@tanstack/ai 0.36.0 → 0.37.0

package/dist/esm/middlewares/otel.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"otel.js","sources":["../../../src/middlewares/otel.ts"],"sourcesContent":["import {\n  SpanKind,\n  SpanStatusCode,\n  context as otelContext,\n  trace as otelTrace,\n} from '@opentelemetry/api'\nimport {\n  MAX_TOKENS_KEYS,\n  NESTED_MAX_TOKENS_KEY,\n} from '../utilities/sampling-keys'\nimport { firstNumber } from '../utilities/numbers'\nimport { errorMessage, errorTypeName } from '../utilities/errors'\nimport { usageAttributes } from './usage-attributes'\nimport type {\n  AttributeValue,\n  Exception,\n  Meter,\n  Span,\n  SpanOptions,\n  Tracer,\n} from '@opentelemetry/api'\nimport type {\n  ChatMiddleware,\n  ChatMiddlewareContext,\n} from '../activities/chat/middleware/types'\nimport type {\n  GenerationActivity,\n  GenerationMiddleware,\n  GenerationMiddlewareContext,\n} from '../activities/middleware/types'\n\n/**\n * Scope (role) of an OTel span emitted by this middleware.\n *\n * - `chat` — the root span for a single `chat()` call\n * - `iteration` — one per agent-loop iteration (one model call)\n * - `tool` — one per tool execution inside an iteration\n * - `generation` — the single span for a media activity call\n *   (`generateImage`, `generateVideo`, `generateSpeech`, …)\n */\nexport type OtelSpanScope = 'chat' | 'iteration' | 'tool' | 'generation'\n\n/**\n * Alias retained for backwards compatibility. Prefer {@link OtelSpanScope}.\n *\n * @deprecated Use `OtelSpanScope` instead — the name shadows OTel's built-in\n * `SpanKind` which is also imported by integrations of this middleware.\n */\nexport type OtelSpanKind = OtelSpanScope\n\n/**\n * Span metadata passed to `spanNameFormatter`, `attributeEnricher`,\n * `onBeforeSpanStart`, and `onSpanEnd`. Discriminated by `kind` so that\n * tool-only fields narrow automatically inside callback bodies.\n */\nexport type OtelSpanInfo<TScope extends OtelSpanScope = OtelSpanScope> =\n  TScope extends 'chat'\n    ? { kind: 'chat'; ctx: ChatMiddlewareContext }\n    : TScope extends 'iteration'\n      ? { kind: 'iteration'; ctx: ChatMiddlewareContext; iteration: number }\n      : TScope extends 'tool'\n        ? {\n            kind: 'tool'\n            ctx: ChatMiddlewareContext\n            iteration: number\n            toolName: string\n            toolCallId: string\n          }\n        : TScope extends 'generation'\n          ? { kind: 'generation'; ctx: GenerationMiddlewareContext }\n          : never\n\n/**\n * `gen_ai.operation.name` per activity. Chat uses the GenAI semconv value;\n * media operations have no semconv entry yet, so these are the de-facto names\n * consumed by GenAI backends (PostHog, Langfuse, …). Documented in\n * `docs/advanced/otel.md`.\n */\nconst OPERATION_NAME: Record<GenerationActivity, string> = {\n  chat: 'chat',\n  image: 'image_generation',\n  video: 'video_generation',\n  audio: 'audio_generation',\n  tts: 'text_to_speech',\n  transcription: 'transcription',\n}\n\nexport interface OtelMiddlewareOptions {\n  /** OTel `Tracer` used to start root, iteration, and tool spans. */\n  tracer: Tracer\n  /**\n   * Optional OTel `Meter`. When provided, the middleware records\n   * `gen_ai.client.operation.duration` and `gen_ai.client.token.usage`\n   * histograms. Omit to disable metrics without disabling tracing.\n   */\n  meter?: Meter\n  /**\n   * When `true`, prompt and completion content is attached to iteration spans\n   * as `gen_ai.*.message` / `gen_ai.choice` events. Defaults to `false` so\n   * that PII never lands on a span by accident.\n   */\n  captureContent?: boolean\n  /**\n   * Invoked on every captured content string before it lands on a span.\n   * Return a redacted version. If this function throws, the middleware emits\n   * the literal sentinel `\"[redaction_failed]\"` instead of the original text\n   * — it never falls back to raw content.\n   */\n  redact?: (text: string) => string\n  /**\n   * Maximum characters kept in the per-iteration assistant text buffer used\n   * to emit `gen_ai.choice` events. Extra characters are truncated with a\n   * trailing `\"…\"` marker. Defaults to 100 000. Set to `0` to disable the\n   * cap. Exporters typically truncate long attribute values anyway.\n   */\n  maxContentLength?: number\n  /** Override the default span name for each `kind`. */\n  spanNameFormatter?: (info: OtelSpanInfo) => string\n  /** Add extra attributes to each span. */\n  attributeEnricher?: (info: OtelSpanInfo) => Record<string, AttributeValue>\n  /** Mutate `SpanOptions` immediately before `tracer.startSpan(...)`. */\n  onBeforeSpanStart?: (info: OtelSpanInfo, options: SpanOptions) => SpanOptions\n  /** Fires just before every `span.end()`. */\n  onSpanEnd?: (info: OtelSpanInfo, span: Span) => void\n}\n\ninterface RequestState {\n  rootSpan: Span\n  currentIterationSpan: Span | null\n  toolSpans: Map<string, { span: Span; toolName: string }>\n  iterationCount: number\n  assistantTextBuffer: string\n  assistantTextBufferTruncated: boolean\n  startTime: number\n  /**\n   * Finish reason from the most recent `RUN_FINISHED` chunk. Captured in\n   * `onChunk` so `onFinish` can stamp it on the root span without reading it\n   * from the (base-shaped) finish info, which doesn't carry it.\n   */\n  lastFinishReason: string | null\n}\n\nconst stateByCtx = new WeakMap<ChatMiddlewareContext, RequestState>()\n\nconst DEFAULT_MAX_CONTENT_LENGTH = 100_000\nconst REDACTION_FAILED_SENTINEL = '[redaction_failed]'\n\nfunction serializeContent(content: unknown): string {\n  if (typeof content === 'string') return content\n  if (!Array.isArray(content)) return ''\n  const parts: Array<string> = []\n  for (const part of content) {\n    if (!part || typeof part !== 'object') continue\n    const type = (part as { type?: string }).type\n    switch (type) {\n      case 'text':\n        parts.push(\n          (\n            (part as { text?: string }).text ??\n            (part as { content?: string }).content ??\n            ''\n          ).toString(),\n        )\n        break\n      case 'image':\n        parts.push('[image]')\n        break\n      case 'audio':\n        parts.push('[audio]')\n        break\n      case 'video':\n        parts.push('[video]')\n        break\n      case 'document':\n        parts.push('[document]')\n        break\n      case undefined:\n        parts.push('[unknown]')\n        break\n      default:\n        parts.push(`[${type}]`)\n    }\n  }\n  return parts.join(' ')\n}\n\nfunction messageEventName(role: string): string {\n  switch (role) {\n    case 'user':\n      return 'gen_ai.user.message'\n    case 'assistant':\n      return 'gen_ai.assistant.message'\n    case 'tool':\n      return 'gen_ai.tool.message'\n    case 'system':\n      return 'gen_ai.system.message'\n    default:\n      return `gen_ai.${role}.message`\n  }\n}\n\nfunction safeCall<T>(label: string, fn: () => T): T | undefined {\n  try {\n    return fn()\n  } catch (err) {\n    // Keep middleware non-fatal, but surface callback failures so that broken\n    // extension points (attributeEnricher, spanNameFormatter, onSpanEnd, ...)\n    // are observable. Matches the guarantee documented in docs/advanced/otel.md.\n    console.warn(`[otelMiddleware] ${label} failed`, err)\n    return undefined\n  }\n}\n\nexport function otelMiddleware(\n  options: OtelMiddlewareOptions,\n): GenerationMiddleware & ChatMiddleware {\n  const {\n    tracer,\n    meter,\n    captureContent = false,\n    redact = (s) => s,\n    maxContentLength = DEFAULT_MAX_CONTENT_LENGTH,\n    spanNameFormatter,\n    attributeEnricher,\n    onBeforeSpanStart,\n    onSpanEnd,\n  } = options\n\n  const durationHistogram = meter?.createHistogram(\n    'gen_ai.client.operation.duration',\n    {\n      description: 'GenAI client operation duration',\n      unit: 's',\n    },\n  )\n  const tokenHistogram = meter?.createHistogram('gen_ai.client.token.usage', {\n    description: 'GenAI client token usage',\n    unit: '{token}',\n  })\n\n  // Redact user content, failing closed to a sentinel string instead of ever\n  // letting raw text through. Callers that pass `captureContent: true` with a\n  // third-party PII redactor depend on this invariant.\n  const redactContent = (text: string): string => {\n    try {\n      return redact(text)\n    } catch (err) {\n      console.warn('[otelMiddleware] otel.redact failed', err)\n      return REDACTION_FAILED_SENTINEL\n    }\n  }\n\n  const appendAssistantText = (state: RequestState, delta: string): void => {\n    if (maxContentLength > 0) {\n      if (state.assistantTextBufferTruncated) return\n      const remaining = maxContentLength - state.assistantTextBuffer.length\n      if (remaining <= 0) {\n        state.assistantTextBufferTruncated = true\n        state.assistantTextBuffer += '…'\n        return\n      }\n      if (delta.length > remaining) {\n        state.assistantTextBuffer += delta.slice(0, remaining) + '…'\n        state.assistantTextBufferTruncated = true\n        return\n      }\n    }\n    state.assistantTextBuffer += delta\n  }\n\n  const closeIterationSpan = (\n    state: RequestState,\n    ctx: ChatMiddlewareContext,\n  ): void => {\n    if (!state.currentIterationSpan) return\n    const span = state.currentIterationSpan\n    const iteration = state.iterationCount - 1\n    safeCall('otel.onSpanEnd', () =>\n      onSpanEnd?.