@code-yeongyu/senpi 2026.9.5 → 2026.9.7

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":"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. When omitted, adapters use provider-specific behavior. */\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\t/** Exact provider-native effort level used for this response. Absent for legacy or unmanaged responses. */\n\tproviderThinkingLevel?: string;\n\tdiagnostics?: AssistantMessageDiagnostic[]; // Redacted provider/runtime diagnostics for failures and recoveries.\n\tusage: Usage;\n\tstopReason: StopReason;\n\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\t/** Whether the provider accepts the `max_output_tokens` parameter. Some Codex-protocol gateways reject it. Default: true. */\n\tsupportsMaxOutputTokens?: boolean;\n\t/**\n\t * vLLM scheduler priority sent as the top-level `priority` request field (lower values are\n\t * handled earlier; server default 0). Only meaningful when vLLM runs with\n\t * `--scheduling-policy priority`; useful for keeping background/batch work from stalling\n\t * interactive sessions. Off by default; not set on the generated catalog.\n\t */\n\tvllmPriority?: number;\n}\n\n/** Compatibility settings for 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/** Whether the exact model transport supports effort-only system messages and thinking binding controls. Default: false. */\n\tsupportsMidConvoEffort?: 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"]}
+{"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. When omitted, adapters use provider-specific behavior. */\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\n/** A durable Responses configuration change projected into the message stream. */\nexport interface ConfigurationUpdateMessage {\n\trole: \"configurationUpdate\";\n\tcontent: (TextContent | ImageContent)[];\n\teffort: string;\n\ttimestamp: number;\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\t/** Exact provider-native effort level used for this response. Absent for legacy or unmanaged responses. */\n\tproviderThinkingLevel?: string;\n\tdiagnostics?: AssistantMessageDiagnostic[]; // Redacted provider/runtime diagnostics for failures and recoveries.\n\tusage: Usage;\n\tstopReason: StopReason;\n\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 | ConfigurationUpdateMessage;\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\t/** Whether the provider accepts the `max_output_tokens` parameter. Some Codex-protocol gateways reject it. Default: true. */\n\tsupportsMaxOutputTokens?: boolean;\n\t/**\n\t * vLLM scheduler priority sent as the top-level `priority` request field (lower values are\n\t * handled earlier; server default 0). Only meaningful when vLLM runs with\n\t * `--scheduling-policy priority`; useful for keeping background/batch work from stalling\n\t * interactive sessions. Off by default; not set on the generated catalog.\n\t */\n\tvllmPriority?: number;\n}\n\n/** Compatibility settings for 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/** Whether the exact model transport supports effort-only system messages and thinking binding controls. Default: false. */\n\tsupportsMidConvoEffort?: 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/estimate.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"estimate.d.ts","sourceRoot":"","sources":["../../src/utils/estimate.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAoB,OAAO,EAAE,YAAY,EAAE,OAAO,EAAE,WAAW,EAAQ,KAAK,EAAE,MAAM,aAAa,CAAC;AAE9G,MAAM,WAAW,oBAAoB;IACpC,sCAAsC;IACtC,MAAM,EAAE,MAAM,CAAC;IACf,2EAA2E;IAC3E,WAAW,EAAE,MAAM,CAAC;IACpB,+EAA+E;IAC/E,cAAc,EAAE,MAAM,CAAC;IACvB,qFAAqF;IACrF,cAAc,EAAE,MAAM,GAAG,IAAI,CAAC;CAC9B;AAKD,wBAAgB,sBAAsB,CAAC,KAAK,EAAE,KAAK,GAAG,MAAM,CAE3D;AAkBD,wBAAgB,kBAAkB,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,CAEvD;AAED,wBAAgB,iCAAiC,CAAC,OAAO,EAAE,MAAM,GAAG,KAAK,CAAC,WAAW,GAAG,YAAY,CAAC,GAAG,MAAM,CAE7G;AAED,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,OAAO,GAAG,MAAM,CAkB9D;AAqDD,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,OAAO,GAAG,SAAS,OAAO,EAAE,GAAG,oBAAoB,CA6BjG"}
