: unknown): string {\n\tif (value instanceof Error) return value.message || value.name;\n\tif (typeof value === \"string\") return value;\n\treturn String(value);\n}\n\nexport function extractDiagnosticError(error: unknown): DiagnosticErrorInfo {\n\tif (!(error instanceof Error)) return { name: \"ThrownValue\", message: formatThrownValue(error) };\n\tconst code = (error as Error & { code?: unknown }).code;\n\treturn {\n\t\tname: error.name || undefined,\n\t\tmessage: error.message || error.name,\n\t\tstack: error.stack,\n\t\tcode: typeof code === \"string\" || typeof code === \"number\" ? code : undefined,\n\t};\n}\n\nexport function createAssistantMessageDiagnostic(\n\ttype: string,\n\terror: unknown,\n\tdetails?: Record<string, unknown>,\n): AssistantMessageDiagnostic {\n\treturn { type, timestamp: Date.now(), error: extractDiagnosticError(error), details };\n}\n\nexport function appendAssistantMessageDiagnostic<T extends { diagnostics?: AssistantMessageDiagnostic[] }>(\n\tmessage: T,\n\tdiagnostic: AssistantMessageDiagnostic,\n): void {\n\tmessage.diagnostics = [...(message.diagnostics ?? []), diagnostic];\n}\n"]}
+{"version":3,"file":"diagnostics.js","sourceRoot":"","sources":["../../src/utils/diagnostics.ts"],"names":[],"mappings":"AAgBA,MAAM,UAAU,iBAAiB,CAAC,KAAc,EAAU;IACzD,IAAI,KAAK,YAAY,KAAK;QAAE,OAAO,KAAK,CAAC,OAAO,IAAI,KAAK,CAAC,IAAI,CAAC;IAC/D,IAAI,OAAO,KAAK,KAAK,QAAQ;QAAE,OAAO,KAAK,CAAC;IAC5C,OAAO,MAAM,CAAC,KAAK,CAAC,CAAC;AAAA,CACrB;AAED,MAAM,UAAU,sBAAsB,CAAC,KAAc,EAAuB;IAC3E,IAAI,CAAC,CAAC,KAAK,YAAY,KAAK,CAAC;QAAE,OAAO,EAAE,IAAI,EAAE,aAAa,EAAE,OAAO,EAAE,iBAAiB,CAAC,KAAK,CAAC,EAAE,CAAC;IACjG,MAAM,IAAI,GAAI,KAAoC,CAAC,IAAI,CAAC;IACxD,OAAO;QACN,IAAI,EAAE,KAAK,CAAC,IAAI,IAAI,SAAS;QAC7B,OAAO,EAAE,KAAK,CAAC,OAAO,IAAI,KAAK,CAAC,IAAI;QACpC,KAAK,EAAE,KAAK,CAAC,KAAK;QAClB,IAAI,EAAE,OAAO,IAAI,KAAK,QAAQ,IAAI,OAAO,IAAI,KAAK,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,SAAS;KAC7E,CAAC;AAAA,CACF;AAED,MAAM,UAAU,gCAAgC,CAC/C,IAAY,EACZ,KAAc,EACd,OAAoB,EACS;IAC7B,OAAO,EAAE,IAAI,EAAE,SAAS,EAAE,IAAI,CAAC,GAAG,EAAE,EAAE,KAAK,EAAE,sBAAsB,CAAC,KAAK,CAAC,EAAE,OAAO,EAAE,CAAC;AAAA,CACtF;AAED,MAAM,UAAU,gCAAgC,CAC/C,OAAU,EACV,UAAsC,EAC/B;IACP,OAAO,CAAC,WAAW,GAAG,CAAC,GAAG,CAAC,OAAO,CAAC,WAAW,IAAI,EAAE,CAAC,EAAE,UAAU,CAAC,CAAC;AAAA,CACnE","sourcesContent":["import type { JsonObject } from \"../types.ts\";\n\nexport interface DiagnosticErrorInfo {\n\tname?: string;\n\tmessage: string;\n\tstack?: string;\n\tcode?: string | number;\n}\n\nexport interface AssistantMessageDiagnostic {\n\ttype: string;\n\ttimestamp: number;\n\terror?: DiagnosticErrorInfo;\n\tdetails?: JsonObject;\n}\n\nexport function formatThrownValue(value: unknown): string {\n\tif (value instanceof Error) return value.message || value.name;\n\tif (typeof value === \"string\") return value;\n\treturn String(value);\n}\n\nexport function extractDiagnosticError(error: unknown): DiagnosticErrorInfo {\n\tif (!(error instanceof Error)) return { name: \"ThrownValue\", message: formatThrownValue(error) };\n\tconst code = (error as Error & { code?: unknown }).code;\n\treturn {\n\t\tname: error.name || undefined,\n\t\tmessage: error.message || error.name,\n\t\tstack: error.stack,\n\t\tcode: typeof code === \"string\" || typeof code === \"number\" ? code : undefined,\n\t};\n}\n\nexport function createAssistantMessageDiagnostic(\n\ttype: string,\n\terror: unknown,\n\tdetails?: JsonObject,\n): AssistantMessageDiagnostic {\n\treturn { type, timestamp: Date.now(), error: extractDiagnosticError(error), details };\n}\n\nexport function appendAssistantMessageDiagnostic<T extends { diagnostics?: AssistantMessageDiagnostic[] }>(\n\tmessage: T,\n\tdiagnostic: AssistantMessageDiagnostic,\n): void {\n\tmessage.diagnostics = [...(message.diagnostics ?? []), diagnostic];\n}\n"]}

package/dist/utils/estimate.d.ts

@@ -1,4 +1,4 @@
-import type { Context, DocumentContent, ImageContent, Message, TextContent, Usage } from "../types.ts";
+import type { DocumentContent, ImageContent, Message, TextContent, TranscriptContext, Usage } from "../types.ts";
 export interface ContextUsageEstimate {
     /** Estimated total context tokens. */
     tokens: number;
@@ -13,5 +13,5 @@ export declare function calculateContextTokens(usage: Usage): number;
 export declare function estimateTextTokens(text: string): number;
 export declare function estimateTextAndImageContentTokens(content: string | Array<TextContent | ImageContent | DocumentContent>): number;
 export declare function estimateMessageTokens(message: Message): number;
-export declare function estimateContextTokens(context: Context | readonly Message[]): ContextUsageEstimate;
+export declare function estimateContextTokens(context: TranscriptContext | readonly Message[]): ContextUsageEstimate;
 //# sourceMappingURL=estimate.d.ts.map

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

@@ -1 +1 @@
-{"version":3,"file":"estimate.d.ts","sourceRoot":"","sources":["../../src/utils/estimate.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAEX,OAAO,EACP,eAAe,EACf,YAAY,EACZ,OAAO,EACP,WAAW,EAEX,KAAK,EACL,MAAM,aAAa,CAAC;AAErB,MAAM,WAAW,oBAAoB;IACpC,sCAAsC;IACtC,MAAM,EAAE,MAAM,CAAC;IACf,2EAA2E;IAC3E,WAAW,EAAE,MAAM,CAAC;IACpB,+EAA+E;IAC/E,cAAc,EAAE,MAAM,CAAC;IACvB,qFAAqF;IACrF,cAAc,EAAE,MAAM,GAAG,IAAI,CAAC;CAC9B;AAKD,wBAAgB,sBAAsB,CAAC,KAAK,EAAE,KAAK,GAAG,MAAM,CAE3D;AA8BD,wBAAgB,kBAAkB,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,CAEvD;AAED,wBAAgB,iCAAiC,CAChD,OAAO,EAAE,MAAM,GAAG,KAAK,CAAC,WAAW,GAAG,YAAY,GAAG,eAAe,CAAC,GACnE,MAAM,CAER;AAED,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,OAAO,GAAG,MAAM,CAiB9D;AAqDD,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,OAAO,GAAG,SAAS,OAAO,EAAE,GAAG,oBAAoB,CA6BjG","sourcesContent":["import type {\n\tAssistantMessage,\n\tContext,\n\tDocumentContent,\n\tImageContent,\n\tMessage,\n\tTextContent,\n\tTool,\n\tUsage,\n} from \"../types.ts\";\n\nexport interface ContextUsageEstimate {\n\t/** Estimated total context tokens. */\n\ttokens: number;\n\t/** Tokens reported by the most recent applicable assistant usage block. */\n\tusageTokens: number;\n\t/** Estimated tokens after the most recent applicable assistant usage block. */\n\ttrailingTokens: number;\n\t/** Index of the applicable message that provided usage, or null when none exists. */\n\tlastUsageIndex: number | null;\n}\n\nconst CHARS_PER_TOKEN = 4;\nconst ESTIMATED_IMAGE_CHARS = 4800;\n\nexport function calculateContextTokens(usage: Usage): number {\n\treturn usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite;\n}\n\nfunction safeJsonStringify(value: unknown): string {\n\ttry {\n\t\treturn JSON.stringify(value) ?? \"undefined\";\n\t} catch {\n\t\treturn \"[unserializable]\";\n\t}\n}\n\nfunction estimateTextAndImageContentChars(\n\tcontent: string | Array<TextContent | ImageContent | DocumentContent>,\n): number {\n\tif (typeof content === \"string\") return content.length;\n\n\tlet chars = 0;\n\tfor (const block of content) {\n\t\tif (block.type === \"text\") {\n\t\t\tchars += block.text.length;\n\t\t} else if (block.type === \"document\") {\n\t\t\t// A base64 PDF is far larger than an image and must not be counted as one, or context\n\t\t\t// accounting under-reports badly. Estimate from the encoded payload actually sent.\n\t\t\tchars += block.data.length;\n\t\t} else {\n\t\t\tchars += ESTIMATED_IMAGE_CHARS;\n\t\t}\n\t}\n\treturn chars;\n}\n\nexport function estimateTextTokens(text: string): number {\n\treturn Math.ceil(text.length / CHARS_PER_TOKEN);\n}\n\nexport function estimateTextAndImageContentTokens(\n\tcontent: string | Array<TextContent | ImageContent | DocumentContent>,\n): number {\n\treturn Math.ceil(estimateTextAndImageContentChars(content) / CHARS_PER_TOKEN);\n}\n\nexport function estimateMessageTokens(message: Message): number {\n\tlet chars = 0;\n\n\tif (message.role === \"user\") return estimateTextAndImageContentTokens(message.content);\n\tif (message.role === \"toolResult\") return estimateTextAndImageContentTokens(message.content);\n\n\tfor (const block of message.content) {\n\t\tif (block.type === \"text\") {\n\t\t\tchars += block.text.length;\n\t\t} else if (block.type === \"thinking\") {\n\t\t\tchars += block.thinking.length;\n\t\t} else if (block.type === \"toolCall\") {\n\t\t\tchars += block.name.length + safeJsonStringify(block.arguments).length;\n\t\t}\n\t\t// Fallback boundary markers carry no text and contribute nothing to the estimate.