({ kind: 'iteration', ctx, iteration }, span),\n    )\n    span.end()\n    state.currentIterationSpan = null\n  }\n\n  // --- Media activities -----------------------------------------------------\n  // Media calls (image/video/audio/tts/transcription) are single request →\n  // response, so they get exactly one CLIENT span — opened in `onStart` and\n  // closed in the terminal hook. Keyed by the per-call context object, which is\n  // distinct from the chat state map above so the two paths never collide.\n  const mediaSpans = new WeakMap<GenerationMiddlewareContext, Span>()\n\n  const recordMediaDuration = (\n    ctx: GenerationMiddlewareContext,\n    durationMs: number,\n    errorType?: string,\n  ): void => {\n    if (!durationHistogram) return\n    durationHistogram.record(durationMs / 1000, {\n      'gen_ai.system': ctx.provider,\n      'gen_ai.operation.name': OPERATION_NAME[ctx.activity],\n      'gen_ai.request.model': ctx.model,\n      ...(errorType ? { 'error.type': errorType } : {}),\n    })\n  }\n\n  const startMediaSpan = (ctx: GenerationMiddlewareContext): void => {\n    safeCall('otel.onStart', () => {\n      const operationName = OPERATION_NAME[ctx.activity]\n      const info: OtelSpanInfo<'generation'> = { kind: 'generation', ctx }\n      const name =\n        safeCall('otel.spanNameFormatter', () => spanNameFormatter?.(info)) ??\n        `${operationName} ${ctx.model}`\n      const baseOptions: SpanOptions = {\n        kind: SpanKind.CLIENT,\n        attributes: {\n          'gen_ai.system': ctx.provider,\n          'gen_ai.operation.name': operationName,\n          'gen_ai.request.model': ctx.model,\n        },\n      }\n      const spanOptions =\n        safeCall('otel.onBeforeSpanStart', () =>\n          onBeforeSpanStart?.(info, baseOptions),\n        ) ?? baseOptions\n      const span = tracer.startSpan(name, spanOptions)\n      const enriched = safeCall('otel.attributeEnricher', () =>\n        attributeEnricher?.(info),\n      )\n      if (enriched) span.setAttributes(enriched)\n      mediaSpans.set(ctx, span)\n    })\n  }\n\n  const endMediaSpan = (\n    ctx: GenerationMiddlewareContext,\n    finalize: (span: Span) => void,\n  ): void => {\n    const span = mediaSpans.get(ctx)\n    mediaSpans.delete(ctx)\n    if (!span) return\n    finalize(span)\n    safeCall('otel.onSpanEnd', () =>\n      onSpanEnd?.({ kind: 'generation', ctx }, span),\n    )\n    span.end()\n  }\n\n  return {\n    name: 'otel',\n\n    onStart(ctx) {\n      // Media activities get one CLIENT span; chat builds the root/iteration\n      // tree below. The cast is sound: the chat runner only ever passes a\n      // ChatMiddlewareContext, which `activity: 'chat'` narrows to at runtime.\n      if (ctx.activity !== 'chat') {\n        startMediaSpan(ctx)\n        return\n      }\n      const chatCtx = ctx as ChatMiddlewareContext\n      safeCall('otel.onStart', () => {\n        const info: OtelSpanInfo<'chat'> = { kind: 'chat', ctx: chatCtx }\n        const name =\n          safeCall('otel.spanNameFormatter', () => spanNameFormatter?.(info)) ??\n          `chat ${chatCtx.model}`\n        const baseOptions: SpanOptions = {\n          kind: SpanKind.INTERNAL,\n          attributes: {\n            'gen_ai.system': chatCtx.provider,\n            'gen_ai.request.model': chatCtx.model,\n            // NOTE: `gen_ai.operation.name` is deliberately NOT set on the\n            // root span. The root represents a `chat()` invocation that may\n            // span multiple model calls; only iteration spans correspond to\n            // a single chat operation. Backends that map `operation.name=chat`\n            // to a \"generation\" event (e.g. PostHog LLM Analytics) would\n            // otherwise emit a duplicate generation for the wrapper span.\n          },\n        }\n        const spanOptions =\n          safeCall('otel.onBeforeSpanStart', () =>\n            onBeforeSpanStart?.(info, baseOptions),\n          ) ?? baseOptions\n        const rootSpan = tracer.startSpan(name, spanOptions)\n\n        const enriched = safeCall('otel.attributeEnricher', () =>\n          attributeEnricher?.(info),\n        )\n        if (enriched) rootSpan.setAttributes(enriched)\n\n        stateByCtx.set(chatCtx, {\n          rootSpan,\n          currentIterationSpan: null,\n          toolSpans: new Map(),\n          iterationCount: 0,\n          assistantTextBuffer: '',\n          assistantTextBufferTruncated: false,\n          startTime: Date.now(),\n          lastFinishReason: null,\n        })\n      })\n    },\n\n    onConfig(ctx, config) {\n      if (ctx.phase !== 'beforeModel') return\n      safeCall('otel.onConfig', () => {\n        const state = stateByCtx.get(ctx)\n        if (!state) return\n\n        // The previous iteration's span stays open through tool execution and\n        // onUsage so that tool spans nest under it and token attributes land\n        // on it. Close it here, just before opening the next iteration.\n        closeIterationSpan(state, ctx)\n\n        const info: OtelSpanInfo<'iteration'> = {\n          kind: 'iteration',\n          ctx,\n          iteration: ctx.iteration,\n        }\n        const name =\n          safeCall('otel.spanNameFormatter', () => spanNameFormatter?.(info)) ??\n          `chat ${ctx.model} #${ctx.iteration}`\n\n        const baseAttrs: Record<string, AttributeValue> = {\n          'gen_ai.system': ctx.provider,\n          'gen_ai.operation.name': 'chat',\n          'gen_ai.request.model': ctx.model,\n          'tanstack.ai.iteration': ctx.iteration,\n        }\n        // Sampling options now live in provider-native `modelOptions`, and\n        // providers spell them differently (e.g. `max_output_tokens`,\n        // `max_completion_tokens`, `maxOutputTokens`, `num_predict`). Read the\n        // first numeric value among the known spellings — including Ollama's\n        // nested `options` — so gen_ai attributes populate across providers.\n        const sampling = config.modelOptions ?? {}\n        const nestedOptions =\n          sampling['options'] && typeof sampling['options'] === 'object'\n            ? (sampling['options'] as Record<string, unknown>)\n            : undefined\n        const samplingTemperature = firstNumber(\n          sampling['temperature'],\n          nestedOptions?.['temperature'],\n        )\n        const samplingTopP = firstNumber(\n          sampling['top_p'],\n          sampling['topP'],\n          nestedOptions?.['top_p'],\n        )\n        // Spellings come from the shared `MAX_TOKENS_KEYS` table so this stays\n        // in lockstep with the summarize wrapper's caller-limit detection.\n        const samplingMaxTokens = firstNumber(\n          ...MAX_TOKENS_KEYS.map((k) => sampling[k]),\n          nestedOptions?.[NESTED_MAX_TOKENS_KEY],\n        )\n        if (samplingTemperature !== undefined)\n          baseAttrs['gen_ai.request.temperature'] = samplingTemperature\n        if (samplingTopP !== undefined)\n          baseAttrs['gen_ai.request.top_p'] = samplingTopP\n        if (samplingMaxTokens !== undefined)\n          baseAttrs['gen_ai.request.max_tokens'] = samplingMaxTokens\n\n        const baseOptions: SpanOptions = {\n          kind: SpanKind.CLIENT,\n          attributes: baseAttrs,\n        }\n        const spanOptions =\n          safeCall('otel.onBeforeSpanStart', () =>\n            onBeforeSpanStart?.(info, baseOptions),\n          ) ?? baseOptions\n\n        const parentCtx = otelTrace.setSpan(\n          otelContext.active(),\n          state.rootSpan,\n        )\n        let iterSpan!: Span\n        otelContext.with(parentCtx, () => {\n          // Pass the parent context explicitly as the 3rd arg — this is a\n          // real-OTel-compatible way to ensure the span is parented to\n          // `rootSpan` even when the host app has not registered a context\n          // manager (e.g. in tests or minimal setups).\n          iterSpan = tracer.startSpan(name, spanOptions, parentCtx)\n        })\n\n        const enriched = safeCall('otel.attributeEnricher', () =>\n          attributeEnricher?.(info),\n        )\n        if (enriched) iterSpan.setAttributes(enriched)\n\n        state.currentIterationSpan = iterSpan\n        state.assistantTextBuffer = ''\n        state.assistantTextBufferTruncated = false\n\n        if (captureContent) {\n          const systemPromptContents = config.systemPrompts.map((p) =>\n            typeof p === 'string' ? p : p.content,\n          )\n          // Anthropic prompt-caching users need to know which prompt carried\n          // `cache_control`: it's the one attribute that explains cache\n          // hit/miss in observability. Serialise per-prompt metadata as a\n          // single JSON span attribute so backends that don't understand\n          // GenAI events can still surface it. Kept off span events to\n          // avoid breaking the one-event-per-message GenAI semconv contract.