@code-yeongyu/senpi 2026.8.23 → 2026.8.25

package/node_modules/@earendil-works/pi-ai/dist/types.js.map

@@ -1 +1 @@
-{"version":3,"file":"types.js","sourceRoot":"","sources":["../src/types.ts"],"names":[],"mappings":"AAybA,sFAAsF;AACtF,MAAM,UAAU,eAAe,CAAC,QAAgB;IAC/C,OAAO,QAAQ,CAAC,WAAW,EAAE,CAAC,UAAU,CAAC,QAAQ,CAAC,CAAC;AACpD,CAAC","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 { CursorAgentOptions } from \"./api/cursor-agent/types.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 { Model } from \"./model.ts\";\nimport type {\n\tOpenAIResponsesCompat as BaseOpenAIResponsesCompat,\n\tSessionAffinityFormat,\n} from \"./openai-responses-compat.ts\";\nimport type { AssistantMessageDiagnostic } from \"./utils/diagnostics.ts\";\nimport type { AssistantMessageEventStream } from \"./utils/event-stream.ts\";\n\nexport type { Model } from \"./model.ts\";\nexport type { SessionAffinityFormat } from \"./openai-responses-compat.ts\";\nexport type OpenAIResponsesCompat = BaseOpenAIResponsesCompat & {\n\t/** Whether the model supports message-anchored `additional_tools` input items. Default: false. */\n\tsupportsAdditionalTools?: boolean;\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| \"cursor-agent\"\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\" | \"openai-images\";\n\nexport type ImagesApi = KnownImagesApi | (string & {});\n\nexport type KnownProvider =\n\t| \"alibaba-token-plan\"\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| \"ollama\"\n\t| \"cursor\"\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| \"opengateway\"\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;\n\n/**\n * Provenance-bearing thinking selection. Distinct from the always-materialized\n * effective level: a selection exists only when the user (or a legacy variant\n * alias) explicitly chose a level. See .omo/plans/cursor-reasoning-levels.md §4.1.\n */\nexport interface ThinkingSelection {\n\treadonly level: ModelThinkingLevel;\n\treadonly source: \"explicit\" | \"legacy-variant\";\n\t/** Required when source is \"legacy-variant\": the exact allowlisted original variant id. */\n\treadonly legacyVariantId?: string;\n}\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\tmax?: 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>;\n/** Effective model and fully transformed headers for a request payload hook. */\nexport type ProviderRequestMetadata = {\n\tmodel: Model<Api>;\n\theaders: ProviderHeaders;\n};\nexport type FetchFunction = typeof globalThis.fetch;\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/**\n\t * Abort the request when the provider reports that a safety classifier\n\t * declined the requested model and a substitute model served the turn\n\t * instead (Anthropic `server-side-fallback-*` betas). The substitute's output\n\t * is billed but was never requested, so aborting keeps model selection with\n\t * the caller. Providers without server-side fallback ignore this.\n\t * Default: undefined (honor the substituted response).\n\t */\n\tabortServerSideFallback?: boolean;\n\t/** Explicit parent context for telemetry produced by this logical request. */\n\ttelemetryContext?: TelemetryContext;\n\tapiKey?: string;\n\t/**\n\t * Stable identity of the originating session for account-affinity providers,\n\t * preserved across auxiliary calls that replace `sessionId` (for example\n\t * compaction). Providers without account affinity ignore this.\n\t */\n\taffinitySessionId?: string;\n\t/**\n\t * Identifies whether the stream belongs to the main agent loop or an auxiliary request.\n\t * An absent value must be treated as auxiliary as a fail-safe.\n\t */\n\tstreamKind?: \"main\" | \"auxiliary\";\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?: (\n\t\tpayload: unknown,\n\t\tmodel: TModel,\n\t\trequest?: ProviderRequestMetadata,\n\t) => 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\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 * Optional custom fields to merge into the outgoing provider request body.\n\t * Applied before `onPayload` so `onPayload` can observe/override them.\n\t * Provider-managed fields (e.g. `model`, `messages`, `stream`) are preserved\n\t * and cannot be overridden to avoid breaking core request semantics.\n\t */\n\textraBody?: Record<string, unknown>;\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 * stream idleness after connection uses timeoutMs.\n\t */\n\twebsocketConnectTimeoutMs?: 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\"cursor-agent\": CursorAgentOptions;\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 type AnthropicRefusalFallback = \"default\" | readonly { model: string }[];\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. Default: \"auto\". */\n\ttoolChoice?: ToolChoice;\n\treasoning?: ThinkingLevel;\n\t/**\n\t * Provenance-bearing thinking selection. Providers that need to distinguish\n\t * an explicit user choice from the always-materialized effective level read\n\t * this; `reasoning` remains the normalized effective level for everyone.\n\t */\n\tthinkingSelection?: ThinkingSelection;\n\t/** Anthropic server-side fallback for eligible refusal stop reasons. Anthropic providers only. */\n\trefusalFallbacks?: AnthropicRefusalFallback;\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// - Once invoked, request/model/runtime failures should be encoded in the\n//   returned 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\t/** Epoch ms, stamped by the agent loop at stream-event receipt, best-effort. */\n\tstartedAt?: number;\n\t/** Epoch ms, stamped by the agent loop at stream-event receipt, best-effort. */\n\tendedAt?: number;\n\tthinkingSignature?: string; // e.g., for OpenAI responses, the reasoning item ID\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\nexport interface ImageContent {\n\ttype: \"image\";\n\tdata: string; // base64 encoded image data\n\tmimeType: string; // e.g., \"image/jpeg\", \"image/png\". Video payloads (e.g. \"video/mp4\") ride this\n\t// same block for models that declare the \"video\" input modality; use isVideoMimeType() to branch.\n}\n\n/** True when an ImageContent block actually carries video data (e.g. \"video/mp4\"). */\nexport function isVideoMimeType(mimeType: string): boolean {\n\treturn mimeType.toLowerCase().startsWith(\"video/\");\n}\n\nexport interface ToolCall {\n\ttype: \"toolCall\";\n\tid: string;\n\tname: string;\n\targuments: Record<string, any>;\n\t/** Set by text tool-call middleware when a truncated call could not be recovered. Carriers of `incomplete` MUST NOT be executed. */\n\tincomplete?: true;\n\t/** Error explaining why an incomplete tool call could not be recovered. */\n\terrorMessage?: string;\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\n/**\n * Provider-emitted content block whose semantics are owned by the originating\n * provider. Surfaces server-side tool invocations (web_search), server-side\n * tool results (web_search_tool_result), code execution traces, grounding\n * metadata fragments, and any other vendor-specific block the core does not\n * normalize.\n *\n * `subtype` mirrors the provider's wire `type` discriminator (e.g.\n * \"server_tool_use\", \"web_search_tool_result\", \"executableCode\"). `raw` is\n * the verbatim wire payload — extensions interpret it per provider+subtype.\n *\n * The owning provider id lives on the parent `AssistantMessage.provider`, not\n * here, to avoid drift on replay.\n *\n * On replay across providers, `convertMessages` implementations MUST drop\n * these blocks (they are not portable). This mirrors the existing\n * `redacted_thinking` skip pattern.\n */\nexport interface ProviderNativeContent {\n\ttype: \"providerNative\";\n\tsubtype: string;\n\traw: unknown;\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 type AssistantStopDetails = { type: \"refusal\"; explanation?: string } | { type: \"sensitive\" };\n\nexport interface UserMessage {\n\trole: \"user\";\n\tcontent: string | (TextContent | ImageContent)[];\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport interface AssistantMessage {\n\trole: \"assistant\";\n\tcontent: (TextContent | ThinkingContent | ToolCall | ProviderNativeContent)[];\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\tdiagnostics?: AssistantMessageDiagnostic[]; // Redacted provider/runtime diagnostics for failures and recoveries.\n\tusage: Usage;\n\tstopReason: StopReason;\n\tstopDetails?: AssistantStopDetails;\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\nexport interface FreeformToolFormat {\n\ttype: \"grammar\";\n\tsyntax: \"lark\";\n\tdefinition: string;\n}\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  };\nexport interface Tool<TParameters extends TSchema = TSchema> {\n\tname: string;\n\tdescription: string;\n\tparameters: TParameters;\n\tfreeform?: FreeformToolFormat;\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 * Streams should emit `start` before partial updates, then terminate with either:\n * - `done` carrying the final successful AssistantMessage, or\n * - `error` carrying the final AssistantMessage with stopReason \"error\" or \"aborted\"\n *   and errorMessage.\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/** Finalized - executable iff `incomplete !== true` on `toolCall`. */\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/** 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/** Whether the provider accepts explicit disabled-thinking markers when thinking is off. Default: true. */\n\tsupportsDisabledThinking?: boolean;\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/**\n\t * Provider-specific JSON Schema flavor for tool parameters. `\"moonshot-mfjs\"`\n\t * normalizes schemas for Moonshot / Kimi backends that enforce a stricter subset.\n\t */\n\ttoolSchemaFlavor?: \"moonshot-mfjs\";\n\t/**\n\t * Tool call format for models that don't natively support tool calling.\n\t * When set, the middleware will intercept tool calls and format them as text.\n\t * Supported values: \"hermes\", \"morph-xml\", \"xml\" (deprecated alias for \"morph-xml\"), \"yaml-xml\", \"gemma4-delimiter\", \"anthropic-xml\", \"antml\", \"kimi-xtml\"\n\t */\n\ttoolCallFormat?: string;\n\t/** Cache control convention for prompt caching. \"anthropic\" applies Anthropic-style `cache_control` markers to the system prompt, last tool definition, and last user, assistant, or tool-result text content. */\n\tcacheControlFormat?: \"anthropic\";\n\t/** Whether to send session-affinity data from `options.sessionId`. Default: false. */\n\tsendSessionAffinityHeaders?: boolean;\n\t/** Provider-specific deferred tool serialization mode. */\n\tdeferredToolsMode?: \"kimi\";\n\t/** Session-affinity header format: `openai` sends `session_id`, `x-client-request-id`, and `x-session-affinity`; `openai-nosession` sends `x-client-request-id` and `x-session-affinity`; `openrouter` sends `x-session-id`. Does not affect the `prompt_cache_key` body param, which is governed by cache retention. Default: auto-detected. */\n\tsessionAffinityFormat?: SessionAffinityFormat;\n\t/** Whether the provider accepts the `prompt_cache_key` request field. Default: auto-detected. */\n\tsupportsPromptCacheKey?: boolean;\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}\n\n/** Compatibility settings for Anthropic Messages-compatible APIs. */\nexport interface AnthropicMessagesCompat {\n\t/**\n\t * Whether the provider accepts per-tool `eager_input_streaming`.\n\t * When false, the Anthropic provider omits `tools[].eager_input_streaming`\n\t * and sends the legacy `fine-grained-tool-streaming-2025-05-14` beta header\n\t * for tool-enabled requests.\n\t * Default: true.\n\t */\n\tsupportsEagerToolInputStreaming?: boolean;\n\t/** Whether the provider supports Anthropic long cache retention (`cache_control.ttl: \"1h\"`). Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n\t/**\n\t * Whether to send the `x-session-affinity` header from `options.sessionId`\n\t * when caching is enabled. Required for providers like Fireworks that use\n\t * session affinity for prompt cache routing (requests to the same replica\n\t * maximize cache hits).\n\t * Default: false.\n\t */\n\tsendSessionAffinityHeaders?: boolean;\n\t/**\n\t * Whether the provider supports Anthropic-style `cache_control` markers on\n\t * tool definitions. When false, `cache_control` is omitted from tool params.\n\t * Some Anthropic-compatible providers (e.g., Fireworks) do not support this\n\t * field on tools and may reject or ignore it.\n\t * Default: true.\n\t */\n\tsupportsCacheControlOnTools?: boolean;\n\t/**\n\t * Whether the provider accepts `thinking: { type: \"disabled\" }` when\n\t * extended thinking is off. Some Anthropic-compatible providers support\n\t * thinking but reject Anthropic's explicit disabled marker.\n\t * Default: true.\n\t */\n\tsupportsDisabledThinking?: 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/** Whether the model accepts Anthropic `tool_choice`. Default: true. */\n\tsupportsToolChoice?: boolean;\n\t/**\n\t * Whether the model accepts forced Anthropic tool choices (`any` or a named\n\t * tool). Default: true, except Claude Fable/Mythos models.\n\t */\n\tsupportsForcedToolChoice?: 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 opt out on\n\t * overridden built-in models and disable model id/name inference.\n\t * Default: infer from adaptive Claude model ids/names when possible.\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/**\n\t * How to replay thinking blocks that have no usable Anthropic signature.\n\t * `\"text\"` demotes them to text; `\"empty-signature\"` preserves Anthropic's\n\t * thinking shape with `signature: \"\"`. Defaults to `\"text\"`; the legacy\n\t * `allowEmptySignature` setting remains a compatibility alias for\n\t * `\"empty-signature\"`.\n\t */\n\tunsignedThinkingReplay?: \"text\" | \"empty-signature\";\n\t/**\n\t * Model ids Anthropic accepts in `fallbacks` for server-side refusal fallback.\n\t * When absent or empty, callers must omit `fallbacks`; Anthropic rejects the\n\t * field for models with no permitted fallback targets.\n\t */\n\tallowedFallbackModels?: string[];\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 provider executes Anthropic server-side `web_search_*` native\n\t * tools and accepts their `server_tool_use` / `web_search_tool_result`\n\t * blocks on replay. Anthropic-compatible endpoints (e.g., kimi-coding) may\n\t * run the search but reject the replayed server-tool blocks on the next\n\t * request. When false, hook-injected `web_search_*` tools are stripped from\n\t * the payload. Default: true only for the first-party Anthropic endpoint.\n\t */\n\tsupportsWebSearch?: 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}\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\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":"AA+bA,sFAAsF;AACtF,MAAM,UAAU,eAAe,CAAC,QAAgB;IAC/C,OAAO,QAAQ,CAAC,WAAW,EAAE,CAAC,UAAU,CAAC,QAAQ,CAAC,CAAC;AACpD,CAAC","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 { CursorAgentOptions } from \"./api/cursor-agent/types.