\n\t}\n\treturn Math.ceil(chars / CHARS_PER_TOKEN);\n}\n\nfunction getLastAssistantUsageInfo(messages: readonly Message[]): { usage: Usage; index: number } | undefined {\n\tlet latestPrefixTimestamp = Number.NEGATIVE_INFINITY;\n\tlet usageInfo: { usage: Usage; index: number } | undefined;\n\n\tfor (let i = 0; i < messages.length; i++) {\n\t\tconst message = messages[i];\n\t\tif (message.role === \"assistant\") {\n\t\t\tconst assistant = message as AssistantMessage;\n\t\t\t// A newer prefix message was inserted after this response (for example, a\n\t\t\t// compaction summary), so its usage cannot describe the current prefix.\n\t\t\tconst usageAppliesToPrefix = assistant.timestamp >= latestPrefixTimestamp;\n\t\t\tif (\n\t\t\t\tusageAppliesToPrefix &&\n\t\t\t\tassistant.stopReason !== \"aborted\" &&\n\t\t\t\tassistant.stopReason !== \"error\" &&\n\t\t\t\tcalculateContextTokens(assistant.usage) > 0\n\t\t\t) {\n\t\t\t\tusageInfo = { usage: assistant.usage, index: i };\n\t\t\t}\n\t\t}\n\t\tlatestPrefixTimestamp = Math.max(latestPrefixTimestamp, message.timestamp);\n\t}\n\n\treturn usageInfo;\n}\n\nfunction estimateMessages(messages: readonly Message[]): ContextUsageEstimate {\n\tconst usageInfo = getLastAssistantUsageInfo(messages);\n\tif (usageInfo) {\n\t\tconst usageTokens = calculateContextTokens(usageInfo.usage);\n\t\tlet trailingTokens = 0;\n\t\tfor (let i = usageInfo.index + 1; i < messages.length; i++) {\n\t\t\ttrailingTokens += estimateMessageTokens(messages[i]);\n\t\t}\n\t\treturn { tokens: usageTokens + trailingTokens, usageTokens, trailingTokens, lastUsageIndex: usageInfo.index };\n\t}\n\n\tlet tokens = 0;\n\tfor (const message of messages) tokens += estimateMessageTokens(message);\n\treturn { tokens, usageTokens: 0, trailingTokens: tokens, lastUsageIndex: null };\n}\n\nfunction estimateToolsTokens(tools: readonly Tool[] | undefined): number {\n\tif (!tools || tools.length === 0) return 0;\n\treturn estimateTextTokens(safeJsonStringify(tools));\n}\n\nfunction isMessageArray(value: Context | readonly Message[]): value is readonly Message[] {\n\treturn Array.isArray(value);\n}\n\nexport function estimateContextTokens(context: Context | readonly Message[]): ContextUsageEstimate {\n\tif (isMessageArray(context)) return estimateMessages(context);\n\n\tconst estimate = estimateMessages(context.messages);\n\tif (estimate.lastUsageIndex !== null) {\n\t\tconst addedNames = new Set(\n\t\t\tcontext.messages\n\t\t\t\t.slice(estimate.lastUsageIndex + 1)\n\t\t\t\t.filter((message) => message.role === \"toolResult\")\n\t\t\t\t.flatMap((message) => message.addedToolNames ?? []),\n\t\t);\n\t\tconst addedToolTokens = estimateToolsTokens(context.tools?.filter((tool) => addedNames.has(tool.name)));\n\t\treturn {\n\t\t\ttokens: estimate.tokens + addedToolTokens,\n\t\t\tusageTokens: estimate.usageTokens,\n\t\t\ttrailingTokens: estimate.trailingTokens + addedToolTokens,\n\t\t\tlastUsageIndex: estimate.lastUsageIndex,\n\t\t};\n\t}\n\n\tconst prefixTokens =\n\t\t(context.systemPrompt ? estimateTextTokens(context.systemPrompt) : 0) + estimateToolsTokens(context.tools);\n\n\treturn {\n\t\ttokens: estimate.tokens + prefixTokens,\n\t\tusageTokens: estimate.usageTokens,\n\t\ttrailingTokens: estimate.trailingTokens + prefixTokens,\n\t\tlastUsageIndex: estimate.lastUsageIndex,\n\t};\n}\n"]}
+{"version":3,"file":"estimate.d.ts","sourceRoot":"","sources":["../../src/utils/estimate.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAEX,eAAe,EACf,YAAY,EACZ,OAAO,EACP,WAAW,EACX,iBAAiB,EACjB,KAAK,EACL,MAAM,aAAa,CAAC;AAGrB,MAAM,WAAW,oBAAoB;IACpC,sCAAsC;IACtC,MAAM,EAAE,MAAM,CAAC;IACf,2EAA2E;IAC3E,WAAW,EAAE,MAAM,CAAC;IACpB,+EAA+E;IAC/E,cAAc,EAAE,MAAM,CAAC;IACvB,qFAAqF;IACrF,cAAc,EAAE,MAAM,GAAG,IAAI,CAAC;CAC9B;AAKD,wBAAgB,sBAAsB,CAAC,KAAK,EAAE,KAAK,GAAG,MAAM,CAE3D;AA8BD,wBAAgB,kBAAkB,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,CAEvD;AAED,wBAAgB,iCAAiC,CAChD,OAAO,EAAE,MAAM,GAAG,KAAK,CAAC,WAAW,GAAG,YAAY,GAAG,eAAe,CAAC,GACnE,MAAM,CAER;AAED,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,OAAO,GAAG,MAAM,CAwB9D;AA4BD,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,iBAAiB,GAAG,SAAS,OAAO,EAAE,GAAG,oBAAoB,CAe3G","sourcesContent":["import type {\n\tAssistantMessage,\n\tDocumentContent,\n\tImageContent,\n\tMessage,\n\tTextContent,\n\tTranscriptContext,\n\tUsage,\n} from \"../types.ts\";\nimport { getSystemMessageText } from \"./text.ts\";\n\nexport interface ContextUsageEstimate {\n\t/** Estimated total context tokens. */\n\ttokens: number;\n\t/** Tokens reported by the most recent applicable assistant usage block. */\n\tusageTokens: number;\n\t/** Estimated tokens after the most recent applicable assistant usage block. */\n\ttrailingTokens: number;\n\t/** Index of the applicable message that provided usage, or null when none exists. */\n\tlastUsageIndex: number | null;\n}\n\nconst CHARS_PER_TOKEN = 4;\nconst ESTIMATED_IMAGE_CHARS = 4800;\n\nexport function calculateContextTokens(usage: Usage): number {\n\treturn usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite;\n}\n\nfunction safeJsonStringify(value: unknown): string {\n\ttry {\n\t\treturn JSON.stringify(value) ?? \"undefined\";\n\t} catch {\n\t\treturn \"[unserializable]\";\n\t}\n}\n\nfunction estimateTextAndImageContentChars(\n\tcontent: string | Array<TextContent | ImageContent | DocumentContent>,\n): number {\n\tif (typeof content === \"string\") return content.length;\n\n\tlet chars = 0;\n\tfor (const block of content) {\n\t\tif (block.type === \"text\") {\n\t\t\tchars += block.text.length;\n\t\t} else if (block.type === \"document\") {\n\t\t\t// A base64 PDF is far larger than an image and must not be counted as one, or context\n\t\t\t// accounting under-reports badly. Estimate from the encoded payload actually sent.\n\t\t\tchars += block.data.length;\n\t\t} else {\n\t\t\tchars += ESTIMATED_IMAGE_CHARS;\n\t\t}\n\t}\n\treturn chars;\n}\n\nexport function estimateTextTokens(text: string): number {\n\treturn Math.ceil(text.length / CHARS_PER_TOKEN);\n}\n\nexport function estimateTextAndImageContentTokens(\n\tcontent: string | Array<TextContent | ImageContent | DocumentContent>,\n): number {\n\treturn Math.ceil(estimateTextAndImageContentChars(content) / CHARS_PER_TOKEN);\n}\n\nexport function estimateMessageTokens(message: Message): number {\n\tlet chars = 0;\n\n\tif (message.role === \"system\") {\n\t\treturn (\n\t\t\testimateTextTokens(getSystemMessageText(message)) +\n\t\t\testimateToolsTokens(message.toolsAdded) +\n\t\t\testimateToolsTokens(message.toolsRemoved)\n\t\t);\n\t}\n\tif (message.role === \"user\") return estimateTextAndImageContentTokens(message.content);\n\tif (message.role === \"toolResult\") return estimateTextAndImageContentTokens(message.content);\n\n\tfor (const block of message.content) {\n\t\tif (block.type === \"text\") {\n\t\t\tchars += block.text.length;\n\t\t} else if (block.type === \"thinking\") {\n\t\t\tchars += block.thinking.length;\n\t\t} else if (block.type === \"toolCall\") {\n\t\t\tchars += block.name.length + safeJsonStringify(block.arguments).length;\n\t\t}\n\t\t// Fallback boundary markers carry no text and contribute nothing to the estimate.\n\t}\n\treturn Math.ceil(chars / CHARS_PER_TOKEN);\n}\n\nfunction getLastAssistantUsageInfo(messages: readonly Message[]): { usage: Usage; index: number } | undefined {\n\tlet latestPrefixTimestamp = Number.NEGATIVE_INFINITY;\n\tlet usageInfo: { usage: Usage; index: number } | undefined;\n\n\tfor (let i = 0; i < messages.length; i++) {\n\t\tconst message = messages[i];\n\t\tif (message.role === \"assistant\") {\n\t\t\tconst assistant = message as AssistantMessage;\n\t\t\t// A newer prefix message was inserted after this response (for example, a\n\t\t\t// compaction summary), so its usage cannot describe the current prefix.\n\t\t\tconst usageAppliesToPrefix = assistant.timestamp >= latestPrefixTimestamp;\n\t\t\tif (\n\t\t\t\tusageAppliesToPrefix &&\n\t\t\t\tassistant.stopReason !== \"aborted\" &&\n\t\t\t\tassistant.stopReason !