\n          const systemPromptMetadata = config.systemPrompts.map((p) =>\n            typeof p === 'string' || p.metadata === undefined\n              ? null\n              : p.metadata,\n          )\n          if (systemPromptMetadata.some((m) => m !== null)) {\n            iterSpan.setAttribute(\n              'tanstack.ai.system_prompt.metadata',\n              JSON.stringify(systemPromptMetadata),\n            )\n          }\n          // Span events follow the original GenAI semconv (one event per\n          // message). Backends that read events get content this way.\n          for (const sys of systemPromptContents) {\n            iterSpan.addEvent('gen_ai.system.message', {\n              content: redactContent(sys),\n            })\n          }\n          for (const m of config.messages) {\n            const body = serializeContent(m.content)\n            if (body.length === 0) continue\n            iterSpan.addEvent(messageEventName(m.role), {\n              content: redactContent(body),\n            })\n          }\n\n          // Also emit the current GenAI-semconv attribute form\n          // (`gen_ai.input.messages`) — backends like PostHog read prompt\n          // content from this attribute, not from span events.\n          const inputMessages: Array<{ role: string; content: string }> = []\n          for (const sys of systemPromptContents) {\n            inputMessages.push({\n              role: 'system',\n              content: redactContent(sys),\n            })\n          }\n          for (const m of config.messages) {\n            const body = serializeContent(m.content)\n            if (body.length === 0) continue\n            inputMessages.push({\n              role: m.role,\n              content: redactContent(body),\n            })\n          }\n          if (inputMessages.length > 0) {\n            const inputJson = JSON.stringify(inputMessages)\n            // Current OTel GenAI semconv — Sentry / PostHog / Datadog read\n            // prompt content from this attribute.\n            iterSpan.setAttribute('gen_ai.input.messages', inputJson)\n            // Langfuse-native attribute. Highest priority in Langfuse's OTLP\n            // ingestion (checked before events and gen_ai.input.messages) so\n            // the Input panel populates reliably. Harmless to other backends —\n            // the attribute is namespaced and unrecognised keys are ignored.\n            iterSpan.setAttribute('langfuse.observation.input', inputJson)\n\n            // Mirror the first iteration's input onto the root span and at\n            // trace level so Langfuse fills Input on the trace card and the\n            // chat-level observation. Later iterations append tool-call /\n            // assistant messages that are useful per-iteration but noise at\n            // the chat / trace level.\n            if (state.iterationCount === 0) {\n              state.rootSpan.setAttribute(\n                'langfuse.observation.input',\n                inputJson,\n              )\n              state.rootSpan.setAttribute('langfuse.trace.input', inputJson)\n            }\n          }\n        }\n\n        state.iterationCount += 1\n      })\n      return undefined\n    },\n\n    onChunk(ctx, chunk) {\n      safeCall('otel.onChunk', () => {\n        const state = stateByCtx.get(ctx)\n        if (!state) return\n\n        if (captureContent && chunk.type === 'TEXT_MESSAGE_CONTENT') {\n          appendAssistantText(state, chunk.delta)\n        }\n\n        if (chunk.type !== 'RUN_FINISHED') return\n        // Capture for the root-span finish_reasons attribute set in onFinish,\n        // which receives base-shaped info without a finishReason field.\n        if (chunk.finishReason) state.lastFinishReason = chunk.finishReason\n        const span = state.currentIterationSpan\n        if (!span) return\n\n        if (chunk.finishReason) {\n          span.setAttribute('gen_ai.response.finish_reasons', [\n            chunk.finishReason,\n          ])\n        }\n        if (chunk.model) span.setAttribute('gen_ai.response.model', chunk.model)\n\n        // Set usage attributes on the iteration span directly from the chunk\n        // so they're available before `onUsage` fires. Histogram recording is\n        // deliberately NOT done here — the chat runner always invokes\n        // `runOnUsage` when `chunk.usage` is present, and `onUsage` is the\n        // canonical place for the metric. Recording in both would double-count.\n        if (chunk.usage) {\n          span.setAttributes(usageAttributes(chunk.usage))\n        }\n\n        if (captureContent && state.assistantTextBuffer.length > 0) {\n          const completion = redactContent(state.assistantTextBuffer)\n          const outputJson = JSON.stringify([\n            { role: 'assistant', content: completion },\n          ])\n          // Event form (older semconv) — kept for backends that consume it.\n          span.addEvent('gen_ai.choice', { content: completion })\n          // Attribute form (current semconv) — required by backends like\n          // PostHog that read completion content from `gen_ai.output.messages`.\n          span.setAttribute('gen_ai.output.messages', outputJson)\n          // Langfuse-native attribute (highest priority in Langfuse mapping).\n          span.setAttribute('langfuse.observation.output', outputJson)\n          // Mirror to the root span and trace card. Each iteration overwrites,\n          // so the final iteration's completion lands on the root — which is\n          // the final answer the user saw, not an intermediate tool-call turn.\n          state.rootSpan.setAttribute('langfuse.observation.output', outputJson)\n          state.rootSpan.setAttribute('langfuse.trace.output', outputJson)\n          state.assistantTextBuffer = ''\n          state.assistantTextBufferTruncated = false\n        }\n\n        // Intentionally leave the iteration span open: tool spans started\n        // after `RUN_FINISHED` (tool_calls finishReason) must nest under it,\n        // and `onUsage` may still fire. The span is closed in `onConfig` when\n        // the next iteration starts, or in `onFinish` / `onError` / `onAbort`.\n      })\n      return undefined\n    },\n\n    onUsage(ctx, usage) {\n      if (ctx.activity !== 'chat') {\n        // Media: stamp usage on the single span. No token histogram — media\n        // unit billing lands as span attributes via usageAttributes, matching\n        // prior media behavior and avoiding chat-shaped token metrics.\n        safeCall('otel.onUsage', () => {\n          const span = mediaSpans.get(ctx)\n          if (span) span.setAttributes(usageAttributes(usage))\n        })\n        return\n      }\n      const chatCtx = ctx as ChatMiddlewareContext\n      safeCall('otel.onUsage', () => {\n        const state = stateByCtx.get(chatCtx)\n        if (!state) return\n\n        // Always record the token histogram — metrics don't depend on having\n        // an iteration span, and skipping here would drop metric data if an\n        // adapter emits `onUsage` outside the iteration window.\n        if (tokenHistogram) {\n          const metricAttrs = {\n            'gen_ai.system': chatCtx.provider,\n            'gen_ai.operation.name': 'chat',\n            'gen_ai.request.model': chatCtx.model,\n          }\n          tokenHistogram.record(usage.promptTokens, {\n            ...metricAttrs,\n            'gen_ai.token.type': 'input',\n          })\n          tokenHistogram.record(usage.completionTokens, {\n            ...metricAttrs,\n            'gen_ai.token.type': 'output',\n          })\n        }\n\n        const span = state.currentIterationSpan ?? state.rootSpan\n        span.setAttributes(usageAttributes(usage))\n      })\n    },\n\n    onBeforeToolCall(ctx, hookCtx) {\n      safeCall('otel.onBeforeToolCall', () => {\n        const state = stateByCtx.get(ctx)\n        if (!state) return\n        const parent = state.currentIterationSpan ?? state.rootSpan\n\n        const info: OtelSpanInfo<'tool'> = {\n          kind: 'tool',\n          ctx,\n          toolName: hookCtx.toolName,\n          toolCallId: hookCtx.toolCallId,\n          iteration: state.iterationCount - 1,\n        }\n        const name =\n          safeCall('otel.spanNameFormatter', () => spanNameFormatter?.(info)) ??\n          `execute_tool ${hookCtx.toolName}`\n\n        const baseAttrs: Record<string, AttributeValue> = {\n          'gen_ai.tool.name': hookCtx.toolName,\n          'gen_ai.tool.call.id': hookCtx.toolCallId,\n          'gen_ai.tool.type': 'function',\n        }\n        const baseOptions: SpanOptions = {\n          kind: SpanKind.INTERNAL,\n          attributes: baseAttrs,\n        }\n        const spanOptions =\n          safeCall('otel.onBeforeSpanStart', () =>\n            onBeforeSpanStart?.