@thanh01.pmt/curriculum-kit 1.4.31 → 1.4.33

package/dist/ai/index.cjs

@@ -933,6 +933,52 @@ async function runCurriculumAIInference(messages, projectContext, options = {},
 }
 
 // src/ai/provider-factory.ts
+function cleanNonStreamingFetch() {
+  return async (input, init) => {
+    const res = await fetch(input, init);
+    let wantsStream = false;
+    try {
+      const body = init?.body;
+      if (typeof body === "string") wantsStream = /"stream"\s*:\s*true/.test(body);
+    } catch {
+    }
+    if (wantsStream) return res;
+    const contentType = res.headers.get("content-type") || "";
+    if (!contentType.includes("text/event-stream") && !contentType.includes("application/json")) return res;
+    const text = await res.text();
+    const trimmed = text.trim();
+    if (trimmed.startsWith("data:")) {
+      let content = "";
+      let lastChunk = null;
+      for (const line of trimmed.split("\n")) {
+        const l = line.trim();
+        if (!l || !l.startsWith("data:")) continue;
+        const jsonStr = l.slice(5).trim();
+        if (jsonStr === "[DONE]") continue;
+        try {
+          const chunk = JSON.parse(jsonStr);
+          lastChunk = chunk;
+          content += chunk.choices?.[0]?.delta?.content || "";
+        } catch {
+        }
+      }
+      const assembled = {
+        id: lastChunk?.id || "chatcmpl-" + Date.now(),
+        object: "chat.completion",
+        created: lastChunk?.created || Math.floor(Date.now() / 1e3),
+        model: lastChunk?.model || "9router",
+        choices: [{ index: 0, message: { role: "assistant", content }, finish_reason: "stop" }],
+        usage: lastChunk?.usage || void 0
+      };
+      const headers = new Headers(res.headers);
+      headers.set("content-type", "application/json");
+      return new Response(JSON.stringify(assembled), { status: res.status, statusText: res.statusText, headers });
+    }
+    const doneIndex = text.indexOf("data: [DONE]");
+    if (doneIndex === -1) return res;
+    return new Response(text.slice(0, doneIndex).trim(), { status: res.status, statusText: res.statusText, headers: res.headers });
+  };
+}
 var DEFAULT_MODELS = {
   "9router": "fast-reasoning",
   ninerouter: "fast-reasoning",
@@ -972,7 +1018,8 @@ function getAIModel(options = {}) {
     const rawUrl = options.baseURL || resolveApiKey("NINEROUTER_BASE_URL") || resolveApiKey("NINE_ROUTER_BASE_URL") || "https://ai-router.orchable.app/v1";
     const ninerouter = openai.createOpenAI({
       apiKey,
-      baseURL: rawUrl.replace(/^["']|["']$/g, "").trim()
+      baseURL: rawUrl.replace(/^["']|["']$/g, "").trim(),
+      fetch: cleanNonStreamingFetch()
     });
     return ninerouter.chat(resolvedModelName || "fast-reasoning");
   }
@@ -5129,6 +5176,16 @@ async function runTurn(endpoint, model, apiKey, messages, tools, idleTimeoutMs,
   return { content: fullRaw, toolCalls };
 }
 async function streamCurriculumAIInferenceWithTools(messages, projectContext, options, loop, onChunk) {
+  if (options.customToolInference) {
+    return runToolLoopWithCustomInference(messages, projectContext, options, loop, onChunk);
+  }
+  if (resolveApiKey("CURRICULUM_KIT_REQUIRE_CUSTOM_INFERENCE") === "1") {
+    throw createAiInferenceError({
+      errorCode: "ERR_CUSTOM_INFERENCE_REQUIRED",
+      message: "customToolInference is REQUIRED (CURRICULUM_KIT_REQUIRE_CUSTOM_INFERENCE=1): pass buildKitToolInference(...) so tool-calling turns flow through the host AI SDK gateway and its unified Langfuse tracing. Raw-fetch tool-loop provider routes are disabled by mandate.",
