@thanh01.pmt/curriculum-kit 1.4.32 → 1.4.34

package/dist/index.cjs

@@ -991,6 +991,52 @@ var init_streamRunner = __esm({
     exports.DEFAULT_ARTIFACT_STREAM_TOTAL_MS = 48e4;
   }
 });
+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 });
+  };
+}
 function getAIModel(options = {}) {
   const designated = getDesignatedFallbackConfig();
   const defaultProvider = process.env.DEFAULT_AI_PROVIDER || process.env.LLM_PROVIDER || detectDefaultProvider();
@@ -1009,7 +1055,8 @@ 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");
   }
@@ -29472,7 +29519,7 @@ function safeParseJson(rawText) {
   }
   return null;
 }
-async function streamLLMWithFallback(systemPrompt, userPrompt, apiKeys = {}, onChunk, preferredModel, preferredProvider, budget, signal) {
+async function streamLLMWithFallback(systemPrompt, userPrompt, apiKeys = {}, onChunk, preferredModel, preferredProvider, budget, signal, onEvent) {
   const alibabaKey = apiKeys.alibaba || apiKeys.dashscope || process.env.ALIBABA_API_KEY || process.env.DASHSCOPE_API_KEY;
   const nvidiaKey = apiKeys.nvidia || process.env.NVIDIA_API_KEY;
   const deepseekKey = apiKeys.deepseek || process.env.DEEPSEEK_API_KEY;
@@ -29553,12 +29600,14 @@ THINKING DIRECTIVE: Do ALL your reasoning internally BEFORE emitting output. Do
     idleMs: budget?.idleMs ?? DEFAULT_STREAM_IDLE_MS,
     totalMs: budget?.totalMs ?? DEFAULT_STREAM_TOTAL_MS
   };
-  for (const candidate of candidates) {
+  for (const [candidateIdx, candidate] of candidates.entries()) {
     if (signal?.aborted) {
       console.log(`[curriculum-kit:VercelAI] AbortSignal already aborted, cancelling candidate loop.`);
       break;
     }
     const t0 = Date.now();
+    const attempt = candidateIdx + 1;
+    onEvent?.({ type: "candidate-start", provider: candidate.provider, model: candidate.model, attempt, total: candidates.length, elapsedMs: 0 });
     const abort = createStreamAbortSignal(resolvedBudget, signal);
     try {
       const modelInstance = getAIModel({
@@ -29566,18 +29615,35 @@ THINKING DIRECTIVE: Do ALL your reasoning internally BEFORE emitting output. Do
         modelName: candidate.model,
         apiKey: candidate.apiKey
       });
+      const wireFetch = (input, init) => fetch(input, init).then((res) => {
+        if (!res.body) return res;
+        const kicked = res.body.pipeThrough(
+          new TransformStream({
+            transform(chunk, ctrl) {
+              abort.kick();
+              ctrl.enqueue(chunk);
+            }
+          })
+        );
+        return new Response(kicked, res);
+      });
       const streamResult = ai.streamText({
         model: modelInstance,
         system: systemInstructions,
         prompt: userPrompt,
         temperature: 0.2,
         includeRawChunks: true,
-        abortSignal: abort.signal
+        abortSignal: abort.signal,
+        fetch: wireFetch
       });
       let fullContent = "";
       const extract = createStreamChunkExtractor();
       for await (const part of streamResult.fullStream) {
         abort.kick();
+        if (part?.type === "error") {
+          const errText = part.errorText ?? String(part.error ?? "unknown stream error");
+          throw new Error(`[${candidate.provider}] stream error part: ${errText}`);
+        }
         const extracted = extract(part);
         if (extracted.thought) onChunk?.(extracted.thought, "thought");
         if (extracted.content) {
@@ -29587,10 +29653,14 @@ THINKING DIRECTIVE: Do ALL your reasoning internally BEFORE emitting output. Do
       }
       if (fullContent.trim()) {
         console.log(`[curriculum-kit:VercelAI] \u2705 ${candidate.provider} (${candidate.model}) streamed in ${Date.now() - t0}ms`);
+        onEvent?.({ type: "candidate-success", provider: candidate.provider, model: candidate.model, attempt, total: candidates.length, elapsedMs: Date.now() - t0 });
         return fullContent.trim();
       }
+      console.warn(`[curriculum-kit:VercelAI] Candidate ${candidate.provider} (${candidate.model}) returned EMPTY content in ${Date.now() - t0}ms \u2192 trying next candidate.`);
+      onEvent?.({ type: "candidate-failed", provider: candidate.provider, model: candidate.model, attempt, total: candidates.length, elapsedMs: Date.now() - t0, error: "empty content" });
     } catch (e) {
       console.warn(`[curriculum-kit:VercelAI] Candidate ${candidate.provider} (${candidate.model}) failed in ${Date.now() - t0}ms:`, e?.message || e);
+      onEvent?.({ type: "candidate-failed", provider: candidate.provider, model: candidate.model, attempt, total: candidates.length, elapsedMs: Date.now() - t0, error: e?.message || String(e) });
       const isClientAborted = Boolean(
         signal?.aborted || e?.name === "AbortError" && signal?.aborted || e?.message && (e.message.includes("Controller is already closed") || e.message.includes("The operation was aborted") || e.message.includes("Aborted by parent signal"))
       );
@@ -29650,7 +29720,8 @@ Return concise JSON matching:
       model: options.model,
       provider: options.provider
     }),
-    options.signal
+    options.signal,
+    options.onProviderEvent
   );
   let parsedResearch = {
     groundTruthSummary: `Verified ecosystem standards for ${projectName}`,
@@ -29732,6 +29803,23 @@ async function streamLayerPrefillWithFallback(options, onChunk) {
     const research = await conductPreliminaryResearch(options, onChunk);
     return { success: true, research };
   }
+  const runWithCustomInference = async (systemInstruction, userPrompt2) => {
+    if (!options.customInference) return "";
+    return options.customInference({
+      systemInstruction,
+      userPrompt: userPrompt2,
+      messages: [{ role: "user", content: userPrompt2 }],
+      temperature: 0.2,
+      // Layer 3 synthesizes full context — give reasoning models the same
+      // generous ceiling the artifact paths use (32k) so JSON is never truncated.
+      maxTokens: 32768,
+      onChunk: (token, type) => {
+        if (type === "content") rawContent += token;
+        if (type === "content" || type === "thought") onChunk?.(token, type);
+      }
+    });
+  };
+  let rawContent = "";
   const {
     projectName = "Curriculum Course",
     projectCode = "curriculum-course",
@@ -29843,27 +29931,31 @@ FULL ACCUMULATED CONTEXT (ALL CONFIRMED PARAMETERS FROM EVERY PREVIOUS STEP \u20
 ${buildRemainingContextBlock(accumulatedData) || "(no additional context)"}
 