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 { Model } from \"./model.ts\";\nimport type {\n\tOpenAIResponsesCompat as BaseOpenAIResponsesCompat,\n\tSessionAffinityFormat,\n} from \"./openai-responses-compat.ts\";\nimport type { AssistantMessageDiagnostic } from \"./utils/diagnostics.ts\";\nimport type { AssistantMessageEventStream } from \"./utils/event-stream.ts\";\n\nexport type { Model } from \"./model.ts\";\nexport type { SessionAffinityFormat } from \"./openai-responses-compat.ts\";\nexport type OpenAIResponsesCompat = BaseOpenAIResponsesCompat & {\n\t/** Whether the model supports message-anchored `additional_tools` input items. Default: false. */\n\tsupportsAdditionalTools?: boolean;\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| \"cursor-agent\"\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\" | \"openai-images\";\n\nexport type ImagesApi = KnownImagesApi | (string & {});\n\nexport type KnownProvider =\n\t| \"alibaba-token-plan\"\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| \"ollama\"\n\t| \"cursor\"\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| \"opengateway\"\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;\n\n/**\n * Provenance-bearing thinking selection. Distinct from the always-materialized\n * effective level: a selection exists only when the user (or a legacy variant\n * alias) explicitly chose a level. See .omo/plans/cursor-reasoning-levels.md §4.1.\n */\nexport interface ThinkingSelection {\n\treadonly level: ModelThinkingLevel;\n\treadonly source: \"explicit\" | \"legacy-variant\";\n\t/** Required when source is \"legacy-variant\": the exact allowlisted original variant id. */\n\treadonly legacyVariantId?: string;\n}\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\tmax?: 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>;\n/** Effective model and fully transformed headers for a request payload hook. */\nexport type ProviderRequestMetadata = {\n\tmodel: Model<Api>;\n\theaders: ProviderHeaders;\n};\nexport type FetchFunction = typeof globalThis.fetch;\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/**\n\t * Abort the request when the provider reports that a safety classifier\n\t * declined the requested model and a substitute model served the turn\n\t * instead (Anthropic `server-side-fallback-*` betas). The substitute's output\n\t * is billed but was never requested, so aborting keeps model selection with\n\t * the caller. Providers without server-side fallback ignore this.\n\t * Default: undefined (honor the substituted response).\n\t */\n\tabortServerSideFallback?: boolean;\n\t/** Explicit parent context for telemetry produced by this logical request. */\n\ttelemetryContext?: TelemetryContext;\n\tapiKey?: string;\n\t/**\n\t * Stable identity of the originating session for account-affinity providers,\n\t * preserved across auxiliary calls that replace `sessionId` (for example\n\t * compaction). Providers without account affinity ignore this.\n\t */\n\taffinitySessionId?: string;\n\t/**\n\t * Identifies whether the stream belongs to the main agent loop or an auxiliary request.\n\t * An absent value must be treated as auxiliary as a fail-safe.\n\t */\n\tstreamKind?: \"main\" | \"auxiliary\";\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?: (\n\t\tpayload: unknown,\n\t\tmodel: TModel,\n\t\trequest?: ProviderRequestMetadata,\n\t) => 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\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 * Optional custom fields to merge into the outgoing provider request body.\n\t * Applied before `onPayload` so `onPayload` can observe/override them.\n\t * Provider-managed fields (e.g. `model`, `messages`, `stream`) are preserved\n\t * and cannot be overridden to avoid breaking core request semantics.\n\t */\n\textraBody?: Record<string, unknown>;\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 * stream idleness after connection uses timeoutMs.\n\t */\n\twebsocketConnectTimeoutMs?: 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\"cursor-agent\": CursorAgentOptions;\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\nexport type AnthropicRefusalFallback = \"default\" | readonly { model: string }[];\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. Default: \"auto\". */\n\ttoolChoice?: ToolChoice;\n\treasoning?: ThinkingLevel;\n\t/**\n\t * Provenance-bearing thinking selection. Providers that need to distinguish\n\t * an explicit user choice from the always-materialized effective level read\n\t * this; `reasoning` remains the normalized effective level for everyone.\n\t */\n\tthinkingSelection?: ThinkingSelection;\n\t/** Anthropic server-side fallback for eligible refusal stop reasons. Anthropic providers only. */\n\trefusalFallbacks?: AnthropicRefusalFallback;\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// - Once invoked, request/model/runtime failures should be encoded in the\n//   returned 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\t/** Epoch ms, stamped by the agent loop at stream-event receipt, best-effort. */\n\tstartedAt?: number;\n\t/** Epoch ms, stamped by the agent loop at stream-event receipt, best-effort. */\n\tendedAt?: number;\n\tthinkingSignature?: string; // e.g., for OpenAI responses, the reasoning item ID\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\nexport interface ImageContent {\n\ttype: \"image\";\n\tdata: string; // base64 encoded image data\n\tmimeType: string; // e.g., \"image/jpeg\", \"image/png\". Video payloads (e.g. \"video/mp4\") ride this\n\t// same block for models that declare the \"video\" input modality; use isVideoMimeType() to branch.\n}\n\n/** True when an ImageContent block actually carries video data (e.g. \"video/mp4\"). */\nexport function isVideoMimeType(mimeType: string): boolean {\n\treturn mimeType.toLowerCase().startsWith(\"video/\");\n}\n\nexport interface ToolCall {\n\ttype: \"toolCall\";\n\tid: string;\n\tname: string;\n\targuments: Record<string, any>;\n\t/** Set by text tool-call middleware when a truncated call could not be recovered. Carriers of `incomplete` MUST NOT be executed. */\n\tincomplete?: true;\n\t/** Error explaining why an incomplete tool call could not be recovered. */\n\terrorMessage?: string;\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\n/**\n * Provider-emitted content block whose semantics are owned by the originating\n * provider. Surfaces server-side tool invocations (web_search), server-side\n * tool results (web_search_tool_result), code execution traces, grounding\n * metadata fragments, and any other vendor-specific block the core does not\n * normalize.\n *\n * `subtype` mirrors the provider's wire `type` discriminator (e.g.\n * \"server_tool_use\", \"web_search_tool_result\", \"executableCode\"). `raw` is\n * the verbatim wire payload — extensions interpret it per provider+subtype.\n *\n * The owning provider id lives on the parent `AssistantMessage.provider`, not\n * here, to avoid drift on replay.\n *\n * On replay across providers, `convertMessages` implementations MUST drop\n * these blocks (they are not portable). This mirrors the existing\n * `redacted_thinking` skip pattern.\n */\nexport interface ProviderNativeContent {\n\ttype: \"providerNative\";\n\tsubtype: string;\n\traw: unknown;\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 type AssistantStopDetails = { type: \"refusal\"; explanation?: string } | { type: \"sensitive\" };\n\nexport interface UserMessage {\n\trole: \"user\";\n\tcontent: string | (TextContent | ImageContent)[];\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport interface AssistantMessage {\n\trole: \"assistant\";\n\tcontent: (TextContent | ThinkingContent | ToolCall | ProviderNativeContent)[];\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\tdiagnostics?: AssistantMessageDiagnostic[]; // Redacted provider/runtime diagnostics for failures and recoveries.\n\tusage: Usage;\n\tstopReason: StopReason;\n\tstopDetails?: AssistantStopDetails;\n\tdeferred?: DeferredHandle;\n\terrorMessage?: string;\n\t/** Explicit owner for an abort initiated by the provider retry watchdog. */\n\tabortSource?: \"provider\";\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\nexport interface FreeformToolFormat {\n\ttype: \"grammar\";\n\tsyntax: \"lark\";\n\tdefinition: string;\n}\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  };\nexport interface Tool<TParameters extends TSchema = TSchema> {\n\tname: string;\n\tdescription: string;\n\tparameters: TParameters;\n\tfreeform?: FreeformToolFormat;\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 * Streams should emit `start` before partial updates, then terminate with either:\n * - `done` carrying the final successful AssistantMessage, or\n * - `error` carrying the final AssistantMessage with stopReason \"error\" or \"aborted\"\n *   and errorMessage.\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/** Finalized - executable iff `incomplete !== true` on `toolCall`. */\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/** 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/** Whether the provider accepts explicit disabled-thinking markers when thinking is off. Default: true. */\n\tsupportsDisabledThinking?: boolean;\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/**\n\t * Provider-specific JSON Schema flavor for tool parameters. `\"moonshot-mfjs\"`\n\t * normalizes schemas for Moonshot / Kimi backends that enforce a stricter subset.\n\t */\n\ttoolSchemaFlavor?: \"moonshot-mfjs\";\n\t/**\n\t * Tool call format for models that don't natively support tool calling.\n\t * When set, the middleware will intercept tool calls and format them as text.\n\t * Supported values: \"hermes\", \"morph-xml\", \"xml\" (deprecated alias for \"morph-xml\"), \"yaml-xml\", \"gemma4-delimiter\", \"anthropic-xml\", \"antml\", \"kimi-xtml\"\n\t */\n\ttoolCallFormat?: string;\n\t/** Cache control convention for prompt caching. \"anthropic\" applies Anthropic-style `cache_control` markers to the system prompt, last tool definition, and last user, assistant, or tool-result text content. */\n\tcacheControlFormat?: \"anthropic\";\n\t/** Whether to send session-affinity data from `options.sessionId`. Default: false. */\n\tsendSessionAffinityHeaders?: boolean;\n\t/** Provider-specific deferred tool serialization mode. */\n\tdeferredToolsMode?: \"kimi\";\n\t/** Session-affinity header format: `openai` sends `session_id`, `x-client-request-id`, and `x-session-affinity`; `openai-nosession` sends `x-client-request-id` and `x-session-affinity`; `openrouter` sends `x-session-id`. Does not affect the `prompt_cache_key` body param, which is governed by cache retention. Default: auto-detected. */\n\tsessionAffinityFormat?: SessionAffinityFormat;\n\t/** Whether the provider accepts the `prompt_cache_key` request field. Default: auto-detected. */\n\tsupportsPromptCacheKey?: boolean;\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}\n\n/** Compatibility settings for Anthropic Messages-compatible APIs. */\nexport interface AnthropicMessagesCompat {\n\t/**\n\t * Whether the provider accepts per-tool `eager_input_streaming`.\n\t * When false, the Anthropic provider omits `tools[].eager_input_streaming`\n\t * and sends the legacy `fine-grained-tool-streaming-2025-05-14` beta header\n\t * for tool-enabled requests.\n\t * Default: true.\n\t */\n\tsupportsEagerToolInputStreaming?: boolean;\n\t/** Whether the provider supports Anthropic long cache retention (`cache_control.ttl: \"1h\"`). Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n\t/**\n\t * Whether to send the `x-session-affinity` header from `options.sessionId`\n\t * when caching is enabled. Required for providers like Fireworks that use\n\t * session affinity for prompt cache routing (requests to the same replica\n\t * maximize cache hits).\n\t * Default: false.\n\t */\n\tsendSessionAffinityHeaders?: boolean;\n\t/**\n\t * Whether the provider supports Anthropic-style `cache_control` markers on\n\t * tool definitions. When false, `cache_control` is omitted from tool params.\n\t * Some Anthropic-compatible providers (e.g., Fireworks) do not support this\n\t * field on tools and may reject or ignore it.\n\t * Default: true.\n\t */\n\tsupportsCacheControlOnTools?: boolean;\n\t/**\n\t * Whether the provider accepts `thinking: { type: \"disabled\" }` when\n\t * extended thinking is off. Some Anthropic-compatible providers support\n\t * thinking but reject Anthropic's explicit disabled marker.\n\t * Default: true.\n\t */\n\tsupportsDisabledThinking?: 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/** Whether the model accepts Anthropic `tool_choice`. Default: true. */\n\tsupportsToolChoice?: boolean;\n\t/**\n\t * Whether the model accepts forced Anthropic tool choices (`any` or a named\n\t * tool). Default: true, except Claude Fable/Mythos models.\n\t */\n\tsupportsForcedToolChoice?: 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 opt out on\n\t * overridden built-in models and disable model id/name inference.\n\t * Default: infer from adaptive Claude model ids/names when possible.\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/**\n\t * How to replay thinking blocks that have no usable Anthropic signature.\n\t * `\"text\"` demotes them to text; `\"empty-signature\"` preserves Anthropic's\n\t * thinking shape with `signature: \"\"`. Defaults to `\"text\"`; the legacy\n\t * `allowEmptySignature` setting remains a compatibility alias for\n\t * `\"empty-signature\"`.\n\t */\n\tunsignedThinkingReplay?: \"text\" | \"empty-signature\";\n\t/**\n\t * Model ids Anthropic accepts in `fallbacks` for server-side refusal fallback.\n\t * When absent or empty, callers must omit `fallbacks`; Anthropic rejects the\n\t * field for models with no permitted fallback targets.\n\t */\n\tallowedFallbackModels?: Array<AnthropicAllowedFallbackModel | string>;\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 provider executes Anthropic server-side `web_search_*` native\n\t * tools and accepts their `server_tool_use` / `web_search_tool_result`\n\t * blocks on replay. Anthropic-compatible endpoints (e.g., kimi-coding) may\n\t * run the search but reject the replayed server-tool blocks on the next\n\t * request. When false, hook-injected `web_search_*` tools are stripped from\n\t * the payload. Default: true only for the first-party Anthropic endpoint.\n\t */\n\tsupportsWebSearch?: 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}\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\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/node_modules/@earendil-works/pi-ai/dist/utils/overflow.d.ts