+{"version":3,"file":"estimate.d.ts","sourceRoot":"","sources":["../../src/utils/estimate.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAoB,OAAO,EAAE,YAAY,EAAE,OAAO,EAAE,WAAW,EAAQ,KAAK,EAAE,MAAM,aAAa,CAAC;AAE9G,MAAM,WAAW,oBAAoB;IACpC,sCAAsC;IACtC,MAAM,EAAE,MAAM,CAAC;IACf,2EAA2E;IAC3E,WAAW,EAAE,MAAM,CAAC;IACpB,+EAA+E;IAC/E,cAAc,EAAE,MAAM,CAAC;IACvB,qFAAqF;IACrF,cAAc,EAAE,MAAM,GAAG,IAAI,CAAC;CAC9B;AAKD,wBAAgB,sBAAsB,CAAC,KAAK,EAAE,KAAK,GAAG,MAAM,CAE3D;AAkBD,wBAAgB,kBAAkB,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,CAEvD;AAED,wBAAgB,iCAAiC,CAAC,OAAO,EAAE,MAAM,GAAG,KAAK,CAAC,WAAW,GAAG,YAAY,CAAC,GAAG,MAAM,CAE7G;AAED,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,OAAO,GAAG,MAAM,CAmB9D;AAqDD,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,OAAO,GAAG,SAAS,OAAO,EAAE,GAAG,oBAAoB,CA6BjG"}

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

@@ -31,6 +31,8 @@ export function estimateMessageTokens(message) {
         return estimateTextAndImageContentTokens(message.content);
     if (message.role === "toolResult")
         return estimateTextAndImageContentTokens(message.content);
+    if (message.role === "configurationUpdate")
+        return 0;
     for (const block of message.content) {
         if (block.type === "text") {
             chars += block.text.length;

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

@@ -1 +1 @@
-{"version":3,"file":"estimate.js","sourceRoot":"","sources":["../../src/utils/estimate.ts"],"names":[],"mappings":"AAaA,MAAM,eAAe,GAAG,CAAC,CAAC;AAC1B,MAAM,qBAAqB,GAAG,IAAI,CAAC;AAEnC,MAAM,UAAU,sBAAsB,CAAC,KAAY;IAClD,OAAO,KAAK,CAAC,WAAW,IAAI,KAAK,CAAC,KAAK,GAAG,KAAK,CAAC,MAAM,GAAG,KAAK,CAAC,SAAS,GAAG,KAAK,CAAC,UAAU,CAAC;AAC7F,CAAC;AAED,SAAS,iBAAiB,CAAC,KAAc;IACxC,IAAI,CAAC;QACJ,OAAO,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,IAAI,WAAW,CAAC;IAC7C,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,kBAAkB,CAAC;IAC3B,CAAC;AACF,CAAC;AAED,SAAS,gCAAgC,CAAC,OAAmD;IAC5F,IAAI,OAAO,OAAO,KAAK,QAAQ;QAAE,OAAO,OAAO,CAAC,MAAM,CAAC;IAEvD,IAAI,KAAK,GAAG,CAAC,CAAC;IACd,KAAK,MAAM,KAAK,IAAI,OAAO;QAAE,KAAK,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM,CAAC,CAAC,CAAC,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,qBAAqB,CAAC;IACxG,OAAO,KAAK,CAAC;AACd,CAAC;AAED,MAAM,UAAU,kBAAkB,CAAC,IAAY;IAC9C,OAAO,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,MAAM,GAAG,eAAe,CAAC,CAAC;AACjD,CAAC;AAED,MAAM,UAAU,iCAAiC,CAAC,OAAmD;IACpG,OAAO,IAAI,CAAC,IAAI,CAAC,gCAAgC,CAAC,OAAO,CAAC,GAAG,eAAe,CAAC,CAAC;AAC/E,CAAC;AAED,MAAM,UAAU,qBAAqB,CAAC,OAAgB;IACrD,IAAI,KAAK,GAAG,CAAC,CAAC;IAEd,IAAI,OAAO,CAAC,IAAI,KAAK,MAAM;QAAE,OAAO,iCAAiC,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC;IACvF,IAAI,OAAO,CAAC,IAAI,KAAK,YAAY;QAAE,OAAO,iCAAiC,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC;IAE7F,KAAK,MAAM,KAAK,IAAI,OAAO,CAAC,OAAO,EAAE,CAAC;QACrC,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;YAC3B,KAAK,IAAI,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC;QAC5B,CAAC;aAAM,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;YACtC,KAAK,IAAI,KAAK,CAAC,QAAQ,CAAC,MAAM,CAAC;QAChC,CAAC;aAAM,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;YACtC,KAAK,IAAI,KAAK,CAAC,IAAI,CAAC,MAAM,GAAG,iBAAiB,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC,MAAM,CAAC;QACxE,CAAC;aAAM,CAAC;YACP,KAAK,IAAI,KAAK,CAAC,OAAO,CAAC,MAAM,GAAG,iBAAiB,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,MAAM,CAAC;QACrE,CAAC;IACF,CAAC;IACD,OAAO,IAAI,CAAC,IAAI,CAAC,KAAK,GAAG,eAAe,CAAC,CAAC;AAC3C,CAAC;AAED,SAAS,yBAAyB,CAAC,QAA4B;IAC9D,IAAI,qBAAqB,GAAG,MAAM,CAAC,iBAAiB,CAAC;IACrD,IAAI,SAAsD,CAAC;IAE3D,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QAC1C,MAAM,OAAO,GAAG,QAAQ,CAAC,CAAC,CAAC,CAAC;QAC5B,IAAI,OAAO,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;YAClC,MAAM,SAAS,GAAG,OAA2B,CAAC;YAC9C,0EAA0E;YAC1E,wEAAwE;YACxE,MAAM,oBAAoB,GAAG,SAAS,CAAC,SAAS,IAAI,qBAAqB,CAAC;YAC1E,IACC,oBAAoB;gBACpB,SAAS,CAAC,UAAU,KAAK,SAAS;gBAClC,SAAS,CAAC,UAAU,KAAK,OAAO;gBAChC,sBAAsB,CAAC,SAAS,CAAC,KAAK,CAAC,GAAG,CAAC,EAC1C,CAAC;gBACF,SAAS,GAAG,EAAE,KAAK,EAAE,SAAS,CAAC,KAAK,EAAE,KAAK,EAAE,CAAC,EAAE,CAAC;YAClD,CAAC;QACF,CAAC;QACD,qBAAqB,GAAG,IAAI,CAAC,GAAG,CAAC,qBAAqB,EAAE,OAAO,CAAC,SAAS,CAAC,CAAC;IAC5E,CAAC;IAED,OAAO,SAAS,CAAC;AAClB,CAAC;AAED,SAAS,gBAAgB,CAAC,QAA4B;IACrD,MAAM,SAAS,GAAG,yBAAyB,CAAC,QAAQ,CAAC,CAAC;IACtD,IAAI,SAAS,EAAE,CAAC;QACf,MAAM,WAAW,GAAG,sBAAsB,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC;QAC5D,IAAI,cAAc,GAAG,CAAC,CAAC;QACvB,KAAK,IAAI,CAAC,GAAG,SAAS,CAAC,KAAK,GAAG,CAAC,EAAE,CAAC,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YAC5D,cAAc,IAAI,qBAAqB,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC;QACtD,CAAC;QACD,OAAO,EAAE,MAAM,EAAE,WAAW,GAAG,cAAc,EAAE,WAAW,EAAE,cAAc,EAAE,cAAc,EAAE,SAAS,CAAC,KAAK,EAAE,CAAC;IAC/G,CAAC;IAED,IAAI,MAAM,GAAG,CAAC,CAAC;IACf,KAAK,MAAM,OAAO,IAAI,QAAQ;QAAE,MAAM,IAAI,qBAAqB,CAAC,OAAO,CAAC,CAAC;IACzE,OAAO,EAAE,MAAM,EAAE,WAAW,EAAE,CAAC,EAAE,cAAc,EAAE,MAAM,EAAE,cAAc,EAAE,IAAI,EAAE,CAAC;AACjF,CAAC;AAED,SAAS,mBAAmB,CAAC,KAAkC;IAC9D,IAAI,CAAC,KAAK,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,CAAC,CAAC;IAC3C,OAAO,kBAAkB,CAAC,iBAAiB,CAAC,KAAK,CAAC,CAAC,CAAC;AACrD,CAAC;AAED,SAAS,cAAc,CAAC,KAAmC;IAC1D,OAAO,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;AAC7B,CAAC;AAED,MAAM,UAAU,qBAAqB,CAAC,OAAqC;IAC1E,IAAI,cAAc,CAAC,OAAO,CAAC;QAAE,OAAO,gBAAgB,CAAC,OAAO,CAAC,CAAC;IAE9D,MAAM,QAAQ,GAAG,gBAAgB,CAAC,OAAO,CAAC,QAAQ,CAAC,CAAC;IACpD,IAAI,QAAQ,CAAC,cAAc,KAAK,IAAI,EAAE,CAAC;QACtC,MAAM,UAAU,GAAG,IAAI,GAAG,CACzB,OAAO,CAAC,QAAQ;aACd,KAAK,CAAC,QAAQ,CAAC,cAAc,GAAG,CAAC,CAAC;aAClC,MAAM,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,OAAO,CAAC,IAAI,KAAK,YAAY,CAAC;aAClD,OAAO,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,OAAO,CAAC,cAAc,IAAI,EAAE,CAAC,CACpD,CAAC;QACF,MAAM,eAAe,GAAG,mBAAmB,CAAC,OAAO,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,UAAU,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;QACxG,OAAO;YACN,MAAM,EAAE,QAAQ,CAAC,MAAM,GAAG,eAAe;YACzC,WAAW,EAAE,QAAQ,CAAC,WAAW;YACjC,cAAc,EAAE,QAAQ,CAAC,cAAc,GAAG,eAAe;YACzD,cAAc,EAAE,QAAQ,CAAC,cAAc;SACvC,CAAC;IACH,CAAC;IAED,MAAM,YAAY,GACjB,CAAC,OAAO,CAAC,YAAY,CAAC,CAAC,CAAC,kBAAkB,CAAC,OAAO,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,mBAAmB,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;IAE5G,OAAO;QACN,MAAM,EAAE,QAAQ,CAAC,MAAM,GAAG,YAAY;QACtC,WAAW,EAAE,QAAQ,CAAC,WAAW;QACjC,cAAc,EAAE,QAAQ,CAAC,cAAc,GAAG,YAAY;QACtD,cAAc,EAAE,QAAQ,CAAC,cAAc;KACvC,CAAC;AACH,CAAC","sourcesContent":["import type { AssistantMessage, Context, ImageContent, Message, TextContent, Tool, Usage } from \"../types.ts\";\n\nexport interface ContextUsageEstimate {\n\t/** Estimated total context tokens. */\n\ttokens: number;\n\t/** Tokens reported by the most recent applicable assistant usage block. */\n\tusageTokens: number;\n\t/** Estimated tokens after the most recent applicable assistant usage block. */\n\ttrailingTokens: number;\n\t/** Index of the applicable message that provided usage, or null when none exists. */\n\tlastUsageIndex: number | null;\n}\n\nconst CHARS_PER_TOKEN = 4;\nconst ESTIMATED_IMAGE_CHARS = 4800;\n\nexport function calculateContextTokens(usage: Usage): number {\n\treturn usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite;\n}\n\nfunction safeJsonStringify(value: unknown): string {\n\ttry {\n\t\treturn