@apo-ai/sdk 0.13.0 → 0.15.0

package/dist/{chunk-X73JGMK3.js → chunk-4BQSCD6T.js}

@@ -4,7 +4,7 @@ import {
   emitGenerationAndTools,
   servingHostFromBaseURL,
   startGeneration
-} from "./chunk-JTWQD27M.js";
+} from "./chunk-JN7RHBTS.js";
 
 // src/agent-task/integrations/anthropic.ts
 function createApoAnthropic(client, options) {
@@ -78,4 +78,4 @@ function createApoAnthropic(client, options) {
 export {
   createApoAnthropic
 };
-//# sourceMappingURL=chunk-X73JGMK3.js.map
+//# sourceMappingURL=chunk-4BQSCD6T.js.map

package/dist/{chunk-JTWQD27M.js → chunk-JN7RHBTS.js}

@@ -28,6 +28,8 @@ function emitGenerationAndTools(trace, genSpanId, genStartedAt, opts) {
     completion_tokens: opts.completionTokens,
     reasoning_tokens: opts.reasoningTokens,
     served_model: opts.servedModel,
+    ...opts.servedProvider ? { provider: opts.servedProvider } : {},
+    ...opts.servedRoute ? { route: opts.servedRoute } : {},
     output: {
       ...opts.text !== void 0 ? { text: opts.text } : {},
       ...opts.error ? { error: opts.error.message } : {}
@@ -60,8 +62,9 @@ function servingHostFromBaseURL(baseURL, defaultVendor) {
   } catch {
     return defaultVendor ? { provider: defaultVendor } : {};
   }
-  if (host === "localhost" || /^(127\.|0\.0\.0\.0|10\.|192\.168\.|172\.(1[6-9]|2\d|3[01])\.)/.test(host)) {
-    return {};
+  if (host === "localhost" || /^(127\.|0\.0.0\.0|10\.|192\.168\.|172\.(1[6-9]|2\d|3[01])\.)/.test(host)) {
+    const named = process.env.APO_SERVING_GATEWAY?.trim();
+    return named ? { provider: named } : {};
   }
   const known = {
     "api.openai.com": "openai",
@@ -82,6 +85,35 @@ function servingHostFromBaseURL(baseURL, defaultVendor) {
   if (labels.length < 2) return {};
   return { provider: labels.slice(-2).join("."), route: host };
 }
+function servingFromResponse(response, headers) {
+  const endpoints = response?.openrouter_metadata?.endpoints?.available;
+  if (Array.isArray(endpoints)) {
+    const selected = endpoints.find(
+      (e) => e && typeof e === "object" && e.selected === true
+    );
+    const provider = selected && typeof selected.provider === "string" ? selected.provider : void 0;
+    if (provider) {
+      return { provider: "openrouter", route: `openrouter:${provider}` };
+    }
+  }
+  const litellmId = headers?.get?.("x-litellm-model-id");
+  if (typeof litellmId === "string" && litellmId) {
+    const upstream = headers?.get?.("x-litellm-model-api-base");
+    const host = typeof upstream === "string" && upstream ? safeUrlHost(upstream) : void 0;
+    return {
+      provider: "litellm",
+      route: host ? `litellm:${litellmId}@${host}` : `litellm:${litellmId}`
+    };
+  }
+  return {};
+}
+function safeUrlHost(url) {
+  try {
+    return new URL(url).hostname;
+  } catch {
+    return void 0;
+  }
+}
 function startGeneration(trace, opts) {
   const spanId = trace.createSpan({
     task_id: opts.taskId ?? "trace",
@@ -107,6 +139,7 @@ export {
   safeParse,
   emitGenerationAndTools,
   servingHostFromBaseURL,
+  servingFromResponse,
   startGeneration
 };
-//# sourceMappingURL=chunk-JTWQD27M.js.map
+//# sourceMappingURL=chunk-JN7RHBTS.js.map

package/dist/chunk-JN7RHBTS.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/agent-task/integrations/span-helpers.ts"],"sourcesContent":["/**\n * Shared span-emission helpers for the tracing integrations.\n *\n * The OpenAI and Anthropic wrappers both need to:\n * 1. Create a GENERATION span before the LLM call\n * 2. End it with text/tokens/latency after the response\n * 3. Emit a TOOL span for each tool call in the response\n *\n * These helpers keep that logic in one place so the wrappers stay thin.\n *\n * @internal\n */\n\nimport type { AgentTaskTraceContext } from \"../tracing.ts\";\n\nfunction monotonicNowMs(): number {\n  if (typeof performance !== \"undefined\" && typeof performance.now === \"function\") {\n    return performance.now();\n  }\n  return Date.now();\n}\n\nfunction round3(n: number): number {\n  return Math.round(n * 1000) / 1000;\n}\n\nfunction safeParse(raw: unknown): unknown {\n  if (typeof raw !== \"string\") return raw;\n  try {\n    return JSON.parse(raw);\n  } catch {\n    return raw;\n  }\n}\n\n/**\n * Emit a GENERATION span + TOOL spans for a completed LLM call.\n *\n * Both wrappers call this after the SDK returns the response. It:\n * 1. Ends the GENERATION span (created by {@link startGeneration}) with\n *    text, token counts, and latency.\n * 2. Emits a TOOL span per tool call, so `t.calledTool(name)` works.\n */\nexport function emitGenerationAndTools(\n  trace: AgentTaskTraceContext,\n  genSpanId: string,\n  genStartedAt: number,\n  opts: {\n    text?: string;\n    promptTokens?: number;\n    completionTokens?: number;\n    /** Reasoning tokens, when the provider reports the dimension. */\n    reasoningTokens?: number;\n    /** The model that actually served the call (response `model` field). */\n    servedModel?: string;\n    /** Serving host learned from the response (gateway metadata/headers);\n     * overrides the start-time baseURL guess when present. */\n    servedProvider?: string;\n    servedRoute?: string;\n    toolCalls?: Array<{ name: string; input?: unknown }>;\n    taskId?: string;\n    turnNumber?: number;\n    error?: { message: string };\n  },\n): void {\n  const latency = round3(monotonicNowMs() - genStartedAt);\n  const isError = !!opts.error;\n\n  // End the GENERATION span\n  trace.endSpan(genSpanId, {\n    latency_ms: latency,\n    prompt_tokens: opts.promptTokens,\n    completion_tokens: opts.completionTokens,\n    reasoning_tokens: opts.reasoningTokens,\n    served_model: opts.servedModel,\n    ...(opts.servedProvider ? { provider: opts.servedProvider } : {}),\n    ...(opts.servedRoute ? { route: opts.servedRoute } : {}),\n    output: {\n      ...(opts.text !== undefined ? { text: opts.text } : {}),\n      ...(opts.error ? { error: opts.error.message } : {}),\n    },\n    ...(isError\n      ? { level: \"ERROR\" as const, status_message: opts.error!.message }\n      : {}),\n  });\n\n  // Emit TOOL spans (only on success — errors are captured on the GENERATION span)\n  if (!isError && opts.toolCalls) {\n    for (const tc of opts.toolCalls) {\n      const toolSpanId = trace.createSpan({\n        task_id: opts.taskId ?? \"trace\",\n        parent_call_id: genSpanId,\n        step_name: tc.name,\n        observation_type: \"TOOL\",\n        ...