190proof 1.0.102 → 1.0.105

package/README.md

@@ -289,6 +289,10 @@ interface ParsedResponseMessage {
   function_call: FunctionCall | null;
   function_calls: FunctionCall[];
   files: File[]; // For models that return files (e.g., image generation)
+  // Who actually served the response: OpenRouter's upstream provider from the
+  // response body (e.g. "Baidu"), or the SDK provider name ("anthropic", ...)
+  // for direct providers. On fallback, reflects the model that answered.
+  provider?: string;
   usage: {
     prompt_tokens: number;
     completion_tokens: number;

package/dist/index.d.mts

@@ -123,6 +123,11 @@ interface GenericMessage {
      * `redacted_thinking` blocks captured in
      * `ParsedResponseMessage.reasoningDetails`). Preserves the signatures /
      * encrypted payloads that providers validate on round-trip.
+     *
+     * Safe to echo regardless of which provider serves the next call: each
+     * serializer keeps only its own provider's block shapes (Anthropic keeps
+     * `thinking`/`redacted_thinking`; OpenAI-compat keeps `reasoning.*`), so a
+     * cross-provider fallback drops foreign blocks instead of 400ing.
      */
     reasoningDetails?: any;
 }
@@ -170,6 +175,15 @@ interface ParsedResponseMessage {
      * validate. Undefined when the model/provider returns none.
      */
     reasoningDetails?: any;
+    /**
+     * Who actually served the response. For OpenRouter this is the upstream
+     * provider from the response body (e.g. "Baidu", "Morph") — the routing
+     * decision OpenRouter made, not the requested model slug. For direct
+     * providers it's the SDK provider name ("anthropic", "openai", "google",
+     * "groq"). On model fallback it reflects the model that answered, so a
+     * mismatch with the requested model's provider reveals the fallback.
+     */
+    provider?: string;
     usage: {
         prompt_tokens: number;
         completion_tokens: number;

package/dist/index.d.ts

@@ -123,6 +123,11 @@ interface GenericMessage {
      * `redacted_thinking` blocks captured in
      * `ParsedResponseMessage.reasoningDetails`). Preserves the signatures /
      * encrypted payloads that providers validate on round-trip.
+     *
+     * Safe to echo regardless of which provider serves the next call: each
+     * serializer keeps only its own provider's block shapes (Anthropic keeps
+     * `thinking`/`redacted_thinking`; OpenAI-compat keeps `reasoning.*`), so a
+     * cross-provider fallback drops foreign blocks instead of 400ing.
      */
     reasoningDetails?: any;
 }
@@ -170,6 +175,15 @@ interface ParsedResponseMessage {
      * validate. Undefined when the model/provider returns none.
      */
     reasoningDetails?: any;
+    /**
+     * Who actually served the response. For OpenRouter this is the upstream
+     * provider from the response body (e.g. "Baidu", "Morph") — the routing
+     * decision OpenRouter made, not the requested model slug. For direct
+     * providers it's the SDK provider name ("anthropic", "openai", "google",
+     * "groq"). On model fallback it reflects the model that answered, so a
+     * mismatch with the requested model's provider reveals the fallback.
+     */
+    provider?: string;
     usage: {
         prompt_tokens: number;
         completion_tokens: number;