@@ -76,6 +76,22 @@ export declare function isCursorPayloadResourceExhausted(message: {
         totalTokens?: number;
     };
 }, _estimateTokens: number): boolean;
+/**
+ * Detects Cursor's verified usage-pool exhaustion signature: a token-bearing
+ * `resource_exhausted` error while the conversation is well below the model
+ * context window.
+ */
+export declare function isCursorQuotaResourceExhausted(message: {
+    stopReason?: string;
+    errorMessage?: string;
+    usage?: {
+        input?: number;
+        output?: number;
+        cacheRead?: number;
+        cacheWrite?: number;
+        totalTokens?: number;
+    };
+}, contextWindow: number): boolean;
 export declare function isCursorZeroTokenResourceExhausted(message: {
     stopReason?: string;
     errorMessage?: string;

package/node_modules/@earendil-works/pi-ai/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;AA4FpD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAqDG;AACH,wBAAgB,iBAAiB,CAAC,OAAO,EAAE,gBAAgB,EAAE,aAAa,CAAC,EAAE,MAAM,GAAG,OAAO,CAmC5F;AACD;;;;;GAKG;AACH,wBAAgB,mBAAmB,CAAC,OAAO,EAAE,gBAAgB,EAAE,gBAAgB,EAAE,MAAM,GAAG,OAAO,CAEhG;AAED;;GAEG;AACH,wBAAgB,mBAAmB,IAAI,MAAM,EAAE,CAE9C;AAYD,wBAAgB,gCAAgC,CAC/C,OAAO,EAAE;IACR,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,KAAK,CAAC,EAAE;QAAE,KAAK,CAAC,EAAE,MAAM,CAAC;QAAC,MAAM,CAAC,EAAE,MAAM,CAAC;QAAC,SAAS,CAAC,EAAE,MAAM,CAAC;QAAC,UAAU,CAAC,EAAE,MAAM,CAAC;QAAC,WAAW,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC;CAC3G,EACD,eAAe,EAAE,MAAM,GACrB,OAAO,CAKT;AAED,wBAAgB,kCAAkC,CAAC,OAAO,EAAE;IAC3D,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,KAAK,CAAC,EAAE;QAAE,KAAK,CAAC,EAAE,MAAM,CAAC;QAAC,MAAM,CAAC,EAAE,MAAM,CAAC;QAAC,SAAS,CAAC,EAAE,MAAM,CAAC;QAAC,UAAU,CAAC,EAAE,MAAM,CAAC;QAAC,WAAW,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC;CAC3G,GAAG,OAAO,CAKV;AAED,wBAAgB,yCAAyC,CAAC,OAAO,EAAE;IAAE,eAAe,CAAC,EAAE,OAAO,CAAA;CAAE,GAAG,SAAS,GAAG,OAAO,CAErH;AAED,wBAAgB,iCAAiC,CAAC,SAAS,EAAE,OAAO,EAAE,YAAY,EAAE,MAAM,GAAG,OAAO,CAGnG;AAED,wBAAgB,gCAAgC,CAAC,CAAC,SAAS;IAAE,gBAAgB,CAAC,EAAE,MAAM,CAAC;IAAC,kBAAkB,CAAC,EAAE,OAAO,CAAA;CAAE,EACrH,QAAQ,EAAE,CAAC,EACX,QAAQ,EAAE,MAAM,GAAG,SAAS,EAC5B,MAAM,EAAE,MAAM,GAAG,SAAS,GACxB,CAAC,CAIH"}
+{"version":3,"file":"overflow.d.ts","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AA4FpD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAqDG;AACH,wBAAgB,iBAAiB,CAAC,OAAO,EAAE,gBAAgB,EAAE,aAAa,CAAC,EAAE,MAAM,GAAG,OAAO,CAwC5F;AACD;;;;;GAKG;AACH,wBAAgB,mBAAmB,CAAC,OAAO,EAAE,gBAAgB,EAAE,gBAAgB,EAAE,MAAM,GAAG,OAAO,CAEhG;AAED;;GAEG;AACH,wBAAgB,mBAAmB,IAAI,MAAM,EAAE,CAE9C;AAYD,wBAAgB,gCAAgC,CAC/C,OAAO,EAAE;IACR,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,KAAK,CAAC,EAAE;QAAE,KAAK,CAAC,EAAE,MAAM,CAAC;QAAC,MAAM,CAAC,EAAE,MAAM,CAAC;QAAC,SAAS,CAAC,EAAE,MAAM,CAAC;QAAC,UAAU,CAAC,EAAE,MAAM,CAAC;QAAC,WAAW,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC;CAC3G,EACD,eAAe,EAAE,MAAM,GACrB,OAAO,CAKT;AAED;;;;GAIG;AACH,wBAAgB,8BAA8B,CAC7C,OAAO,EAAE;IACR,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,KAAK,CAAC,EAAE;QAAE,KAAK,CAAC,EAAE,MAAM,CAAC;QAAC,MAAM,CAAC,EAAE,MAAM,CAAC;QAAC,SAAS,CAAC,EAAE,MAAM,CAAC;QAAC,UAAU,CAAC,EAAE,MAAM,CAAC;QAAC,WAAW,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC;CAC3G,EACD,aAAa,EAAE,MAAM,GACnB,OAAO,CAST;AAED,wBAAgB,kCAAkC,CAAC,OAAO,EAAE;IAC3D,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,KAAK,CAAC,EAAE;QAAE,KAAK,CAAC,EAAE,MAAM,CAAC;QAAC,MAAM,CAAC,EAAE,MAAM,CAAC;QAAC,SAAS,CAAC,EAAE,MAAM,CAAC;QAAC,UAAU,CAAC,EAAE,MAAM,CAAC;QAAC,WAAW,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC;CAC3G,GAAG,OAAO,CAKV;AAED,wBAAgB,yCAAyC,CAAC,OAAO,EAAE;IAAE,eAAe,CAAC,EAAE,OAAO,CAAA;CAAE,GAAG,SAAS,GAAG,OAAO,CAErH;AAED,wBAAgB,iCAAiC,CAAC,SAAS,EAAE,OAAO,EAAE,YAAY,EAAE,MAAM,GAAG,OAAO,CAGnG;AAED,wBAAgB,gCAAgC,CAAC,CAAC,SAAS;IAAE,gBAAgB,CAAC,EAAE,MAAM,CAAC;IAAC,kBAAkB,CAAC,EAAE,OAAO,CAAA;CAAE,EACrH,QAAQ,EAAE,CAAC,EACX,QAAQ,EAAE,MAAM,GAAG,SAAS,EAC5B,MAAM,EAAE,MAAM,GAAG,SAAS,GACxB,CAAC,CAIH"}