(tc.input !== undefined\n          ? {\n              input:\n                tc.input && typeof tc.input === \"object\"\n                  ? (tc.input as Record<string, unknown>)\n                  : { value: tc.input },\n            }\n          : {}),\n        metadata: { toolName: tc.name },\n      });\n      trace.endSpan(toolSpanId, {});\n    }\n  }\n}\n\n/**\n * Resolve the serving host a client actually talks to, from its baseURL.\n *\n * The vendor default (no baseURL) means the SDK's own endpoint. A custom\n * baseURL is the serving host itself — for a known gateway the provider is\n * its name; for anything else the registered domain is the provider and the\n * full host the route, which is exactly the split the backend stores\n * (`gen_ai.provider.name` / `apo.llm.route`, issue #307). Local/private\n * hosts are proxies in front of an unknown backend — report nothing rather\n * than a wrong host.\n */\nexport function servingHostFromBaseURL(\n  baseURL: unknown,\n  defaultVendor?: string,\n): { provider?: string; route?: string } {\n  if (typeof baseURL !== \"string\" || baseURL.trim() === \"\") {\n    return defaultVendor ? { provider: defaultVendor } : {};\n  }\n  let host: string;\n  try {\n    host = new URL(baseURL).hostname;\n  } catch {\n    return defaultVendor ? { provider: defaultVendor } : {};\n  }\n  if (\n    host === \"localhost\" ||\n    /^(127\\.|0\\.0.0\\.0|10\\.|192\\.168\\.|172\\.(1[6-9]|2\\d|3[01])\\.)/.test(host)\n  ) {\n    // A private host is a gateway in front of an unknown backend. An\n    // operator can name it once via APO_SERVING_GATEWAY (e.g. \"litellm\")\n    // instead of every run reporting \"unknown\".\n    const named = process.env.APO_SERVING_GATEWAY?.trim();\n    return named ? { provider: named } : {};\n  }\n  const known: Record<string, string> = {\n    \"api.openai.com\": \"openai\",\n    \"api.anthropic.com\": \"anthropic\",\n    \"openrouter.ai\": \"openrouter\",\n    \"api.fireworks.ai\": \"fireworks\",\n    \"api.groq.com\": \"groq\",\n    \"api.cerebras.ai\": \"cerebras\",\n    \"api.deepseek.com\": \"deepseek\",\n    \"api.mistral.ai\": \"mistral\",\n    \"api.x.ai\": \"xai\",\n    \"api.together.xyz\": \"together\",\n    \"generativelanguage.googleapis.com\": \"google\",\n  };\n  const provider = known[host];\n  if (provider) return { provider };\n  const labels = host.split(\".\");\n  if (labels.length < 2) return {};\n  return { provider: labels.slice(-2).join(\".\"), route: host };\n}\n\n/**\n * Derive the serving host a gateway reports about itself, from the response.\n *\n * Gateways in front of many backends know exactly which one served a call;\n * the baseURL only ever names the gateway. Priority:\n *   1. OpenRouter's opt-in `openrouter_metadata.endpoints.available[selected]`\n *      (send `X-OpenRouter-Metadata: enabled` to receive it).\n *   2. LiteLLM's response headers: `x-litellm-model-id` (the deployment)\n *      and `x-litellm-model-api-base` (the upstream actually called).\n * Returns {} when the gateway reports nothing — unknown stays unknown.\n */\nexport function servingFromResponse(\n  response: {\n    openrouter_metadata?: {\n      endpoints?: {\n        available?: Array<{ provider?: unknown; selected?: unknown }>;\n      };\n    };\n  },\n  headers?: { get(name: string): string | null } | null,\n): { provider?: string; route?: string } {\n  const endpoints = response?.openrouter_metadata?.endpoints?.available;\n  if (Array.isArray(endpoints)) {\n    const selected = endpoints.find(\n      (e) => e && typeof e === \"object\" && e.selected === true,\n    );\n    const provider =\n      selected && typeof (selected as { provider?: unknown }).provider === \"string\"\n        ? (selected as { provider: string }).provider\n        : undefined;\n    if (provider) {\n      return { provider: \"openrouter\", route: `openrouter:${provider}` };\n    }\n  }\n  const litellmId = headers?.get?.(\"x-litellm-model-id\");\n  if (typeof litellmId === \"string\" && litellmId) {\n    const upstream = headers?.get?.(\"x-litellm-model-api-base\");\n    const host =\n      typeof upstream === \"string\" && upstream\n        ? safeUrlHost(upstream)\n        : undefined;\n    return {\n      provider: \"litellm\",\n      route: host ? `litellm:${litellmId}@${host}` : `litellm:${litellmId}`,\n    };\n  }\n  return {};\n}\n\nfunction safeUrlHost(url: string): string | undefined {\n  try {\n    return new URL(url).hostname;\n  } catch {\n    return undefined;\n  }\n}\n\n/**\n * Create a GENERATION span before the LLM call.\n * Returns `{ spanId, startedAt }` to pass to {@link emitGenerationAndTools}.\n */\nexport function startGeneration(\n  trace: AgentTaskTraceContext,\n  opts: {\n    model: string;\n    system?: string;\n    messages?: unknown;\n    parentSpanId?: string;\n    taskId?: string;\n    turnNumber?: number;\n    /** Serving host attributes for this generation (issue #307). */\n    provider?: string;\n    route?: string;\n  },\n): { spanId: string; startedAt: number } {\n  const spanId = trace.createSpan({\n    task_id: opts.taskId ?? \"trace\",\n    parent_call_id: opts.parentSpanId ?? trace.rootSpanId,\n    step_name: \"agent.generate\",\n    model: opts.model,\n    provider: opts.provider,\n    route: opts.route,\n    observation_type: \"GENERATION\",\n    input: {\n      ...(opts.system !== undefined ? { system: opts.system } : {}),\n      ...(opts.messages !== undefined ? { messages: opts.messages } : {}),\n    },\n    metadata: {\n      ...(opts.turnNumber !== undefined ? { turnNumber: opts.turnNumber } : {}),\n      ...