== \"error\" &&\n\t\t\t\tcalculateContextTokens(assistant.usage) > 0\n\t\t\t) {\n\t\t\t\tusageInfo = { usage: assistant.usage, index: i };\n\t\t\t}\n\t\t}\n\t\tlatestPrefixTimestamp = Math.max(latestPrefixTimestamp, message.timestamp);\n\t}\n\n\treturn usageInfo;\n}\n\nexport function estimateContextTokens(context: TranscriptContext | readonly Message[]): ContextUsageEstimate {\n\tconst messages = \"messages\" in context ? context.messages : context;\n\tconst usageInfo = getLastAssistantUsageInfo(messages);\n\tif (usageInfo) {\n\t\tconst usageTokens = calculateContextTokens(usageInfo.usage);\n\t\tlet trailingTokens = 0;\n\t\tfor (let i = usageInfo.index + 1; i < messages.length; i++) {\n\t\t\ttrailingTokens += estimateMessageTokens(messages[i]);\n\t\t}\n\t\treturn { tokens: usageTokens + trailingTokens, usageTokens, trailingTokens, lastUsageIndex: usageInfo.index };\n\t}\n\n\tlet tokens = 0;\n\tfor (const message of messages) tokens += estimateMessageTokens(message);\n\treturn { tokens, usageTokens: 0, trailingTokens: tokens, lastUsageIndex: null };\n}\n\nfunction estimateToolsTokens(tools: readonly unknown[] | undefined): number {\n\tif (!tools || tools.length === 0) return 0;\n\treturn estimateTextTokens(safeJsonStringify(tools));\n}\n"]}

package/dist/utils/estimate.js

@@ -1,3 +1,4 @@
+import { getSystemMessageText } from "./text.js";
 const CHARS_PER_TOKEN = 4;
 const ESTIMATED_IMAGE_CHARS = 4800;
 export function calculateContextTokens(usage) {
@@ -38,6 +39,11 @@ export function estimateTextAndImageContentTokens(content) {
 }
 export function estimateMessageTokens(message) {
     let chars = 0;
+    if (message.role === "system") {
+        return (estimateTextTokens(getSystemMessageText(message)) +
+            estimateToolsTokens(message.toolsAdded) +
+            estimateToolsTokens(message.toolsRemoved));
+    }
     if (message.role === "user")
         return estimateTextAndImageContentTokens(message.content);
     if (message.role === "toolResult")
@@ -77,7 +83,8 @@ function getLastAssistantUsageInfo(messages) {
     }
     return usageInfo;
 }
-function estimateMessages(messages) {
+export function estimateContextTokens(context) {
+    const messages = "messages" in context ? context.messages : context;
     const usageInfo = getLastAssistantUsageInfo(messages);
     if (usageInfo) {
         const usageTokens = calculateContextTokens(usageInfo.usage);
@@ -97,32 +104,4 @@ function estimateToolsTokens(tools) {
         return 0;
     return estimateTextTokens(safeJsonStringify(tools));
 }
-function isMessageArray(value) {
-    return Array.isArray(value);
-}
-export function estimateContextTokens(context) {
-    if (isMessageArray(context))
-        return estimateMessages(context);
-    const estimate = estimateMessages(context.messages);
-    if (estimate.lastUsageIndex !== null) {
-        const addedNames = new Set(context.messages
-            .slice(estimate.lastUsageIndex + 1)
-            .filter((message) => message.role === "toolResult")
-            .flatMap((message) => message.addedToolNames ?? []));
-        const addedToolTokens = estimateToolsTokens(context.tools?.filter((tool) => addedNames.has(tool.name)));
-        return {
-            tokens: estimate.tokens + addedToolTokens,
-            usageTokens: estimate.usageTokens,
-            trailingTokens: estimate.trailingTokens + addedToolTokens,
-            lastUsageIndex: estimate.lastUsageIndex,
-        };
-    }
-    const prefixTokens = (context.systemPrompt ? estimateTextTokens(context.systemPrompt) : 0) + estimateToolsTokens(context.tools);
-    return {
-        tokens: estimate.tokens + prefixTokens,
-        usageTokens: estimate.usageTokens,
-        trailingTokens: estimate.trailingTokens + prefixTokens,
-        lastUsageIndex: estimate.lastUsageIndex,
-    };
-}
 //# sourceMappingURL=estimate.js.map

package/dist/utils/estimate.js.map

@@ -1 +1 @@
-{"version":3,"file":"estimate.js","sourceRoot":"","sources":["../../src/utils/estimate.ts"],"names":[],"mappings":"AAsBA,MAAM,eAAe,GAAG,CAAC,CAAC;AAC1B,MAAM,qBAAqB,GAAG,IAAI,CAAC;AAEnC,MAAM,UAAU,sBAAsB,CAAC,KAAY,EAAU;IAC5D,OAAO,KAAK,CAAC,WAAW,IAAI,KAAK,CAAC,KAAK,GAAG,KAAK,CAAC,MAAM,GAAG,KAAK,CAAC,SAAS,GAAG,KAAK,CAAC,UAAU,CAAC;AAAA,CAC5F;AAED,SAAS,iBAAiB,CAAC,KAAc,EAAU;IAClD,IAAI,CAAC;QACJ,OAAO,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,IAAI,WAAW,CAAC;IAC7C,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,kBAAkB,CAAC;IAC3B,CAAC;AAAA,CACD;AAED,SAAS,gCAAgC,CACxC,OAAqE,EAC5D;IACT,IAAI,OAAO,OAAO,KAAK,QAAQ;QAAE,OAAO,OAAO,CAAC,MAAM,CAAC;IAEvD,IAAI,KAAK,GAAG,CAAC,CAAC;IACd,KAAK,MAAM,KAAK,IAAI,OAAO,EAAE,CAAC;QAC7B,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;YAC3B,KAAK,IAAI,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC;QAC5B,CAAC;aAAM,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;YACtC,sFAAsF;YACtF,mFAAmF;YACnF,KAAK,IAAI,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC;QAC5B,CAAC;aAAM,CAAC;YACP,KAAK,IAAI,qBAAqB,CAAC;QAChC,CAAC;IACF,CAAC;IACD,OAAO,KAAK,CAAC;AAAA,CACb;AAED,MAAM,UAAU,kBAAkB,CAAC,IAAY,EAAU;IACxD,OAAO,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,MAAM,GAAG,eAAe,CAAC,CAAC;AAAA,CAChD;AAED,MAAM,UAAU,iCAAiC,CAChD,OAAqE,EAC5D;IACT,OAAO,IAAI,CAAC,IAAI,CAAC,gCAAgC,CAAC,OAAO,CAAC,GAAG,eAAe,CAAC,CAAC;AAAA,CAC9E;AAED,MAAM,UAAU,qBAAqB,CAAC,OAAgB,EAAU;IAC/D,IAAI,KAAK,GAAG,CAAC,CAAC;IAEd,IAAI,OAAO,CAAC,IAAI,KAAK,MAAM;QAAE,OAAO,iCAAiC,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC;IACvF,IAAI,OAAO,CAAC,IAAI,KAAK,YAAY;QAAE,OAAO,iCAAiC,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC;IAE7F,KAAK,MAAM,KAAK,IAAI,OAAO,CAAC,OAAO,EAAE,CAAC;QACrC,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;YAC3B,KAAK,IAAI,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC;QAC5B,CAAC;aAAM,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;YACtC,KAAK,IAAI,KAAK,CAAC,QAAQ,CAAC,MAAM,CAAC;QAChC,CAAC;aAAM,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;YACtC,KAAK,IAAI,KAAK,CAAC,IAAI,CAAC,MAAM,GAAG,iBAAiB,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC,MAAM,CAAC;QACxE,CAAC;QACD,kFAAkF;IACnF,CAAC;IACD,OAAO,IAAI,CAAC,IAAI,CAAC,KAAK,GAAG,eAAe,CAAC,CAAC;AAAA,CAC1C;AAED,SAAS,yBAAyB,CAAC,QAA4B,EAA+C;IAC7G,IAAI,qBAAqB,GAAG,MAAM,CAAC,iBAAiB,CAAC;IACrD,IAAI,SAAsD,CAAC;IAE3D,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QAC1C,MAAM,OAAO,GAAG,QAAQ,CAAC,CAAC,CAAC,CAAC;QAC5B,IAAI,OAAO,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;YAClC,MAAM,SAAS,GAAG,OAA2B,CAAC;YAC9C,0EAA0E;YAC1E,wEAAwE;YACxE,MAAM,oBAAoB,GAAG,SAAS,CAAC,SAAS,IAAI,qBAAqB,CAAC;YAC1E,IACC,oBAAoB;gBACpB,SAAS,CAAC,UAAU,KAAK,SAAS;gBAClC,SAAS,CAAC,UAAU,KAAK,OAAO;gBAChC,sBAAsB,CAAC,SAAS,CAAC,KAAK,CAAC,GAAG,CAAC,EAC1C,CAAC;gBACF,SAAS,GAAG,EAAE,KAAK,EAAE,SAAS,CAAC,KAAK,EAAE,KAAK,EAAE,CAAC,EAAE,CAAC;YAClD,CAAC;QACF,CAAC;QACD,qBAAqB,GAAG,IAAI,CAAC,GAAG,CAAC,qBAAqB,EAAE,OAAO,CAAC,SAAS,CAAC,CAAC;IAC5E,CAAC;IAED,OAAO,SAAS,CAAC;AAAA,CACjB;AAED,SAAS,gBAAgB,CAAC,QAA4B,EAAwB;IAC7E,MAAM,SAAS,GAAG,yBAAyB,CAAC,QAAQ,CAAC,CAAC;IACtD,IAAI,SAAS,EAAE,CAAC;QACf,MAAM,WAAW,GAAG,sBAAsB,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC;QAC5D,IAAI,cAAc,GAAG,CAAC,CAAC;QACvB,KAAK,IAAI,CAAC,GAAG,SAAS,CAAC,KAAK,GAAG,CAAC,EAAE,CAAC,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YAC5D,cAAc,IAAI,qBAAqB,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC;QACtD,CAAC;QACD,OAAO,EAAE,MAAM,EAAE,WAAW,GAAG,cAAc,EAAE,WAAW,EAAE,cAAc,EAAE,cAAc,EAAE,SAAS,CAAC,KAAK,EAAE,CAAC;IAC/G,CAAC;IAED,IAAI,MAAM,GAAG,CAAC,CAAC;IACf,KAAK,MAAM,OAAO,IAAI,QAAQ;QAAE,MAAM,IAAI,qBAAqB,CAAC,OAAO,CAAC,CAAC;IACzE,OAAO,EAAE,MAAM,EAAE,WAAW,EAAE,CAAC,EAAE,cAAc,EAAE,MAAM,EAAE,cAAc,EAAE,IAAI,EAAE,CAAC;AAAA,CAChF;AAED,SAAS,mBAAmB,CAAC,KAAkC,EAAU;IACxE,IAAI,CAAC,KAAK,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,CAAC,CAAC;IAC3C,OAAO,kBAAkB,CAAC,iBAAiB,CAAC,KAAK,CAAC,CAAC,CAAC;AAAA,CACpD;AAED,SAAS,cAAc,CAAC,KAAmC,EAA+B;IACzF,OAAO,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;AAAA,CAC5B;AAED,MAAM,UAAU,qBAAqB,CAAC,OAAqC,EAAwB;IAClG,IAAI,cAAc,CAAC,OAAO,CAAC;QAAE,OAAO,gBAAgB,CAAC,OAAO,CAAC,CAAC;IAE9D,MAAM,QAAQ,GAAG,gBAAgB,CAAC,OAAO,CAAC,QAAQ,CAAC,CAAC;IACpD,IAAI,QAAQ,CAAC,cAAc,KAAK,IAAI,EAAE,CAAC;QACtC,MAAM,UAAU,GAAG,IAAI,GAAG,CACzB,OAAO,CAAC,QAAQ;aACd,KAAK,CAAC,QAAQ,CAAC,cAAc,GAAG,CAAC,CAAC;aAClC,MAAM,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,OAAO,CAAC,IAAI,KAAK,YAAY,CAAC;aAClD,OAAO,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,OAAO,CAAC,cAAc,IAAI,EAAE,CAAC,CACpD,CAAC;QACF,MAAM,eAAe,GAAG,mBAAmB,CAAC,OAAO,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,UAAU,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;QACxG,OAAO;YACN,MAAM,EAAE,QAAQ,CAAC,MAAM,GAAG,eAAe;YACzC,WAAW,EAAE,QAAQ,CAAC,WAAW;YACjC,cAAc,EAAE,QAAQ,CAAC,cAAc,GAAG,eAAe;YACzD,cAAc,EAAE,QAAQ,CAAC,cAAc;SACvC,CAAC;IACH,CAAC;IAED,MAAM,YAAY,GACjB,CAAC,OAAO,CAAC,YAAY,CAAC,CAAC,CAAC,kBAAkB,CAAC,OAAO,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,mBAAmB,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;IAE5G,OAAO;QACN,MAAM,EAAE,QAAQ,CAAC,MAAM,GAAG,YAAY;QACtC,WAAW,EAAE,QAAQ,CAAC,WAAW;QACjC,cAAc,EAAE,QAAQ,CAAC,cAAc,GAAG,YAAY;QACtD,cAAc,EAAE,QAAQ,CAAC,cAAc;KACvC,CAAC;AAAA,CACF","sourcesContent":["import