(info, baseOptions),\n          ) ?? baseOptions\n\n        const parentCtx = otelTrace.setSpan(otelContext.active(), parent)\n        let toolSpan!: Span\n        otelContext.with(parentCtx, () => {\n          toolSpan = tracer.startSpan(name, spanOptions, parentCtx)\n        })\n\n        const enriched = safeCall('otel.attributeEnricher', () =>\n          attributeEnricher?.(info),\n        )\n        if (enriched) toolSpan.setAttributes(enriched)\n\n        // Stamp the tool args onto the tool span so backends that render an\n        // input panel per span (e.g. PostHog) have something to show.\n        if (captureContent) {\n          const argsBody =\n            typeof hookCtx.args === 'string'\n              ? hookCtx.args\n              : (safeCall('otel.serializeToolArgs', () =>\n                  JSON.stringify(hookCtx.args ?? null),\n                ) ?? '[unserializable_tool_args]')\n          const redactedArgs = redactContent(argsBody)\n          const toolInputJson = JSON.stringify([\n            { role: 'tool', content: redactedArgs },\n          ])\n          toolSpan.setAttribute('gen_ai.input.messages', toolInputJson)\n          // Langfuse-native (highest priority in Langfuse mapping).\n          toolSpan.setAttribute('langfuse.observation.input', toolInputJson)\n        }\n\n        state.toolSpans.set(hookCtx.toolCallId, {\n          span: toolSpan,\n          toolName: hookCtx.toolName,\n        })\n      })\n      return undefined\n    },\n\n    onAfterToolCall(ctx, info) {\n      safeCall('otel.onAfterToolCall', () => {\n        const state = stateByCtx.get(ctx)\n        if (!state) return\n        const entry = state.toolSpans.get(info.toolCallId)\n        if (!entry) return\n        const { span: toolSpan } = entry\n\n        const outcome = info.ok ? 'success' : 'error'\n        toolSpan.setAttribute('tanstack.ai.tool.outcome', outcome)\n\n        if (!info.ok && info.error !== undefined) {\n          toolSpan.recordException(info.error as Exception)\n          const msg = errorMessage(info.error)\n          toolSpan.setStatus({\n            code: SpanStatusCode.ERROR,\n            ...(msg !== undefined && { message: msg }),\n          })\n        }\n\n        if (captureContent) {\n          // Serialization can throw on circular refs or `BigInt` values. If it\n          // does, fall back to a sentinel so the rest of this handler (span\n          // end, onSpanEnd, toolSpans cleanup) still runs — otherwise the tool\n          // span would dangle until the onFinish/onError sweep.\n          const body =\n            typeof info.result === 'string'\n              ? info.result\n              : (safeCall('otel.serializeToolResult', () =>\n                  JSON.stringify(info.result ?? null),\n                ) ?? '[unserializable_tool_result]')\n          const redactedBody = redactContent(body)\n          if (state.currentIterationSpan) {\n            state.currentIterationSpan.addEvent('gen_ai.tool.message', {\n              content: redactedBody,\n              tool_call_id: info.toolCallId,\n            })\n          }\n          // Output panel of the tool span itself — `gen_ai.output.messages` is\n          // what current GenAI semconv consumers (e.g. PostHog) read.\n          const toolOutputJson = JSON.stringify([\n            { role: 'tool', content: redactedBody },\n          ])\n          toolSpan.setAttribute('gen_ai.output.messages', toolOutputJson)\n          // Langfuse-native (highest priority in Langfuse mapping).\n          toolSpan.setAttribute('langfuse.observation.output', toolOutputJson)\n        }\n\n        safeCall('otel.onSpanEnd', () =>\n          onSpanEnd?.(\n            {\n              kind: 'tool',\n              ctx,\n              toolName: info.toolName,\n              toolCallId: info.toolCallId,\n              iteration: state.iterationCount - 1,\n            },\n            toolSpan,\n          ),\n        )\n        toolSpan.end()\n        state.toolSpans.delete(info.toolCallId)\n      })\n    },\n\n    onError(ctx, info) {\n      if (ctx.activity !== 'chat') {\n        safeCall('otel.onError', () => {\n          const message = errorMessage(info.error)\n          endMediaSpan(ctx, (span) => {\n            span.recordException(info.error as Exception)\n            span.setStatus({\n              code: SpanStatusCode.ERROR,\n              ...(message !== undefined ? { message } : {}),\n            })\n          })\n          recordMediaDuration(ctx, info.duration, errorTypeName(info.error))\n        })\n        return\n      }\n      const chatCtx = ctx as ChatMiddlewareContext\n      safeCall('otel.onError', () => {\n        const state = stateByCtx.get(chatCtx)\n        if (!state) return\n\n        const errType = errorTypeName(info.error)\n        const message = errorMessage(info.error)\n        const statusMessage =\n          message !== undefined ? { message } : ({} as const)\n        const exception = info.error as Exception\n\n        const iterationSpan = state.currentIterationSpan\n        if (iterationSpan) {\n          iterationSpan.recordException(exception)\n          iterationSpan.setStatus({\n            code: SpanStatusCode.ERROR,\n            ...statusMessage,\n          })\n          safeCall('otel.onSpanEnd', () =>\n            onSpanEnd?.(\n              {\n                kind: 'iteration',\n                ctx: chatCtx,\n                iteration: state.iterationCount - 1,\n              },\n              iterationSpan,\n            ),\n          )\n          iterationSpan.end()\n          state.currentIterationSpan = null\n        }\n\n        for (const [id, entry] of state.toolSpans) {\n          const { span, toolName } = entry\n          span.recordException(exception)\n          span.setStatus({ code: SpanStatusCode.ERROR, ...statusMessage })\n          safeCall('otel.onSpanEnd', () =>\n            onSpanEnd?.(\n              {\n                kind: 'tool',\n                ctx: chatCtx,\n                toolCallId: id,\n                toolName,\n                iteration: state.iterationCount - 1,\n              },\n              span,\n            ),\n          )\n          span.end()\n          state.toolSpans.delete(id)\n        }\n\n        state.rootSpan.recordException(exception)\n        state.rootSpan.setStatus({\n          code: SpanStatusCode.ERROR,\n          ...statusMessage,\n        })\n\n        if (durationHistogram) {\n          durationHistogram.record(info.duration / 1000, {\n            'gen_ai.system': chatCtx.provider,\n            'gen_ai.operation.name': 'chat',\n            'gen_ai.request.model': chatCtx.model,\n            'error.type': errType,\n          })\n        }\n\n        safeCall('otel.onSpanEnd', () =>\n          onSpanEnd?.({ kind: 'chat', ctx: chatCtx }, state.rootSpan),\n        )\n        state.rootSpan.end()\n        stateByCtx.delete(chatCtx)\n      })\n    },\n\n    onAbort(ctx, info) {\n      if (ctx.activity !== 'chat') {\n        // Media abandonment (e.g. a video stream dropped before completion).\n        safeCall('otel.onAbort', () => {\n          endMediaSpan(ctx, (span) => {\n            span.setAttribute('tanstack.ai.completion.reason', 'cancelled')\n            span.setStatus({\n              code: SpanStatusCode.ERROR,\n              message: info.reason ?? 'cancelled',\n            })\n          })\n          recordMediaDuration(ctx, info.duration, 'cancelled')\n        })\n        return\n      }\n      const chatCtx = ctx as ChatMiddlewareContext\n      safeCall('otel.onAbort', () => {\n        const state = stateByCtx.get(chatCtx)\n        if (!state) return\n\n        const closeCancelled = (span: Span): void => {\n          // `gen_ai.completion.reason` is not part of the GenAI semconv; use a\n          // TanStack-namespaced attribute so downstream exporters don't treat\n          // it as standard. The span status still carries the error code.\n          span.setAttribute('tanstack.ai.completion.reason', 'cancelled')\n          span.setStatus({ code: SpanStatusCode.ERROR, message: 'cancelled' })\n        }\n\n        const iterationSpan = state.currentIterationSpan\n        if (iterationSpan) {\n          closeCancelled(iterationSpan)\n          safeCall('otel.onSpanEnd', () =>\n            onSpanEnd?.(\n              {\n                kind: 'iteration',\n                ctx: chatCtx,\n                iteration: state.iterationCount - 1,\n              },\n              iterationSpan,\n            ),\n          )\n          iterationSpan.end()\n          state.currentIterationSpan = null\n        }\n        for (const [id, entry] of state.toolSpans) {\n          const { span, toolName } = entry\n          closeCancelled(span)\n          safeCall('otel.onSpanEnd', () =>\n            onSpanEnd?.