JSON.stringify(value) ?? \"undefined\";\n\t} catch {\n\t\treturn \"[unserializable]\";\n\t}\n}\n\nfunction estimateTextAndImageContentChars(content: string | Array<TextContent | ImageContent>): number {\n\tif (typeof content === \"string\") return content.length;\n\n\tlet chars = 0;\n\tfor (const block of content) chars += block.type === \"text\" ? block.text.length : ESTIMATED_IMAGE_CHARS;\n\treturn chars;\n}\n\nexport function estimateTextTokens(text: string): number {\n\treturn Math.ceil(text.length / CHARS_PER_TOKEN);\n}\n\nexport function estimateTextAndImageContentTokens(content: string | Array<TextContent | ImageContent>): number {\n\treturn Math.ceil(estimateTextAndImageContentChars(content) / CHARS_PER_TOKEN);\n}\n\nexport function estimateMessageTokens(message: Message): number {\n\tlet chars = 0;\n\n\tif (message.role === \"user\") return estimateTextAndImageContentTokens(message.content);\n\tif (message.role === \"toolResult\") return estimateTextAndImageContentTokens(message.content);\n\n\tfor (const block of message.content) {\n\t\tif (block.type === \"text\") {\n\t\t\tchars += block.text.length;\n\t\t} else if (block.type === \"thinking\") {\n\t\t\tchars += block.thinking.length;\n\t\t} else if (block.type === \"toolCall\") {\n\t\t\tchars += block.name.length + safeJsonStringify(block.arguments).length;\n\t\t} else {\n\t\t\tchars += block.subtype.length + safeJsonStringify(block.raw).length;\n\t\t}\n\t}\n\treturn Math.ceil(chars / CHARS_PER_TOKEN);\n}\n\nfunction getLastAssistantUsageInfo(messages: readonly Message[]): { usage: Usage; index: number } | undefined {\n\tlet latestPrefixTimestamp = Number.NEGATIVE_INFINITY;\n\tlet usageInfo: { usage: Usage; index: number } | undefined;\n\n\tfor (let i = 0; i < messages.length; i++) {\n\t\tconst message = messages[i];\n\t\tif (message.role === \"assistant\") {\n\t\t\tconst assistant = message as AssistantMessage;\n\t\t\t// A newer prefix message was inserted after this response (for example, a\n\t\t\t// compaction summary), so its usage cannot describe the current prefix.\n\t\t\tconst usageAppliesToPrefix = assistant.timestamp >= latestPrefixTimestamp;\n\t\t\tif (\n\t\t\t\tusageAppliesToPrefix &&\n\t\t\t\tassistant.stopReason !== \"aborted\" &&\n\t\t\t\tassistant.stopReason !== \"error\" &&\n\t\t\t\tcalculateContextTokens(assistant.usage) > 0\n\t\t\t) {\n\t\t\t\tusageInfo = { usage: assistant.usage, index: i };\n\t\t\t}\n\t\t}\n\t\tlatestPrefixTimestamp = Math.max(latestPrefixTimestamp, message.timestamp);\n\t}\n\n\treturn usageInfo;\n}\n\nfunction estimateMessages(messages: readonly Message[]): ContextUsageEstimate {\n\tconst usageInfo = getLastAssistantUsageInfo(messages);\n\tif (usageInfo) {\n\t\tconst usageTokens = calculateContextTokens(usageInfo.usage);\n\t\tlet trailingTokens = 0;\n\t\tfor (let i = usageInfo.index + 1; i < messages.length; i++) {\n\t\t\ttrailingTokens += estimateMessageTokens(messages[i]);\n\t\t}\n\t\treturn { tokens: usageTokens + trailingTokens, usageTokens, trailingTokens, lastUsageIndex: usageInfo.index };\n\t}\n\n\tlet tokens = 0;\n\tfor (const message of messages) tokens += estimateMessageTokens(message);\n\treturn { tokens, usageTokens: 0, trailingTokens: tokens, lastUsageIndex: null };\n}\n\nfunction estimateToolsTokens(tools: readonly Tool[] | undefined): number {\n\tif (!tools || tools.length === 0) return 0;\n\treturn estimateTextTokens(safeJsonStringify(tools));\n}\n\nfunction isMessageArray(value: Context | readonly Message[]): value is readonly Message[] {\n\treturn Array.isArray(value);\n}\n\nexport function estimateContextTokens(context: Context | readonly Message[]): ContextUsageEstimate {\n\tif (isMessageArray(context)) return estimateMessages(context);\n\n\tconst estimate = estimateMessages(context.messages);\n\tif (estimate.lastUsageIndex !