type {\n\tAssistantMessage,\n\tContext,\n\tDocumentContent,\n\tImageContent,\n\tMessage,\n\tTextContent,\n\tTool,\n\tUsage,\n} from \"../types.ts\";\n\nexport interface ContextUsageEstimate {\n\t/** Estimated total context tokens. */\n\ttokens: number;\n\t/** Tokens reported by the most recent applicable assistant usage block. */\n\tusageTokens: number;\n\t/** Estimated tokens after the most recent applicable assistant usage block. */\n\ttrailingTokens: number;\n\t/** Index of the applicable message that provided usage, or null when none exists. */\n\tlastUsageIndex: number | null;\n}\n\nconst CHARS_PER_TOKEN = 4;\nconst ESTIMATED_IMAGE_CHARS = 4800;\n\nexport function calculateContextTokens(usage: Usage): number {\n\treturn usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite;\n}\n\nfunction safeJsonStringify(value: unknown): string {\n\ttry {\n\t\treturn JSON.stringify(value) ?? \"undefined\";\n\t} catch {\n\t\treturn \"[unserializable]\";\n\t}\n}\n\nfunction estimateTextAndImageContentChars(\n\tcontent: string | Array<TextContent | ImageContent | DocumentContent>,\n): number {\n\tif (typeof content === \"string\") return content.length;\n\n\tlet chars = 0;\n\tfor (const block of content) {\n\t\tif (block.type === \"text\") {\n\t\t\tchars += block.text.length;\n\t\t} else if (block.type === \"document\") {\n\t\t\t// A base64 PDF is far larger than an image and must not be counted as one, or context\n\t\t\t// accounting under-reports badly. Estimate from the encoded payload actually sent.\n\t\t\tchars += block.data.length;\n\t\t} else {\n\t\t\tchars += ESTIMATED_IMAGE_CHARS;\n\t\t}\n\t}\n\treturn chars;\n}\n\nexport function estimateTextTokens(text: string): number {\n\treturn Math.ceil(text.length / CHARS_PER_TOKEN);\n}\n\nexport function estimateTextAndImageContentTokens(\n\tcontent: string | Array<TextContent | ImageContent | DocumentContent>,\n): number {\n\treturn Math.ceil(estimateTextAndImageContentChars(content) / CHARS_PER_TOKEN);\n}\n\nexport function estimateMessageTokens(message: Message): number {\n\tlet chars = 0;\n\n\tif (message.role === \"user\") return estimateTextAndImageContentTokens(message.content);\n\tif (message.role === \"toolResult\") return estimateTextAndImageContentTokens(message.content);\n\n\tfor (const block of message.content) {\n\t\tif (block.type === \"text\") {\n\t\t\tchars += block.text.length;\n\t\t} else if (block.type === \"thinking\") {\n\t\t\tchars += block.thinking.length;\n\t\t} else if (block.type === \"toolCall\") {\n\t\t\tchars += block.name.length + safeJsonStringify(block.arguments).length;\n\t\t}\n\t\t// Fallback boundary markers carry no text and contribute nothing to the estimate.\n\t}\n\treturn Math.ceil(chars / CHARS_PER_TOKEN);\n}\n\nfunction getLastAssistantUsageInfo(messages: readonly Message[]): { usage: Usage; index: number } | undefined {\n\tlet latestPrefixTimestamp = Number.NEGATIVE_INFINITY;\n\tlet usageInfo: { usage: Usage; index: number } | undefined;\n\n\tfor (let i = 0; i < messages.length; i++) {\n\t\tconst message = messages[i];\n\t\tif (message.role === \"assistant\") {\n\t\t\tconst assistant = message as AssistantMessage;\n\t\t\t// A newer prefix message was inserted after this response (for example, a\n\t\t\t// compaction summary), so its usage cannot describe the current prefix.\n\t\t\tconst usageAppliesToPrefix = assistant.timestamp >= latestPrefixTimestamp;\n\t\t\tif (\n\t\t\t\tusageAppliesToPrefix &&\n\t\t\t\tassistant.stopReason !== \"aborted\" &&\n\t\t\t\tassistant.stopReason !== \"error\" &&\n\t\t\t\tcalculateContextTokens(assistant.usage) > 0\n\t\t\t) {\n\t\t\t\tusageInfo = { usage: assistant.usage, index: i };\n\t\t\t}\n\t\t}\n\t\tlatestPrefixTimestamp = Math.max(latestPrefixTimestamp, message.timestamp);\n\t}\n\n\treturn usageInfo;\n}\n\nfunction estimateMessages(messages: readonly Message[]): ContextUsageEstimate {\n\tconst usageInfo = getLastAssistantUsageInfo(messages);\n\tif (usageInfo) {\n\t\tconst usageTokens = calculateContextTokens(usageInfo.usage);\n\t\tlet trailingTokens = 0;\n\t\tfor (let i = usageInfo.index + 1; i < messages.length; i++) {\n\t\t\ttrailingTokens += estimateMessageTokens(messages[i]);\n\t\t}\n\t\treturn { tokens: usageTokens + trailingTokens, usageTokens, trailingTokens, lastUsageIndex: usageInfo.index };\n\t}\n\n\tlet tokens = 0;\n\tfor (const message of messages) tokens += estimateMessageTokens(message);\n\treturn { tokens, usageTokens: 0, trailingTokens: tokens, lastUsageIndex: null };\n}\n\nfunction estimateToolsTokens(tools: readonly Tool[] | undefined): number {\n\tif (!tools || tools.length === 0) return 0;\n\treturn estimateTextTokens(safeJsonStringify(tools));\n}\n\nfunction isMessageArray(value: Context | readonly Message[]): value is readonly Message[] {\n\treturn Array.isArray(value);\n}\n\nexport function estimateContextTokens(context: Context | readonly Message[]): ContextUsageEstimate {\n\tif (isMessageArray(context)) return estimateMessages(context);\n\n\tconst estimate = estimateMessages(context.messages);\n\tif (estimate.lastUsageIndex !== null) {\n\t\tconst addedNames = new Set(\n\t\t\tcontext.messages\n\t\t\t\t.slice(estimate.lastUsageIndex + 1)\n\t\t\t\t.filter((message) => message.role === \"toolResult\")\n\t\t\t\t.flatMap((message) => message.addedToolNames ?? []),\n\t\t);\n\t\tconst addedToolTokens = estimateToolsTokens(context.tools?.filter((tool) => addedNames.has(tool.name)));\n\t\treturn {\n\t\t\ttokens: estimate.tokens + addedToolTokens,\n\t\t\tusageTokens: estimate.usageTokens,\n\t\t\ttrailingTokens: estimate.trailingTokens + addedToolTokens,\n\t\t\tlastUsageIndex: estimate.lastUsageIndex,\n\t\t};\n\t}\n\n\tconst prefixTokens =\n\t\t(context.systemPrompt ? estimateTextTokens(context.systemPrompt) : 0) + estimateToolsTokens(context.tools);\n\n\treturn {\n\t\ttokens: estimate.tokens + prefixTokens,\n\t\tusageTokens: estimate.usageTokens,\n\t\ttrailingTokens: estimate.trailingTokens + prefixTokens,\n\t\tlastUsageIndex: estimate.lastUsageIndex,\n\t};\n}\n"]}
+{"version":3,"file":"estimate.js","sourceRoot":"","sources":["../../src/utils/estimate.ts"],"names":[],"mappings":"AASA,OAAO,EAAE,oBAAoB,EAAE,MAAM,WAAW,CAAC;AAajD,MAAM,eAAe,GAAG,CAAC,CAAC;AAC1B,MAAM,qBAAqB,GAAG,IAAI,CAAC;AAEnC,MAAM,UAAU,sBAAsB,CAAC,KAAY,EAAU;IAC5D,OAAO,KAAK,CAAC,WAAW,IAAI,KAAK,CAAC,KAAK,GAAG,KAAK,CAAC,MAAM,GAAG,KAAK,CAAC,SAAS,GAAG,KAAK,CAAC,UAAU,CAAC;AAAA,CAC5F;AAED,SAAS,iBAAiB,CAAC,KAAc,EAAU;IAClD,IAAI,CAAC;QACJ,OAAO,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,IAAI,WAAW,CAAC;IAC7C,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,kBAAkB,CAAC;IAC3B,CAAC;AAAA,CACD;AAED,SAAS,gCAAgC,CACxC,OAAqE,EAC5D;IACT,IAAI,OAAO,OAAO,KAAK,QAAQ;QAAE,OAAO,OAAO,CAAC,MAAM,CAAC;IAEvD,IAAI,KAAK,GAAG,CAAC,CAAC;IACd,KAAK,MAAM,KAAK,IAAI,OAAO,EAAE,CAAC;QAC7B,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;YAC3B,KAAK,IAAI,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC;QAC5B,CAAC;aAAM,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;YACtC,sFAAsF;YACtF,mFAAmF;YACnF,KAAK,IAAI,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC;QAC5B,CAAC;aAAM,CAAC;YACP,KAAK,IAAI,qBAAqB,CAAC;QAChC,CAAC;IACF,CAAC;IACD,OAAO,KAAK,CAAC;AAAA,CACb;AAED,MAAM,UAAU,kBAAkB,CAAC,IAAY,EAAU;IACxD,OAAO,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,MAAM,GAAG,eAAe,CAAC,CAAC;AAAA,CAChD;AAED,MAAM,UAAU,iCAAiC,CAChD,OAAqE,EAC5D;IACT,OAAO,IAAI,CAAC,IAAI,CAAC,gCAAgC,CAAC,OAAO,CAAC,GAAG,eAAe,CAAC,CAAC;AAAA,CAC9E;AAED,MAAM,UAAU,qBAAqB,CAAC,OAAgB,EAAU;IAC/D,IAAI,KAAK,GAAG,CAAC,CAAC;IAEd,IAAI,OAAO,CAAC,IAAI,KAAK,QAAQ,EAAE,CAAC;QAC/B,OAAO,CACN,kBAAkB,CAAC,oBAAoB,CAAC,OAAO,CAAC,CAAC;YACjD,mBAAmB,CAAC,OAAO,CAAC,UAAU,CAAC;YACvC,mBAAmB,CAAC,OAAO,CAAC,YAAY,CAAC,CACzC,CAAC;IACH,CAAC;IACD,IAAI,OAAO,CAAC,IAAI,KAAK,MAAM;QAAE,OAAO,iCAAiC,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC;IACvF,IAAI,OAAO,CAAC,IAAI,KAAK,YAAY;QAAE,OAAO,iCAAiC,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC;IAE7F,KAAK,MAAM,KAAK,IAAI,OAAO,CAAC,OAAO,EAAE,CAAC;QACrC,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;YAC3B,KAAK,IAAI,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC;QAC5B,CAAC;aAAM,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;YACtC,KAAK,IAAI,KAAK,CAAC,QAAQ,CAAC,MAAM,CAAC;QAChC,CAAC;aAAM,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;YACtC,KAAK,IAAI,KAAK,CAAC,IAAI,CAAC,MAAM,GAAG,iBAAiB,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC,MAAM,CAAC;QACxE,CAAC;QACD,kFAAkF;IACnF,CAAC;IACD,OAAO,IAAI,CAAC,IAAI,CAAC,KAAK,GAAG,eAAe,CAAC,CAAC;AAAA,CAC1C;AAED,SAAS,yBAAyB,CAAC,QAA4B,EAA+C;IAC7G,IAAI,qBAAqB,GAAG,MAAM,CAAC,iBAAiB,CAAC;IACrD,IAAI,SAAsD,CAAC;IAE3D,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QAC1C,MAAM,OAAO,GAAG,QAAQ,CAAC,CAAC,CAAC,CAAC;QAC5B,IAAI,OAAO,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;YAClC,MAAM,SAAS,GAAG,OAA2B,CAAC;YAC9C,0EAA0E;YAC1E,wEAAwE;YACxE,MAAM,oBAAoB,GAAG,SAAS,CAAC,SAAS,IAAI,qBAAqB,CAAC;YAC1E,IACC,oBAAoB;gBACpB,SAAS,CAAC,UAAU,KAAK,SAAS;gBAClC,SAAS,CAAC,UAAU,KAAK,OAAO;gBAChC,sBAAsB,CAAC,SAAS,CAAC,KAAK,CAAC,GAAG,CAAC,EAC1C,CAAC;gBACF,SAAS,GAAG,EAAE,KAAK,EAAE,SAAS,CAAC,KAAK,EAAE,KAAK,EAAE,CAAC,EAAE,CAAC;YAClD,CAAC;QACF,CAAC;QACD,qBAAqB,GAAG,IAAI,CAAC,GAAG,CAAC,qBAAqB,EAAE,OAAO,CAAC,SAAS,CAAC,CAAC;IAC5E,CAAC;IAED,OAAO,SAAS,CAAC;AAAA,CACjB;AAED,MAAM,UAAU,qBAAqB,CAAC,OAA+C,EAAwB;IAC5G,MAAM,QAAQ,GAAG,UAAU,IAAI,OAAO,CAAC,CAAC,CAAC,OAAO,CAAC,QAAQ,CAAC,CAAC,CAAC,OAAO,CAAC;IACpE,MAAM,SAAS,GAAG,yBAAyB,CAAC,QAAQ,CAAC,CAAC;IACtD,IAAI,SAAS,EAAE,CAAC;QACf,MAAM,WAAW,GAAG,sBAAsB,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC;QAC5D,IAAI,cAAc,GAAG,CAAC,CAAC;QACvB,KAAK,IAAI,CAAC,GAAG,SAAS,CAAC,KAAK,GAAG,CAAC,EAAE,CAAC,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YAC5D,cAAc,IAAI,qBAAqB,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC;QACtD,CAAC;QACD,OAAO,EAAE,MAAM,EAAE,WAAW,GAAG,cAAc,EAAE,WAAW,EAAE,cAAc,EAAE,cAAc,EAAE,SAAS,CAAC,KAAK,EAAE,CAAC;IAC/G,CAAC;IAED,IAAI,MAAM,GAAG,CAAC,CAAC;IACf,KAAK,MAAM,OAAO,IAAI,QAAQ;QAAE,MAAM,IAAI,qBAAqB,CAAC,OAAO,CAAC,CAAC;IACzE,OAAO,EAAE,MAAM,EAAE,WAAW,EAAE,CAAC,EAAE,cAAc,EAAE,MAAM,EAAE,cAAc,EAAE,IAAI,EAAE,CAAC;AAAA,CAChF;AAED,SAAS,mBAAmB,CAAC,KAAqC,EAAU;IAC3E,IAAI,CAAC,KAAK,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,CAAC,CAAC;IAC3C,OAAO,kBAAkB,CAAC,iBAAiB,CAAC,KAAK,CAAC,CAAC,CAAC;AAAA,CACpD","sourcesContent":["import type {\n\tAssistantMessage,\n\tDocumentContent,\n\tImageContent,\n\tMessage,\n\tTextContent,\n\tTranscriptContext,\n\tUsage,\n} from \"../types.ts\";\nimport { getSystemMessageText } from \"./text.ts\";\n\nexport interface ContextUsageEstimate {\n\t/** Estimated total context tokens. */\n\ttokens: number;\n\t/** Tokens reported by the most recent applicable assistant usage block. */\n\tusageTokens: number;\n\t/** Estimated tokens after the most recent applicable assistant usage block. */\n\ttrailingTokens: number;\n\t/** Index of the applicable message that provided usage, or null when none exists. */\n\tlastUsageIndex: number | null;\n}\n\nconst CHARS_PER_TOKEN = 4;\nconst ESTIMATED_IMAGE_CHARS = 4800;\n\nexport function calculateContextTokens(usage: Usage): number {\n\treturn usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite;\n}\n\nfunction safeJsonStringify(value: unknown): string {\n\ttry {\n\t\treturn JSON.stringify(value) ?? \"undefined\";\n\t} catch {\n\t\treturn \"[unserializable]\";\n\t}\n}\n\nfunction estimateTextAndImageContentChars(\n\tcontent: string | Array<TextContent | ImageContent | DocumentContent>,\n): number {\n\tif (typeof content === \"string\") return content.length;\n\n\tlet chars = 0;\n\tfor (const block of content) {\n\t\tif (block.type === \"text\") {\n\t\t\tchars += block.text.length;\n\t\t} else if (block.type === \"document\") {\n\t\t\t// A base64 PDF is far larger than an image and must not be counted as one, or context\n\t\t\t// accounting under-reports badly. Estimate from the encoded payload actually sent.\n\t\t\tchars += block.data.length;\n\t\t} else {\n\t\t\tchars += ESTIMATED_IMAGE_CHARS;\n\t\t}\n\t}\n\treturn chars;\n}\n\nexport function estimateTextTokens(text: string): number {\n\treturn Math.ceil(text.length / CHARS_PER_TOKEN);\n}\n\nexport function estimateTextAndImageContentTokens(\n\tcontent: string | Array<TextContent | ImageContent | DocumentContent>,\n): number {\n\treturn Math.ceil(estimateTextAndImageContentChars(content) / CHARS_PER_TOKEN);\n}\n\nexport function estimateMessageTokens(message: Message): number {\n\tlet chars = 0;\n\n\tif (message.role === \"system\") {\n\t\treturn (\n\t\t\testimateTextTokens(getSystemMessageText(message)) +\n\t\t\testimateToolsTokens(message.toolsAdded) +\n\t\t\testimateToolsTokens(message.toolsRemoved)\n\t\t);\n\t}\n\tif (message.role === \"user\") return estimateTextAndImageContentTokens(message.content);\n\tif (message.role === \"toolResult\") return estimateTextAndImageContentTokens(message.content);\n\n\tfor (const block of message.content) {\n\t\tif (block.type === \"text\") {\n\t\t\tchars += block.text.length;\n\t\t} else if (block.type === \"thinking\") {\n\t\t\tchars += block.thinking.length;\n\t\t} else if (block.type === \"toolCall\") {\n\t\t\tchars += block.name.length + safeJsonStringify(block.arguments).length;\n\t\t}\n\t\t// Fallback boundary markers carry no text and contribute nothing to the estimate.\n\t}\n\treturn Math.ceil(chars / CHARS_PER_TOKEN);\n}\n\nfunction getLastAssistantUsageInfo(messages: readonly Message[]): { usage: Usage; index: number } | undefined {\n\tlet latestPrefixTimestamp = Number.NEGATIVE_INFINITY;\n\tlet usageInfo: { usage: Usage; index: number } | undefined;\n\n\tfor (let i = 0; i < messages.length; i++) {\n\t\tconst message = messages[i];\n\t\tif (message.role === \"assistant\") {\n\t\t\tconst assistant = message as AssistantMessage;\n\t\t\t// A newer prefix message was inserted after this response (for example, a\n\t\t\t// compaction summary), so its usage cannot describe the current prefix.\n\t\t\tconst usageAppliesToPrefix = assistant.timestamp >= latestPrefixTimestamp;\n\t\t\tif (\n\t\t\t\tusageAppliesToPrefix &&\n\t\t\t\tassistant.stopReason !== \"aborted\" &&\n\t\t\t\tassistant.stopReason !== \"error\" &&\n\t\t\t\tcalculateContextTokens(assistant.usage) > 0\n\t\t\t) {\n\t\t\t\tusageInfo = { usage: assistant.usage, index: i };\n\t\t\t}\n\t\t}\n\t\tlatestPrefixTimestamp = Math.max(latestPrefixTimestamp, message.timestamp);\n\t}\n\n\treturn usageInfo;\n}\n\nexport function estimateContextTokens(context: TranscriptContext | readonly Message[]): ContextUsageEstimate {\n\tconst messages = \"messages\" in context ? context.messages : context;\n\tconst usageInfo = getLastAssistantUsageInfo(messages);\n\tif (usageInfo) {\n\t\tconst usageTokens = calculateContextTokens(usageInfo.usage);\n\t\tlet trailingTokens = 0;\n\t\tfor (let i = usageInfo.index + 1; i < messages.length; i++) {\n\t\t\ttrailingTokens += estimateMessageTokens(messages[i]);\n\t\t}\n\t\treturn { tokens: usageTokens + trailingTokens, usageTokens, trailingTokens, lastUsageIndex: usageInfo.index };\n\t}\n\n\tlet tokens = 0;\n\tfor (const message of messages) tokens += estimateMessageTokens(message);\n\treturn { tokens, usageTokens: 0, trailingTokens: tokens, lastUsageIndex: null };\n}\n\nfunction estimateToolsTokens(tools: readonly unknown[] | undefined): number {\n\tif (!tools || tools.length === 0) return 0;\n\treturn estimateTextTokens(safeJsonStringify(tools));\n}\n"]}

package/dist/utils/overflow.d.ts

@@ -28,6 +28,7 @@ import type { AssistantMessage } from "../types.ts";
  * - Kimi For Coding: "exceeded model token limit: X (requested: Y)"