(\n              {\n                kind: 'tool',\n                ctx: chatCtx,\n                toolCallId: id,\n                toolName,\n                iteration: state.iterationCount - 1,\n              },\n              span,\n            ),\n          )\n          span.end()\n          state.toolSpans.delete(id)\n        }\n        closeCancelled(state.rootSpan)\n\n        if (durationHistogram) {\n          durationHistogram.record(info.duration / 1000, {\n            'gen_ai.system': chatCtx.provider,\n            'gen_ai.operation.name': 'chat',\n            'gen_ai.request.model': chatCtx.model,\n            'error.type': 'cancelled',\n          })\n        }\n\n        safeCall('otel.onSpanEnd', () =>\n          onSpanEnd?.({ kind: 'chat', ctx: chatCtx }, state.rootSpan),\n        )\n        state.rootSpan.end()\n        stateByCtx.delete(chatCtx)\n      })\n    },\n\n    onFinish(ctx, info) {\n      if (ctx.activity !== 'chat') {\n        safeCall('otel.onFinish', () => {\n          endMediaSpan(ctx, (span) => {\n            if (info.usage) span.setAttributes(usageAttributes(info.usage))\n          })\n          recordMediaDuration(ctx, info.duration)\n        })\n        return\n      }\n      const chatCtx = ctx as ChatMiddlewareContext\n      safeCall('otel.onFinish', () => {\n        const state = stateByCtx.get(chatCtx)\n        if (!state) return\n\n        // Close any tool spans that never received `onAfterToolCall` (adapter\n        // quirk). Done before the iteration span so the hierarchy is closed\n        // in depth-first order.\n        for (const [id, entry] of state.toolSpans) {\n          const { span, toolName } = entry\n          span.setAttribute('tanstack.ai.tool.outcome', 'unknown')\n          safeCall('otel.onSpanEnd', () =>\n            onSpanEnd?.(\n              {\n                kind: 'tool',\n                ctx: chatCtx,\n                toolCallId: id,\n                toolName,\n                iteration: state.iterationCount - 1,\n              },\n              span,\n            ),\n          )\n          span.end()\n          state.toolSpans.delete(id)\n        }\n\n        // The final iteration's span is still open because we keep it open\n        // through tool execution and `onUsage`. Close it now.\n        closeIterationSpan(state, chatCtx)\n\n        if (durationHistogram) {\n          durationHistogram.record(info.duration / 1000, {\n            'gen_ai.system': chatCtx.provider,\n            'gen_ai.operation.name': 'chat',\n            'gen_ai.request.model': chatCtx.model,\n          })\n        }\n\n        if (info.usage) {\n          state.rootSpan.setAttributes(usageAttributes(info.usage))\n        }\n        if (state.lastFinishReason) {\n          state.rootSpan.setAttribute('gen_ai.response.finish_reasons', [\n            state.lastFinishReason,\n          ])\n        }\n        state.rootSpan.setAttribute(\n          'tanstack.ai.iterations',\n          state.iterationCount,\n        )\n\n        safeCall('otel.onSpanEnd', () =>\n          onSpanEnd?.({ kind: 'chat', ctx: chatCtx }, state.rootSpan),\n        )\n        state.rootSpan.end()\n        stateByCtx.delete(chatCtx)\n      })\n    },\n  }\n}\n"],"names":["otelTrace","otelContext"],"mappings":";;;;;AA8EA,MAAM,iBAAqD;AAAA,EACzD,MAAM;AAAA,EACN,OAAO;AAAA,EACP,OAAO;AAAA,EACP,OAAO;AAAA,EACP,KAAK;AAAA,EACL,eAAe;AACjB;AAyDA,MAAM,iCAAiB,QAAA;AAEvB,MAAM,6BAA6B;AACnC,MAAM,4BAA4B;AAElC,SAAS,iBAAiB,SAA0B;AAClD,MAAI,OAAO,YAAY,SAAU,QAAO;AACxC,MAAI,CAAC,MAAM,QAAQ,OAAO,EAAG,QAAO;AACpC,QAAM,QAAuB,CAAA;AAC7B,aAAW,QAAQ,SAAS;AAC1B,QAAI,CAAC,QAAQ,OAAO,SAAS,SAAU;AACvC,UAAM,OAAQ,KAA2B;AACzC,YAAQ,MAAA;AAAA,MACN,KAAK;AACH,cAAM;AAAA,WAED,KAA2B,QAC3B,KAA8B,WAC/B,IACA,SAAA;AAAA,QAAS;AAEb;AAAA,MACF,KAAK;AACH,cAAM,KAAK,SAAS;AACpB;AAAA,MACF,KAAK;AACH,cAAM,KAAK,SAAS;AACpB;AAAA,MACF,KAAK;AACH,cAAM,KAAK,SAAS;AACpB;AAAA,MACF,KAAK;AACH,cAAM,KAAK,YAAY;AACvB;AAAA,MACF,KAAK;AACH,cAAM,KAAK,WAAW;AACtB;AAAA,MACF;AACE,cAAM,KAAK,IAAI,IAAI,GAAG;AAAA,IAAA;AAAA,EAE5B;AACA,SAAO,MAAM,KAAK,GAAG;AACvB;AAEA,SAAS,iBAAiB,MAAsB;AAC9C,UAAQ,MAAA;AAAA,IACN,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT;AACE,aAAO,UAAU,IAAI;AAAA,EAAA;AAE3B;AAEA,SAAS,SAAY,OAAe,IAA4B;AAC9D,MAAI;AACF,WAAO,GAAA;AAAA,EACT,SAAS,KAAK;AAIZ,YAAQ,KAAK,oBAAoB,KAAK,WAAW,GAAG;AACpD,WAAO;AAAA,EACT;AACF;AAEO,SAAS,eACd,SACuC;AACvC,QAAM;AAAA,IACJ;AAAA,IACA;AAAA,IACA,iBAAiB;AAAA,IACjB,SAAS,CAAC,MAAM;AAAA,IAChB,mBAAmB;AAAA,IACnB;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EAAA,IACE;AAEJ,QAAM,oBAAoB,OAAO;AAAA,IAC/B;AAAA,IACA;AAAA,MACE,aAAa;AAAA,MACb,MAAM;AAAA,IAAA;AAAA,EACR;AAEF,QAAM,iBAAiB,OAAO,gBAAgB,6BAA6B;AAAA,IACzE,aAAa;AAAA,IACb,MAAM;AAAA,EAAA,CACP;AAKD,QAAM,gBAAgB,CAAC,SAAyB;AAC9C,QAAI;AACF,aAAO,OAAO,IAAI;AAAA,IACpB,SAAS,KAAK;AACZ,cAAQ,KAAK,uCAAuC,GAAG;AACvD,aAAO;AAAA,IACT;AAAA,EACF;AAEA,QAAM,sBAAsB,CAAC,OAAqB,UAAwB;AACxE,QAAI,mBAAmB,GAAG;AACxB,UAAI,MAAM,6BAA8B;AACxC,YAAM,YAAY,mBAAmB,MAAM,oBAAoB;AAC/D,UAAI,aAAa,GAAG;AAClB,cAAM,+BAA+B;AACrC,cAAM,uBAAuB;AAC7B;AAAA,MACF;AACA,UAAI,MAAM,SAAS,WAAW;AAC5B,cAAM,uBAAuB,MAAM,MAAM,GAAG,SAAS,IAAI;AACzD,cAAM,+BAA+B;AACrC;AAAA,MACF;AAAA,IACF;AACA,UAAM,uBAAuB;AAAA,EAC/B;AAEA,QAAM,qBAAqB,CACzB,OACA,QACS;AACT,QAAI,CAAC,MAAM,qBAAsB;AACjC,UAAM,OAAO,MAAM;AACnB,UAAM,YAAY,MAAM,iBAAiB;AACzC;AAAA,MAAS;AAAA,MAAkB,MACzB,YAAY,EAAE,MAAM,aAAa,KAAK,UAAA,GAAa,IAAI;AAAA,IAAA;AAEzD,SAAK,IAAA;AACL,UAAM,uBAAuB;AAAA,EAC/B;AAOA,QAAM,iCAAiB,QAAA;AAEvB,QAAM,sBAAsB,CAC1B,KACA,YACA,cACS;AACT,QAAI,CAAC,kBAAmB;AACxB,sBAAkB,OAAO,aAAa,KAAM;AAAA,MAC1C,iBAAiB,IAAI;AAAA,MACrB,yBAAyB,eAAe,IAAI,QAAQ;AAAA,MACpD,wBAAwB,IAAI;AAAA,MAC5B,GAAI,YAAY,EAAE,cAAc,cAAc,CAAA;AAAA,IAAC,CAChD;AAAA,EACH;AAEA,QAAM,iBAAiB,CAAC,QAA2C;AACjE,aAAS,gBAAgB,MAAM;AAC7B,YAAM,gBAAgB,eAAe,IAAI,QAAQ;AACjD,YAAM,OAAmC,EAAE,MAAM,cAAc,IAAA;AAC/D,YAAM,OACJ,SAAS,0BAA0B,MAAM,oBAAoB,IAAI,CAAC,KAClE,GAAG,aAAa,IAAI,IAAI,KAAK;AAC/B,YAAM,cAA2B;AAAA,QAC/B,MAAM,SAAS;AAAA,QACf,YAAY;AAAA,UACV,iBAAiB,IAAI;AAAA,UACrB,yBAAyB;AAAA,UACzB,wBAAwB,IAAI;AAAA,QAAA;AAAA,MAC9B;AAEF,YAAM,cACJ;AAAA,QAAS;AAAA,QAA0B,MACjC,oBAAoB,MAAM,WAAW;AAAA,MAAA,KAClC;AACP,YAAM,OAAO,OAAO,UAAU,MAAM,WAAW;AAC/C,YAAM,WAAW;AAAA,QAAS;AAAA,QAA0B,MAClD,oBAAoB,IAAI;AAAA,MAAA;AAE1B,UAAI,SAAU,MAAK,cAAc,QAAQ;AACzC,iBAAW,IAAI,KAAK,IAAI;AAAA,IAC1B,CAAC;AAAA,EACH;AAEA,QAAM,eAAe,CACnB,KACA,aACS;AACT,UAAM,OAAO,WAAW,IAAI,GAAG;AAC/B,eAAW,OAAO,GAAG;AACrB,QAAI,CAAC,KAAM;AACX,aAAS,IAAI;AACb;AAAA,MAAS;AAAA,MAAkB,MACzB,YAAY,EAAE,MAAM,cAAc,IAAA,GAAO,IAAI;AAAA,IAAA;AAE/C,SAAK,IAAA;AAAA,EACP;AAEA,SAAO;AAAA,IACL,MAAM;AAAA,IAEN,QAAQ,KAAK;AAIX,UAAI,IAAI,aAAa,QAAQ;AAC3B,uBAAe,GAAG;AAClB;AAAA,MACF;AACA,YAAM,UAAU;AAChB,eAAS,gBAAgB,MAAM;AAC7B,cAAM,OAA6B,EAAE,MAAM,QAAQ,KAAK,QAAA;AACxD,cAAM,OACJ,SAAS,0BAA0B,MAAM,oBAAoB,IAAI,CAAC,KAClE,QAAQ,QAAQ,KAAK;AACvB,cAAM,cAA2B;AAAA,UAC/B,MAAM,SAAS;AAAA,UACf,YAAY;AAAA,YACV,iBAAiB,QAAQ;AAAA,YACzB,wBAAwB,QAAQ;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,UAAA;AAAA,QAOlC;AAEF,cAAM,cACJ;AAAA,UAAS;AAAA,UAA0B,MACjC,oBAAoB,MAAM,WAAW;AAAA,QAAA,KAClC;AACP,cAAM,WAAW,OAAO,UAAU,MAAM,WAAW;AAEnD,cAAM,WAAW;AAAA,UAAS;AAAA,UAA0B,MAClD,oBAAoB,IAAI;AAAA,QAAA;AAE1B,YAAI,SAAU,UAAS,cAAc,QAAQ;AAE7C,mBAAW,IAAI,SAAS;AAAA,UACtB;AAAA,UACA,sBAAsB;AAAA,UACtB,+BAAe,IAAA;AAAA,UACf,gBAAgB;AAAA,UAChB,qBAAqB;AAAA,UACrB,8BAA8B;AAAA,UAC9B,WAAW,KAAK,IAAA;AAAA,UAChB,kBAAkB;AAAA,QAAA,CACnB;AAAA,MACH,CAAC;AAAA,IACH;AAAA,IAEA,SAAS,KAAK,QAAQ;AACpB,UAAI,IAAI,UAAU,cAAe;AACjC,eAAS,iBAAiB,MAAM;AAC9B,cAAM,QAAQ,WAAW,IAAI,GAAG;AAChC,YAAI,CAAC,MAAO;AAKZ,2BAAmB,OAAO,GAAG;AAE7B,cAAM,OAAkC;AAAA,UACtC,MAAM;AAAA,UACN;AAAA,UACA,WAAW,IAAI;AAAA,QAAA;AAEjB,cAAM,OACJ,SAAS,0BAA0B,MAAM,oBAAoB,IAAI,CAAC,KAClE,QAAQ,IAAI,KAAK,KAAK,IAAI,SAAS;AAErC,cAAM,YAA4C;AAAA,UAChD,iBAAiB,IAAI;AAAA,UACrB,yBAAyB;AAAA,UACzB,wBAAwB,IAAI;AAAA,UAC5B,yBAAyB,IAAI;AAAA,QAAA;AAO/B,cAAM,WAAW,OAAO,gBAAgB,CAAA;AACxC,cAAM,gBACJ,SAAS,SAAS,KAAK,OAAO,SAAS,SAAS,MAAM,WACjD,SAAS,SAAS,IACnB;AACN,cAAM,sBAAsB;AAAA,UAC1B,SAAS,aAAa;AAAA,UACtB,gBAAgB,aAAa;AAAA,QAAA;AAE/B,cAAM,eAAe;AAAA,UACnB,SAAS,OAAO;AAAA,UAChB,SAAS,MAAM;AAAA,UACf,gBAAgB,OAAO;AAAA,QAAA;AAIzB,cAAM,oBAAoB;AAAA,UACxB,GAAG,gBAAgB,IAAI,CAAC,MAAM,SAAS,CAAC,CAAC;AAAA,UACzC,gBAAgB,qBAAqB;AAAA,QAAA;AAEvC,YAAI,wBAAwB;AAC1B,oBAAU,4BAA4B,IAAI;AAC5C,YAAI,iBAAiB;AACnB,oBAAU,sBAAsB,IAAI;AACtC,YAAI,sBAAsB;AACxB,oBAAU,2BAA2B,IAAI;AAE3C,cAAM,cAA2B;AAAA,UAC/B,MAAM,SAAS;AAAA,UACf,YAAY;AAAA,QAAA;AAEd,cAAM,cACJ;AAAA,UAAS;AAAA,UAA0B,MACjC,oBAAoB,MAAM,WAAW;AAAA,QAAA,KAClC;AAEP,cAAM,YAAYA,MAAU;AAAA,UAC1BC,QAAY,OAAA;AAAA,UACZ,MAAM;AAAA,QAAA;AAER,YAAI;AACJA,gBAAY,KAAK,WAAW,MAAM;AAKhC,qBAAW,OAAO,UAAU,MAAM,aAAa,SAAS;AAAA,QAC1D,CAAC;AAED,cAAM,WAAW;AAAA,UAAS;AAAA,UAA0B,MAClD,oBAAoB,IAAI;AAAA,QAAA;AAE1B,YAAI,SAAU,UAAS,cAAc,QAAQ;AAE7C,cAAM,uBAAuB;AAC7B,cAAM,sBAAsB;AAC5B,cAAM,+BAA+B;AAErC,YAAI,gBAAgB;AAClB,gBAAM,uBAAuB,OAAO,cAAc;AAAA,YAAI,CAAC,MACrD,OAAO,MAAM,WAAW,IAAI,EAAE;AAAA,UAAA;AAQhC,gBAAM,uBAAuB,OAAO,cAAc;AAAA,YAAI,CAAC,MACrD,OAAO,MAAM,YAAY,EAAE,aAAa,SACpC,OACA,EAAE;AAAA,UAAA;AAER,cAAI,qBAAqB,KAAK,CAAC,MAAM,MAAM,IAAI,GAAG;AAChD,qBAAS;AAAA,cACP;AAAA,cACA,KAAK,UAAU,oBAAoB;AAAA,YAAA;AAAA,