== null) {\n\t\tconst addedNames = new Set(\n\t\t\tcontext.messages\n\t\t\t\t.slice(estimate.lastUsageIndex + 1)\n\t\t\t\t.filter((message) => message.role === \"toolResult\")\n\t\t\t\t.flatMap((message) => message.addedToolNames ?? []),\n\t\t);\n\t\tconst addedToolTokens = estimateToolsTokens(context.tools?.filter((tool) => addedNames.has(tool.name)));\n\t\treturn {\n\t\t\ttokens: estimate.tokens + addedToolTokens,\n\t\t\tusageTokens: estimate.usageTokens,\n\t\t\ttrailingTokens: estimate.trailingTokens + addedToolTokens,\n\t\t\tlastUsageIndex: estimate.lastUsageIndex,\n\t\t};\n\t}\n\n\tconst prefixTokens =\n\t\t(context.systemPrompt ? estimateTextTokens(context.systemPrompt) : 0) + estimateToolsTokens(context.tools);\n\n\treturn {\n\t\ttokens: estimate.tokens + prefixTokens,\n\t\tusageTokens: estimate.usageTokens,\n\t\ttrailingTokens: estimate.trailingTokens + prefixTokens,\n\t\tlastUsageIndex: estimate.lastUsageIndex,\n\t};\n}\n"]}
+{"version":3,"file":"estimate.js","sourceRoot":"","sources":["../../src/utils/estimate.ts"],"names":[],"mappings":"AAaA,MAAM,eAAe,GAAG,CAAC,CAAC;AAC1B,MAAM,qBAAqB,GAAG,IAAI,CAAC;AAEnC,MAAM,UAAU,sBAAsB,CAAC,KAAY;IAClD,OAAO,KAAK,CAAC,WAAW,IAAI,KAAK,CAAC,KAAK,GAAG,KAAK,CAAC,MAAM,GAAG,KAAK,CAAC,SAAS,GAAG,KAAK,CAAC,UAAU,CAAC;AAC7F,CAAC;AAED,SAAS,iBAAiB,CAAC,KAAc;IACxC,IAAI,CAAC;QACJ,OAAO,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,IAAI,WAAW,CAAC;IAC7C,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,kBAAkB,CAAC;IAC3B,CAAC;AACF,CAAC;AAED,SAAS,gCAAgC,CAAC,OAAmD;IAC5F,IAAI,OAAO,OAAO,KAAK,QAAQ;QAAE,OAAO,OAAO,CAAC,MAAM,CAAC;IAEvD,IAAI,KAAK,GAAG,CAAC,CAAC;IACd,KAAK,MAAM,KAAK,IAAI,OAAO;QAAE,KAAK,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM,CAAC,CAAC,CAAC,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,qBAAqB,CAAC;IACxG,OAAO,KAAK,CAAC;AACd,CAAC;AAED,MAAM,UAAU,kBAAkB,CAAC,IAAY;IAC9C,OAAO,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,MAAM,GAAG,eAAe,CAAC,CAAC;AACjD,CAAC;AAED,MAAM,UAAU,iCAAiC,CAAC,OAAmD;IACpG,OAAO,IAAI,CAAC,IAAI,CAAC,gCAAgC,CAAC,OAAO,CAAC,GAAG,eAAe,CAAC,CAAC;AAC/E,CAAC;AAED,MAAM,UAAU,qBAAqB,CAAC,OAAgB;IACrD,IAAI,KAAK,GAAG,CAAC,CAAC;IAEd,IAAI,OAAO,CAAC,IAAI,KAAK,MAAM;QAAE,OAAO,iCAAiC,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC;IACvF,IAAI,OAAO,CAAC,IAAI,KAAK,YAAY;QAAE,OAAO,iCAAiC,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC;IAC7F,IAAI,OAAO,CAAC,IAAI,KAAK,qBAAqB;QAAE,OAAO,CAAC,CAAC;IAErD,KAAK,MAAM,KAAK,IAAI,OAAO,CAAC,OAAO,EAAE,CAAC;QACrC,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;YAC3B,KAAK,IAAI,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC;QAC5B,CAAC;aAAM,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;YACtC,KAAK,IAAI,KAAK,CAAC,QAAQ,CAAC,MAAM,CAAC;QAChC,CAAC;aAAM,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;YACtC,KAAK,IAAI,KAAK,CAAC,IAAI,CAAC,MAAM,GAAG,iBAAiB,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC,MAAM,CAAC;QACxE,CAAC;aAAM,CAAC;YACP,KAAK,IAAI,KAAK,CAAC,OAAO,CAAC,MAAM,GAAG,iBAAiB,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,MAAM,CAAC;QACrE,CAAC;IACF,CAAC;IACD,OAAO,IAAI,CAAC,IAAI,CAAC,KAAK,GAAG,eAAe,CAAC,CAAC;AAC3C,CAAC;AAED,SAAS,yBAAyB,CAAC,QAA4B;IAC9D,IAAI,qBAAqB,GAAG,MAAM,CAAC,iBAAiB,CAAC;IACrD,IAAI,SAAsD,CAAC;IAE3D,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QAC1C,MAAM,OAAO,GAAG,QAAQ,CAAC,CAAC,CAAC,CAAC;QAC5B,IAAI,OAAO,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;YAClC,MAAM,SAAS,GAAG,OAA2B,CAAC;YAC9C,0EAA0E;YAC1E,wEAAwE;YACxE,MAAM,oBAAoB,GAAG,SAAS,CAAC,SAAS,IAAI,qBAAqB,CAAC;YAC1E,IACC,oBAAoB;gBACpB,SAAS,CAAC,UAAU,KAAK,SAAS;gBAClC,SAAS,CAAC,UAAU,KAAK,OAAO;gBAChC,sBAAsB,CAAC,SAAS,CAAC,KAAK,CAAC,GAAG,CAAC,EAC1C,CAAC;gBACF,SAAS,GAAG,EAAE,KAAK,EAAE,SAAS,CAAC,KAAK,EAAE,KAAK,EAAE,CAAC,EAAE,CAAC;YAClD,CAAC;QACF,CAAC;QACD,qBAAqB,GAAG,IAAI,CAAC,GAAG,CAAC,qBAAqB,EAAE,OAAO,CAAC,SAAS,CAAC,CAAC;IAC5E,CAAC;IAED,OAAO,SAAS,CAAC;AAClB,CAAC;AAED,SAAS,gBAAgB,CAAC,QAA4B;IACrD,MAAM,SAAS,GAAG,yBAAyB,CAAC,QAAQ,CAAC,CAAC;IACtD,IAAI,SAAS,EAAE,CAAC;QACf,MAAM,WAAW,GAAG,sBAAsB,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC;QAC5D,IAAI,cAAc,GAAG,CAAC,CAAC;QACvB,KAAK,IAAI,CAAC,GAAG,SAAS,CAAC,KAAK,GAAG,CAAC,EAAE,CAAC,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YAC5D,cAAc,IAAI,qBAAqB,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC;QACtD,CAAC;QACD,OAAO,EAAE,MAAM,EAAE,WAAW,GAAG,cAAc,EAAE,WAAW,EAAE,cAAc,EAAE,cAAc,EAAE,SAAS,CAAC,KAAK,EAAE,CAAC;IAC/G,CAAC;IAED,IAAI,MAAM,GAAG,CAAC,CAAC;IACf,KAAK,MAAM,OAAO,IAAI,QAAQ;QAAE,MAAM,IAAI,qBAAqB,CAAC,OAAO,CAAC,CAAC;IACzE,OAAO,EAAE,MAAM,EAAE,WAAW,EAAE,CAAC,EAAE,cAAc,EAAE,MAAM,EAAE,cAAc,EAAE,IAAI,EAAE,CAAC;AACjF,CAAC;AAED,SAAS,mBAAmB,CAAC,KAAkC;IAC9D,IAAI,CAAC,KAAK,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,CAAC,CAAC;IAC3C,OAAO,kBAAkB,CAAC,iBAAiB,CAAC,KAAK,CAAC,CAAC,CAAC;AACrD,CAAC;AAED,SAAS,cAAc,CAAC,KAAmC;IAC1D,OAAO,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;AAC7B,CAAC;AAED,MAAM,UAAU,qBAAqB,CAAC,OAAqC;IAC1E,IAAI,cAAc,CAAC,OAAO,CAAC;QAAE,OAAO,gBAAgB,CAAC,OAAO,CAAC,CAAC;IAE9D,MAAM,QAAQ,GAAG,gBAAgB,CAAC,OAAO,CAAC,QAAQ,CAAC,CAAC;IACpD,IAAI,QAAQ,CAAC,cAAc,KAAK,IAAI,EAAE,CAAC;QACtC,MAAM,UAAU,GAAG,IAAI,GAAG,CACzB,OAAO,CAAC,QAAQ;aACd,KAAK,CAAC,QAAQ,CAAC,cAAc,GAAG,CAAC,CAAC;aAClC,MAAM,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,OAAO,CAAC,IAAI,KAAK,YAAY,CAAC;aAClD,OAAO,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,OAAO,CAAC,cAAc,IAAI,EAAE,CAAC,CACpD,CAAC;QACF,MAAM,eAAe,GAAG,mBAAmB,CAAC,OAAO,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,UAAU,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;QACxG,OAAO;YACN,MAAM,EAAE,QAAQ,CAAC,MAAM,GAAG,eAAe;YACzC,WAAW,EAAE,QAAQ,CAAC,WAAW;YACjC,cAAc,EAAE,QAAQ,CAAC,cAAc,GAAG,eAAe;YACzD,cAAc,EAAE,QAAQ,CAAC,cAAc;SACvC,CAAC;IACH,CAAC;IAED,MAAM,YAAY,GACjB,CAAC,OAAO,CAAC,YAAY,CAAC,CAAC,CAAC,kBAAkB,CAAC,OAAO,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,mBAAmB,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;IAE5G,OAAO;QACN,MAAM,EAAE,QAAQ,CAAC,MAAM,GAAG,YAAY;QACtC,WAAW,EAAE,QAAQ,CAAC,WAAW;QACjC,cAAc,EAAE,QAAQ,CAAC,cAAc,GAAG,YAAY;QACtD,cAAc,EAAE,QAAQ,CAAC,cAAc;KACvC,CAAC;AACH,CAAC","sourcesContent":["import type { AssistantMessage, Context, ImageContent, Message, TextContent, Tool, Usage } from \"../types.ts\";\n\nexport interface ContextUsageEstimate {\n\t/** Estimated total context tokens. */\n\ttokens: number;\n\t/** Tokens reported by the most recent applicable assistant usage block. */\n\tusageTokens: number;\n\t/** Estimated tokens after the most recent applicable assistant usage block. */\n\ttrailingTokens: number;\n\t/** Index of the applicable message that provided usage, or null when none exists. */\n\tlastUsageIndex: number | null;\n}\n\nconst CHARS_PER_TOKEN = 4;\nconst ESTIMATED_IMAGE_CHARS = 4800;\n\nexport function calculateContextTokens(usage: Usage): number {\n\treturn usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite;\n}\n\nfunction safeJsonStringify(value: unknown): string {\n\ttry {\n\t\treturn JSON.stringify(value) ?? \"undefined\";\n\t} catch {\n\t\treturn \"[unserializable]\";\n\t}\n}\n\nfunction estimateTextAndImageContentChars(content: string | Array<TextContent | ImageContent>): number {\n\tif (typeof content === \"string\") return content.length;\n\n\tlet chars = 0;\n\tfor (const block of content) chars += block.type === \"text\" ? block.text.length : ESTIMATED_IMAGE_CHARS;\n\treturn chars;\n}\n\nexport function estimateTextTokens(text: string): number {\n\treturn Math.ceil(text.length / CHARS_PER_TOKEN);\n}\n\nexport function estimateTextAndImageContentTokens(content: string | Array<TextContent | ImageContent>): number {\n\treturn Math.ceil(estimateTextAndImageContentChars(content) / CHARS_PER_TOKEN);\n}\n\nexport function estimateMessageTokens(message: Message): number {\n\tlet chars = 0;\n\n\tif (message.role === \"user\") return estimateTextAndImageContentTokens(message.content);\n\tif (message.role === \"toolResult\") return estimateTextAndImageContentTokens(message.content);\n\tif (message.role === \"configurationUpdate\") return 0;\n\n\tfor (const block of message.content) {\n\t\tif (block.type === \"text\") {\n\t\t\tchars += block.text.length;\n\t\t} else if (block.type === \"thinking\") {\n\t\t\tchars += block.thinking.length;\n\t\t} else if (block.type === \"toolCall\") {\n\t\t\tchars += block.name.length + safeJsonStringify(block.arguments).length;\n\t\t} else {\n\t\t\tchars += block.subtype.length + safeJsonStringify(block.raw).length;\n\t\t}\n\t}\n\treturn Math.ceil(chars / CHARS_PER_TOKEN);\n}\n\nfunction getLastAssistantUsageInfo(messages: readonly Message[]): { usage: Usage; index: number } | undefined {\n\tlet latestPrefixTimestamp = Number.NEGATIVE_INFINITY;\n\tlet usageInfo: { usage: Usage; index: number } | undefined;\n\n\tfor (let i = 0; i < messages.length; i++) {\n\t\tconst message = messages[i];\n\t\tif (message.role === \"assistant\") {\n\t\t\tconst assistant = message as AssistantMessage;\n\t\t\t// A newer prefix message was inserted after this response (for example, a\n\t\t\t// compaction summary), so its usage cannot describe the current prefix.\n\t\t\tconst usageAppliesToPrefix = assistant.timestamp >= latestPrefixTimestamp;\n\t\t\tif (\n\t\t\t\tusageAppliesToPrefix &&\n\t\t\t\tassistant.stopReason !== \"aborted\" &&\n\t\t\t\tassistant.stopReason !== \"error\" &&\n\t\t\t\tcalculateContextTokens(assistant.usage) > 0\n\t\t\t) {\n\t\t\t\tusageInfo = { usage: assistant.usage, index: i };\n\t\t\t}\n\t\t}\n\t\tlatestPrefixTimestamp = Math.max(latestPrefixTimestamp, message.timestamp);\n\t}\n\n\treturn usageInfo;\n}\n\nfunction estimateMessages(messages: readonly Message[]): ContextUsageEstimate {\n\tconst usageInfo = getLastAssistantUsageInfo(messages);\n\tif (usageInfo) {\n\t\tconst usageTokens = calculateContextTokens(usageInfo.usage);\n\t\tlet trailingTokens = 0;\n\t\tfor (let i = usageInfo.index + 1; i < messages.length; i++) {\n\t\t\ttrailingTokens += estimateMessageTokens(messages[i]);\n\t\t}\n\t\treturn { tokens: usageTokens + trailingTokens, usageTokens, trailingTokens, lastUsageIndex: usageInfo.index };\n\t}\n\n\tlet tokens = 0;\n\tfor (const message of messages) tokens += estimateMessageTokens(message);\n\treturn { tokens, usageTokens: 0, trailingTokens: tokens, lastUsageIndex: null };\n}\n\nfunction estimateToolsTokens(tools: readonly Tool[] | undefined): number {\n\tif (!tools || tools.length === 0) return 0;\n\treturn estimateTextTokens(safeJsonStringify(tools));\n}\n\nfunction isMessageArray(value: Context | readonly Message[]): value is readonly Message[] {\n\treturn Array.isArray(value);\n}\n\nexport function estimateContextTokens(context: Context | readonly Message[]): ContextUsageEstimate {\n\tif (isMessageArray(context)) return estimateMessages(context);\n\n\tconst estimate = estimateMessages(context.messages);\n\tif (estimate.lastUsageIndex !== null) {\n\t\tconst addedNames = new Set(\n\t\t\tcontext.messages\n\t\t\t\t.slice(estimate.lastUsageIndex + 1)\n\t\t\t\t.filter((message) => message.role === \"toolResult\")\n\t\t\t\t.flatMap((message) => message.addedToolNames ?? []),\n\t\t);\n\t\tconst addedToolTokens = estimateToolsTokens(context.tools?.filter((tool) => addedNames.has(tool.name)));\n\t\treturn {\n\t\t\ttokens: estimate.tokens + addedToolTokens,\n\t\t\tusageTokens: estimate.usageTokens,\n\t\t\ttrailingTokens: estimate.trailingTokens + addedToolTokens,\n\t\t\tlastUsageIndex: estimate.lastUsageIndex,\n\t\t};\n\t}\n\n\tconst prefixTokens =\n\t\t(context.systemPrompt ? estimateTextTokens(context.systemPrompt) : 0) + estimateToolsTokens(context.tools);\n\n\treturn {\n\t\ttokens: estimate.tokens + prefixTokens,\n\t\tusageTokens: estimate.usageTokens,\n\t\ttrailingTokens: estimate.trailingTokens + prefixTokens,\n\t\tlastUsageIndex: estimate.lastUsageIndex,\n\t};\n}\n"]}