  * - DS4: "Prompt has X tokens, but the configured context size is Y tokens"
  * - DashScope/Qwen: "Range of input length should be [1, X]"
+ * - z.ai: "Prompt too long"
  *
  * **Unreliable detection:**
  * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),

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

@@ -1 +1 @@
-{"version":3,"file":"overflow.d.ts","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AA+EpD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAqDG;AACH,wBAAgB,iBAAiB,CAAC,OAAO,EAAE,gBAAgB,EAAE,aAAa,CAAC,EAAE,MAAM,GAAG,OAAO,CA6B5F;AAED;;;;;GAKG;AACH,wBAAgB,mBAAmB,CAAC,OAAO,EAAE,gBAAgB,EAAE,gBAAgB,EAAE,MAAM,GAAG,OAAO,CAEhG;AAED;;GAEG;AACH,wBAAgB,mBAAmB,IAAI,MAAM,EAAE,CAE9C","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\n\n/**\n * Regex patterns to detect context overflow errors from different providers.\n *\n * These patterns match error messages returned when the input exceeds\n * the model's context window.\n *\n * Provider-specific patterns (with example error messages):\n *\n * - Anthropic: \"prompt is too long: 213462 tokens > 200000 maximum\"\n * - Anthropic: \"413 {\\\"error\\\":{\\\"type\\\":\\\"request_too_large\\\",\\\"message\\\":\\\"Request exceeds the maximum size\\\"}}\"\n * - OpenAI: \"Your input exceeds the context window of this model\"\n * - OpenAI/LiteLLM: \"Requested token count exceeds the model's maximum context length of 131072 tokens\"\n * - OpenAI-compatible: \"Input length (265330) exceeds model's maximum context length (262144).\"\n * - Google: \"The input token count (1196265) exceeds the maximum number of tokens allowed (1048575)\"\n * - xAI: \"This model's maximum prompt length is 131072 but the request contains 537812 tokens\"\n * - Groq: \"Please reduce the length of the messages or completion\"\n * - OpenRouter: \"This endpoint's maximum context length is X tokens. However, you requested about Y tokens\"\n * - OpenRouter/Poolside: \"Input length X exceeds the maximum allowed input length of Y tokens.\"\n * - Together AI: \"The input (X tokens) is longer than the model's context length (Y tokens).\"\n * - llama.cpp: \"the request exceeds the available context size, try increasing it\"\n * - LM Studio: \"tokens to keep from the initial prompt is greater than the context length\"\n * - GitHub Copilot: \"prompt token count of X exceeds the limit of Y\"\n * - MiniMax: \"invalid params, context window exceeds limit\"\n * - Kimi For Coding: \"Your request exceeded model token limit: X (requested: Y)\"\n * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - Cerebras: \"400/413 status code (no body)\"\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - z.ai: Does NOT error, accepts overflow silently - handled via usage.input > contextWindow\n * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason \"length\"\n *   with output=0 (no room left to generate). Detected via stopReason \"length\" + zero output +\n *   input filling the context window.\n * - DashScope/Qwen: \"Range of input length should be [1, X]\" (HTTP 400 invalid_parameter_error)\n * - Ollama: Some deployments truncate silently, others return errors like \"prompt too long; exceeded max context length by X tokens\"\n */\nconst OVERFLOW_PATTERNS = [\n\t/prompt is too long/i, // Anthropic token overflow\n\t/request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)\n\t/input is too long for requested model/i, // Amazon Bedrock\n\t/exceeds the context window/i, // OpenAI (Completions & Responses API)\n\t/exceeds (?:the )?(?:model'?s )?maximum context length(?: of [\\d,]+ tokens?|\\s*\\([\\d,]+\\))/i, // OpenAI-compatible proxies (LiteLLM)\n\t/input token count.*exceeds the maximum/i, // Google (Gemini)\n\t/maximum prompt length is \\d+/i, // xAI (Grok)\n\t/reduce the length of the messages/i, // Groq\n\t/maximum context length is \\d+ tokens/i, // OpenRouter (most backends)\n\t/exceeds (?:the )?maximum allowed input length of [\\d,]+ tokens?/i, // OpenRouter/Poolside\n\t/input \\(\\d+ tokens\\) is longer than the model'?s context length \\(\\d+ tokens\\)/i, // Together AI\n\t/exceeds the limit of \\d+/i, // GitHub Copilot\n\t/exceeds the available context size/i, // llama.cpp server\n\t/greater than the context length/i, // LM Studio\n\t/context window exceeds limit/i, // MiniMax\n\t/exceeded model token limit/i, // Kimi For Coding\n\t/too large for model with \\d+ maximum context length/i, // Mistral\n\t/prompt has [\\d,]+ tokens?, but the configured context size is [\\d,]+ tokens?/i, // DS4 server\n\t/model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text\n\t/prompt too long; exceeded (?:max )?context length/i, // Ollama explicit overflow error\n\t/range of input length should be/i, // DashScope / Qwen Token Plan\n\t/context[_ ]length[_ ]exceeded/i, // Generic fallback\n\t/too many tokens/i, // Generic fallback\n\t/token limit exceeded/i, // Generic fallback\n\t/^4(?:00|13)\\s*(?:status code)?\\s*\\(no body\\)/i, // Cerebras: 400/413 with no body\n];\n\n/**\n * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).\n * Error messages matching any of these are excluded from overflow detection\n * even if they also match an OVERFLOW_PATTERN.\n *\n * Example: Bedrock formats throttling errors as \"ThrottlingException: Too many tokens,\n * please wait before trying again.\" which would match the /too many tokens/i overflow\n * pattern without this exclusion.\n */\nconst NON_OVERFLOW_PATTERNS = [\n\t/^(Throttling error|Service unavailable):/i, // AWS Bedrock non-overflow errors (human-readable prefixes from formatBedrockError)\n\t/rate limit/i, // Generic rate limiting\n\t/too many requests/i, // Generic HTTP 429 style\n];\n\n/**\n * Check if an assistant message represents a context overflow error.\n *\n * This handles three cases:\n * 1. Error-based overflow: Most providers return stopReason \"error\" with a\n *    specific error message pattern.\n * 2. Silent overflow: Some providers accept overflow requests and return\n *    successfully. For these, we check if usage.input exceeds the context window.\n * 3. Length-stop overflow: Xiaomi MiMo can return \"length\" with zero output when\n *    the input fills the context window.\n *\n * ## Reliability by Provider\n *\n * **Reliable detection (returns error with detectable message):**\n * - Anthropic: \"prompt is too long: X tokens > Y maximum\" or \"request_too_large\"\n * - OpenAI (Completions & Responses): \"exceeds the context window\", \"exceeds the model's maximum context length of X tokens\", or \"exceeds model's maximum context length (X)\"\n * - Google Gemini: \"input token count exceeds the maximum\"\n * - xAI (Grok): \"maximum prompt length is X but request contains Y\"\n * - Groq: \"reduce the length of the messages\"\n * - Cerebras: 400/413 status code (no body)\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - OpenRouter (most backends): \"maximum context length is X tokens\"\n * - OpenRouter/Poolside: \"Input length X exceeds the maximum allowed input length of Y tokens.\"\n * - Together AI: \"The input (X tokens) is longer than the model's context length (Y tokens).\"\n * - llama.cpp: \"exceeds the available context size\"\n * - LM Studio: \"greater than the context length\"\n * - Kimi For Coding: \"exceeded model token limit: X (requested: Y)\"\n * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - DashScope/Qwen: \"Range of input length should be [1, X]\"\n *\n * **Unreliable detection:**\n * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),\n *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.\n * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason \"length\" with\n *   output=0. Pass contextWindow param to detect via the \"filled context + zero output\" signal.\n * - Ollama: May truncate input silently for some setups, but may also return explicit\n *   overflow errors that match the patterns above. Silent truncation still cannot be\n *   detected here because we do not know the expected token count.