UAEvC;AAGA,qBAAW,OAAO,sBAAsB;AACtC,qBAAS,SAAS,yBAAyB;AAAA,cACzC,SAAS,cAAc,GAAG;AAAA,YAAA,CAC3B;AAAA,UACH;AACA,qBAAW,KAAK,OAAO,UAAU;AAC/B,kBAAM,OAAO,iBAAiB,EAAE,OAAO;AACvC,gBAAI,KAAK,WAAW,EAAG;AACvB,qBAAS,SAAS,iBAAiB,EAAE,IAAI,GAAG;AAAA,cAC1C,SAAS,cAAc,IAAI;AAAA,YAAA,CAC5B;AAAA,UACH;AAKA,gBAAM,gBAA0D,CAAA;AAChE,qBAAW,OAAO,sBAAsB;AACtC,0BAAc,KAAK;AAAA,cACjB,MAAM;AAAA,cACN,SAAS,cAAc,GAAG;AAAA,YAAA,CAC3B;AAAA,UACH;AACA,qBAAW,KAAK,OAAO,UAAU;AAC/B,kBAAM,OAAO,iBAAiB,EAAE,OAAO;AACvC,gBAAI,KAAK,WAAW,EAAG;AACvB,0BAAc,KAAK;AAAA,cACjB,MAAM,EAAE;AAAA,cACR,SAAS,cAAc,IAAI;AAAA,YAAA,CAC5B;AAAA,UACH;AACA,cAAI,cAAc,SAAS,GAAG;AAC5B,kBAAM,YAAY,KAAK,UAAU,aAAa;AAG9C,qBAAS,aAAa,yBAAyB,SAAS;AAKxD,qBAAS,aAAa,8BAA8B,SAAS;AAO7D,gBAAI,MAAM,mBAAmB,GAAG;AAC9B,oBAAM,SAAS;AAAA,gBACb;AAAA,gBACA;AAAA,cAAA;AAEF,oBAAM,SAAS,aAAa,wBAAwB,SAAS;AAAA,YAC/D;AAAA,UACF;AAAA,QACF;AAEA,cAAM,kBAAkB;AAAA,MAC1B,CAAC;AACD,aAAO;AAAA,IACT;AAAA,IAEA,QAAQ,KAAK,OAAO;AAClB,eAAS,gBAAgB,MAAM;AAC7B,cAAM,QAAQ,WAAW,IAAI,GAAG;AAChC,YAAI,CAAC,MAAO;AAEZ,YAAI,kBAAkB,MAAM,SAAS,wBAAwB;AAC3D,8BAAoB,OAAO,MAAM,KAAK;AAAA,QACxC;AAEA,YAAI,MAAM,SAAS,eAAgB;AAGnC,YAAI,MAAM,aAAc,OAAM,mBAAmB,MAAM;AACvD,cAAM,OAAO,MAAM;AACnB,YAAI,CAAC,KAAM;AAEX,YAAI,MAAM,cAAc;AACtB,eAAK,aAAa,kCAAkC;AAAA,YAClD,MAAM;AAAA,UAAA,CACP;AAAA,QACH;AACA,YAAI,MAAM,MAAO,MAAK,aAAa,yBAAyB,MAAM,KAAK;AAOvE,YAAI,MAAM,OAAO;AACf,eAAK,cAAc,gBAAgB,MAAM,KAAK,CAAC;AAAA,QACjD;AAEA,YAAI,kBAAkB,MAAM,oBAAoB,SAAS,GAAG;AAC1D,gBAAM,aAAa,cAAc,MAAM,mBAAmB;AAC1D,gBAAM,aAAa,KAAK,UAAU;AAAA,YAChC,EAAE,MAAM,aAAa,SAAS,WAAA;AAAA,UAAW,CAC1C;AAED,eAAK,SAAS,iBAAiB,EAAE,SAAS,YAAY;AAGtD,eAAK,aAAa,0BAA0B,UAAU;AAEtD,eAAK,aAAa,+BAA+B,UAAU;AAI3D,gBAAM,SAAS,aAAa,+BAA+B,UAAU;AACrE,gBAAM,SAAS,aAAa,yBAAyB,UAAU;AAC/D,gBAAM,sBAAsB;AAC5B,gBAAM,+BAA+B;AAAA,QACvC;AAAA,MAMF,CAAC;AACD,aAAO;AAAA,IACT;AAAA,IAEA,QAAQ,KAAK,OAAO;AAClB,UAAI,IAAI,aAAa,QAAQ;AAI3B,iBAAS,gBAAgB,MAAM;AAC7B,gBAAM,OAAO,WAAW,IAAI,GAAG;AAC/B,cAAI,KAAM,MAAK,cAAc,gBAAgB,KAAK,CAAC;AAAA,QACrD,CAAC;AACD;AAAA,MACF;AACA,YAAM,UAAU;AAChB,eAAS,gBAAgB,MAAM;AAC7B,cAAM,QAAQ,WAAW,IAAI,OAAO;AACpC,YAAI,CAAC,MAAO;AAKZ,YAAI,gBAAgB;AAClB,gBAAM,cAAc;AAAA,YAClB,iBAAiB,QAAQ;AAAA,YACzB,yBAAyB;AAAA,YACzB,wBAAwB,QAAQ;AAAA,UAAA;AAElC,yBAAe,OAAO,MAAM,cAAc;AAAA,YACxC,GAAG;AAAA,YACH,qBAAqB;AAAA,UAAA,CACtB;AACD,yBAAe,OAAO,MAAM,kBAAkB;AAAA,YAC5C,GAAG;AAAA,YACH,qBAAqB;AAAA,UAAA,CACtB;AAAA,QACH;AAEA,cAAM,OAAO,MAAM,wBAAwB,MAAM;AACjD,aAAK,cAAc,gBAAgB,KAAK,CAAC;AAAA,MAC3C,CAAC;AAAA,IACH;AAAA,IAEA,iBAAiB,KAAK,SAAS;AAC7B,eAAS,yBAAyB,MAAM;AACtC,cAAM,QAAQ,WAAW,IAAI,GAAG;AAChC,YAAI,CAAC,MAAO;AACZ,cAAM,SAAS,MAAM,wBAAwB,MAAM;AAEnD,cAAM,OAA6B;AAAA,UACjC,MAAM;AAAA,UACN;AAAA,UACA,UAAU,QAAQ;AAAA,UAClB,YAAY,QAAQ;AAAA,UACpB,WAAW,MAAM,iBAAiB;AAAA,QAAA;AAEpC,cAAM,OACJ,SAAS,0BAA0B,MAAM,oBAAoB,IAAI,CAAC,KAClE,gBAAgB,QAAQ,QAAQ;AAElC,cAAM,YAA4C;AAAA,UAChD,oBAAoB,QAAQ;AAAA,UAC5B,uBAAuB,QAAQ;AAAA,UAC/B,oBAAoB;AAAA,QAAA;AAEtB,cAAM,cAA2B;AAAA,UAC/B,MAAM,SAAS;AAAA,UACf,YAAY;AAAA,QAAA;AAEd,cAAM,cACJ;AAAA,UAAS;AAAA,UAA0B,MACjC,oBAAoB,MAAM,WAAW;AAAA,QAAA,KAClC;AAEP,cAAM,YAAYD,MAAU,QAAQC,QAAY,OAAA,GAAU,MAAM;AAChE,YAAI;AACJA,gBAAY,KAAK,WAAW,MAAM;AAChC,qBAAW,OAAO,UAAU,MAAM,aAAa,SAAS;AAAA,QAC1D,CAAC;AAED,cAAM,WAAW;AAAA,UAAS;AAAA,UAA0B,MAClD,oBAAoB,IAAI;AAAA,QAAA;AAE1B,YAAI,SAAU,UAAS,cAAc,QAAQ;AAI7C,YAAI,gBAAgB;AAClB,gBAAM,WACJ,OAAO,QAAQ,SAAS,WACpB,QAAQ,OACP;AAAA,YAAS;AAAA,YAA0B,MAClC,KAAK,UAAU,QAAQ,QAAQ,IAAI;AAAA,UAAA,KAChC;AACX,gBAAM,eAAe,cAAc,QAAQ;AAC3C,gBAAM,gBAAgB,KAAK,UAAU;AAAA,YACnC,EAAE,MAAM,QAAQ,SAAS,aAAA;AAAA,UAAa,CACvC;AACD,mBAAS,aAAa,yBAAyB,aAAa;AAE5D,mBAAS,aAAa,8BAA8B,aAAa;AAAA,QACnE;AAEA,cAAM,UAAU,IAAI,QAAQ,YAAY;AAAA,UACtC,MAAM;AAAA,UACN,UAAU,QAAQ;AAAA,QAAA,CACnB;AAAA,MACH,CAAC;AACD,aAAO;AAAA,IACT;AAAA,IAEA,gBAAgB,KAAK,MAAM;AACzB,eAAS,wBAAwB,MAAM;AACrC,cAAM,QAAQ,WAAW,IAAI,GAAG;AAChC,YAAI,CAAC,MAAO;AACZ,cAAM,QAAQ,MAAM,UAAU,IAAI,KAAK,UAAU;AACjD,YAAI,CAAC,MAAO;AACZ,cAAM,EAAE,MAAM,SAAA,IAAa;AAE3B,cAAM,UAAU,KAAK,KAAK,YAAY;AACtC,iBAAS,aAAa,4BAA4B,OAAO;AAEzD,YAAI,CAAC,KAAK,MAAM,KAAK,UAAU,QAAW;AACxC,mBAAS,gBAAgB,KAAK,KAAkB;AAChD,gBAAM,MAAM,aAAa,KAAK,KAAK;AACnC,mBAAS,UAAU;AAAA,YACjB,MAAM,eAAe;AAAA,YACrB,GAAI,QAAQ,UAAa,EAAE,SAAS,IAAA;AAAA,UAAI,CACzC;AAAA,QACH;AAEA,YAAI,gBAAgB;AAKlB,gBAAM,OACJ,OAAO,KAAK,WAAW,WACnB,KAAK,SACJ;AAAA,YAAS;AAAA,YAA4B,MACpC,KAAK,UAAU,KAAK,UAAU,IAAI;AAAA,UAAA,KAC/B;AACX,gBAAM,eAAe,cAAc,IAAI;AACvC,cAAI,MAAM,sBAAsB;AAC9B,kBAAM,qBAAqB,SAAS,uBAAuB;AAAA,cACzD,SAAS;AAAA,cACT,cAAc,KAAK;AAAA,YAAA,CACpB;AAAA,UACH;AAGA,gBAAM,iBAAiB,KAAK,UAAU;AAAA,YACpC,EAAE,MAAM,QAAQ,SAAS,aAAA;AAAA,UAAa,CACvC;AACD,mBAAS,aAAa,0BAA0B,cAAc;AAE9D,mBAAS,aAAa,+BAA+B,cAAc;AAAA,QACrE;AAEA;AAAA,UAAS;AAAA,UAAkB,MACzB;AAAA,YACE;AAAA,cACE,MAAM;AAAA,cACN;AAAA,cACA,UAAU,KAAK;AAAA,cACf,YAAY,KAAK;AAAA,cACjB,WAAW,MAAM,iBAAiB;AAAA,YAAA;AAAA,YAEpC;AAAA,UAAA;AAAA,QACF;AAEF,iBAAS,IAAA;AACT,cAAM,UAAU,OAAO,KAAK,UAAU;AAAA,MACxC,CAAC;AAAA,IACH;AAAA,IAEA,QAAQ,KAAK,MAAM;AACjB,UAAI,IAAI,aAAa,QAAQ;AAC3B,iBAAS,gBAAgB,MAAM;AAC7B,gBAAM,UAAU,aAAa,KAAK,KAAK;AACvC,uBAAa,KAAK,CAAC,SAAS;AAC1B,iBAAK,gBAAgB,KAAK,KAAkB;AAC5C,iBAAK,UAAU;AAAA,cACb,MAAM,eAAe;AAAA,cACrB,GAAI,YAAY,SAAY,EAAE,QAAA,IAAY,CAAA;AAAA,YAAC,CAC5C;AAAA,UACH,CAAC;AACD,8BAAoB,KAAK,KAAK,UAAU,cAAc,KAAK,KAAK,CAAC;AAAA,QACnE,CAAC;AACD;AAAA,MACF;AACA,YAAM,UAAU;AAChB,eAAS,gBAAgB,MAAM;AAC7B,cAAM,QAAQ,WAAW,IAAI,OAAO;AACpC,YAAI,CAAC,MAAO;AAEZ,cAAM,UAAU,cAAc,KAAK,KAAK;AACxC,cAAM,UAAU,aAAa,KAAK,KAAK;AACvC,cAAM,gBACJ,YAAY,SAAY,EAAE,QAAA,IAAa,CAAA;AACzC,cAAM,YAAY,KAAK;AAEvB,cAAM,gBAAgB,MAAM;AAC5B,YAAI,eAAe;AACjB,wBAAc,gBAAgB,SAAS;AACvC,wBAAc,UAAU;AAAA,YACtB,MAAM,eAAe;AAAA,YACrB,GAAG;AAAA,UAAA,CACJ;AACD;AAAA,YAAS;AAAA,YAAkB,MACzB;AAAA,cACE;AAAA,gBACE,MAAM;AAAA,gBACN,KAAK;AAAA,gBACL,WAAW,MAAM,iBAAiB;AAAA,cAAA;AAAA,cAEpC;AAAA,YAAA;AAAA,UACF;AAEF,wBAAc,IAAA;AACd,gBAAM,uBAAuB;AAAA,QAC/B;AAEA,mBAAW,CAAC,IAAI,KAAK,KAAK,MAAM,WAAW;AACzC,gBAAM,EAAE,MAAM,SAAA,IAAa;AAC3B,eAAK,gBAAgB,SAAS;AAC9B,eAAK,UAAU,EAAE,MAAM,eAAe,OAAO,GAAG,eAAe;AAC/D;AAAA,YAAS;AAAA,YAAkB,MACzB;AAAA,cACE;AAAA,gBACE,MAAM;AAAA,gBACN,KAAK;AAAA,gBACL,YAAY;AAAA,gBACZ;AAAA,gBACA,WAAW,MAAM,iBAAiB;AAAA,cAAA;AAAA,cAEpC;AAAA,YAAA;AAAA,UACF;AAEF,eAAK,IAAA;AACL,gBAAM,UAAU,OAAO,EAAE;AAAA,QAC3B;AAEA,cAAM,SAAS,gBAAgB,SAAS;AACxC,cAAM,SAAS,UAAU;AAAA,UACvB,MAAM,eAAe;AAAA,UACrB,GAAG;AAAA,QAAA,CACJ;AAED,YAAI,mBAAmB;AACrB,4BAAkB,OAAO,KAAK,WAAW,KAAM;AAAA,YAC7C,iBAAiB,QAAQ;AAAA,YACzB,yBAAyB;AAAA,YACzB,wBAAwB,QAAQ;AAAA,YAChC,cAAc;AAAA,UAAA,CACf;AAAA,QACH;AAEA;AAAA,UAAS;AAAA,UAAkB,MACzB,YAAY,EAAE,MAAM,QAAQ,KAAK,QAAA,GAAW,MAAM,QAAQ;AAAA,QAAA;AAE5D,cAAM,SAAS,IAAA;AACf,mBAAW,OAAO,OAAO;AAAA,MAC3B,CAAC;AAAA,IACH;AAAA,IAEA,QAAQ,KAAK,MAAM;AACjB,UAAI,IAAI,aAAa,QAAQ;AAE3B,iBAAS,gBAAgB,MAAM;AAC7B,uBAAa,KAAK,CAAC,SAAS;AAC1B,iBAAK,aAAa,iCAAiC,WAAW;AAC9D,iBAAK,UAAU;AAAA,cACb,MAAM,eAAe;AAAA,cACrB,SAAS,KAAK,UAAU;AAAA,YAAA,CACzB;AAAA,UACH,CAAC;AACD,8BAAoB,KAAK,KAAK,UAAU,WAAW;AAAA,QACrD,CAAC;AACD;AAAA,MACF;AACA,YAAM,UAAU;AAChB,eAAS,gBAAgB,MAAM;AAC7B,cAAM,QAAQ,WAAW,IAAI,OAAO;AACpC,YAAI,CAAC,MAAO;AAEZ,cAAM,iBAAiB,CAAC,SAAqB;AAI3C,eAAK,aAAa,iCAAiC,WAAW;AAC9D,eAAK,UAAU,EAAE,MAAM,eAAe,OAAO,SAAS,aAAa;AAAA,QACrE;AAEA,cAAM,gBAAgB,MAAM;AAC5B,YAAI,eAAe;AACjB,yBAAe,aAAa;AAC5B;AAAA,YAAS;AAAA,YAAkB,MACzB;AAAA,cACE;AAAA,gBACE,MAAM;AAAA,gBACN,KAAK;AAAA,gBACL,WAAW,MAAM,iBAAiB;AAAA,cAAA;AAAA,cAEpC;AAAA,YAAA;AAAA,UACF;AAEF,wBAAc,IAAA;AACd,gBAAM,uBAAuB;AAAA,QAC/B;AACA,mBAAW,CAAC,IAAI,KAAK,KAAK,MAAM,WAAW;AACzC,gBAAM,EAAE,MAAM,SAAA,IAAa;AAC3B,yBAAe,IAAI;AACnB;AAAA,YAAS;AAAA,YAAkB,MACzB;AAAA,cACE;AAAA,gBACE,MAAM;AAAA,gBACN,KAAK;AAAA,gBACL,YAAY;AAAA,gBACZ;AAAA,gBACA,WAAW,MAAM,iBAAiB;AAAA,cAAA;AAAA,cAEpC;AAAA,YAAA;AAAA,UACF;AAEF,eAAK,IAAA;AACL,gBAAM,UAAU,OAAO,EAAE;AAAA,QAC3B;AACA,uBAAe,MAAM,QAAQ;AAE7B,YAAI,mBAAmB;AACrB,4BAAkB,OAAO,KAAK,WAAW,KAAM;AAAA,YAC7C,iBAAiB,QAAQ;AAAA,YACzB,yBAAyB;AAAA,YACzB,wBAAwB,QAAQ;AAAA,YAChC,cAAc;AAAA,UAAA,CACf;AAAA,QACH;AAEA;AAAA,UAAS;AAAA,UAAkB,MACzB,YAAY,EAAE,MAAM,QAAQ,KAAK,QAAA,GAAW,MAAM,QAAQ;AAAA,QAAA;AAE5D,cAAM,SAAS,IAAA;AACf,mBAAW,OAAO,OAAO;AAAA,MAC3B,CAAC;AAAA,IACH;AAAA,IAEA,SAAS,KAAK,MAAM;AAClB,UAAI,IAAI,aAAa,QAAQ;AAC3B,iBAAS,iBAAiB,MAAM;AAC9B,uBAAa,KAAK,CAAC,SAAS;AAC1B,gBAAI,KAAK,MAAO,MAAK,cAAc,gBAAgB,KAAK,KAAK,CAAC;AAAA,UAChE,CAAC;AACD,8BAAoB,KAAK,KAAK,QAAQ;AAAA,QACxC,CAAC;AACD;AAAA,MACF;AACA,YAAM,UAAU;AAChB,eAAS,iBAAiB,MAAM;AAC9B,cAAM,QAAQ,WAAW,IAAI,OAAO;AACpC,YAAI,CAAC,MAAO;AAKZ,mBAAW,CAAC,IAAI,KAAK,KAAK,MAAM,WAAW;AACzC,gBAAM,EAAE,MAAM,SAAA,IAAa;AAC3B,eAAK,aAAa,4BAA4B,SAAS;AACvD;AAAA,YAAS;AAAA,YAAkB,MACzB;AAAA,cACE;AAAA,gBACE,MAAM;AAAA,gBACN,KAAK;AAAA,gBACL,YAAY;AAAA,gBACZ;AAAA,gBACA,WAAW,MAAM,iBAAiB;AAAA,cAAA;AAAA,cAEpC;AAAA,YAAA;AAAA,UACF;AAEF,eAAK,IAAA;AACL,gBAAM,UAAU,OAAO,EAAE;AAAA,QAC3B;AAIA,2BAAmB,OAAO,OAAO;AAEjC,YAAI,mBAAmB;AACrB,4BAAkB,OAAO,KAAK,WAAW,KAAM;AAAA,YAC7C,iBAAiB,QAAQ;AAAA,YACzB,yBAAyB;AAAA,YACzB,wBAAwB,QAAQ;AAAA,UAAA,CACjC;AAAA,QACH;AAEA,YAAI,KAAK,OAAO;AACd,gBAAM,SAAS,cAAc,gBAAgB,KAAK,KAAK,CAAC;AAAA,QAC1D;AACA,YAAI,MAAM,kBAAkB;AAC1B,gBAAM,SAAS,aAAa,kCAAkC;AAAA,YAC5D,MAAM;AAAA,UAAA,CACP;AAAA,QACH;AACA,cAAM,SAAS;AAAA,UACb;AAAA,UACA,MAAM;AAAA,QAAA;AAGR;AAAA,UAAS;AAAA,UAAkB,MACzB,YAAY,EAAE,MAAM,QAAQ,KAAK,QAAA,GAAW,MAAM,QAAQ;AAAA,QAAA;AAE5D,cAAM,SAAS,IAAA;AACf,mBAAW,OAAO,OAAO;AAAA,MAC3B,CAAC;AAAA,IACH;AAAA,EAAA;AAEJ;"}