package/node_modules/@earendil-works/pi-ai/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@earendil-works/pi-ai",
-	"version": "2026.9.5",
+	"version": "2026.9.7",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"type": "module",
 	"main": "./dist/index.js",
@@ -76,7 +76,7 @@
 	"dependencies": {
 		"@anthropic-ai/sdk": "0.123.0",
 		"@aws-sdk/client-bedrock-runtime": "3.1116.0",
-		"@earendil-works/pi-telemetry": "npm:@code-yeongyu/senpi-telemetry@2026.9.5",
+		"@earendil-works/pi-telemetry": "npm:@code-yeongyu/senpi-telemetry@2026.9.7",
 		"@google/genai": "2.18.0",
 		"@smithy/node-http-handler": "4.11.3",
 		"http-proxy-agent": "9.1.0",

package/node_modules/@earendil-works/pi-pty/package.json

@@ -1,7 +1,7 @@
 {
 	"name": "@earendil-works/pi-pty",
 	"private": true,
-	"version": "2026.9.5",
+	"version": "2026.9.7",
 	"description": "Typed PTY session loader for Senpi persistent terminals",
 	"type": "module",
 	"main": "./dist/index.js",

package/node_modules/@earendil-works/pi-telemetry/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@earendil-works/pi-telemetry",
-	"version": "2026.9.5",
+	"version": "2026.9.7",
 	"description": "Vendor-neutral telemetry contracts and typed schema utilities for pi",
 	"type": "module",
 	"main": "./dist/index.js",

package/node_modules/@earendil-works/pi-tui/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@earendil-works/pi-tui",
-	"version": "2026.9.5",
+	"version": "2026.9.7",
 	"description": "Terminal User Interface library with differential rendering for efficient text-based applications",
 	"type": "module",
 	"main": "dist/index.js",

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@code-yeongyu/senpi",
-	"version": "2026.9.5",
+	"version": "2026.9.7",
 	"description": "Coding agent CLI with read, bash, edit, write tools and session management",
 	"type": "module",
 	"piConfig": {
@@ -48,9 +48,9 @@
 		"qa:app-server": "node scripts/qa-app-server/run-all.mjs"
 	},
 	"dependencies": {
-		"@earendil-works/pi-agent-core": "npm:@code-yeongyu/senpi-agent-core@2026.9.5",
-		"@earendil-works/pi-ai": "npm:@code-yeongyu/senpi-ai@2026.9.5",
-		"@earendil-works/pi-tui": "npm:@code-yeongyu/senpi-tui@2026.9.5",
+		"@earendil-works/pi-agent-core": "npm:@code-yeongyu/senpi-agent-core@2026.9.7",
+		"@earendil-works/pi-ai": "npm:@code-yeongyu/senpi-ai@2026.9.7",
+		"@earendil-works/pi-tui": "npm:@code-yeongyu/senpi-tui@2026.9.7",
 		"@silvia-odwyer/photon-node": "0.3.4",
 		"chalk": "6.0.0",
 		"cross-spawn": "7.0.6",
@@ -70,8 +70,8 @@
 		"@anthropic-ai/sdk": "0.123.0",
 		"@aws-sdk/client-bedrock-runtime": "3.1116.0",
 		"@bufbuild/protobuf": "2.14.0",
-		"@code-yeongyu/senpi-codemode": "2026.9.5",
-		"@earendil-works/pi-pty": "npm:@code-yeongyu/senpi-pty@2026.9.5",
+		"@code-yeongyu/senpi-codemode": "2026.9.7",
+		"@earendil-works/pi-pty": "npm:@code-yeongyu/senpi-pty@2026.9.7",
 		"@modelcontextprotocol/sdk": "1.30.0",
 		"@mozilla/readability": "0.6.0",
 		"@opentelemetry/api": "1.9.1",
@@ -119,7 +119,7 @@
 		"@csstools/css-parser-algorithms": "4.0.0",
 		"@csstools/css-syntax-patches-for-csstree": "1.1.8",
 		"@csstools/css-tokenizer": "4.0.0",
-		"@earendil-works/pi-telemetry": "npm:@code-yeongyu/senpi-telemetry@2026.9.5",
+		"@earendil-works/pi-telemetry": "npm:@code-yeongyu/senpi-telemetry@2026.9.7",
 		"@exodus/bytes": "1.15.1",
 		"@google/genai": "2.18.0",
 		"@hono/node-server": "2.1.1",