(opts.taskId ? { taskId: opts.taskId } : {}),\n    },\n  });\n  return { spanId, startedAt: monotonicNowMs() };\n}\n\n// Re-export safeParse for the wrappers (OpenAI arguments is a JSON string)\nexport { safeParse };\n"],"mappings":";;;;AAeA,SAAS,iBAAyB;AAChC,MAAI,OAAO,gBAAgB,eAAe,OAAO,YAAY,QAAQ,YAAY;AAC/E,WAAO,YAAY,IAAI;AAAA,EACzB;AACA,SAAO,KAAK,IAAI;AAClB;AAEA,SAAS,OAAO,GAAmB;AACjC,SAAO,KAAK,MAAM,IAAI,GAAI,IAAI;AAChC;AAEA,SAAS,UAAU,KAAuB;AACxC,MAAI,OAAO,QAAQ,SAAU,QAAO;AACpC,MAAI;AACF,WAAO,KAAK,MAAM,GAAG;AAAA,EACvB,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAUO,SAAS,uBACd,OACA,WACA,cACA,MAiBM;AACN,QAAM,UAAU,OAAO,eAAe,IAAI,YAAY;AACtD,QAAM,UAAU,CAAC,CAAC,KAAK;AAGvB,QAAM,QAAQ,WAAW;AAAA,IACvB,YAAY;AAAA,IACZ,eAAe,KAAK;AAAA,IACpB,mBAAmB,KAAK;AAAA,IACxB,kBAAkB,KAAK;AAAA,IACvB,cAAc,KAAK;AAAA,IACnB,GAAI,KAAK,iBAAiB,EAAE,UAAU,KAAK,eAAe,IAAI,CAAC;AAAA,IAC/D,GAAI,KAAK,cAAc,EAAE,OAAO,KAAK,YAAY,IAAI,CAAC;AAAA,IACtD,QAAQ;AAAA,MACN,GAAI,KAAK,SAAS,SAAY,EAAE,MAAM,KAAK,KAAK,IAAI,CAAC;AAAA,MACrD,GAAI,KAAK,QAAQ,EAAE,OAAO,KAAK,MAAM,QAAQ,IAAI,CAAC;AAAA,IACpD;AAAA,IACA,GAAI,UACA,EAAE,OAAO,SAAkB,gBAAgB,KAAK,MAAO,QAAQ,IAC/D,CAAC;AAAA,EACP,CAAC;AAGD,MAAI,CAAC,WAAW,KAAK,WAAW;AAC9B,eAAW,MAAM,KAAK,WAAW;AAC/B,YAAM,aAAa,MAAM,WAAW;AAAA,QAClC,SAAS,KAAK,UAAU;AAAA,QACxB,gBAAgB;AAAA,QAChB,WAAW,GAAG;AAAA,QACd,kBAAkB;AAAA,QAClB,GAAI,GAAG,UAAU,SACb;AAAA,UACE,OACE,GAAG,SAAS,OAAO,GAAG,UAAU,WAC3B,GAAG,QACJ,EAAE,OAAO,GAAG,MAAM;AAAA,QAC1B,IACA,CAAC;AAAA,QACL,UAAU,EAAE,UAAU,GAAG,KAAK;AAAA,MAChC,CAAC;AACD,YAAM,QAAQ,YAAY,CAAC,CAAC;AAAA,IAC9B;AAAA,EACF;AACF;AAaO,SAAS,uBACd,SACA,eACuC;AACvC,MAAI,OAAO,YAAY,YAAY,QAAQ,KAAK,MAAM,IAAI;AACxD,WAAO,gBAAgB,EAAE,UAAU,cAAc,IAAI,CAAC;AAAA,EACxD;AACA,MAAI;AACJ,MAAI;AACF,WAAO,IAAI,IAAI,OAAO,EAAE;AAAA,EAC1B,QAAQ;AACN,WAAO,gBAAgB,EAAE,UAAU,cAAc,IAAI,CAAC;AAAA,EACxD;AACA,MACE,SAAS,eACT,+DAA+D,KAAK,IAAI,GACxE;AAIA,UAAM,QAAQ,QAAQ,IAAI,qBAAqB,KAAK;AACpD,WAAO,QAAQ,EAAE,UAAU,MAAM,IAAI,CAAC;AAAA,EACxC;AACA,QAAM,QAAgC;AAAA,IACpC,kBAAkB;AAAA,IAClB,qBAAqB;AAAA,IACrB,iBAAiB;AAAA,IACjB,oBAAoB;AAAA,IACpB,gBAAgB;AAAA,IAChB,mBAAmB;AAAA,IACnB,oBAAoB;AAAA,IACpB,kBAAkB;AAAA,IAClB,YAAY;AAAA,IACZ,oBAAoB;AAAA,IACpB,qCAAqC;AAAA,EACvC;AACA,QAAM,WAAW,MAAM,IAAI;AAC3B,MAAI,SAAU,QAAO,EAAE,SAAS;AAChC,QAAM,SAAS,KAAK,MAAM,GAAG;AAC7B,MAAI,OAAO,SAAS,EAAG,QAAO,CAAC;AAC/B,SAAO,EAAE,UAAU,OAAO,MAAM,EAAE,EAAE,KAAK,GAAG,GAAG,OAAO,KAAK;AAC7D;AAaO,SAAS,oBACd,UAOA,SACuC;AACvC,QAAM,YAAY,UAAU,qBAAqB,WAAW;AAC5D,MAAI,MAAM,QAAQ,SAAS,GAAG;AAC5B,UAAM,WAAW,UAAU;AAAA,MACzB,CAAC,MAAM,KAAK,OAAO,MAAM,YAAY,EAAE,aAAa;AAAA,IACtD;AACA,UAAM,WACJ,YAAY,OAAQ,SAAoC,aAAa,WAChE,SAAkC,WACnC;AACN,QAAI,UAAU;AACZ,aAAO,EAAE,UAAU,cAAc,OAAO,cAAc,QAAQ,GAAG;AAAA,IACnE;AAAA,EACF;AACA,QAAM,YAAY,SAAS,MAAM,oBAAoB;AACrD,MAAI,OAAO,cAAc,YAAY,WAAW;AAC9C,UAAM,WAAW,SAAS,MAAM,0BAA0B;AAC1D,UAAM,OACJ,OAAO,aAAa,YAAY,WAC5B,YAAY,QAAQ,IACpB;AACN,WAAO;AAAA,MACL,UAAU;AAAA,MACV,OAAO,OAAO,WAAW,SAAS,IAAI,IAAI,KAAK,WAAW,SAAS;AAAA,IACrE;AAAA,EACF;AACA,SAAO,CAAC;AACV;AAEA,SAAS,YAAY,KAAiC;AACpD,MAAI;AACF,WAAO,IAAI,IAAI,GAAG,EAAE;AAAA,EACtB,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAMO,SAAS,gBACd,OACA,MAWuC;AACvC,QAAM,SAAS,MAAM,WAAW;AAAA,IAC9B,SAAS,KAAK,UAAU;AAAA,IACxB,gBAAgB,KAAK,gBAAgB,MAAM;AAAA,IAC3C,WAAW;AAAA,IACX,OAAO,KAAK;AAAA,IACZ,UAAU,KAAK;AAAA,IACf,OAAO,KAAK;AAAA,IACZ,kBAAkB;AAAA,IAClB,OAAO;AAAA,MACL,GAAI,KAAK,WAAW,SAAY,EAAE,QAAQ,KAAK,OAAO,IAAI,CAAC;AAAA,MAC3D,GAAI,KAAK,aAAa,SAAY,EAAE,UAAU,KAAK,SAAS,IAAI,CAAC;AAAA,IACnE;AAAA,IACA,UAAU;AAAA,MACR,GAAI,KAAK,eAAe,SAAY,EAAE,YAAY,KAAK,WAAW,IAAI,CAAC;AAAA,MACvE,GAAI,KAAK,SAAS,EAAE,QAAQ,KAAK,OAAO,IAAI,CAAC;AAAA,IAC/C;AAAA,EACF,CAAC;AACD,SAAO,EAAE,QAAQ,WAAW,eAAe,EAAE;AAC/C;","names":[]}

package/dist/{chunk-A3UYQ7M4.js → chunk-NSMKSPBU.js}

@@ -3,20 +3,26 @@
 import {
   emitGenerationAndTools,
   safeParse,
+  servingFromResponse,
   servingHostFromBaseURL,
   startGeneration
-} from "./chunk-JTWQD27M.js";
+} from "./chunk-JN7RHBTS.js";
 
 // src/agent-task/integrations/openai.ts
 function createApoOpenAI(client, options) {
   const { trace, parentSpanId, taskId, turnNumber } = options;
   const realCreate = client.chat.completions.create.bind(client.chat.completions);
+  const rawCreate = client.chat.completions.withRawResponse?.create?.bind(
+    client.chat.completions.withRawResponse
+  );
   const tracedCreate = async (...args) => {
     const params = args[0] ?? {};
     const model = String(params.model ?? "unknown");
     if (params.stream === true) {
       return realCreate(params);
     }
+    const baseURL = client.baseURL;
+    const requestParams = typeof baseURL === "string" && baseURL.includes("openrouter.ai") ? { ...params, headers: { ...params.headers, "X-OpenRouter-Metadata": "enabled" } } : params;
     const serving = servingHostFromBaseURL(
       client.baseURL,
       "openai"
@@ -31,16 +37,27 @@ function createApoOpenAI(client, options) {
       ...serving
     });
     try {
-      const response = await realCreate(params);
+      let response;
+      let headers = null;
+      if (rawCreate) {
+        const raw = await rawCreate(requestParams);
+        headers = raw.headers;
+        response = await raw.parse();
+      } else {
+        response = await realCreate(requestParams);
+      }
       const message = response.choices?.[0]?.message;
       const text = message?.content ?? void 0;
       const toolCalls = (message?.tool_calls ?? []).map((tc) => ({
         name: tc.function?.name ?? "unknown",
         input: tc.function?.arguments ? safeParse(tc.function.arguments) : void 0
       }));
+      const reported = servingFromResponse(response, headers);
       emitGenerationAndTools(trace, spanId, startedAt, {
         text,
         servedModel: response.model,
+        servedProvider: reported.provider,
+        servedRoute: reported.route,
         promptTokens: response.usage?.prompt_tokens,
         completionTokens: response.usage?.completion_tokens,
         // OpenAI-compatible providers report thinking tokens here (o-series,
@@ -87,4 +104,4 @@ function createApoOpenAI(client, options) {
 export {
   createApoOpenAI
 };
-//# sourceMappingURL=chunk-A3UYQ7M4.js.map
+//# sourceMappingURL=chunk-NSMKSPBU.js.map