package/node_modules/@earendil-works/pi-ai/dist/utils/overflow.js

@@ -149,7 +149,10 @@ export function isContextOverflow(message, contextWindow) {
         }
         const usage = message.usage;
         const hasTokenEvidence = (usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite) > 0;
-        if (!isNonOverflow && hasTokenEvidence && RESOURCE_EXHAUSTED_PATTERN.test(message.errorMessage)) {
+        if (!isNonOverflow &&
+            hasTokenEvidence &&
+            RESOURCE_EXHAUSTED_PATTERN.test(message.errorMessage) &&
+            (contextWindow === undefined || contextWindow <= 0 || cursorZeroTokenCount(message) >= contextWindow * 0.5)) {
             return true;
         }
     }
@@ -198,6 +201,19 @@ export function isCursorPayloadResourceExhausted(message, _estimateTokens) {
     }
     return !(cursorZeroTokenCount(message) > 0);
 }
+/**
+ * Detects Cursor's verified usage-pool exhaustion signature: a token-bearing
+ * `resource_exhausted` error while the conversation is well below the model
+ * context window.
+ */
+export function isCursorQuotaResourceExhausted(message, contextWindow) {
+    const tokens = cursorZeroTokenCount(message);
+    return (message.stopReason === "error" &&
+        RESOURCE_EXHAUSTED_PATTERN.test(message.errorMessage || "") &&
+        contextWindow > 0 &&
+        tokens > 0 &&
+        tokens < contextWindow * 0.5);
+}
 export function isCursorZeroTokenResourceExhausted(message) {
     if (message.stopReason !== "error" || !/resource.?exhausted/i.test(message.errorMessage || "")) {
         return false;