 REQUIREMENT: Synthesize ALL parameters above \u2014 every single line in FULL ACCUMULATED CONTEXT is a confirmed user decision or system constraint and MUST influence the generated configuration. Produce the high-fidelity pre-fill configuration for LAYER ${targetLayer}.`;
-  let rawContent = "";
-  await streamLLMWithFallback(
-    systemPrompt,
-    userPrompt,
-    apiKeys,
-    (chunk, type) => {
-      if (type === "content") rawContent += chunk;
-      onChunk?.(chunk, type);
-    },
-    options.model,
-    options.provider,
-    // RC-W1: layer-aware budget — Layer 3 synthesizes full context and free-tier
-    // models (especially NVIDIA NIM) may think/stream for minutes. Idle window still catches hangs.
-    resolveStreamBudget(targetLayer, {
-      idleMs: options.idleTimeoutMs,
-      totalMs: options.timeoutMs,
-      model: options.model,
-      provider: options.provider
-    }),
-    options.signal
-  );
+  if (options.customInference) {
+    rawContent = await runWithCustomInference(systemPrompt, userPrompt) || "";
+  } else {
+    await streamLLMWithFallback(
+      systemPrompt,
+      userPrompt,
+      apiKeys,
+      (chunk, type) => {
+        if (type === "content") rawContent += chunk;
+        onChunk?.(chunk, type);
+      },
+      options.model,
+      options.provider,
+      // RC-W1: layer-aware budget — Layer 3 synthesizes full context and free-tier
+      // models (especially NVIDIA NIM) may think/stream for minutes. Idle window still catches hangs.
+      resolveStreamBudget(targetLayer, {
+        idleMs: options.idleTimeoutMs,
+        totalMs: options.timeoutMs,
+        model: options.model,
+        provider: options.provider
+      }),
+      options.signal,
+      options.onProviderEvent
+    );
+  }
   const parsed = safeParseJson(rawContent);
   return {
     success: !!parsed,