package/dist/chunk-NSMKSPBU.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/agent-task/integrations/openai.ts"],"sourcesContent":["/**\n * OpenAI JS SDK tracing integration for apo.\n *\n * Wraps an `OpenAI` client so `chat.completions.create()` calls are\n * automatically traced. Tool calls and generations are captured without\n * any manual span code in the adapter.\n *\n * Uses a Proxy to intercept the `create` method, then emits GENERATION +\n * TOOL spans through the trace context, where they become the observations\n * that `t.calledTool` reads.\n *\n * @example\n * ```ts\n * import OpenAI from \"openai\";\n * import { createApoOpenAI } from \"@apo-ai/sdk/agent-task\";\n *\n * async sendUserTurn(turn, { trace, parentSpanId }) {\n *   const client = createApoOpenAI(\n *     new OpenAI({ apiKey, baseURL: \"https://openrouter.ai/api/v1\" }),\n *     { trace, parentSpanId, taskId: ctx.task.id, turnNumber }\n *   );\n *   const response = await client.chat.completions.create({\n *     model: \"deepseek/deepseek-v4.1-flash\",\n *     messages: [{ role: \"user\", content: String(turn) }],\n *     tools: [...],\n *   });\n *   return { response: response.choices[0].message.content ?? \"\" };\n * }\n * ```\n *\n * @module\n */\n\nimport type { AgentTaskTraceContext } from \"../tracing.ts\";\nimport {\n  startGeneration,\n  emitGenerationAndTools,\n  safeParse,\n  servingFromResponse,\n  servingHostFromBaseURL,\n} from \"./span-helpers.ts\";\n\nexport interface CreateApoOpenAIOptions {\n  /** The tee'd trace context from `sendUserTurn`'s second argument. */\n  trace: AgentTaskTraceContext;\n  /** The parent span id for this turn (from `sendUserTurn`'s context). */\n  parentSpanId?: string;\n  /** Optional task id, surfaced in span metadata. */\n  taskId?: string;\n  /** Optional turn number, surfaced in span metadata. */\n  turnNumber?: number;\n}\n\n/**\n * Minimal structural shape of the OpenAI client we need to intercept.\n * We avoid importing `openai` at the type level so this module doesn't\n * take a runtime dependency — the wrapper works with any client that has\n * `chat.completions.create`.\n *\n * The `create` signature uses a broad type so the OpenAI SDK's overloaded\n * `create` method (which has specific param types) is assignable to this.\n */\ninterface OpenAIClientLike {\n  chat: {\n    completions: {\n      // eslint-disable-next-line @typescript-eslint/no-explicit-any\n      create: (...args: any[]) => Promise<any>;\n      // openai SDK ≥4 exposes the raw-response variant; optional so custom\n      // clients without it still work.\n      withRawResponse?: {\n        // eslint-disable-next-line @typescript-eslint/no-explicit-any\n        create: (...args: any[]) => Promise<any>;\n      };\n    };\n  };\n  // eslint-disable-next-line @typescript-eslint/no-explicit-any\n  [key: string]: any;\n}\n\n/**\n * The response shape we read from `chat.completions.create()`.\n */\ninterface OpenAIResponseLike {\n  model?: string;\n  /** OpenRouter's routing report, opt-in via X-OpenRouter-Metadata. */\n  openrouter_metadata?: {\n    endpoints?: {\n      available?: Array<{ provider?: string; selected?: boolean }>;\n    };\n  };\n  choices?: Array<{\n    message?: {\n      content?: string | null;\n      tool_calls?: Array<{\n        function?: { name?: string; arguments?: string };\n      }>;\n    };\n  }>;\n  usage?: {\n    prompt_tokens?: number;\n    completion_tokens?: number;\n    completion_tokens_details?: {\n      /** Thinking tokens, reported by o-series/DeepSeek/GLM-style providers. */\n      reasoning_tokens?: number;\n    };\n  };\n}\n\n/**\n * Wrap an OpenAI client so `chat.completions.create()` is traced.\n * Returns a Proxy-wrapped client — call it exactly as you would the\n * original. Non-streaming only; streaming responses pass through\n * untraced.\n *\n * For SDKs that emit OpenTelemetry natively (OpenAI Agents SDK, Vercel AI\n * SDK, Claude Agent SDK), prefer `configureApoTelemetry()` instead — no\n * wrapper needed. This manual wrapper is for the raw `openai` package or\n * custom clients that don't emit OTel on their own.\n */\nexport function createApoOpenAI<T extends OpenAIClientLike>(\n  client: T,\n  options: CreateApoOpenAIOptions,\n): T {\n  const { trace, parentSpanId, taskId, turnNumber } = options;\n  const realCreate = client.chat.completions.create.bind(client.chat.completions);\n  // The raw-response variant (openai SDK ≥4) is how response headers reach\n  // us — LiteLLM reports the serving deployment there. Optional: custom\n  // clients without it fall back to the plain call.\n  const rawCreate:\n    | ((...args: unknown[]) => Promise<{\n        headers: { get(name: string): string | null };\n        parse(): Promise<unknown>;\n      }>)\n    | undefined = client.chat.completions.withRawResponse?.create?.bind(\n      client.chat.completions.withRawResponse,\n    );\n\n  // eslint-disable-next-line @typescript-eslint/no-explicit-any\n  const tracedCreate = async (...args: any[]): Promise<any> => {\n    const params = (args[0] ?? {}) as Record<string, unknown>;\n    const model = String(params.model ?? \"unknown\");\n\n    // Non-streaming: trace fully. Streaming (stream: true): pass through.\n    if (params.stream === true) {\n      return realCreate(params);\n    }\n\n    // OpenRouter's routing metadata is opt-in per request; asking for it on\n    // every openrouter call costs nothing and yields the selected provider\n    // (plus the fallback audit trail OpenRouter then includes).\n    const baseURL = (client as { baseURL?: unknown }).baseURL;\n    const requestParams =\n      typeof baseURL === \"string\" && baseURL.includes(\"openrouter.ai\")\n        ? { ...params, headers: { ...(params.headers as object | undefined), \"X-OpenRouter-Metadata\": \"enabled\" } }\n        : params;\n\n    // The OpenAI SDK is often pointed at a gateway via baseURL — the host\n    // it actually calls is the run's serving host (issue #307).\n    const serving = servingHostFromBaseURL(\n      (client as { baseURL?: unknown }).baseURL,\n      \"openai\",\n    );\n    const { spanId, startedAt } = startGeneration(trace, {\n      model,\n      system: typeof params.system === \"string\" ? params.system : undefined,\n      messages: params.messages,\n      parentSpanId,\n      taskId,\n      turnNumber,\n      ...serving,\n    });\n\n    try {\n      let response: OpenAIResponseLike;\n      let headers: { get(name: string): string | null } | null = null;\n      if (rawCreate) {\n        const raw = await rawCreate(requestParams);\n        headers = raw.headers;\n        response = (await raw.parse()) as OpenAIResponseLike;\n      } else {\n        response = (await realCreate(requestParams)) as OpenAIResponseLike;\n      }\n      const message = response.choices?.[0]?.message;\n      const text = message?.content ?? undefined;\n      const toolCalls = (message?.tool_calls ?? []).map((tc) => ({\n        name: tc.function?.name ?? \"unknown\",\n        input: tc.function?.arguments ? safeParse(tc.function.arguments) : undefined,\n      }));\n      // The gateway's own report of who served the call beats the baseURL\n      // guess: OpenRouter metadata / LiteLLM response headers.