package/dist/esm/middlewares/tool-cache.d.ts

@@ -0,0 +1 @@
+export { toolCacheMiddleware, type ToolCacheMiddlewareOptions, type ToolCacheStorage, type ToolCacheEntry, } from '../activities/chat/middleware/tool-cache-middleware.js';

package/dist/esm/middlewares/usage-attributes.d.ts

@@ -0,0 +1,24 @@
+import { AttributeValue } from '@opentelemetry/api';
+import { TokenUsage } from '../types.js';
+/**
+ * Build the full set of `gen_ai.usage.*` span attributes from a `TokenUsage`.
+ *
+ * Beyond input/output tokens, this emits provider-reported cost, total tokens,
+ * cache and reasoning breakdowns, duration-based billing, and media unit counts
+ * — every field is guarded so spans stay clean when a provider doesn't report
+ * it. Cache and reasoning use the official GenAI semconv names;
+ * `gen_ai.usage.cost` and `gen_ai.usage.total_tokens` are de-facto extensions
+ * consumed by backends like PostHog (which otherwise re-derive cost from their
+ * own price tables, losing cache discounts and gateway markup). Fields with no
+ * semconv or de-facto convention (`costDetails`, `durationSeconds`,
+ * `unitsBilled`) are TanStack-namespaced.
+ *
+ * Shared by `otelMiddleware` across every activity (chat and the media
+ * activities) so usage lands identically whichever activity produced the span.
+ *
+ * Deliberately not emitted: `providerUsageDetails` (a provider-shaped bag,
+ * unsafe to spread onto spans) and the per-modality token breakdowns
+ * (`promptTokensDetails.audioTokens`, etc.) — those can balloon the attribute
+ * set and have no agreed convention yet.
+ */
+export declare function usageAttributes(usage: TokenUsage): Record<string, AttributeValue>;

package/dist/esm/middlewares/usage-attributes.js

@@ -0,0 +1,43 @@
+import { firstNumber } from "../utilities/numbers.js";
+function usageAttributes(usage) {
+  const attrs = {
+    "gen_ai.usage.input_tokens": usage.promptTokens,
+    "gen_ai.usage.output_tokens": usage.completionTokens
+  };
+  const optional = [
+    ["gen_ai.usage.total_tokens", usage.totalTokens],
+    ["gen_ai.usage.cost", usage.cost],
+    [
+      "gen_ai.usage.cache_read.input_tokens",
+      usage.promptTokensDetails?.cachedTokens
+    ],
+    [
+      "gen_ai.usage.cache_creation.input_tokens",
+      usage.promptTokensDetails?.cacheWriteTokens
+    ],
+    [
+      "gen_ai.usage.reasoning.output_tokens",
+      usage.completionTokensDetails?.reasoningTokens
+    ],
+    ["tanstack.ai.usage.duration_seconds", usage.durationSeconds],
+    ["tanstack.ai.usage.units_billed", usage.unitsBilled],
+    ["tanstack.ai.usage.upstream_cost", usage.costDetails?.upstreamCost],
+    [
+      "tanstack.ai.usage.upstream_input_cost",
+      usage.costDetails?.upstreamInputCost
+    ],
+    [
+      "tanstack.ai.usage.upstream_output_cost",
+      usage.costDetails?.upstreamOutputCost
+    ]
+  ];
+  for (const [key, value] of optional) {
+    const num = firstNumber(value);
+    if (num !== void 0) attrs[key] = num;
+  }
+  return attrs;
+}
+export {
+  usageAttributes
+};
+//# sourceMappingURL=usage-attributes.js.map

package/dist/esm/middlewares/usage-attributes.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"usage-attributes.js","sources":["../../../src/middlewares/usage-attributes.ts"],"sourcesContent":["import { firstNumber } from '../utilities/numbers'\nimport type { AttributeValue } from '@opentelemetry/api'\nimport type { TokenUsage } from '../types'\n\n/**\n * Build the full set of `gen_ai.usage.*` span attributes from a `TokenUsage`.\n *\n * Beyond input/output tokens, this emits provider-reported cost, total tokens,\n * cache and reasoning breakdowns, duration-based billing, and media unit counts\n * — every field is guarded so spans stay clean when a provider doesn't report\n * it. Cache and reasoning use the official GenAI semconv names;\n * `gen_ai.usage.cost` and `gen_ai.usage.total_tokens` are de-facto extensions\n * consumed by backends like PostHog (which otherwise re-derive cost from their\n * own price tables, losing cache discounts and gateway markup). Fields with no\n * semconv or de-facto convention (`costDetails`, `durationSeconds`,\n * `unitsBilled`) are TanStack-namespaced.\n *\n * Shared by `otelMiddleware` across every activity (chat and the media\n * activities) so usage lands identically whichever activity produced the span.\n *\n * Deliberately not emitted: `providerUsageDetails` (a provider-shaped bag,\n * unsafe to spread onto spans) and the per-modality token breakdowns\n * (`promptTokensDetails.audioTokens`, etc.) — those can balloon the attribute\n * set and have no agreed convention yet.\n */\nexport function usageAttributes(\n  usage: TokenUsage,\n): Record<string, AttributeValue> {\n  const attrs: Record<string, AttributeValue> = {\n    'gen_ai.usage.input_tokens': usage.promptTokens,\n    'gen_ai.usage.output_tokens': usage.completionTokens,\n  }\n  const optional: Array<[key: string, value: unknown]> = [\n    ['gen_ai.usage.total_tokens', usage.totalTokens],\n    ['gen_ai.usage.cost', usage.cost],\n    [\n      'gen_ai.usage.cache_read.input_tokens',\n      usage.promptTokensDetails?.cachedTokens,\n    ],\n    [\n      'gen_ai.usage.cache_creation.input_tokens',\n      usage.promptTokensDetails?.cacheWriteTokens,\n    ],\n    [\n      'gen_ai.usage.reasoning.output_tokens',\n      usage.completionTokensDetails?.reasoningTokens,\n    ],\n    ['tanstack.ai.usage.duration_seconds', usage.durationSeconds],\n    ['tanstack.ai.usage.units_billed', usage.unitsBilled],\n    ['tanstack.ai.usage.upstream_cost', usage.costDetails?.upstreamCost],\n    [\n      'tanstack.ai.usage.upstream_input_cost',\n      usage.costDetails?.upstreamInputCost,\n    ],\n    [\n      'tanstack.ai.usage.upstream_output_cost',\n      usage.costDetails?.upstreamOutputCost,\n    ],\n  ]\n  for (const [key, value] of optional) {\n    const num = firstNumber(value)\n    if (num !== undefined) attrs[key] = num\n  }\n  return attrs\n}\n"],"names":[],"mappings":";AAyBO,SAAS,gBACd,OACgC;AAChC,QAAM,QAAwC;AAAA,IAC5C,6BAA6B,MAAM;AAAA,IACnC,8BAA8B,MAAM;AAAA,EAAA;AAEtC,QAAM,WAAiD;AAAA,IACrD,CAAC,6BAA6B,MAAM,WAAW;AAAA,IAC/C,CAAC,qBAAqB,MAAM,IAAI;AAAA,IAChC;AAAA,MACE;AAAA,MACA,MAAM,qBAAqB;AAAA,IAAA;AAAA,IAE7B;AAAA,MACE;AAAA,MACA,MAAM,qBAAqB;AAAA,IAAA;AAAA,IAE7B;AAAA,MACE;AAAA,MACA,MAAM,yBAAyB;AAAA,IAAA;AAAA,IAEjC,CAAC,sCAAsC,MAAM,eAAe;AAAA,IAC5D,CAAC,kCAAkC,MAAM,WAAW;AAAA,IACpD,CAAC,mCAAmC,MAAM,aAAa,YAAY;AAAA,IACnE;AAAA,MACE;AAAA,MACA,MAAM,aAAa;AAAA,IAAA;AAAA,IAErB;AAAA,MACE;AAAA,MACA,MAAM,aAAa;AAAA,IAAA;AAAA,EACrB;AAEF,aAAW,CAAC,KAAK,KAAK,KAAK,UAAU;AACnC,UAAM,MAAM,YAAY,KAAK;AAC7B,QAAI,QAAQ,OAAW,OAAM,GAAG,IAAI;AAAA,EACtC;AACA,SAAO;AACT;"}

package/dist/esm/realtime/index.d.ts

@@ -0,0 +1,28 @@
+import { RealtimeToken, RealtimeTokenOptions } from './types.js';
+export type * from './types.js';
+/**
+ * Generate a realtime token using the provided adapter.
+ *
+ * This function is used on the server to generate ephemeral tokens
+ * that clients can use to establish realtime connections.
+ *
+ * @param options - Token generation options including the adapter
+ * @returns Promise resolving to a RealtimeToken
+ *
+ * @example
+ * ```typescript
+ * import { realtimeToken } from '@tanstack/ai'
+ * import { openaiRealtimeToken } from '@tanstack/ai-openai'
+ *
+ * // Server function (TanStack Start example)
+ * export const getRealtimeToken = createServerFn()
+ *   .handler(async () => {
+ *     return realtimeToken({
+ *       adapter: openaiRealtimeToken({
+ *         model: 'gpt-realtime',
+ *       }),
+ *     })
+ *   })
+ * ```
+ */
+export declare function realtimeToken(options: RealtimeTokenOptions): Promise<RealtimeToken>;

package/dist/esm/realtime/index.js

@@ -0,0 +1,8 @@
+async function realtimeToken(options) {
+  const { adapter } = options;
+  return adapter.generateToken();
+}
+export {
+  realtimeToken
+};
+//# sourceMappingURL=index.js.map

package/dist/esm/realtime/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sources":["../../../src/realtime/index.ts"],"sourcesContent":["import type { RealtimeToken, RealtimeTokenOptions } from './types'\n\n// Re-export all types\nexport type * from './types'\n\n/**\n * Generate a realtime token using the provided adapter.\n *\n * This function is used on the server to generate ephemeral tokens\n * that clients can use to establish realtime connections.\n *\n * @param options - Token generation options including the adapter\n * @returns Promise resolving to a RealtimeToken\n *\n * @example\n * ```typescript\n * import { realtimeToken } from '@tanstack/ai'\n * import { openaiRealtimeToken } from '@tanstack/ai-openai'\n *\n * // Server function (TanStack Start example)\n * export const getRealtimeToken = createServerFn()\n *   .handler(async () => {\n *     return realtimeToken({\n *       adapter: openaiRealtimeToken({\n *         model: 'gpt-realtime',\n *       }),\n *     })\n *   })\n * ```\n */\nexport async function realtimeToken(\n  options: RealtimeTokenOptions,\n): Promise<RealtimeToken> {\n  const { adapter } = options\n  return adapter.generateToken()\n}\n"],"names":[],"mappings":"AA8BA,eAAsB,cACpB,SACwB;AACxB,QAAM,EAAE,YAAY;AACpB,SAAO,QAAQ,cAAA;AACjB;"}