package/dist/index.js

@@ -320,6 +320,14 @@ function buildOpenAIRequestConfig(identifier, model, config) {
     }
   };
 }
+function filterOpenAICompatReasoningDetails(details) {
+  if (!Array.isArray(details))
+    return details || void 0;
+  const blocks = details.filter(
+    (block) => typeof (block == null ? void 0 : block.type) === "string" && block.type.startsWith("reasoning.")
+  );
+  return blocks.length ? blocks : void 0;
+}
 async function prepareOpenAIPayload(identifier, payload) {
   var _a, _b;
   const preparedPayload = {
@@ -387,8 +395,11 @@ async function prepareOpenAIPayload(identifier, payload) {
     }
     if (message.reasoning)
       outMessage.reasoning = message.reasoning;
-    if (message.reasoningDetails)
-      outMessage.reasoning_details = message.reasoningDetails;
+    const reasoningDetails = filterOpenAICompatReasoningDetails(
+      message.reasoningDetails
+    );
+    if (reasoningDetails)
+      outMessage.reasoning_details = reasoningDetails;
     preparedPayload.messages.push(outMessage);
   }
   return preparedPayload;
@@ -673,7 +684,11 @@ async function prepareAnthropicPayload(_identifier, payload) {
     model: payload.model,
     messages: [],
     functions: payload.functions,
-    temperature: payload.temperature
+    temperature: payload.temperature,
+    // Map the generic function_call to Anthropic tool_choice ("none" forces a
+    // text-only turn). Only meaningful alongside tools — callAnthropic sends
+    // it only when tools are present (tool_choice without tools 400s).
+    tool_choice: payload.function_call ? typeof payload.function_call === "string" ? { type: payload.function_call } : { type: "tool", name: payload.function_call.name } : void 0
   };
   for (const message of payload.messages) {
     if (message.role === "system") {
@@ -728,7 +743,9 @@ async function prepareAnthropicPayload(_identifier, payload) {
         });
       }
     }
-    const leadingBlocks = message.role === "assistant" && Array.isArray(message.reasoningDetails) ? message.reasoningDetails : [];
+    const leadingBlocks = message.role === "assistant" && Array.isArray(message.reasoningDetails) ? message.reasoningDetails.filter(
+      (block) => (block == null ? void 0 : block.type) === "thinking" || (block == null ? void 0 : block.type) === "redacted_thinking"
+    ) : [];
     const toolUseBlocks = (message.functionCalls || []).map((fc, i) => {
       var _a;
       return {
@@ -785,6 +802,8 @@ async function callAnthropic(id, payload, config, requestTimeoutMs = 12e4, signa
         model: payload.model,
         messages: anthropicMessages,
         tools: cachedTools,
+        // tool_choice requires tools in the request; drop it otherwise.
+        tool_choice: (cachedTools == null ? void 0 : cachedTools.length) ? payload.tool_choice : void 0,
         temperature: payload.temperature,
         system: payload.system,
         max_tokens: 4096
@@ -916,6 +935,13 @@ async function prepareGoogleAIPayload(_identifier, payload) {
           ...fn.parameters
         }
       }))
+    } : void 0,
+    // Map the generic function_call to Gemini's functionCallingConfig ("none"
+    // → mode NONE forces a text-only turn). Only meaningful alongside tools.
+    toolConfig: payload.functions && payload.function_call ? {
+      functionCallingConfig: {
+        mode: typeof payload.function_call === "string" ? payload.function_call === "none" ? "NONE" : "AUTO" : "ANY"
+      }
     } : void 0
   };
   const toolNameById = /* @__PURE__ */ new Map();
@@ -1001,6 +1027,9 @@ async function callGoogleAI(id, payload, requestTimeoutMs = 12e4, signal) {
   };
   if (payload.tools)
     requestBody.tools = [payload.tools];
+  if (payload.tools && payload.toolConfig) {
+    requestBody.toolConfig = payload.toolConfig;
+  }
   if (payload.systemInstruction) {
     requestBody.systemInstruction = {
       parts: [{ text: payload.systemInstruction }]
@@ -1227,8 +1256,11 @@ function prepareOpenAICompatMessages(messages) {
     }
     if (message.reasoning)
       outMessage.reasoning = message.reasoning;
-    if (message.reasoningDetails)