\n      const reported = servingFromResponse(response, headers);\n\n      emitGenerationAndTools(trace, spanId, startedAt, {\n        text,\n        servedModel: response.model,\n        servedProvider: reported.provider,\n        servedRoute: reported.route,\n        promptTokens: response.usage?.prompt_tokens,\n        completionTokens: response.usage?.completion_tokens,\n        // OpenAI-compatible providers report thinking tokens here (o-series,\n        // DeepSeek, GLM). Absent detail = unreported, not zero.\n        reasoningTokens:\n          response.usage?.completion_tokens_details?.reasoning_tokens,\n        toolCalls,\n        taskId,\n        turnNumber,\n      });\n\n      return response;\n    } catch (error) {\n      emitGenerationAndTools(trace, spanId, startedAt, {\n        taskId,\n        turnNumber,\n        error: { message: error instanceof Error ? error.message : String(error) },\n      });\n      throw error;\n    }\n  };\n\n  // Proxy that intercepts client.chat.completions.create but passes\n  // everything else through unchanged.\n  return new Proxy(client, {\n    get(target, prop, receiver) {\n      if (prop === \"chat\") {\n        return new Proxy(target.chat, {\n          get(chatTarget, chatProp) {\n            if (chatProp === \"completions\") {\n              return new Proxy(chatTarget.completions, {\n                get(completionsTarget, completionsProp) {\n                  if (completionsProp === \"create\") {\n                    return tracedCreate;\n                  }\n                  return Reflect.get(completionsTarget, completionsProp);\n                },\n              });\n            }\n            return Reflect.get(chatTarget, chatProp);\n          },\n        });\n      }\n      return Reflect.get(target, prop, receiver);\n    },\n  }) as T;\n}\n"],"mappings":";;;;;;;;;;;AAuHO,SAAS,gBACd,QACA,SACG;AACH,QAAM,EAAE,OAAO,cAAc,QAAQ,WAAW,IAAI;AACpD,QAAM,aAAa,OAAO,KAAK,YAAY,OAAO,KAAK,OAAO,KAAK,WAAW;AAI9E,QAAM,YAKU,OAAO,KAAK,YAAY,iBAAiB,QAAQ;AAAA,IAC7D,OAAO,KAAK,YAAY;AAAA,EAC1B;AAGF,QAAM,eAAe,UAAU,SAA8B;AAC3D,UAAM,SAAU,KAAK,CAAC,KAAK,CAAC;AAC5B,UAAM,QAAQ,OAAO,OAAO,SAAS,SAAS;AAG9C,QAAI,OAAO,WAAW,MAAM;AAC1B,aAAO,WAAW,MAAM;AAAA,IAC1B;AAKA,UAAM,UAAW,OAAiC;AAClD,UAAM,gBACJ,OAAO,YAAY,YAAY,QAAQ,SAAS,eAAe,IAC3D,EAAE,GAAG,QAAQ,SAAS,EAAE,GAAI,OAAO,SAAgC,yBAAyB,UAAU,EAAE,IACxG;AAIN,UAAM,UAAU;AAAA,MACb,OAAiC;AAAA,MAClC;AAAA,IACF;AACA,UAAM,EAAE,QAAQ,UAAU,IAAI,gBAAgB,OAAO;AAAA,MACnD;AAAA,MACA,QAAQ,OAAO,OAAO,WAAW,WAAW,OAAO,SAAS;AAAA,MAC5D,UAAU,OAAO;AAAA,MACjB;AAAA,MACA;AAAA,MACA;AAAA,MACA,GAAG;AAAA,IACL,CAAC;AAED,QAAI;AACF,UAAI;AACJ,UAAI,UAAuD;AAC3D,UAAI,WAAW;AACb,cAAM,MAAM,MAAM,UAAU,aAAa;AACzC,kBAAU,IAAI;AACd,mBAAY,MAAM,IAAI,MAAM;AAAA,MAC9B,OAAO;AACL,mBAAY,MAAM,WAAW,aAAa;AAAA,MAC5C;AACA,YAAM,UAAU,SAAS,UAAU,CAAC,GAAG;AACvC,YAAM,OAAO,SAAS,WAAW;AACjC,YAAM,aAAa,SAAS,cAAc,CAAC,GAAG,IAAI,CAAC,QAAQ;AAAA,QACzD,MAAM,GAAG,UAAU,QAAQ;AAAA,QAC3B,OAAO,GAAG,UAAU,YAAY,UAAU,GAAG,SAAS,SAAS,IAAI;AAAA,MACrE,EAAE;AAGF,YAAM,WAAW,oBAAoB,UAAU,OAAO;AAEtD,6BAAuB,OAAO,QAAQ,WAAW;AAAA,QAC/C;AAAA,QACA,aAAa,SAAS;AAAA,QACtB,gBAAgB,SAAS;AAAA,QACzB,aAAa,SAAS;AAAA,QACtB,cAAc,SAAS,OAAO;AAAA,QAC9B,kBAAkB,SAAS,OAAO;AAAA;AAAA;AAAA,QAGlC,iBACE,SAAS,OAAO,2BAA2B;AAAA,QAC7C;AAAA,QACA;AAAA,QACA;AAAA,MACF,CAAC;AAED,aAAO;AAAA,IACT,SAAS,OAAO;AACd,6BAAuB,OAAO,QAAQ,WAAW;AAAA,QAC/C;AAAA,QACA;AAAA,QACA,OAAO,EAAE,SAAS,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK,EAAE;AAAA,MAC3E,CAAC;AACD,YAAM;AAAA,IACR;AAAA,EACF;AAIA,SAAO,IAAI,MAAM,QAAQ;AAAA,IACvB,IAAI,QAAQ,MAAM,UAAU;AAC1B,UAAI,SAAS,QAAQ;AACnB,eAAO,IAAI,MAAM,OAAO,MAAM;AAAA,UAC5B,IAAI,YAAY,UAAU;AACxB,gBAAI,aAAa,eAAe;AAC9B,qBAAO,IAAI,MAAM,WAAW,aAAa;AAAA,gBACvC,IAAI,mBAAmB,iBAAiB;AACtC,sBAAI,oBAAoB,UAAU;AAChC,2BAAO;AAAA,kBACT;AACA,yBAAO,QAAQ,IAAI,mBAAmB,eAAe;AAAA,gBACvD;AAAA,cACF,CAAC;AAAA,YACH;AACA,mBAAO,QAAQ,IAAI,YAAY,QAAQ;AAAA,UACzC;AAAA,QACF,CAAC;AAAA,MACH;AACA,aAAO,QAAQ,IAAI,QAAQ,MAAM,QAAQ;AAAA,IAC3C;AAAA,EACF,CAAC;AACH;","names":[]}

package/dist/{tracing-FqFObrmt.d.ts → tracing-CEucJqQI.d.ts}

@@ -49,11 +49,26 @@ interface EndSpanParams {
     /** Reasoning/thinking tokens, when the provider reports them as a distinct
      * dimension. Omit when unreported — unknown must stay unknown, not zero. */
     reasoning_tokens?: number;
+    /** Prompt tokens read from the provider's prompt cache. Setting either cache
+     * field declares `prompt_tokens` the whole prompt, cached tokens included
+     * (OTel GenAI semconv), whatever the model's own API reports; the span then
+     * carries `apo.usage.input_includes_cache` so the backend prices it that way. */
+    cache_read_tokens?: number;
+    /** Prompt tokens written to the provider's prompt cache; part of
+     * `prompt_tokens` like `cache_read_tokens`. */
+    cache_creation_tokens?: number;
     /** The model the provider reports as having served the call (the response's
      * `model` field), emitted as `gen_ai.response.model`. A gateway fallback
      * (LiteLLM router, OpenRouter provider fallback) can serve a different
      * model than requested — this is the only field that tells them apart. */
     served_model?: string;
+    /** Serving host learned from the response itself (gateway metadata or
+     * response headers), emitted as `gen_ai.provider.name`. Set at end time
+     * when the gateway reports it — more truthful than any baseURL guess. */
+    provider?: string;
+    /** Finer-grained serving route learned from the response (e.g. a LiteLLM
+     * deployment id), emitted as `apo.llm.route`. */
+    route?: string;
     status_message?: string;
     level?: "DEBUG" | "DEFAULT" | "WARNING" | "ERROR";
     metadata?: Record<string, unknown>;
@@ -72,10 +87,17 @@ interface TraceStepOptions extends Omit<CreateSpanParams, "project" | "task_id"
      * spans whose token counts are only known after the work (e.g. judge
      * spans, whose usage rides the judged result). Applied when the span ends;
      * the extracted counts become `gen_ai.usage.*` span attributes.
+     * `served_model` (the model the provider reports as having served the
+     * call) becomes `gen_ai.response.model` — a gateway fallback diverges
+     * from the requested model, and this is the only field that tells them
+     * apart.