package/node_modules/@earendil-works/pi-ai/dist/utils/overflow.js.map

@@ -1 +1 @@
-{"version":3,"file":"overflow.js","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkCG;AACH,MAAM,iBAAiB,GAAG;IACzB,qBAAqB,EAAE,2BAA2B;IAClD,oBAAoB,EAAE,kDAAkD;IACxE,wCAAwC,EAAE,iBAAiB;IAC3D,6BAA6B,EAAE,uCAAuC;IACtE,4FAA4F,EAAE,sCAAsC;IACpI,yCAAyC,EAAE,kBAAkB;IAC7D,+BAA+B,EAAE,aAAa;IAC9C,oCAAoC,EAAE,OAAO;IAC7C,uCAAuC,EAAE,6BAA6B;IACtE,kEAAkE,EAAE,sBAAsB;IAC1F,iFAAiF,EAAE,cAAc;IACjG,2BAA2B,EAAE,iBAAiB;IAC9C,qCAAqC,EAAE,mBAAmB;IAC1D,kCAAkC,EAAE,YAAY;IAChD,+BAA+B,EAAE,UAAU;IAC3C,6BAA6B,EAAE,kBAAkB;IACjD,sDAAsD,EAAE,UAAU;IAClE,+EAA+E,EAAE,aAAa;IAC9F,gCAAgC,EAAE,yDAAyD;IAC3F,oDAAoD,EAAE,iCAAiC;IACvF,kCAAkC,EAAE,8BAA8B;IAClE,gCAAgC,EAAE,mBAAmB;IACrD,kBAAkB,EAAE,mBAAmB;IACvC,uBAAuB,EAAE,mBAAmB;IAC5C,+CAA+C,EAAE,iCAAiC;IAClF,0DAA0D,EAAE,4UAA4U;CACxY,CAAC;AAEF;;;;;;;;GAQG;AACH,MAAM,qBAAqB,GAAG;IAC7B,2CAA2C,EAAE,oFAAoF;IACjI,aAAa,EAAE,wBAAwB;IACvC,oBAAoB,EAAE,yBAAyB;CAC/C,CAAC;AAEF;;;;;;;;GAQG;AACH,MAAM,0BAA0B,GAAG,sBAAsB,CAAC;AAE1D;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAqDG;AACH,MAAM,UAAU,iBAAiB,CAAC,OAAyB,EAAE,aAAsB;IAClF,uCAAuC;IACvC,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,OAAO,CAAC,YAAY,EAAE,CAAC;QAC5D,oFAAoF;QACpF,MAAM,aAAa,GAAG,qBAAqB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,CAAC;QACvF,IAAI,CAAC,aAAa,IAAI,iBAAiB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,EAAE,CAAC;YACpF,OAAO,IAAI,CAAC;QACb,CAAC;QACD,MAAM,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC;QAC5B,MAAM,gBAAgB,GACrB,CAAC,KAAK,CAAC,WAAW,IAAI,KAAK,CAAC,KAAK,GAAG,KAAK,CAAC,MAAM,GAAG,KAAK,CAAC,SAAS,GAAG,KAAK,CAAC,UAAU,CAAC,GAAG,CAAC,CAAC;QAC5F,IAAI,CAAC,aAAa,IAAI,gBAAgB,IAAI,0BAA0B,CAAC,IAAI,CAAC,OAAO,CAAC,YAAY,CAAC,EAAE,CAAC;YACjG,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,8EAA8E;IAC9E,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,MAAM,EAAE,CAAC;QACpD,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,GAAG,aAAa,EAAE,CAAC;YACjC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,sFAAsF;IACtF,qFAAqF;IACrF,gEAAgE;IAChE,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,QAAQ,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACpF,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,IAAI,aAAa,GAAG,IAAI,EAAE,CAAC;YACzC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,OAAO,KAAK,CAAC;AACd,CAAC;AACD;;;;;GAKG;AACH,MAAM,UAAU,mBAAmB,CAAC,OAAyB,EAAE,gBAAwB;IACtF,OAAO,OAAO,CAAC,UAAU,KAAK,QAAQ,IAAI,gBAAgB,GAAG,CAAC,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,GAAG,gBAAgB,CAAC;AAC3G,CAAC;AAED;;GAEG;AACH,MAAM,UAAU,mBAAmB;IAClC,OAAO,CAAC,GAAG,iBAAiB,CAAC,CAAC;AAC/B,CAAC;AAED,SAAS,oBAAoB,CAAC,OAE7B;IACA,MAAM,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC;IAC5B,IAAI,CAAC,KAAK;QAAE,OAAO,CAAC,CAAC;IACrB,OAAO,CACN,KAAK,CAAC,WAAW,IAAI,CAAC,KAAK,CAAC,KAAK,IAAI,CAAC,CAAC,GAAG,CAAC,KAAK,CAAC,MAAM,IAAI,CAAC,CAAC,GAAG,CAAC,KAAK,CAAC,SAAS,IAAI,CAAC,CAAC,GAAG,CAAC,KAAK,CAAC,UAAU,IAAI,CAAC,CAAC,CAChH,CAAC;AACH,CAAC;AAED,MAAM,UAAU,gCAAgC,CAC/C,OAIC,EACD,eAAuB;IAEvB,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,CAAC,sBAAsB,CAAC,IAAI,CAAC,OAAO,CAAC,YAAY,IAAI,EAAE,CAAC,EAAE,CAAC;QAChG,OAAO,KAAK,CAAC;IACd,CAAC;IACD,OAAO,CAAC,CAAC,oBAAoB,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC;AAC7C,CAAC;AAED,MAAM,UAAU,kCAAkC,CAAC,OAIlD;IACA,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,CAAC,sBAAsB,CAAC,IAAI,CAAC,OAAO,CAAC,YAAY,IAAI,EAAE,CAAC,EAAE,CAAC;QAChG,OAAO,KAAK,CAAC;IACd,CAAC;IACD,OAAO,CAAC,CAAC,oBAAoB,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC;AAC7C,CAAC;AAED,MAAM,UAAU,yCAAyC,CAAC,OAAkD;IAC3G,OAAO,OAAO,EAAE,eAAe,KAAK,IAAI,CAAC;AAC1C,CAAC;AAED,MAAM,UAAU,iCAAiC,CAAC,SAAkB,EAAE,YAAoB;IACzF,IAAI,SAAS;QAAE,OAAO,KAAK,CAAC;IAC5B,OAAO,qBAAqB,CAAC,IAAI,CAAC,YAAY,IAAI,EAAE,CAAC,CAAC;AACvD,CAAC;AAED,MAAM,UAAU,gCAAgC,CAC/C,QAAW,EACX,QAA4B,EAC5B,MAA0B;IAE1B,IAAI,MAAM,KAAK,UAAU;QAAE,OAAO,QAAQ,CAAC;IAC3C,IAAI,QAAQ,KAAK,QAAQ,IAAI,QAAQ,KAAK,kBAAkB;QAAE,OAAO,QAAQ,CAAC;IAC9E,OAAO,EAAE,GAAG,QAAQ,EAAE,gBAAgB,EAAE,CAAC,EAAE,kBAAkB,EAAE,KAAK,EAAE,CAAC;AACxE,CAAC","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 * - Gateways: \"413 Request body too large\" / \"Request Entity Too Large\" / \"Payload Too Large\" (byte-size overflow)\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\t/(?:request[ _])?(?:body|entity|payload)[_ ]too[_ ]large/i, // Gateway HTTP 413 byte-size rejections (\"Request body too large\", \"Request Entity Too Large\", \"body_too_large\", \"Payload Too Large\"). Substring-anchored by design (JSON bodies lack an adjacent status code); a non-context size rejection (e.g. an oversized image) can over-match, which costs one bounded shrink-retry, never a wedge.\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 * Cursor's api2.cursor.sh surfaces a context overflow as a bare gRPC\n * `resource_exhausted` end-stream — the same wording its backend uses for\n * quota and poisoned-conversation rejections. Token evidence disambiguates:\n * a request that already streamed or billed tokens overflowed mid-flight,\n * while a zero-token rejection is a quota/conversation failure that must stay\n * on the rate-limit path (the cursor client rotates the conversation id for\n * those and retry handling supplies the backoff).\n */\nconst RESOURCE_EXHAUSTED_PATTERN = /resource.?exhausted/i;\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\tconst usage = message.usage;\n\t\tconst hasTokenEvidence =\n\t\t\t(usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite) > 0;\n\t\tif (!isNonOverflow && hasTokenEvidence && RESOURCE_EXHAUSTED_PATTERN.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 * 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\nfunction cursorZeroTokenCount(message: {\n\tusage?: { input?: number; output?: number; cacheRead?: number; cacheWrite?: number; totalTokens?: number };\n}): number {\n\tconst usage = message.usage;\n\tif (!usage) return 0;\n\treturn (\n\t\tusage.totalTokens || (usage.input ?? 0) + (usage.output ?? 0) + (usage.cacheRead ?? 0) + (usage.cacheWrite ?? 0)\n\t);\n}\n\nexport function isCursorPayloadResourceExhausted(\n\tmessage: {\n\t\tstopReason?: string;\n\t\terrorMessage?: string;\n\t\tusage?: { input?: number; output?: number; cacheRead?: number; cacheWrite?: number; totalTokens?: number };\n\t},\n\t_estimateTokens: number,\n): boolean {\n\tif (message.stopReason !== \"error\" || !/resource.?exhausted/i.test(message.errorMessage || \"\")) {\n\t\treturn false;\n\t}\n\treturn !(cursorZeroTokenCount(message) > 0);\n}\n\nexport function isCursorZeroTokenResourceExhausted(message: {\n\tstopReason?: string;\n\terrorMessage?: string;\n\tusage?: { input?: number; output?: number; cacheRead?: number; cacheWrite?: number; totalTokens?: number };\n}): boolean {\n\tif (message.stopReason !== \"error\" || !/resource.?exhausted/i.test(message.errorMessage || \"\")) {\n\t\treturn false;\n\t}\n\treturn !(cursorZeroTokenCount(message) > 0);\n}\n\nexport function shouldSkipProviderFallbackForCursorZeroRe(options: { sameModelRemint?: boolean } | undefined): boolean {\n\treturn options?.sameModelRemint === true;\n}\n\nexport function shouldRetryOverflowWithoutCompact(compacted: boolean, errorMessage: string): boolean {\n\tif (compacted) return false;\n\treturn /Nothing to compact/i.test(errorMessage || \"\");\n}\n\nexport function cursorOverflowCompactionSettings<T extends { keepRecentTokens?: number; restorationEnabled?: boolean }>(\n\tsettings: T,\n\tprovider: string | undefined,\n\treason: string | undefined,\n): T {\n\tif (reason !== \"overflow\") return settings;\n\tif (provider !== \"cursor\" && provider !== \"cursor-cli-oauth\") return settings;\n\treturn { ...settings, keepRecentTokens: 0, restorationEnabled: false };\n}\n"]}
+{"version":3,"file":"overflow.js","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkCG;AACH,MAAM,iBAAiB,GAAG;IACzB,qBAAqB,EAAE,2BAA2B;IAClD,oBAAoB,EAAE,kDAAkD;IACxE,wCAAwC,EAAE,iBAAiB;IAC3D,6BAA6B,EAAE,uCAAuC;IACtE,4FAA4F,EAAE,sCAAsC;IACpI,yCAAyC,EAAE,kBAAkB;IAC7D,+BAA+B,EAAE,aAAa;IAC9C,oCAAoC,EAAE,OAAO;IAC7C,uCAAuC,EAAE,6BAA6B;IACtE,kEAAkE,EAAE,sBAAsB;IAC1F,iFAAiF,EAAE,cAAc;IACjG,2BAA2B,EAAE,iBAAiB;IAC9C,qCAAqC,EAAE,mBAAmB;IAC1D,kCAAkC,EAAE,YAAY;IAChD,+BAA+B,EAAE,UAAU;IAC3C,6BAA6B,EAAE,kBAAkB;IACjD,sDAAsD,EAAE,UAAU;IAClE,+EAA+E,EAAE,aAAa;IAC9F,gCAAgC,EAAE,yDAAyD;IAC3F,oDAAoD,EAAE,iCAAiC;IACvF,kCAAkC,EAAE,8BAA8B;IAClE,gCAAgC,EAAE,mBAAmB;IACrD,kBAAkB,EAAE,mBAAmB;IACvC,uBAAuB,EAAE,mBAAmB;IAC5C,+CAA+C,EAAE,iCAAiC;IAClF,0DAA0D,EAAE,4UAA4U;CACxY,CAAC;AAEF;;;;;;;;GAQG;AACH,MAAM,qBAAqB,GAAG;IAC7B,2CAA2C,EAAE,oFAAoF;IACjI,aAAa,EAAE,wBAAwB;IACvC,oBAAoB,EAAE,yBAAyB;CAC/C,CAAC;AAEF;;;;;;;;GAQG;AACH,MAAM,0BAA0B,GAAG,sBAAsB,CAAC;AAE1D;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAqDG;AACH,MAAM,UAAU,iBAAiB,CAAC,OAAyB,EAAE,aAAsB;IAClF,uCAAuC;IACvC,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,OAAO,CAAC,YAAY,EAAE,CAAC;QAC5D,oFAAoF;QACpF,MAAM,aAAa,GAAG,qBAAqB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,CAAC;QACvF,IAAI,CAAC,aAAa,IAAI,iBAAiB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,EAAE,CAAC;YACpF,OAAO,IAAI,CAAC;QACb,CAAC;QACD,MAAM,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC;QAC5B,MAAM,gBAAgB,GACrB,CAAC,KAAK,CAAC,WAAW,IAAI,KAAK,CAAC,KAAK,GAAG,KAAK,CAAC,MAAM,GAAG,KAAK,CAAC,SAAS,GAAG,KAAK,CAAC,UAAU,CAAC,GAAG,CAAC,CAAC;QAC5F,IACC,CAAC,aAAa;YACd,gBAAgB;YAChB,0BAA0B,CAAC,IAAI,CAAC,OAAO,CAAC,YAAY,CAAC;YACrD,CAAC,aAAa,KAAK,SAAS,IAAI,aAAa,IAAI,CAAC,IAAI,oBAAoB,CAAC,OAAO,CAAC,IAAI,aAAa,GAAG,GAAG,CAAC,EAC1G,CAAC;YACF,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,8EAA8E;IAC9E,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,MAAM,EAAE,CAAC;QACpD,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,GAAG,aAAa,EAAE,CAAC;YACjC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,sFAAsF;IACtF,qFAAqF;IACrF,gEAAgE;IAChE,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,QAAQ,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACpF,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,IAAI,aAAa,GAAG,IAAI,EAAE,CAAC;YACzC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,OAAO,KAAK,CAAC;AACd,CAAC;AACD;;;;;GAKG;AACH,MAAM,UAAU,mBAAmB,CAAC,OAAyB,EAAE,gBAAwB;IACtF,OAAO,OAAO,CAAC,UAAU,KAAK,QAAQ,IAAI,gBAAgB,GAAG,CAAC,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,GAAG,gBAAgB,CAAC;AAC3G,CAAC;AAED;;GAEG;AACH,MAAM,UAAU,mBAAmB;IAClC,OAAO,CAAC,GAAG,iBAAiB,CAAC,CAAC;AAC/B,CAAC;AAED,SAAS,oBAAoB,CAAC,OAE7B;IACA,MAAM,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC;IAC5B,IAAI,CAAC,KAAK;QAAE,OAAO,CAAC,CAAC;IACrB,OAAO,CACN,KAAK,CAAC,WAAW,IAAI,CAAC,KAAK,CAAC,KAAK,IAAI,CAAC,CAAC,GAAG,CAAC,KAAK,CAAC,MAAM,IAAI,CAAC,CAAC,GAAG,CAAC,KAAK,CAAC,SAAS,IAAI,CAAC,CAAC,GAAG,CAAC,KAAK,CAAC,UAAU,IAAI,CAAC,CAAC,CAChH,CAAC;AACH,CAAC;AAED,MAAM,UAAU,gCAAgC,CAC/C,OAIC,EACD,eAAuB;IAEvB,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,CAAC,sBAAsB,CAAC,IAAI,CAAC,OAAO,CAAC,YAAY,IAAI,EAAE,CAAC,EAAE,CAAC;QAChG,OAAO,KAAK,CAAC;IACd,CAAC;IACD,OAAO,CAAC,CAAC,oBAAoB,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC;AAC7C,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,8BAA8B,CAC7C,OAIC,EACD,aAAqB;IAErB,MAAM,MAAM,GAAG,oBAAoB,CAAC,OAAO,CAAC,CAAC;IAC7C,OAAO,CACN,OAAO,CAAC,UAAU,KAAK,OAAO;QAC9B,0BAA0B,CAAC,IAAI,CAAC,OAAO,CAAC,YAAY,IAAI,EAAE,CAAC;QAC3D,aAAa,GAAG,CAAC;QACjB,MAAM,GAAG,CAAC;QACV,MAAM,GAAG,aAAa,GAAG,GAAG,CAC5B,CAAC;AACH,CAAC;AAED,MAAM,UAAU,kCAAkC,CAAC,OAIlD;IACA,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,CAAC,sBAAsB,CAAC,IAAI,CAAC,OAAO,CAAC,YAAY,IAAI,EAAE,CAAC,EAAE,CAAC;QAChG,OAAO,KAAK,CAAC;IACd,CAAC;IACD,OAAO,CAAC,CAAC,oBAAoB,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC;AAC7C,CAAC;AAED,MAAM,UAAU,yCAAyC,CAAC,OAAkD;IAC3G,OAAO,OAAO,EAAE,eAAe,KAAK,IAAI,CAAC;AAC1C,CAAC;AAED,MAAM,UAAU,iCAAiC,CAAC,SAAkB,EAAE,YAAoB;IACzF,IAAI,SAAS;QAAE,OAAO,KAAK,CAAC;IAC5B,OAAO,qBAAqB,CAAC,IAAI,CAAC,YAAY,IAAI,EAAE,CAAC,CAAC;AACvD,CAAC;AAED,MAAM,UAAU,gCAAgC,CAC/C,QAAW,EACX,QAA4B,EAC5B,MAA0B;IAE1B,IAAI,MAAM,KAAK,UAAU;QAAE,OAAO,QAAQ,CAAC;IAC3C,IAAI,QAAQ,KAAK,QAAQ,IAAI,QAAQ,KAAK,kBAAkB;QAAE,OAAO,QAAQ,CAAC;IAC9E,OAAO,EAAE,GAAG,QAAQ,EAAE,gBAAgB,EAAE,CAAC,EAAE,kBAAkB,EAAE,KAAK,EAAE,CAAC;AACxE,CAAC","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 * - Gateways: \"413 Request body too large\" / \"Request Entity Too Large\" / \"Payload Too Large\" (byte-size overflow)\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\t/(?:request[ _])?(?:body|entity|payload)[_ ]too[_ ]large/i, // Gateway HTTP 413 byte-size rejections (\"Request body too large\", \"Request Entity Too Large\", \"body_too_large\", \"Payload Too Large\"). Substring-anchored by design (JSON bodies lack an adjacent status code); a non-context size rejection (e.g. an oversized image) can over-match, which costs one bounded shrink-retry, never a wedge.\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 * Cursor's api2.cursor.sh surfaces a context overflow as a bare gRPC\n * `resource_exhausted` end-stream — the same wording its backend uses for\n * quota and poisoned-conversation rejections. Token evidence disambiguates:\n * a request that already streamed or billed tokens overflowed mid-flight,\n * while a zero-token rejection is a quota/conversation failure that must stay\n * on the rate-limit path (the cursor client rotates the conversation id for\n * those and retry handling supplies the backoff).\n */\nconst RESOURCE_EXHAUSTED_PATTERN = /resource.?exhausted/i;\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\tconst usage = message.usage;\n\t\tconst hasTokenEvidence =\n\t\t\t(usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite) > 0;\n\t\tif (\n\t\t\t!isNonOverflow &&\n\t\t\thasTokenEvidence &&\n\t\t\tRESOURCE_EXHAUSTED_PATTERN.test(message.errorMessage) &&\n\t\t\t(contextWindow === undefined || contextWindow <= 0 || cursorZeroTokenCount(message) >= contextWindow * 0.5)\n\t\t) {\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 * 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\nfunction cursorZeroTokenCount(message: {\n\tusage?: { input?: number; output?: number; cacheRead?: number; cacheWrite?: number; totalTokens?: number };\n}): number {\n\tconst usage = message.usage;\n\tif (!usage) return 0;\n\treturn (\n\t\tusage.totalTokens || (usage.input ?? 0) + (usage.output ?? 0) + (usage.cacheRead ?? 0) + (usage.cacheWrite ?? 0)\n\t);\n}\n\nexport function isCursorPayloadResourceExhausted(\n\tmessage: {\n\t\tstopReason?: string;\n\t\terrorMessage?: string;\n\t\tusage?: { input?: number; output?: number; cacheRead?: number; cacheWrite?: number; totalTokens?: number };\n\t},\n\t_estimateTokens: number,\n): boolean {\n\tif (message.stopReason !== \"error\" || !/resource.?exhausted/i.test(message.errorMessage || \"\")) {\n\t\treturn false;\n\t}\n\treturn !(cursorZeroTokenCount(message) > 0);\n}\n\n/**\n * Detects Cursor's verified usage-pool exhaustion signature: a token-bearing\n * `resource_exhausted` error while the conversation is well below the model\n * context window.\n */\nexport function isCursorQuotaResourceExhausted(\n\tmessage: {\n\t\tstopReason?: string;\n\t\terrorMessage?: string;\n\t\tusage?: { input?: number; output?: number; cacheRead?: number; cacheWrite?: number; totalTokens?: number };\n\t},\n\tcontextWindow: number,\n): boolean {\n\tconst tokens = cursorZeroTokenCount(message);\n\treturn (\n\t\tmessage.stopReason === \"error\" &&\n\t\tRESOURCE_EXHAUSTED_PATTERN.test(message.errorMessage || \"\") &&\n\t\tcontextWindow > 0 &&\n\t\ttokens > 0 &&\n\t\ttokens < contextWindow * 0.5\n\t);\n}\n\nexport function isCursorZeroTokenResourceExhausted(message: {\n\tstopReason?: string;\n\terrorMessage?: string;\n\tusage?: { input?: number; output?: number; cacheRead?: number; cacheWrite?: number; totalTokens?: number };\n}): boolean {\n\tif (message.stopReason !== \"error\" || !/resource.?exhausted/i.test(message.errorMessage || \"\")) {\n\t\treturn false;\n\t}\n\treturn !(cursorZeroTokenCount(message) > 0);\n}\n\nexport function shouldSkipProviderFallbackForCursorZeroRe(options: { sameModelRemint?: boolean } | undefined): boolean {\n\treturn options?.sameModelRemint === true;\n}\n\nexport function shouldRetryOverflowWithoutCompact(compacted: boolean, errorMessage: string): boolean {\n\tif (compacted) return false;\n\treturn /Nothing to compact/i.test(errorMessage || \"\");\n}\n\nexport function cursorOverflowCompactionSettings<T extends { keepRecentTokens?: number; restorationEnabled?: boolean }>(\n\tsettings: T,\n\tprovider: string | undefined,\n\treason: string | undefined,\n): T {\n\tif (reason !== \"overflow\") return settings;\n\tif (provider !== \"cursor\" && provider !== \"cursor-cli-oauth\") return settings;\n\treturn { ...settings, keepRecentTokens: 0, restorationEnabled: false };\n}\n"]}