\n *\n * ## Custom Providers\n *\n * If you've added custom models via settings.json, this function may not detect\n * overflow errors from those providers. To add support:\n *\n * 1. Send a request that exceeds the model's context window\n * 2. Check the errorMessage in the response\n * 3. Create a regex pattern that matches the error\n * 4. The pattern should be added to OVERFLOW_PATTERNS in this file, or\n *    check the errorMessage yourself before calling this function\n *\n * @param message - The assistant message to check\n * @param contextWindow - Optional context window size for detecting silent overflow (z.ai)\n * @returns true if the message indicates a context overflow\n */\nexport function isContextOverflow(message: AssistantMessage, contextWindow?: number): boolean {\n\t// Case 1: Check error message patterns\n\tif (message.stopReason === \"error\" && message.errorMessage) {\n\t\t// Skip messages matching known non-overflow patterns (e.g. throttling / rate-limit)\n\t\tconst isNonOverflow = NON_OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!));\n\t\tif (!isNonOverflow && OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!))) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 2: Silent overflow (z.ai style) - successful but usage exceeds context\n\tif (contextWindow && message.stopReason === \"stop\") {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens > contextWindow) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 3: Length-stop overflow (Xiaomi MiMo style) - server truncates oversized input\n\t// to fit the context window, leaving no room for output. Returns stopReason \"length\"\n\t// with output=0 and input+cacheRead filling the context window.\n\tif (contextWindow && message.stopReason === \"length\" && message.usage.output === 0) {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens >= contextWindow * 0.99) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n\n/**\n * Check whether a length stop ended below the caller or model's intended output limit.\n * Such responses may be caused by context pressure or provider-side truncation, so callers\n * can make one bounded compact-and-retry attempt. `desiredMaxOutput` must be the original\n * limit before any context-based clamping.\n */\nexport function isRecoverableLength(message: AssistantMessage, desiredMaxOutput: number): boolean {\n\treturn message.stopReason === \"length\" && desiredMaxOutput > 0 && message.usage.output < desiredMaxOutput;\n}\n\n/**\n * Get the overflow patterns for testing purposes.\n */\nexport function getOverflowPatterns(): RegExp[] {\n\treturn [...OVERFLOW_PATTERNS];\n}\n"]}
+{"version":3,"file":"overflow.d.ts","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AAwFpD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAsDG;AACH,wBAAgB,iBAAiB,CAAC,OAAO,EAAE,gBAAgB,EAAE,aAAa,CAAC,EAAE,MAAM,GAAG,OAAO,CAkC5F;AAED;;;;;GAKG;AACH,wBAAgB,mBAAmB,CAAC,OAAO,EAAE,gBAAgB,EAAE,gBAAgB,EAAE,MAAM,GAAG,OAAO,CAEhG;AAED;;GAEG;AACH,wBAAgB,mBAAmB,IAAI,MAAM,EAAE,CAE9C","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\n\n/**\n * Regex patterns to detect context overflow errors from different providers.\n *\n * These patterns match error messages returned when the input exceeds\n * the model's context window.\n *\n * Provider-specific patterns (with example error messages):\n *\n * - Anthropic: \"prompt is too long: 213462 tokens > 200000 maximum\"\n * - Anthropic: \"413 {\\\"error\\\":{\\\"type\\\":\\\"request_too_large\\\",\\\"message\\\":\\\"Request exceeds the maximum size\\\"}}\"\n * - OpenAI: \"Your input exceeds the context window of this model\"\n * - OpenAI/LiteLLM: \"Requested token count exceeds the model's maximum context length of 131072 tokens\"\n * - OpenAI-compatible: \"Input length (265330) exceeds model's maximum context length (262144).\"\n * - Google: \"The input token count (1196265) exceeds the maximum number of tokens allowed (1048575)\"\n * - xAI: \"This model's maximum prompt length is 131072 but the request contains 537812 tokens\"\n * - Groq: \"Please reduce the length of the messages or completion\"\n * - OpenRouter: \"This endpoint's maximum context length is X tokens. However, you requested about Y tokens\"\n * - OpenRouter/Poolside: \"Input length X exceeds the maximum allowed input length of Y tokens.\"\n * - Together AI: \"The input (X tokens) is longer than the model's context length (Y tokens).\"\n * - llama.cpp: \"the request exceeds the available context size, try increasing it\"\n * - LM Studio: \"tokens to keep from the initial prompt is greater than the context length\"\n * - GitHub Copilot: \"prompt token count of X exceeds the limit of Y\"\n * - MiniMax: \"invalid params, context window exceeds limit\"\n * - Kimi For Coding: \"Your request exceeded model token limit: X (requested: Y)\"\n * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - Cerebras: \"400/413 status code (no body)\"\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - z.ai: `{\"code\":\"1261\",\"message\":\"Prompt too long\"}` or silent overflow via usage.input > contextWindow\n * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason \"length\"\n *   with output=0 (no room left to generate). Detected via stopReason \"length\" + zero output +\n *   input filling the context window.\n * - DashScope/Qwen: \"Range of input length should be [1, X]\" (HTTP 400 invalid_parameter_error)\n * - Ollama: Some deployments truncate silently, others return errors like \"prompt too long; exceeded max context length by X tokens\"\n */\nconst OVERFLOW_PATTERNS = [\n\t/prompt (?:is )?too long/i, // Anthropic and z.ai token overflow\n\t/request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)\n\t/input is too long for requested model/i, // Amazon Bedrock\n\t/exceeds the context window/i, // OpenAI (Completions & Responses API)\n\t/exceeds (?:the )?(?:model'?s )?maximum context length(?: of [\\d,]+ tokens?|\\s*\\([\\d,]+\\))/i, // OpenAI-compatible proxies (LiteLLM)\n\t/input token count.*exceeds the maximum/i, // Google (Gemini)\n\t/maximum prompt length is \\d+/i, // xAI (Grok)\n\t/reduce the length of the messages/i, // Groq\n\t/maximum context length is \\d+ tokens/i, // OpenRouter (most backends)\n\t/exceeds (?:the )?maximum allowed input length of [\\d,]+ tokens?/i, // OpenRouter/Poolside\n\t/input \\(\\d+ tokens\\) is longer than the model'?s context length \\(\\d+ tokens\\)/i, // Together AI\n\t/exceeds the limit of \\d+/i, // GitHub Copilot\n\t/exceeds the available context size/i, // llama.cpp server\n\t/greater than the context length/i, // LM Studio\n\t/context window exceeds limit/i, // MiniMax\n\t/exceeded model token limit/i, // Kimi For Coding\n\t/too large for model with \\d+ maximum context length/i, // Mistral\n\t/prompt has [\\d,]+ tokens?, but the configured context size is [\\d,]+ tokens?/i, // DS4 server\n\t/model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text\n\t/prompt too long; exceeded (?:max )?context length/i, // Ollama explicit overflow error\n\t/range of input length should be/i, // DashScope / Qwen Token Plan\n\t/context[_ ]length[_ ]exceeded/i, // Generic fallback\n\t/too many tokens/i, // Generic fallback\n\t/token limit exceeded/i, // Generic fallback\n];\n\nfunction isCerebrasBodylessOverflow(errorMessage: string): boolean {\n\tconst normalized = errorMessage.trim().toLowerCase();\n\treturn (\n\t\tnormalized === \"400 status code (no body)\" ||\n\t\tnormalized === \"413 status code (no body)\" ||\n\t\tnormalized === \"400 (no body)\" ||\n\t\tnormalized === \"413 (no body)\"\n\t);\n}\n\n/**\n * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).\n * Error messages matching any of these are excluded from overflow detection\n * even if they also match an OVERFLOW_PATTERN.\n *\n * Example: Bedrock formats throttling errors as \"ThrottlingException: Too many tokens,\n * please wait before trying again.\" which would match the /too many tokens/i overflow\n * pattern without this exclusion.\n */\nconst NON_OVERFLOW_PATTERNS = [\n\t/^(Throttling error|Service unavailable):/i, // AWS Bedrock non-overflow errors (human-readable prefixes from formatBedrockError)\n\t/rate limit/i, // Generic rate limiting\n\t/too many requests/i, // Generic HTTP 429 style\n];\n\n/**\n * Check if an assistant message represents a context overflow error.\n *\n * This handles three cases:\n * 1. Error-based overflow: Most providers return stopReason \"error\" with a\n *    specific error message pattern.