      */
     usage?: (result: unknown) => {
         prompt_tokens?: number;
         completion_tokens?: number;
+        cache_read_tokens?: number;
+        cache_creation_tokens?: number;
+        served_model?: string;
     } | undefined;
 }
 interface TraceEventOptions extends Omit<CreateSpanParams, "project" | "task_id" | "run_id"> {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@apo-ai/sdk",
-  "version": "0.13.0",
+  "version": "0.15.0",
   "description": "TypeScript SDK for executable agent capability specifications and harness engineering with Apo",
   "keywords": [
     "ai-agents",
@@ -79,6 +79,7 @@
     "vitest": "^4.1.9"
   },
   "dependencies": {
+    "@ai-sdk/anthropic": "^4.0.74",
     "@ai-sdk/mcp": "^2.0.62",
     "@ai-sdk/openai-compatible": "^3.0.48",
     "@opentelemetry/api": "^1.9.1",

package/dist/chunk-A3UYQ7M4.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/agent-task/integrations/openai.ts"],"sourcesContent":["/**\n * OpenAI JS SDK tracing integration for apo.\n *\n * Wraps an `OpenAI` client so `chat.completions.create()` calls are\n * automatically traced. Tool calls and generations are captured without\n * any manual span code in the adapter.\n *\n * Uses a Proxy to intercept the `create` method, then emits GENERATION +\n * TOOL spans through the trace context, where they become the observations\n * that `t.calledTool` reads.\n *\n * @example\n * ```ts\n * import OpenAI from \"openai\";\n * import { createApoOpenAI } from \"@apo-ai/sdk/agent-task\";\n *\n * async sendUserTurn(turn, { trace, parentSpanId }) {\n *   const client = createApoOpenAI(\n *     new OpenAI({ apiKey, baseURL: \"https://openrouter.ai/api/v1\" }),\n *     { trace, parentSpanId, taskId: ctx.task.id, turnNumber }\n *   );\n *   const response = await client.chat.completions.create({\n *     model: \"deepseek/deepseek-v4.1-flash\",\n *     messages: [{ role: \"user\", content: String(turn) }],\n *     tools: [...],\n *   });\n *   return { response: response.choices[0].message.content ?? \"\" };\n * }\n * ```\n *\n * @module\n */\n\nimport type { AgentTaskTraceContext } from \"../tracing.ts\";\nimport {\n  startGeneration,\n  emitGenerationAndTools,\n  safeParse,\n  servingHostFromBaseURL,\n} from \"./span-helpers.ts\";\n\nexport interface CreateApoOpenAIOptions {\n  /** The tee'd trace context from `sendUserTurn`'s second argument. */\n  trace: AgentTaskTraceContext;\n  /** The parent span id for this turn (from `sendUserTurn`'s context). */\n  parentSpanId?: string;\n  /** Optional task id, surfaced in span metadata. */\n  taskId?: string;\n  /** Optional turn number, surfaced in span metadata. */\n  turnNumber?: number;\n}\n\n/**\n * Minimal structural shape of the OpenAI client we need to intercept.\n * We avoid importing `openai` at the type level so this module doesn't\n * take a runtime dependency — the wrapper works with any client that has\n * `chat.completions.create`.\n *\n * The `create` signature uses a broad type so the OpenAI SDK's overloaded\n * `create` method (which has specific param types) is assignable to this.\n */\ninterface OpenAIClientLike {\n  chat: {\n    completions: {\n      // eslint-disable-next-line @typescript-eslint/no-explicit-any\n      create: (...args: any[]) => Promise<any>;\n    };\n  };\n  // eslint-disable-next-line @typescript-eslint/no-explicit-any\n  [key: string]: any;\n}\n\n/**\n * The response shape we read from `chat.completions.create()`.\n */\ninterface OpenAIResponseLike {\n  model?: string;\n  choices?: Array<{\n    message?: {\n      content?: string | null;\n      tool_calls?: Array<{\n        function?: { name?: string; arguments?: string };\n      }>;\n    };\n  }>;\n  usage?: {\n    prompt_tokens?: number;\n    completion_tokens?: number;\n    completion_tokens_details?: {\n      /** Thinking tokens, reported by o-series/DeepSeek/GLM-style providers. */\n      reasoning_tokens?: number;\n    };\n  };\n}\n\n/**\n * Wrap an OpenAI client so `chat.completions.create()` is traced.\n * Returns a Proxy-wrapped client — call it exactly as you would the\n * original. Non-streaming only; streaming responses pass through\n * untraced.\n *\n * For SDKs that emit OpenTelemetry natively (OpenAI Agents SDK, Vercel AI\n * SDK, Claude Agent SDK), prefer `configureApoTelemetry()` instead — no\n * wrapper needed. This manual wrapper is for the raw `openai` package or\n * custom clients that don't emit OTel on their own.\n */\nexport function createApoOpenAI<T extends OpenAIClientLike>(\n  client: T,\n  options: CreateApoOpenAIOptions,\n): T {\n  const { trace, parentSpanId, taskId, turnNumber } = options;\n  const realCreate = client.chat.completions.create.bind(client.chat.completions);\n\n  // eslint-disable-next-line @typescript-eslint/no-explicit-any\n  const tracedCreate = async (...args: any[]): Promise<any> => {\n    const params = (args[0] ?? {}) as Record<string, unknown>;\n    const model = String(params.model ?? \"unknown\");\n\n    // Non-streaming: trace fully. Streaming (stream: true): pass through.\n    if (params.stream === true) {\n      return realCreate(params);\n    }\n\n    // The OpenAI SDK is often pointed at a gateway via baseURL — the host\n    // it actually calls is the run's serving host (issue #307).\n    const serving = servingHostFromBaseURL(\n      (client as { baseURL?: unknown }).baseURL,\n      \"openai\",\n    );\n    const { spanId, startedAt } = startGeneration(trace, {\n      model,\n      system: typeof params.system === \"string\" ? params.system : undefined,\n      messages: params.messages,\n      parentSpanId,\n      taskId,\n      turnNumber,\n      ...serving,\n    });\n\n    try {\n      const response = (await realCreate(params)) as OpenAIResponseLike;\n      const message = response.choices?.[0]?.message;\n      const text = message?.content ?? undefined;\n      const toolCalls = (message?.tool_calls ?? []).map((tc) => ({\n        name: tc.function?.name ?? \"unknown\",\n        input: tc.function?.arguments ? safeParse(tc.function.arguments) : undefined,\n      }));\n\n      emitGenerationAndTools(trace, spanId, startedAt, {\n        text,\n        servedModel: response.model,\n        promptTokens: response.usage?.prompt_tokens,\n        completionTokens: response.usage?.completion_tokens,\n        // OpenAI-compatible providers report thinking tokens here (o-series,\n        // DeepSeek, GLM). Absent detail = unreported, not zero.