-      outMessage.reasoning_details = message.reasoningDetails;
+    const reasoningDetails = filterOpenAICompatReasoningDetails(
+      message.reasoningDetails
+    );
+    if (reasoningDetails)
+      outMessage.reasoning_details = reasoningDetails;
     out.push(outMessage);
   }
   return out;
@@ -1374,7 +1406,7 @@ function parseDsmlToolCalls(content) {
   return { calls, remainingContent: remaining.length ? remaining : null };
 }
 async function callOpenRouter(id, payload, requestTimeoutMs = 12e4, signal) {
-  var _a, _b, _c, _d, _e, _f, _g, _h, _i;
+  var _a, _b, _c, _d, _e, _f, _g, _h, _i, _j;
   const response = await import_axios.default.post(
     "https://openrouter.ai/api/v1/chat/completions",
     payload,
@@ -1434,11 +1466,14 @@ async function callOpenRouter(id, payload, requestTimeoutMs = 12e4, signal) {
     files: [],
     reasoning: (_f = answer.reasoning) != null ? _f : void 0,
     reasoningDetails: (_g = answer.reasoning_details) != null ? _g : void 0,
+    // The upstream provider OpenRouter routed to (e.g. "Baidu") — finer-grained
+    // than the "openrouter" stamp callWithRetries would apply.
+    provider: (_h = response.data.provider) != null ? _h : void 0,
     usage: response.data.usage ? {
       prompt_tokens: response.data.usage.prompt_tokens,
       completion_tokens: response.data.usage.completion_tokens,
       total_tokens: response.data.usage.total_tokens,
-      cached_tokens: (_i = (_h = response.data.usage.prompt_tokens_details) == null ? void 0 : _h.cached_tokens) != null ? _i : 0
+      cached_tokens: (_j = (_i = response.data.usage.prompt_tokens_details) == null ? void 0 : _i.cached_tokens) != null ? _j : 0
     } : null
   };
 }
@@ -1488,15 +1523,16 @@ function parseModelString(model) {
   );
 }
 async function callWithRetries(id, aiPayload, aiConfig, retries = 5, chunkTimeoutMs = 15e3) {
-  var _a, _b;
+  var _a, _b, _c;
   try {
     const { provider, modelId } = parseModelString(aiPayload.model);
     const routingPayload = { ...aiPayload, model: modelId };
     const requestTimeoutMs = (_a = aiPayload.requestTimeoutMs) != null ? _a : 12e4;
     const signal = aiPayload.signal;
+    let result;
     switch (provider) {
       case "anthropic":
-        return await callAnthropicWithRetries(
+        result = await callAnthropicWithRetries(
           id,
           await prepareAnthropicPayload(id, routingPayload),
           aiConfig,
@@ -1504,8 +1540,9 @@ async function callWithRetries(id, aiPayload, aiConfig, retries = 5, chunkTimeou
           requestTimeoutMs,
           signal
         );
+        break;
       case "openai":
-        return await callOpenAiWithRetries(
+        result = await callOpenAiWithRetries(
           id,
           await prepareOpenAIPayload(id, routingPayload),
           aiConfig,
@@ -1514,33 +1551,39 @@ async function callWithRetries(id, aiPayload, aiConfig, retries = 5, chunkTimeou
           requestTimeoutMs,
           signal
         );
+        break;
       case "groq":
-        return await callGroqWithRetries(
+        result = await callGroqWithRetries(
           id,
           prepareGroqPayload(routingPayload),
           retries,
           requestTimeoutMs,
           signal
         );
+        break;
       case "google":
-        return await callGoogleAIWithRetries(
+        result = await callGoogleAIWithRetries(
           id,
           await prepareGoogleAIPayload(id, routingPayload),
           retries,
           requestTimeoutMs,
           signal
         );
+        break;
       case "openrouter":
-        return await callOpenRouterWithRetries(
+        result = await callOpenRouterWithRetries(
           id,
           prepareOpenRouterPayload(routingPayload),
           retries,
           requestTimeoutMs,
           signal
         );
+        break;
     }
+    (_b = result.provider) != null ? _b : result.provider = provider;
+    return result;
   } catch (error2) {
-    if ((_b = aiPayload.signal) == null ? void 0 : _b.aborted)
+    if ((_c = aiPayload.signal) == null ? void 0 : _c.aborted)
       throw error2;
     if (aiPayload.fallbackModel) {
       logger_default.error(