\n * 2. Silent overflow: Some providers accept overflow requests and return\n *    successfully. For these, we check if usage.input exceeds the context window.\n * 3. Length-stop overflow: Xiaomi MiMo can return \"length\" with zero output when\n *    the input fills the context window.\n *\n * ## Reliability by Provider\n *\n * **Reliable detection (returns error with detectable message):**\n * - Anthropic: \"prompt is too long: X tokens > Y maximum\" or \"request_too_large\"\n * - OpenAI (Completions & Responses): \"exceeds the context window\", \"exceeds the model's maximum context length of X tokens\", or \"exceeds model's maximum context length (X)\"\n * - Google Gemini: \"input token count exceeds the maximum\"\n * - xAI (Grok): \"maximum prompt length is X but request contains Y\"\n * - Groq: \"reduce the length of the messages\"\n * - Cerebras: 400/413 status code (no body)\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - OpenRouter (most backends): \"maximum context length is X tokens\"\n * - OpenRouter/Poolside: \"Input length X exceeds the maximum allowed input length of Y tokens.\"\n * - Together AI: \"The input (X tokens) is longer than the model's context length (Y tokens).\"\n * - llama.cpp: \"exceeds the available context size\"\n * - LM Studio: \"greater than the context length\"\n * - Kimi For Coding: \"exceeded model token limit: X (requested: Y)\"\n * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - DashScope/Qwen: \"Range of input length should be [1, X]\"\n * - z.ai: \"Prompt too long\"\n *\n * **Unreliable detection:**\n * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),\n *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.\n * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason \"length\" with\n *   output=0. Pass contextWindow param to detect via the \"filled context + zero output\" signal.\n * - Ollama: May truncate input silently for some setups, but may also return explicit\n *   overflow errors that match the patterns above. Silent truncation still cannot be\n *   detected here because we do not know the expected token count.\n *\n * ## Custom Providers\n *\n * If you've added custom models via settings.json, this function may not detect\n * overflow errors from those providers. To add support:\n *\n * 1. Send a request that exceeds the model's context window\n * 2. Check the errorMessage in the response\n * 3. Create a regex pattern that matches the error\n * 4. The pattern should be added to OVERFLOW_PATTERNS in this file, or\n *    check the errorMessage yourself before calling this function\n *\n * @param message - The assistant message to check\n * @param contextWindow - Optional context window size for detecting silent overflow (z.ai)\n * @returns true if the message indicates a context overflow\n */\nexport function isContextOverflow(message: AssistantMessage, contextWindow?: number): boolean {\n\t// Case 1: Check error message patterns\n\tif (message.stopReason === \"error\" && message.errorMessage) {\n\t\t// Skip messages matching known non-overflow patterns (e.g. throttling / rate-limit)\n\t\tconst isNonOverflow = NON_OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!));\n\t\tif (!isNonOverflow) {\n\t\t\tif (OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!))) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t\tif (message.provider === \"cerebras\" && isCerebrasBodylessOverflow(message.errorMessage)) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t}\n\t}\n\n\t// Case 2: Silent overflow (z.ai style) - successful but usage exceeds context\n\tif (contextWindow && message.stopReason === \"stop\") {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens > contextWindow) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 3: Length-stop overflow (Xiaomi MiMo style) - server truncates oversized input\n\t// to fit the context window, leaving no room for output. Returns stopReason \"length\"\n\t// with output=0 and input+cacheRead filling the context window.\n\tif (contextWindow && message.stopReason === \"length\" && message.usage.output === 0) {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens >= contextWindow * 0.99) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n\n/**\n * Check whether a length stop ended below the caller or model's intended output limit.\n * Such responses may be caused by context pressure or provider-side truncation, so callers\n * can make one bounded compact-and-retry attempt. `desiredMaxOutput` must be the original\n * limit before any context-based clamping.\n */\nexport function isRecoverableLength(message: AssistantMessage, desiredMaxOutput: number): boolean {\n\treturn message.stopReason === \"length\" && desiredMaxOutput > 0 && message.usage.output < desiredMaxOutput;\n}\n\n/**\n * Get the overflow patterns for testing purposes.\n */\nexport function getOverflowPatterns(): RegExp[] {\n\treturn [...OVERFLOW_PATTERNS];\n}\n"]}

package/dist/utils/overflow.js

@@ -25,7 +25,7 @@
  * - DS4: "Prompt has X tokens, but the configured context size is Y tokens"
  * - Cerebras: "400/413 status code (no body)"
  * - Mistral: "Prompt contains X tokens ... too large for model with Y maximum context length"
- * - z.ai: Does NOT error, accepts overflow silently - handled via usage.input > contextWindow
+ * - z.ai: `{"code":"1261","message":"Prompt too long"}` or silent overflow via usage.input > contextWindow
  * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason "length"
  *   with output=0 (no room left to generate). Detected via stopReason "length" + zero output +
  *   input filling the context window.
@@ -33,7 +33,7 @@
  * - Ollama: Some deployments truncate silently, others return errors like "prompt too long; exceeded max context length by X tokens"
  */
 const OVERFLOW_PATTERNS = [
-    /prompt is too long/i, // Anthropic token overflow
+    /prompt (?:is )?too long/i, // Anthropic and z.ai token overflow
     /request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)
     /input is too long for requested model/i, // Amazon Bedrock
     /exceeds the context window/i, // OpenAI (Completions & Responses API)
@@ -57,8 +57,14 @@ const OVERFLOW_PATTERNS = [
     /context[_ ]length[_ ]exceeded/i, // Generic fallback
     /too many tokens/i, // Generic fallback
     /token limit exceeded/i, // Generic fallback
-    /^4(?:00|13)\s*(?:status code)?\s*\(no body\)/i, // Cerebras: 400/413 with no body
 ];
+function isCerebrasBodylessOverflow(errorMessage) {
+    const normalized = errorMessage.trim().toLowerCase();
+    return (normalized === "400 status code (no body)" ||
+        normalized === "413 status code (no body)" ||
+        normalized === "400 (no body)" ||
+        normalized === "413 (no body)");
+}
 /**
  * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).
  * Error messages matching any of these are excluded from overflow detection
@@ -102,6 +108,7 @@ const NON_OVERFLOW_PATTERNS = [
  * - Kimi For Coding: "exceeded model token limit: X (requested: Y)"
  * - DS4: "Prompt has X tokens, but the configured context size is Y tokens"
  * - DashScope/Qwen: "Range of input length should be [1, X]"
+ * - z.ai: "Prompt too long"
  *
  * **Unreliable detection:**
  * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),
@@ -132,8 +139,13 @@ export function isContextOverflow(message, contextWindow) {
     if (message.stopReason === "error" && message.errorMessage) {
         // Skip messages matching known non-overflow patterns (e.g. throttling / rate-limit)
         const isNonOverflow = NON_OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage));
-        if (!isNonOverflow && OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage))) {
-            return true;
+        if (!isNonOverflow) {
+            if (OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage))) {
+                return true;
+            }
+            if (message.provider === "cerebras" && isCerebrasBodylessOverflow(message.errorMessage)) {
+                return true;
+            }
         }
     }
     // Case 2: Silent overflow (z.ai style) - successful but usage exceeds context