\n        reasoningTokens:\n          response.usage?.completion_tokens_details?.reasoning_tokens,\n        toolCalls,\n        taskId,\n        turnNumber,\n      });\n\n      return response;\n    } catch (error) {\n      emitGenerationAndTools(trace, spanId, startedAt, {\n        taskId,\n        turnNumber,\n        error: { message: error instanceof Error ? error.message : String(error) },\n      });\n      throw error;\n    }\n  };\n\n  // Proxy that intercepts client.chat.completions.create but passes\n  // everything else through unchanged.\n  return new Proxy(client, {\n    get(target, prop, receiver) {\n      if (prop === \"chat\") {\n        return new Proxy(target.chat, {\n          get(chatTarget, chatProp) {\n            if (chatProp === \"completions\") {\n              return new Proxy(chatTarget.completions, {\n                get(completionsTarget, completionsProp) {\n                  if (completionsProp === \"create\") {\n                    return tracedCreate;\n                  }\n                  return Reflect.get(completionsTarget, completionsProp);\n                },\n              });\n            }\n            return Reflect.get(chatTarget, chatProp);\n          },\n        });\n      }\n      return Reflect.get(target, prop, receiver);\n    },\n  }) as T;\n}\n"],"mappings":";;;;;;;;;;AA0GO,SAAS,gBACd,QACA,SACG;AACH,QAAM,EAAE,OAAO,cAAc,QAAQ,WAAW,IAAI;AACpD,QAAM,aAAa,OAAO,KAAK,YAAY,OAAO,KAAK,OAAO,KAAK,WAAW;AAG9E,QAAM,eAAe,UAAU,SAA8B;AAC3D,UAAM,SAAU,KAAK,CAAC,KAAK,CAAC;AAC5B,UAAM,QAAQ,OAAO,OAAO,SAAS,SAAS;AAG9C,QAAI,OAAO,WAAW,MAAM;AAC1B,aAAO,WAAW,MAAM;AAAA,IAC1B;AAIA,UAAM,UAAU;AAAA,MACb,OAAiC;AAAA,MAClC;AAAA,IACF;AACA,UAAM,EAAE,QAAQ,UAAU,IAAI,gBAAgB,OAAO;AAAA,MACnD;AAAA,MACA,QAAQ,OAAO,OAAO,WAAW,WAAW,OAAO,SAAS;AAAA,MAC5D,UAAU,OAAO;AAAA,MACjB;AAAA,MACA;AAAA,MACA;AAAA,MACA,GAAG;AAAA,IACL,CAAC;AAED,QAAI;AACF,YAAM,WAAY,MAAM,WAAW,MAAM;AACzC,YAAM,UAAU,SAAS,UAAU,CAAC,GAAG;AACvC,YAAM,OAAO,SAAS,WAAW;AACjC,YAAM,aAAa,SAAS,cAAc,CAAC,GAAG,IAAI,CAAC,QAAQ;AAAA,QACzD,MAAM,GAAG,UAAU,QAAQ;AAAA,QAC3B,OAAO,GAAG,UAAU,YAAY,UAAU,GAAG,SAAS,SAAS,IAAI;AAAA,MACrE,EAAE;AAEF,6BAAuB,OAAO,QAAQ,WAAW;AAAA,QAC/C;AAAA,QACA,aAAa,SAAS;AAAA,QACtB,cAAc,SAAS,OAAO;AAAA,QAC9B,kBAAkB,SAAS,OAAO;AAAA;AAAA;AAAA,QAGlC,iBACE,SAAS,OAAO,2BAA2B;AAAA,QAC7C;AAAA,QACA;AAAA,QACA;AAAA,MACF,CAAC;AAED,aAAO;AAAA,IACT,SAAS,OAAO;AACd,6BAAuB,OAAO,QAAQ,WAAW;AAAA,QAC/C;AAAA,QACA;AAAA,QACA,OAAO,EAAE,SAAS,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK,EAAE;AAAA,MAC3E,CAAC;AACD,YAAM;AAAA,IACR;AAAA,EACF;AAIA,SAAO,IAAI,MAAM,QAAQ;AAAA,IACvB,IAAI,QAAQ,MAAM,UAAU;AAC1B,UAAI,SAAS,QAAQ;AACnB,eAAO,IAAI,MAAM,OAAO,MAAM;AAAA,UAC5B,IAAI,YAAY,UAAU;AACxB,gBAAI,aAAa,eAAe;AAC9B,qBAAO,IAAI,MAAM,WAAW,aAAa;AAAA,gBACvC,IAAI,mBAAmB,iBAAiB;AACtC,sBAAI,oBAAoB,UAAU;AAChC,2BAAO;AAAA,kBACT;AACA,yBAAO,QAAQ,IAAI,mBAAmB,eAAe;AAAA,gBACvD;AAAA,cACF,CAAC;AAAA,YACH;AACA,mBAAO,QAAQ,IAAI,YAAY,QAAQ;AAAA,UACzC;AAAA,QACF,CAAC;AAAA,MACH;AACA,aAAO,QAAQ,IAAI,QAAQ,MAAM,QAAQ;AAAA,IAC3C;AAAA,EACF,CAAC;AACH;","names":[]}

package/dist/chunk-JTWQD27M.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/agent-task/integrations/span-helpers.ts"],"sourcesContent":["/**\n * Shared span-emission helpers for the tracing integrations.\n *\n * The OpenAI and Anthropic wrappers both need to:\n * 1. Create a GENERATION span before the LLM call\n * 2. End it with text/tokens/latency after the response\n * 3. Emit a TOOL span for each tool call in the response\n *\n * These helpers keep that logic in one place so the wrappers stay thin.\n *\n * @internal\n */\n\nimport type { AgentTaskTraceContext } from \"../tracing.ts\";\n\nfunction monotonicNowMs(): number {\n  if (typeof performance !== \"undefined\" && typeof performance.now === \"function\") {\n    return performance.now();\n  }\n  return Date.now();\n}\n\nfunction round3(n: number): number {\n  return Math.round(n * 1000) / 1000;\n}\n\nfunction safeParse(raw: unknown): unknown {\n  if (typeof raw !== \"string\") return raw;\n  try {\n    return JSON.parse(raw);\n  } catch {\n    return raw;\n  }\n}\n\n/**\n * Emit a GENERATION span + TOOL spans for a completed LLM call.\n *\n * Both wrappers call this after the SDK returns the response. It:\n * 1. Ends the GENERATION span (created by {@link startGeneration}) with\n *    text, token counts, and latency.\n * 2. Emits a TOOL span per tool call, so `t.calledTool(name)` works.\n */\nexport function emitGenerationAndTools(\n  trace: AgentTaskTraceContext,\n  genSpanId: string,\n  genStartedAt: number,\n  opts: {\n    text?: string;\n    promptTokens?: number;\n    completionTokens?: number;\n    /** Reasoning tokens, when the provider reports the dimension. */\n    reasoningTokens?: number;\n    /** The model that actually served the call (response `model` field). */\n    servedModel?: string;\n    toolCalls?: Array<{ name: string; input?: unknown }>;\n    taskId?: string;\n    turnNumber?: number;\n    error?: { message: string };\n  },\n): void {\n  const latency = round3(monotonicNowMs() - genStartedAt);\n  const isError = !!opts.error;\n\n  // End the GENERATION span\n  trace.endSpan(genSpanId, {\n    latency_ms: latency,\n    prompt_tokens: opts.promptTokens,\n    completion_tokens: opts.completionTokens,\n    reasoning_tokens: opts.reasoningTokens,\n    served_model: opts.servedModel,\n    output: {\n      ...(opts.text !== undefined ? { text: opts.text } : {}),\n      ...(opts.error ? { error: opts.error.message } : {}),\n    },\n    ...(isError\n      ? { level: \"ERROR\" as const, status_message: opts.error!.message }\n      : {}),\n  });\n\n  // Emit TOOL spans (only on success — errors are captured on the GENERATION span)\n  if (!isError && opts.toolCalls) {\n    for (const tc of opts.toolCalls) {\n      const toolSpanId = trace.createSpan({\n        task_id: opts.taskId ?? \"trace\",\n        parent_call_id: genSpanId,\n        step_name: tc.name,\n        observation_type: \"TOOL\",\n        ...