package/node_modules/@earendil-works/pi-ai/dist/utils/retry-profile/backoff.d.ts

@@ -0,0 +1,14 @@
+import type { RetryBackoffPolicy } from "./types.ts";
+/**
+ * Pure per-attempt retry delay for a stage's backoff policy.
+ *
+ * The exponential base is `baseDelayMs * growthFactor ** (retryNumber - 1)`
+ * with `retryNumber` 1-based. The per-attempt cap is applied to that base
+ * BEFORE jitter (`null` = no cap, `0` = literal zero cap), then the jitter
+ * mode scales the capped value by the injected `random` sample in `[0, 1)`:
+ * `none` returns the capped delay, `additive` grows it by up to `ratio`,
+ * `subtractive` shrinks it by up to `ratio`. No `Math.random` here — callers
+ * inject the sample so the schedule stays deterministic and testable.
+ */
+export declare function retryBackoffDelayMs(policy: RetryBackoffPolicy, retryNumber: number, random: number): number;
+//# sourceMappingURL=backoff.d.ts.map

package/node_modules/@earendil-works/pi-ai/dist/utils/retry-profile/backoff.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"backoff.d.ts","sourceRoot":"","sources":["../../../src/utils/retry-profile/backoff.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,kBAAkB,EAAE,MAAM,YAAY,CAAC;AAErD;;;;;;;;;;GAUG;AACH,wBAAgB,mBAAmB,CAAC,MAAM,EAAE,kBAAkB,EAAE,WAAW,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,GAAG,MAAM,CAW3G"}

package/node_modules/@earendil-works/pi-ai/dist/utils/retry-profile/backoff.js

@@ -0,0 +1,24 @@
+/**
+ * Pure per-attempt retry delay for a stage's backoff policy.
+ *
+ * The exponential base is `baseDelayMs * growthFactor ** (retryNumber - 1)`
+ * with `retryNumber` 1-based. The per-attempt cap is applied to that base
+ * BEFORE jitter (`null` = no cap, `0` = literal zero cap), then the jitter
+ * mode scales the capped value by the injected `random` sample in `[0, 1)`:
+ * `none` returns the capped delay, `additive` grows it by up to `ratio`,
+ * `subtractive` shrinks it by up to `ratio`. No `Math.random` here — callers
+ * inject the sample so the schedule stays deterministic and testable.
+ */
+export function retryBackoffDelayMs(policy, retryNumber, random) {
+    const exponential = policy.baseDelayMs * policy.growthFactor ** (retryNumber - 1);
+    const capped = policy.perAttemptCapMs === null ? exponential : Math.min(exponential, policy.perAttemptCapMs);
+    switch (policy.jitter.mode) {
+        case "none":
+            return capped;
+        case "additive":
+            return capped + random * policy.jitter.ratio * capped;
+        case "subtractive":
+            return capped - random * policy.jitter.ratio * capped;
+    }
+}
+//# sourceMappingURL=backoff.js.map