+      rawError: { mandate: "CURRICULUM_KIT_REQUIRE_CUSTOM_INFERENCE", hasCustomToolInference: false }
+    });
+  }
   const route = resolveProviderRoute(options);
   const maxTokens = options.maxTokens ?? 32768;
   const maxRounds = loop.maxRounds ?? 4;
@@ -5193,6 +5250,74 @@ Guidelines:
   }
   return finalContent;
 }
+async function runToolLoopWithCustomInference(messages, projectContext, options, loop, onChunk) {
+  const infer = options.customToolInference;
+  const maxTokens = options.maxTokens ?? 32768;
+  const maxRounds = loop.maxRounds ?? 4;
+  const systemInstruction = `${options.systemPersona || "You are Curriculum OS Assistant."}
+
+Project Context:
+${projectContext}
+
+Guidelines:
+1. Answer precisely using the provided tools when they help (project status, deep research, production declaration, gate approval).
+2. Keep Markdown formatting. Never invent tool results.`;
+  const convo = [
+    { role: "system", content: systemInstruction },
+    ...messages.map((m) => ({ role: m.role === "assistant" || m.role === "model" ? "assistant" : "user", content: m.content }))
+  ];
+  let finalContent = "";
+  for (let round = 0; round < maxRounds; round++) {
+    let turn;
+    try {
+      turn = await infer({ messages: convo, tools: loop.tools, temperature: options.temperature ?? 0.3, maxTokens, onChunk: onChunk ?? (() => {
+      }) });
+    } catch (err) {
+      const msg = String(err?.message || err);
+      if (/tool/i.test(msg) && /40[04]|400|not support|invalid/i.test(msg)) {
+        console.warn("[toolLoop] custom gateway rejected tools \u2014 degrading to plain chat:", msg.slice(0, 120));
+        const plain = await infer({ messages: convo, tools: [], temperature: options.temperature ?? 0.3, maxTokens, onChunk: onChunk ?? (() => {
+        }) });
+        finalContent = plain.content;
+        break;
+      }
+      throw err;
+    }
+    const { content, toolCalls } = turn;
+    if (toolCalls.length === 0) {
+      finalContent = content;
+      break;
+    }
+    if (content) finalContent = content;
+    convo.push({ role: "assistant", content: content || null, tool_calls: toolCalls.map((t) => ({ id: t.id, type: "function", function: { name: t.name, arguments: t.arguments } })) });
+    for (const call of toolCalls) {
+      let parsedArgs = {};
+      try {
+        parsedArgs = JSON.parse(call.arguments || "{}");
+      } catch {
+      }
+      loop.onToolStart?.(call.name, parsedArgs);
+      let resultPayload;
+      let isError = false;
+      try {
+        resultPayload = await loop.executor(call.name, parsedArgs);
+      } catch (err) {
+        isError = true;
+        resultPayload = { error: err?.message || String(err) };
+        loop.onToolError?.(call.name, err?.message || String(err));
+      }
+      const exec = { toolCallId: call.id, name: call.name, result: resultPayload, isError };
+      loop.onToolResult?.(exec);
+      convo.push({ role: "tool", tool_call_id: call.id, content: JSON.stringify(resultPayload).slice(0, 12e3) });
+    }
+    if (round === maxRounds - 1) {
+      const last = await infer({ messages: convo, tools: [], temperature: options.temperature ?? 0.3, maxTokens, onChunk: onChunk ?? (() => {
+      }) });
+      finalContent = last.content;
+    }
+  }
+  return finalContent;
+}
 
 exports.DEFAULT_ARTIFACT_STREAM_IDLE_MS = DEFAULT_ARTIFACT_STREAM_IDLE_MS;
 exports.DEFAULT_ARTIFACT_STREAM_TOTAL_MS = DEFAULT_ARTIFACT_STREAM_TOTAL_MS;