package/dist/esm/realtime/types.d.ts

@@ -0,0 +1,282 @@
+import { AnyClientTool } from '../activities/chat/tools/tool-definition.js';
+/**
+ * Voice activity detection configuration
+ */
+export interface VADConfig {
+    /** Sensitivity threshold (0.0-1.0) */
+    threshold?: number;
+    /** Audio to include before speech detection (ms) */
+    prefixPaddingMs?: number;
+    /** Silence duration to end turn (ms) */
+    silenceDurationMs?: number;
+}
+/**
+ * Serializable tool descriptor for realtime session configuration.
+ * Contains only the metadata needed by providers, not Zod schemas or execute functions.
+ */
+export interface RealtimeToolConfig {
+    name: string;
+    description: string;
+    inputSchema?: Record<string, any>;
+}
+/**
+ * Configuration for a realtime session
+ */
+export interface RealtimeSessionConfig {
+    /** Model to use for the session */
+    model?: string;
+    /** Voice to use for audio output */
+    voice?: string;
+    /** System instructions for the assistant */
+    instructions?: string;
+    /** Tools available in the session */
+    tools?: Array<RealtimeToolConfig>;
+    /** VAD mode */
+    vadMode?: 'server' | 'semantic' | 'manual';
+    /** VAD configuration */
+    vadConfig?: VADConfig;
+    /** Output modalities for responses (e.g., ['audio', 'text'], ['text']) */
+    outputModalities?: Array<'audio' | 'text'>;
+    /** Temperature for generation (provider-specific range, e.g., 0.6-1.2 for OpenAI) */
+    temperature?: number;
+    /** Maximum number of tokens in a response */
+    maxOutputTokens?: number | 'inf';
+    /** Eagerness level for semantic VAD ('low', 'medium', 'high') */
+    semanticEagerness?: 'low' | 'medium' | 'high';
+    /** Provider-specific options */
+    providerOptions?: Record<string, any>;
+}
+/**
+ * Token returned by the server for client authentication
+ */
+export interface RealtimeToken {
+    /** Provider identifier */
+    provider: string;
+    /** The ephemeral token value */
+    token: string;
+    /** Token expiration timestamp (ms since epoch) */
+    expiresAt: number;
+    /** Session configuration embedded in the token */
+    config: RealtimeSessionConfig;
+}
+/**
+ * Adapter interface for generating provider-specific tokens
+ */
+export interface RealtimeTokenAdapter {
+    /** Provider identifier */
+    provider: string;
+    /** Generate an ephemeral token for client use */
+    generateToken: () => Promise<RealtimeToken>;
+}
+/**
+ * Options for the realtimeToken function
+ */
+export interface RealtimeTokenOptions {
+    /** The token adapter to use */
+    adapter: RealtimeTokenAdapter;
+}
+/**
+ * Text content part in a realtime message
+ */
+export interface RealtimeTextPart {
+    type: 'text';
+    content: string;
+}
+/**
+ * Audio content part in a realtime message
+ */
+export interface RealtimeAudioPart {
+    type: 'audio';
+    /** Transcription of the audio */
+    transcript: string;
+    /** Raw audio data (optional, if stored) */
+    audioData?: ArrayBuffer;
+    /** Duration of the audio in milliseconds */
+    durationMs?: number;
+}
+/**
+ * Tool call part in a realtime message
+ */
+export interface RealtimeToolCallPart {
+    type: 'tool-call';
+    id: string;
+    name: string;
+    arguments: string;
+    input?: unknown;
+    output?: unknown;
+}
+/**
+ * Tool result part in a realtime message
+ */
+export interface RealtimeToolResultPart {
+    type: 'tool-result';
+    toolCallId: string;
+    content: string;
+}
+/**
+ * Image content part in a realtime message
+ */
+export interface RealtimeImagePart {
+    type: 'image';
+    /** Base64-encoded image data or a URL */
+    data: string;
+    /** MIME type of the image (e.g., 'image/png', 'image/jpeg') */
+    mimeType: string;
+}
+/**
+ * Union of all realtime message parts
+ */
+export type RealtimeMessagePart = RealtimeTextPart | RealtimeAudioPart | RealtimeToolCallPart | RealtimeToolResultPart | RealtimeImagePart;
+/**
+ * A message in a realtime conversation
+ */
+export interface RealtimeMessage {
+    /** Unique message identifier */
+    id: string;
+    /** Message role */
+    role: 'user' | 'assistant';
+    /** Timestamp when the message was created */
+    timestamp: number;
+    /** Content parts of the message */
+    parts: Array<RealtimeMessagePart>;
+    /** Whether this message was interrupted */
+    interrupted?: boolean;
+    /** Reference to audio buffer if stored */
+    audioId?: string;
+    /** Duration of the audio in milliseconds */
+    durationMs?: number;
+}
+/**
+ * Connection status of the realtime client
+ */
+export type RealtimeStatus = 'idle' | 'connecting' | 'connected' | 'reconnecting' | 'error';
+/**
+ * Current mode of the realtime session
+ */
+export type RealtimeMode = 'idle' | 'listening' | 'thinking' | 'speaking';
+/**
+ * Interface for accessing audio visualization data
+ */
+export interface AudioVisualization {
+    /** Input volume level (0-1 normalized) */
+    readonly inputLevel: number;
+    /** Output volume level (0-1 normalized) */
+    readonly outputLevel: number;
+    /** Get frequency data for input audio visualization */
+    getInputFrequencyData: () => Uint8Array;
+    /** Get frequency data for output audio visualization */
+    getOutputFrequencyData: () => Uint8Array;
+    /** Get time domain data for input waveform */
+    getInputTimeDomainData: () => Uint8Array;
+    /** Get time domain data for output waveform */
+    getOutputTimeDomainData: () => Uint8Array;
+    /** Input sample rate */
+    readonly inputSampleRate: number;
+    /** Output sample rate */
+    readonly outputSampleRate: number;
+    /** Subscribe to raw input audio samples */
+    onInputAudio?: (callback: (samples: Float32Array, sampleRate: number) => void) => () => void;
+    /** Subscribe to raw output audio samples */
+    onOutputAudio?: (callback: (samples: Float32Array, sampleRate: number) => void) => () => void;
+}
+/**
+ * Events emitted by the realtime connection
+ */
+export type RealtimeEvent = 'status_change' | 'mode_change' | 'transcript' | 'audio_chunk' | 'tool_call' | 'message_complete' | 'interrupted' | 'error';
+/**
+ * Event payloads for realtime events
+ */
+export interface RealtimeEventPayloads {
+    status_change: {
+        status: RealtimeStatus;
+    };
+    mode_change: {
+        mode: RealtimeMode;
+    };
+    transcript: {
+        role: 'user' | 'assistant';
+        transcript: string;
+        isFinal: boolean;
+    };
+    audio_chunk: {
+        data: ArrayBuffer;
+        sampleRate: number;
+    };
+    tool_call: {
+        toolCallId: string;
+        toolName: string;
+        input: unknown;
+    };
+    message_complete: {
+        message: RealtimeMessage;
+    };
+    interrupted: {
+        messageId?: string;
+    };
+    error: {
+        error: Error;
+    };
+}
+/**
+ * Handler type for realtime events
+ */
+export type RealtimeEventHandler<TEvent extends RealtimeEvent> = (payload: RealtimeEventPayloads[TEvent]) => void;
+/**
+ * Error codes for realtime errors
+ */
+export type RealtimeErrorCode = 'TOKEN_EXPIRED' | 'CONNECTION_FAILED' | 'PERMISSION_DENIED' | 'PROVIDER_ERROR' | 'UNKNOWN';
+/**
+ * Extended error with realtime-specific information
+ */
+export interface RealtimeError extends Error {
+    code: RealtimeErrorCode;
+    provider?: string;
+    details?: unknown;
+}
+/**
+ * Adapter interface for connecting to realtime providers.
+ * Each provider (OpenAI, ElevenLabs, etc.) implements this interface.
+ *
+ * Defined here in `@tanstack/ai` — the shared layer that both provider adapter
+ * packages and the client runtime (`@tanstack/ai-client`) already depend on —
+ * so a provider package can describe its realtime adapter without taking a
+ * dependency on the client-only `@tanstack/ai-client`. `@tanstack/ai-client`
+ * re-exports this type for backwards compatibility.
+ */
+export interface RealtimeAdapter {
+    /** Provider identifier */
+    provider: string;
+    /**
+     * Create a connection using the provided token
+     * @param token - The ephemeral token from the server
+     * @param clientTools - Optional client-side tools to register with the provider
+     * @returns A connection instance
+     */
+    connect: (token: RealtimeToken, clientTools?: ReadonlyArray<AnyClientTool>) => Promise<RealtimeConnection>;
+}
+/**
+ * Connection interface representing an active realtime session.
+ * Handles audio I/O, events, and session management.
+ */
+export interface RealtimeConnection {
+    /** Disconnect from the realtime session */
+    disconnect: () => Promise<void>;
+    /** Start capturing audio from the microphone */
+    startAudioCapture: () => Promise<void>;
+    /** Stop capturing audio */
+    stopAudioCapture: () => void;
+    /** Send a text message (fallback for when voice isn't available) */
+    sendText: (text: string) => void;
+    /** Send an image to the conversation */
+    sendImage: (imageData: string, mimeType: string) => void;
+    /** Send a tool execution result back to the provider */
+    sendToolResult: (callId: string, result: string) => void;
+    /** Update session configuration */
+    updateSession: (config: Partial<RealtimeSessionConfig>) => void;
+    /** Interrupt the current response */
+    interrupt: () => void;
+    /** Subscribe to connection events */
+    on: <TEvent extends RealtimeEvent>(event: TEvent, handler: RealtimeEventHandler<TEvent>) => () => void;
+    /** Get audio visualization data */
+    getAudioVisualization: () => AudioVisualization;
+}

package/dist/esm/stream-to-response.d.ts

@@ -0,0 +1,102 @@
+import { StreamChunk } from './types.js';
+/**
+ * Collect all text content from a StreamChunk async iterable and return as a string.
+ *
+ * This function consumes the entire stream, accumulating content from TEXT_MESSAGE_CONTENT events,
+ * and returns the final concatenated text.
+ *
+ * @param stream - AsyncIterable of StreamChunks from chat()
+ * @returns Promise<string> - The accumulated text content
+ *
+ * @example
+ * ```typescript
+ * const stream = chat({
+ *   adapter: openaiText(),
+ *   model: 'gpt-4o',
+ *   messages: [{ role: 'user', content: 'Hello!' }]
+ * });
+ * const text = await streamToText(stream);
+ * console.log(text); // "Hello! How can I help you today?"
+ * ```
+ */
+export declare function streamToText(stream: AsyncIterable<StreamChunk>): Promise<string>;
+/**
+ * Convert a StreamChunk async iterable to a ReadableStream in Server-Sent Events format
+ *
+ * This creates a ReadableStream that emits chunks in SSE format:
+ * - Each chunk is prefixed with "data: "
+ * - Each chunk is followed by "\n\n"
+ * - Stream ends when the underlying iterable is exhausted (RUN_FINISHED is the terminal event)
+ *
+ * @param stream - AsyncIterable of StreamChunks from chat()
+ * @param abortController - Optional AbortController to abort when stream is cancelled
+ * @returns ReadableStream in Server-Sent Events format
+ */
+export declare function toServerSentEventsStream(stream: AsyncIterable<StreamChunk>, abortController?: AbortController): ReadableStream<Uint8Array>;
+/**
+ * Convert a StreamChunk async iterable to a Response in Server-Sent Events format
+ *
+ * This creates a Response that emits chunks in SSE format:
+ * - Each chunk is prefixed with "data: "
+ * - Each chunk is followed by "\n\n"
+ * - Stream ends when the underlying iterable is exhausted (RUN_FINISHED is the terminal event)
+ *
+ * @param stream - AsyncIterable of StreamChunks from chat()
+ * @param init - Optional Response initialization options (including `abortController`)
+ * @returns Response in Server-Sent Events format
+ *
+ * @example
+ * ```typescript
+ * const stream = chat({ adapter: openaiText(), model: "gpt-4o", messages: [...] });
+ * return toServerSentEventsResponse(stream, { abortController });
+ * ```
+ */
+export declare function toServerSentEventsResponse(stream: AsyncIterable<StreamChunk>, init?: ResponseInit & {
+    abortController?: AbortController;
+}): Response;
+/**
+ * Convert a StreamChunk async iterable to a ReadableStream in HTTP stream format (newline-delimited JSON)
+ *
+ * This creates a ReadableStream that emits chunks as newline-delimited JSON:
+ * - Each chunk is JSON.stringify'd and followed by "\n"
+ * - No SSE formatting (no "data: " prefix)
+ *
+ * This format is compatible with `fetchHttpStream` connection adapter.
+ *
+ * @param stream - AsyncIterable of StreamChunks from chat()
+ * @param abortController - Optional AbortController to abort when stream is cancelled
+ * @returns ReadableStream in HTTP stream format (newline-delimited JSON)
+ *
+ * @example
+ * ```typescript
+ * const stream = chat({ adapter: openaiText(), model: "gpt-4o", messages: [...] });
+ * const readableStream = toHttpStream(stream);
+ * // Use with Response for HTTP streaming (not SSE)
+ * return new Response(readableStream, {
+ *   headers: { 'Content-Type': 'application/x-ndjson' }
+ * });
+ * ```
+ */
+export declare function toHttpStream(stream: AsyncIterable<StreamChunk>, abortController?: AbortController): ReadableStream<Uint8Array>;
+/**
+ * Convert a StreamChunk async iterable to a Response in HTTP stream format (newline-delimited JSON)
+ *
+ * This creates a Response that emits chunks in HTTP stream format:
+ * - Each chunk is JSON.stringify'd and followed by "\n"
+ * - No SSE formatting (no "data: " prefix)
+ *
+ * This format is compatible with `fetchHttpStream` connection adapter.
+ *
+ * @param stream - AsyncIterable of StreamChunks from chat()
+ * @param init - Optional Response initialization options (including `abortController`)
+ * @returns Response in HTTP stream format (newline-delimited JSON)
+ *
+ * @example
+ * ```typescript
+ * const stream = chat({ adapter: openaiText(), model: "gpt-4o", messages: [...] });
+ * return toHttpResponse(stream, { abortController });
+ * ```
+ */
+export declare function toHttpResponse(stream: AsyncIterable<StreamChunk>, init?: ResponseInit & {
+    abortController?: AbortController;
+}): Response;