package/node_modules/@earendil-works/pi-ai/dist/utils/retry-profile/backoff.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"backoff.js","sourceRoot":"","sources":["../../../src/utils/retry-profile/backoff.ts"],"names":[],"mappings":"AAEA;;;;;;;;;;GAUG;AACH,MAAM,UAAU,mBAAmB,CAAC,MAA0B,EAAE,WAAmB,EAAE,MAAc;IAClG,MAAM,WAAW,GAAG,MAAM,CAAC,WAAW,GAAG,MAAM,CAAC,YAAY,IAAI,CAAC,WAAW,GAAG,CAAC,CAAC,CAAC;IAClF,MAAM,MAAM,GAAG,MAAM,CAAC,eAAe,KAAK,IAAI,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,WAAW,EAAE,MAAM,CAAC,eAAe,CAAC,CAAC;IAC7G,QAAQ,MAAM,CAAC,MAAM,CAAC,IAAI,EAAE,CAAC;QAC5B,KAAK,MAAM;YACV,OAAO,MAAM,CAAC;QACf,KAAK,UAAU;YACd,OAAO,MAAM,GAAG,MAAM,GAAG,MAAM,CAAC,MAAM,CAAC,KAAK,GAAG,MAAM,CAAC;QACvD,KAAK,aAAa;YACjB,OAAO,MAAM,GAAG,MAAM,GAAG,MAAM,CAAC,MAAM,CAAC,KAAK,GAAG,MAAM,CAAC;IACxD,CAAC;AACF,CAAC","sourcesContent":["import type { RetryBackoffPolicy } from \"./types.ts\";\n\n/**\n * Pure per-attempt retry delay for a stage's backoff policy.\n *\n * The exponential base is `baseDelayMs * growthFactor ** (retryNumber - 1)`\n * with `retryNumber` 1-based. The per-attempt cap is applied to that base\n * BEFORE jitter (`null` = no cap, `0` = literal zero cap), then the jitter\n * mode scales the capped value by the injected `random` sample in `[0, 1)`:\n * `none` returns the capped delay, `additive` grows it by up to `ratio`,\n * `subtractive` shrinks it by up to `ratio`. No `Math.random` here — callers\n * inject the sample so the schedule stays deterministic and testable.\n */\nexport function retryBackoffDelayMs(policy: RetryBackoffPolicy, retryNumber: number, random: number): number {\n\tconst exponential = policy.baseDelayMs * policy.growthFactor ** (retryNumber - 1);\n\tconst capped = policy.perAttemptCapMs === null ? exponential : Math.min(exponential, policy.perAttemptCapMs);\n\tswitch (policy.jitter.mode) {\n\t\tcase \"none\":\n\t\t\treturn capped;\n\t\tcase \"additive\":\n\t\t\treturn capped + random * policy.jitter.ratio * capped;\n\t\tcase \"subtractive\":\n\t\t\treturn capped - random * policy.jitter.ratio * capped;\n\t}\n}\n"]}

package/node_modules/@earendil-works/pi-ai/dist/utils/retry-profile/classifiers.d.ts

@@ -0,0 +1,36 @@
+import type { RetryClassifier } from "./types.ts";
+/**
+ * Kimi provider-request classifier, ported from the source provider's retry
+ * policy (policy only, not its code). The switch is exhaustive over
+ * {@link RetryFailure}["kind"]:
+ *
+ * - abort / refusal / sensitive / quota-exhausted / image-format / unknown are
+ *   terminal: user cancellations, content refusals, dead accounts, rejected
+ *   payloads, and unrecognized failures never recover by re-sending.
+ * - connection / timeout / provider failures are transient transport noise.
+ * - empty-response is transient unless the model said `finishReason` was
+ *   "filtered", which is a deterministic content rejection.
+ * - http-status is retryable only for the whitelisted codes above.
+ */
+export declare const classifyKimiFailure: RetryClassifier;
+/**
+ * Senpi assistant-turn classifier: the existing message regexes decide first;
+ * structured failure facts are consulted only where the regexes are silent.
+ *
+ * Precedence (terminal wins at every tie):
+ * 1. Deterministic failure kinds (abort/refusal/sensitive/quota/image-format)
+ *    are terminal before any text inspection.
+ * 2. A non-retryable regex match is terminal — request-shape and quota wording
+ *    outrank any structured status (a 500 carrying invalid-tool-schema text
+ *    stays terminal, unlike under {@link classifyKimiFailure}).
+ * 3. Structured terminal facts (`shouldRetry: false`, quota provider codes).
+ * 4. A retryable regex match is transient — e.g. a 400 carrying Anthropic's
+ *    server-tool pairing text stays retryable because the request builder
+ *    repairs the replayed history.
+ * 5. Only when the regexes have no opinion does the structured status decide:
+ *    whitelisted transient codes retry, everything else is terminal.
+ * 6. No regex match and no structured facts: terminal, exactly the legacy
+ *    boolean fallback.
+ */
+export declare const classifySenpiAssistantFailure: RetryClassifier;
+//# sourceMappingURL=classifiers.d.ts.map

package/node_modules/@earendil-works/pi-ai/dist/utils/retry-profile/classifiers.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"classifiers.d.ts","sourceRoot":"","sources":["../../../src/utils/retry-profile/classifiers.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAuB,eAAe,EAAgB,MAAM,YAAY,CAAC;AAqBrF;;;;;;;;;;;;GAYG;AACH,eAAO,MAAM,mBAAmB,EAAE,eAoBjC,CAAC;AAkBF;;;;;;;;;;;;;;;;;;GAkBG;AACH,eAAO,MAAM,6BAA6B,EAAE,eAwB3C,CAAC"}

package/node_modules/@earendil-works/pi-ai/dist/utils/retry-profile/classifiers.js

@@ -0,0 +1,109 @@
+import { classifyErrorMessage } from "../retry.js";
+/**
+ * Status codes Kimi's provider-request stage treats as retryable. 429 is its
+ * own `rate-limited` verdict (the turn budget treats it through hint tiers);
+ * the other whitelisted codes are ordinary `transient`. Every other status —
+ * client errors and non-whitelisted 5xx alike — is `terminal`.
+ */
+const KIMI_RETRYABLE_STATUS_CODES = new Set([408, 409, 429, 500, 502, 503, 504, 529]);
+function classifyKimiHttpStatus(statusCode) {
+    if (statusCode === undefined || !KIMI_RETRYABLE_STATUS_CODES.has(statusCode)) {
+        return { verdict: "terminal" };
+    }
+    return statusCode === 429 ? { verdict: "rate-limited" } : { verdict: "transient" };
+}
+function assertNever(kind) {
+    throw new Error(`Unhandled retry failure kind: ${kind}`);
+}
+/**
+ * Kimi provider-request classifier, ported from the source provider's retry
+ * policy (policy only, not its code). The switch is exhaustive over
+ * {@link RetryFailure}["kind"]:
+ *
+ * - abort / refusal / sensitive / quota-exhausted / image-format / unknown are
+ *   terminal: user cancellations, content refusals, dead accounts, rejected
+ *   payloads, and unrecognized failures never recover by re-sending.
+ * - connection / timeout / provider failures are transient transport noise.
+ * - empty-response is transient unless the model said `finishReason` was
+ *   "filtered", which is a deterministic content rejection.
+ * - http-status is retryable only for the whitelisted codes above.
+ */
+export const classifyKimiFailure = (failure) => {
+    switch (failure.kind) {
+        case "abort":
+        case "refusal":
+        case "sensitive":
+        case "quota-exhausted":
+        case "image-format":
+        case "unknown":
+            return { verdict: "terminal" };
+        case "connection":
+        case "timeout":
+        case "provider":
+            return { verdict: "transient" };
+        case "empty-response":
+            return failure.finishReason === "filtered" ? { verdict: "terminal" } : { verdict: "transient" };
+        case "http-status":
+            return classifyKimiHttpStatus(failure.statusCode);
+        default:
+            return assertNever(failure.kind);
+    }
+};
+/**
+ * Structured status codes the senpi default classifier treats as retryable
+ * when — and only when — the message regexes have no opinion. Mirrors the
+ * transient half of Kimi's table plus the Cloudflare gateway codes senpi
+ * already retries by regex.
+ */
+const SENPI_STRUCTURED_RETRYABLE_STATUS_CODES = new Set([
+    408, 409, 429, 500, 502, 503, 504, 522, 524, 529,
+]);
+/** Provider error codes that prove a dead account regardless of message text. */
+const SENPI_STRUCTURED_TERMINAL_PROVIDER_CODES = new Set([
+    "insufficient_quota",
+    "credits_required",
+]);
+/**
+ * Senpi assistant-turn classifier: the existing message regexes decide first;
+ * structured failure facts are consulted only where the regexes are silent.
+ *
+ * Precedence (terminal wins at every tie):
+ * 1. Deterministic failure kinds (abort/refusal/sensitive/quota/image-format)
+ *    are terminal before any text inspection.
+ * 2. A non-retryable regex match is terminal — request-shape and quota wording
+ *    outrank any structured status (a 500 carrying invalid-tool-schema text
+ *    stays terminal, unlike under {@link classifyKimiFailure}).
+ * 3. Structured terminal facts (`shouldRetry: false`, quota provider codes).
+ * 4. A retryable regex match is transient — e.g. a 400 carrying Anthropic's
+ *    server-tool pairing text stays retryable because the request builder
+ *    repairs the replayed history.
+ * 5. Only when the regexes have no opinion does the structured status decide:
+ *    whitelisted transient codes retry, everything else is terminal.
+ * 6. No regex match and no structured facts: terminal, exactly the legacy
+ *    boolean fallback.
+ */
+export const classifySenpiAssistantFailure = (failure) => {
+    switch (failure.kind) {
+        case "abort":
+        case "refusal":
+        case "sensitive":
+        case "quota-exhausted":
+        case "image-format":
+            return { verdict: "terminal" };
+    }
+    const regexVerdict = classifyErrorMessage(failure.message);
+    if (regexVerdict === "non-retryable")
+        return { verdict: "terminal" };
+    if (failure.shouldRetry === false)
+        return { verdict: "terminal" };
+    if (failure.providerCodes?.some((code) => SENPI_STRUCTURED_TERMINAL_PROVIDER_CODES.has(code))) {
+        return { verdict: "terminal" };
+    }
+    if (regexVerdict === "retryable")
+        return { verdict: "transient" };
+    if (failure.statusCode !== undefined && SENPI_STRUCTURED_RETRYABLE_STATUS_CODES.has(failure.statusCode)) {
+        return failure.statusCode === 429 ? { verdict: "rate-limited" } : { verdict: "transient" };
+    }
+    return { verdict: "terminal" };
+};
+//# sourceMappingURL=classifiers.js.map