(tc.input !== undefined\n          ? {\n              input:\n                tc.input && typeof tc.input === \"object\"\n                  ? (tc.input as Record<string, unknown>)\n                  : { value: tc.input },\n            }\n          : {}),\n        metadata: { toolName: tc.name },\n      });\n      trace.endSpan(toolSpanId, {});\n    }\n  }\n}\n\n/**\n * Resolve the serving host a client actually talks to, from its baseURL.\n *\n * The vendor default (no baseURL) means the SDK's own endpoint. A custom\n * baseURL is the serving host itself — for a known gateway the provider is\n * its name; for anything else the registered domain is the provider and the\n * full host the route, which is exactly the split the backend stores\n * (`gen_ai.provider.name` / `apo.llm.route`, issue #307). Local/private\n * hosts are proxies in front of an unknown backend — report nothing rather\n * than a wrong host.\n */\nexport function servingHostFromBaseURL(\n  baseURL: unknown,\n  defaultVendor?: string,\n): { provider?: string; route?: string } {\n  if (typeof baseURL !== \"string\" || baseURL.trim() === \"\") {\n    return defaultVendor ? { provider: defaultVendor } : {};\n  }\n  let host: string;\n  try {\n    host = new URL(baseURL).hostname;\n  } catch {\n    return defaultVendor ? { provider: defaultVendor } : {};\n  }\n  if (\n    host === \"localhost\" ||\n    /^(127\\.|0\\.0\\.0\\.0|10\\.|192\\.168\\.|172\\.(1[6-9]|2\\d|3[01])\\.)/.test(host)\n  ) {\n    return {};\n  }\n  const known: Record<string, string> = {\n    \"api.openai.com\": \"openai\",\n    \"api.anthropic.com\": \"anthropic\",\n    \"openrouter.ai\": \"openrouter\",\n    \"api.fireworks.ai\": \"fireworks\",\n    \"api.groq.com\": \"groq\",\n    \"api.cerebras.ai\": \"cerebras\",\n    \"api.deepseek.com\": \"deepseek\",\n    \"api.mistral.ai\": \"mistral\",\n    \"api.x.ai\": \"xai\",\n    \"api.together.xyz\": \"together\",\n    \"generativelanguage.googleapis.com\": \"google\",\n  };\n  const provider = known[host];\n  if (provider) return { provider };\n  const labels = host.split(\".\");\n  if (labels.length < 2) return {};\n  return { provider: labels.slice(-2).join(\".\"), route: host };\n}\n\n/**\n * Create a GENERATION span before the LLM call.\n * Returns `{ spanId, startedAt }` to pass to {@link emitGenerationAndTools}.\n */\nexport function startGeneration(\n  trace: AgentTaskTraceContext,\n  opts: {\n    model: string;\n    system?: string;\n    messages?: unknown;\n    parentSpanId?: string;\n    taskId?: string;\n    turnNumber?: number;\n    /** Serving host attributes for this generation (issue #307). */\n    provider?: string;\n    route?: string;\n  },\n): { spanId: string; startedAt: number } {\n  const spanId = trace.createSpan({\n    task_id: opts.taskId ?? \"trace\",\n    parent_call_id: opts.parentSpanId ?? trace.rootSpanId,\n    step_name: \"agent.generate\",\n    model: opts.model,\n    provider: opts.provider,\n    route: opts.route,\n    observation_type: \"GENERATION\",\n    input: {\n      ...(opts.system !== undefined ? { system: opts.system } : {}),\n      ...(opts.messages !== undefined ? { messages: opts.messages } : {}),\n    },\n    metadata: {\n      ...(opts.turnNumber !== undefined ? { turnNumber: opts.turnNumber } : {}),\n      ...(opts.taskId ? { taskId: opts.taskId } : {}),\n    },\n  });\n  return { spanId, startedAt: monotonicNowMs() };\n}\n\n// Re-export safeParse for the wrappers (OpenAI arguments is a JSON string)\nexport { safeParse };\n"],"mappings":";;;;AAeA,SAAS,iBAAyB;AAChC,MAAI,OAAO,gBAAgB,eAAe,OAAO,YAAY,QAAQ,YAAY;AAC/E,WAAO,YAAY,IAAI;AAAA,EACzB;AACA,SAAO,KAAK,IAAI;AAClB;AAEA,SAAS,OAAO,GAAmB;AACjC,SAAO,KAAK,MAAM,IAAI,GAAI,IAAI;AAChC;AAEA,SAAS,UAAU,KAAuB;AACxC,MAAI,OAAO,QAAQ,SAAU,QAAO;AACpC,MAAI;AACF,WAAO,KAAK,MAAM,GAAG;AAAA,EACvB,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAUO,SAAS,uBACd,OACA,WACA,cACA,MAaM;AACN,QAAM,UAAU,OAAO,eAAe,IAAI,YAAY;AACtD,QAAM,UAAU,CAAC,CAAC,KAAK;AAGvB,QAAM,QAAQ,WAAW;AAAA,IACvB,YAAY;AAAA,IACZ,eAAe,KAAK;AAAA,IACpB,mBAAmB,KAAK;AAAA,IACxB,kBAAkB,KAAK;AAAA,IACvB,cAAc,KAAK;AAAA,IACnB,QAAQ;AAAA,MACN,GAAI,KAAK,SAAS,SAAY,EAAE,MAAM,KAAK,KAAK,IAAI,CAAC;AAAA,MACrD,GAAI,KAAK,QAAQ,EAAE,OAAO,KAAK,MAAM,QAAQ,IAAI,CAAC;AAAA,IACpD;AAAA,IACA,GAAI,UACA,EAAE,OAAO,SAAkB,gBAAgB,KAAK,MAAO,QAAQ,IAC/D,CAAC;AAAA,EACP,CAAC;AAGD,MAAI,CAAC,WAAW,KAAK,WAAW;AAC9B,eAAW,MAAM,KAAK,WAAW;AAC/B,YAAM,aAAa,MAAM,WAAW;AAAA,QAClC,SAAS,KAAK,UAAU;AAAA,QACxB,gBAAgB;AAAA,QAChB,WAAW,GAAG;AAAA,QACd,kBAAkB;AAAA,QAClB,GAAI,GAAG,UAAU,SACb;AAAA,UACE,OACE,GAAG,SAAS,OAAO,GAAG,UAAU,WAC3B,GAAG,QACJ,EAAE,OAAO,GAAG,MAAM;AAAA,QAC1B,IACA,CAAC;AAAA,QACL,UAAU,EAAE,UAAU,GAAG,KAAK;AAAA,MAChC,CAAC;AACD,YAAM,QAAQ,YAAY,CAAC,CAAC;AAAA,IAC9B;AAAA,EACF;AACF;AAaO,SAAS,uBACd,SACA,eACuC;AACvC,MAAI,OAAO,YAAY,YAAY,QAAQ,KAAK,MAAM,IAAI;AACxD,WAAO,gBAAgB,EAAE,UAAU,cAAc,IAAI,CAAC;AAAA,EACxD;AACA,MAAI;AACJ,MAAI;AACF,WAAO,IAAI,IAAI,OAAO,EAAE;AAAA,EAC1B,QAAQ;AACN,WAAO,gBAAgB,EAAE,UAAU,cAAc,IAAI,CAAC;AAAA,EACxD;AACA,MACE,SAAS,eACT,gEAAgE,KAAK,IAAI,GACzE;AACA,WAAO,CAAC;AAAA,EACV;AACA,QAAM,QAAgC;AAAA,IACpC,kBAAkB;AAAA,IAClB,qBAAqB;AAAA,IACrB,iBAAiB;AAAA,IACjB,oBAAoB;AAAA,IACpB,gBAAgB;AAAA,IAChB,mBAAmB;AAAA,IACnB,oBAAoB;AAAA,IACpB,kBAAkB;AAAA,IAClB,YAAY;AAAA,IACZ,oBAAoB;AAAA,IACpB,qCAAqC;AAAA,EACvC;AACA,QAAM,WAAW,MAAM,IAAI;AAC3B,MAAI,SAAU,QAAO,EAAE,SAAS;AAChC,QAAM,SAAS,KAAK,MAAM,GAAG;AAC7B,MAAI,OAAO,SAAS,EAAG,QAAO,CAAC;AAC/B,SAAO,EAAE,UAAU,OAAO,MAAM,EAAE,EAAE,KAAK,GAAG,GAAG,OAAO,KAAK;AAC7D;AAMO,SAAS,gBACd,OACA,MAWuC;AACvC,QAAM,SAAS,MAAM,WAAW;AAAA,IAC9B,SAAS,KAAK,UAAU;AAAA,IACxB,gBAAgB,KAAK,gBAAgB,MAAM;AAAA,IAC3C,WAAW;AAAA,IACX,OAAO,KAAK;AAAA,IACZ,UAAU,KAAK;AAAA,IACf,OAAO,KAAK;AAAA,IACZ,kBAAkB;AAAA,IAClB,OAAO;AAAA,MACL,GAAI,KAAK,WAAW,SAAY,EAAE,QAAQ,KAAK,OAAO,IAAI,CAAC;AAAA,MAC3D,GAAI,KAAK,aAAa,SAAY,EAAE,UAAU,KAAK,SAAS,IAAI,CAAC;AAAA,IACnE;AAAA,IACA,UAAU;AAAA,MACR,GAAI,KAAK,eAAe,SAAY,EAAE,YAAY,KAAK,WAAW,IAAI,CAAC;AAAA,MACvE,GAAI,KAAK,SAAS,EAAE,QAAQ,KAAK,OAAO,IAAI,CAAC;AAAA,IAC/C;AAAA,EACF,CAAC;AACD,SAAO,EAAE,QAAQ,WAAW,eAAe,EAAE;AAC/C;","names":[]}