@thanh01.pmt/curriculum-kit 1.0.8 → 1.0.9

package/dist/{streamRunner-C7j5LKon.d.cts → streamRunner-BTTzbb-l.d.cts}

@@ -15,6 +15,25 @@ interface StreamRunnerOptions {
     maxTokens?: number;
     temperature?: number;
     reasoningEffort?: 'off' | 'low' | 'medium' | 'high' | 'max';
+    /**
+     * Host-provided inference override (Vercel AI SDK gateway, test stubs, …).
+     * When set, ALL built-in provider routes (raw-fetch NVIDIA/OpenRouter/
+     * DeepSeek/DashScope/Gemini + designated fallback chain) are bypassed and
+     * every inference call goes through this function instead. It receives the
+     * fully-formed system instruction (persona + project context + guidelines)
+     * and the user prompt, plus the same onChunk contract as the built-in
+     * routes: stream text deltas via ('content'|'thought'), and report token
+     * spend via onChunk(JSON.stringify({ promptTokens, completionTokens,
+     * totalTokens, reasoningTokens }), 'usage'). Return the full completion text.
+     */
+    customInference?: (request: {
+        systemInstruction: string;
+        userPrompt: string;
+        messages: ChatMessage[];
+        temperature: number;
+        maxTokens: number;
+        onChunk: (token: string, type: ChunkType) => void;
+    }) => Promise<string>;
 }
 /**
  * Checks if a specific AI provider is enabled via environment variables.

package/dist/{streamRunner-C7j5LKon.d.ts → streamRunner-BTTzbb-l.d.ts}

@@ -15,6 +15,25 @@ interface StreamRunnerOptions {
     maxTokens?: number;
     temperature?: number;
     reasoningEffort?: 'off' | 'low' | 'medium' | 'high' | 'max';
+    /**
+     * Host-provided inference override (Vercel AI SDK gateway, test stubs, …).
+     * When set, ALL built-in provider routes (raw-fetch NVIDIA/OpenRouter/
+     * DeepSeek/DashScope/Gemini + designated fallback chain) are bypassed and
+     * every inference call goes through this function instead. It receives the
+     * fully-formed system instruction (persona + project context + guidelines)
+     * and the user prompt, plus the same onChunk contract as the built-in
+     * routes: stream text deltas via ('content'|'thought'), and report token
+     * spend via onChunk(JSON.stringify({ promptTokens, completionTokens,
+     * totalTokens, reasoningTokens }), 'usage'). Return the full completion text.
+     */
+    customInference?: (request: {
+        systemInstruction: string;
+        userPrompt: string;
+        messages: ChatMessage[];
+        temperature: number;
+        maxTokens: number;
+        onChunk: (token: string, type: ChunkType) => void;
+    }) => Promise<string>;
 }
 /**
  * Checks if a specific AI provider is enabled via environment variables.

package/dist/workflow/index.cjs

@@ -429,6 +429,32 @@ Guidelines:
 3. Keep technical terminology and official API symbols intact.
 4. Format output using clean Markdown, clear sections, bullet points, and code blocks where required.
 5. Do NOT use emojis in headings, tab titles, or UI navigation labels.`;
+  if (options.customInference) {
+    try {
+      const text = await options.customInference({
+        systemInstruction,
+        userPrompt: messages.map((m) => m.content).join("\n\n"),
+        messages,
+        temperature,
+        maxTokens,
+        onChunk: onChunk ?? (() => {
+        })
+      });
+      if (text && text.trim()) return text;
+      throw createAiInferenceError({
+        errorCode: "ERR_AI_ALL_PROVIDERS_FAILED",
+        message: "customInference returned empty output",
+        rawError: { customInference: "empty response" }
+      });
+    } catch (e) {
+      if (e?.errorCode || e?.name === "AiInferenceError") throw e;
+      throw createAiInferenceError({
+        errorCode: "ERR_AI_ALL_PROVIDERS_FAILED",
+        message: `customInference failed: ${e?.message || e}`,
+        rawError: { customInference: String(e?.message || e) }
+      });
+    }
+  }
   const formattedMessages = [
     { role: "system", content: systemInstruction },
     ...messages.map((m) => ({