package/node_modules/@earendil-works/pi-ai/dist/utils/retry-profile/classifiers.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"classifiers.js","sourceRoot":"","sources":["../../../src/utils/retry-profile/classifiers.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,oBAAoB,EAAE,MAAM,aAAa,CAAC;AAGnD;;;;;GAKG;AACH,MAAM,2BAA2B,GAAwB,IAAI,GAAG,CAAC,CAAC,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,CAAC,CAAC,CAAC;AAE3G,SAAS,sBAAsB,CAAC,UAA8B;IAC7D,IAAI,UAAU,KAAK,SAAS,IAAI,CAAC,2BAA2B,CAAC,GAAG,CAAC,UAAU,CAAC,EAAE,CAAC;QAC9E,OAAO,EAAE,OAAO,EAAE,UAAU,EAAE,CAAC;IAChC,CAAC;IACD,OAAO,UAAU,KAAK,GAAG,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,cAAc,EAAE,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,WAAW,EAAE,CAAC;AACpF,CAAC;AAED,SAAS,WAAW,CAAC,IAAW;IAC/B,MAAM,IAAI,KAAK,CAAC,iCAAiC,IAAI,EAAE,CAAC,CAAC;AAC1D,CAAC;AAED;;;;;;;;;;;;GAYG;AACH,MAAM,CAAC,MAAM,mBAAmB,GAAoB,CAAC,OAAO,EAAE,EAAE;IAC/D,QAAQ,OAAO,CAAC,IAAI,EAAE,CAAC;QACtB,KAAK,OAAO,CAAC;QACb,KAAK,SAAS,CAAC;QACf,KAAK,WAAW,CAAC;QACjB,KAAK,iBAAiB,CAAC;QACvB,KAAK,cAAc,CAAC;QACpB,KAAK,SAAS;YACb,OAAO,EAAE,OAAO,EAAE,UAAU,EAAE,CAAC;QAChC,KAAK,YAAY,CAAC;QAClB,KAAK,SAAS,CAAC;QACf,KAAK,UAAU;YACd,OAAO,EAAE,OAAO,EAAE,WAAW,EAAE,CAAC;QACjC,KAAK,gBAAgB;YACpB,OAAO,OAAO,CAAC,YAAY,KAAK,UAAU,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,UAAU,EAAE,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,WAAW,EAAE,CAAC;QACjG,KAAK,aAAa;YACjB,OAAO,sBAAsB,CAAC,OAAO,CAAC,UAAU,CAAC,CAAC;QACnD;YACC,OAAO,WAAW,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC;IACnC,CAAC;AACF,CAAC,CAAC;AAEF;;;;;GAKG;AACH,MAAM,uCAAuC,GAAwB,IAAI,GAAG,CAAC;IAC5E,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG;CAChD,CAAC,CAAC;AAEH,iFAAiF;AACjF,MAAM,wCAAwC,GAAwB,IAAI,GAAG,CAAC;IAC7E,oBAAoB;IACpB,kBAAkB;CAClB,CAAC,CAAC;AAEH;;;;;;;;;;;;;;;;;;GAkBG;AACH,MAAM,CAAC,MAAM,6BAA6B,GAAoB,CAAC,OAAO,EAAE,EAAE;IACzE,QAAQ,OAAO,CAAC,IAAI,EAAE,CAAC;QACtB,KAAK,OAAO,CAAC;QACb,KAAK,SAAS,CAAC;QACf,KAAK,WAAW,CAAC;QACjB,KAAK,iBAAiB,CAAC;QACvB,KAAK,cAAc;YAClB,OAAO,EAAE,OAAO,EAAE,UAAU,EAAE,CAAC;IACjC,CAAC;IAED,MAAM,YAAY,GAAG,oBAAoB,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC;IAC3D,IAAI,YAAY,KAAK,eAAe;QAAE,OAAO,EAAE,OAAO,EAAE,UAAU,EAAE,CAAC;IAErE,IAAI,OAAO,CAAC,WAAW,KAAK,KAAK;QAAE,OAAO,EAAE,OAAO,EAAE,UAAU,EAAE,CAAC;IAClE,IAAI,OAAO,CAAC,aAAa,EAAE,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,wCAAwC,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,EAAE,CAAC;QAC/F,OAAO,EAAE,OAAO,EAAE,UAAU,EAAE,CAAC;IAChC,CAAC;IAED,IAAI,YAAY,KAAK,WAAW;QAAE,OAAO,EAAE,OAAO,EAAE,WAAW,EAAE,CAAC;IAElE,IAAI,OAAO,CAAC,UAAU,KAAK,SAAS,IAAI,uCAAuC,CAAC,GAAG,CAAC,OAAO,CAAC,UAAU,CAAC,EAAE,CAAC;QACzG,OAAO,OAAO,CAAC,UAAU,KAAK,GAAG,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,cAAc,EAAE,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,WAAW,EAAE,CAAC;IAC5F,CAAC;IACD,OAAO,EAAE,OAAO,EAAE,UAAU,EAAE,CAAC;AAChC,CAAC,CAAC","sourcesContent":["import { classifyErrorMessage } from \"../retry.ts\";\nimport type { RetryClassification, RetryClassifier, RetryFailure } from \"./types.ts\";\n\n/**\n * Status codes Kimi's provider-request stage treats as retryable. 429 is its\n * own `rate-limited` verdict (the turn budget treats it through hint tiers);\n * the other whitelisted codes are ordinary `transient`. Every other status —\n * client errors and non-whitelisted 5xx alike — is `terminal`.\n */\nconst KIMI_RETRYABLE_STATUS_CODES: ReadonlySet<number> = new Set([408, 409, 429, 500, 502, 503, 504, 529]);\n\nfunction classifyKimiHttpStatus(statusCode: number | undefined): RetryClassification {\n\tif (statusCode === undefined || !KIMI_RETRYABLE_STATUS_CODES.has(statusCode)) {\n\t\treturn { verdict: \"terminal\" };\n\t}\n\treturn statusCode === 429 ? { verdict: \"rate-limited\" } : { verdict: \"transient\" };\n}\n\nfunction assertNever(kind: never): never {\n\tthrow new Error(`Unhandled retry failure kind: ${kind}`);\n}\n\n/**\n * Kimi provider-request classifier, ported from the source provider's retry\n * policy (policy only, not its code). The switch is exhaustive over\n * {@link RetryFailure}[\"kind\"]:\n *\n * - abort / refusal / sensitive / quota-exhausted / image-format / unknown are\n *   terminal: user cancellations, content refusals, dead accounts, rejected\n *   payloads, and unrecognized failures never recover by re-sending.\n * - connection / timeout / provider failures are transient transport noise.\n * - empty-response is transient unless the model said `finishReason` was\n *   \"filtered\", which is a deterministic content rejection.\n * - http-status is retryable only for the whitelisted codes above.\n */\nexport const classifyKimiFailure: RetryClassifier = (failure) => {\n\tswitch (failure.kind) {\n\t\tcase \"abort\":\n\t\tcase \"refusal\":\n\t\tcase \"sensitive\":\n\t\tcase \"quota-exhausted\":\n\t\tcase \"image-format\":\n\t\tcase \"unknown\":\n\t\t\treturn { verdict: \"terminal\" };\n\t\tcase \"connection\":\n\t\tcase \"timeout\":\n\t\tcase \"provider\":\n\t\t\treturn { verdict: \"transient\" };\n\t\tcase \"empty-response\":\n\t\t\treturn failure.finishReason === \"filtered\" ? { verdict: \"terminal\" } : { verdict: \"transient\" };\n\t\tcase \"http-status\":\n\t\t\treturn classifyKimiHttpStatus(failure.statusCode);\n\t\tdefault:\n\t\t\treturn assertNever(failure.kind);\n\t}\n};\n\n/**\n * Structured status codes the senpi default classifier treats as retryable\n * when — and only when — the message regexes have no opinion. Mirrors the\n * transient half of Kimi's table plus the Cloudflare gateway codes senpi\n * already retries by regex.\n */\nconst SENPI_STRUCTURED_RETRYABLE_STATUS_CODES: ReadonlySet<number> = new Set([\n\t408, 409, 429, 500, 502, 503, 504, 522, 524, 529,\n]);\n\n/** Provider error codes that prove a dead account regardless of message text. */\nconst SENPI_STRUCTURED_TERMINAL_PROVIDER_CODES: ReadonlySet<string> = new Set([\n\t\"insufficient_quota\",\n\t\"credits_required\",\n]);\n\n/**\n * Senpi assistant-turn classifier: the existing message regexes decide first;\n * structured failure facts are consulted only where the regexes are silent.\n *\n * Precedence (terminal wins at every tie):\n * 1. Deterministic failure kinds (abort/refusal/sensitive/quota/image-format)\n *    are terminal before any text inspection.\n * 2. A non-retryable regex match is terminal — request-shape and quota wording\n *    outrank any structured status (a 500 carrying invalid-tool-schema text\n *    stays terminal, unlike under {@link classifyKimiFailure}).\n * 3. Structured terminal facts (`shouldRetry: false`, quota provider codes).\n * 4. A retryable regex match is transient — e.g. a 400 carrying Anthropic's\n *    server-tool pairing text stays retryable because the request builder\n *    repairs the replayed history.\n * 5. Only when the regexes have no opinion does the structured status decide:\n *    whitelisted transient codes retry, everything else is terminal.\n * 6. No regex match and no structured facts: terminal, exactly the legacy\n *    boolean fallback.\n */\nexport const classifySenpiAssistantFailure: RetryClassifier = (failure) => {\n\tswitch (failure.kind) {\n\t\tcase \"abort\":\n\t\tcase \"refusal\":\n\t\tcase \"sensitive\":\n\t\tcase \"quota-exhausted\":\n\t\tcase \"image-format\":\n\t\t\treturn { verdict: \"terminal\" };\n\t}\n\n\tconst regexVerdict = classifyErrorMessage(failure.message);\n\tif (regexVerdict === \"non-retryable\") return { verdict: \"terminal\" };\n\n\tif (failure.shouldRetry === false) return { verdict: \"terminal\" };\n\tif (failure.providerCodes?.some((code) => SENPI_STRUCTURED_TERMINAL_PROVIDER_CODES.has(code))) {\n\t\treturn { verdict: \"terminal\" };\n\t}\n\n\tif (regexVerdict === \"retryable\") return { verdict: \"transient\" };\n\n\tif (failure.statusCode !== undefined && SENPI_STRUCTURED_RETRYABLE_STATUS_CODES.has(failure.statusCode)) {\n\t\treturn failure.statusCode === 429 ? { verdict: \"rate-limited\" } : { verdict: \"transient\" };\n\t}\n\treturn { verdict: \"terminal\" };\n};\n"]}