@thanh01.pmt/curriculum-kit 1.0.16 → 1.0.17

package/dist/index.d.cts

@@ -407,12 +407,15 @@ interface ExtractedStreamChunk {
  * Pure function so the extraction contract is unit-testable.
  *
  * Reasoning arrives through provider-specific shapes:
- * - `reasoning-delta` / legacy `reasoning` parts (mapped by SDK providers),
- * - `raw` parts (emitted when `includeRawChunks: true`) whose `rawValue` is the
- *   provider's OpenAI-compatible SSE chunk: NVIDIA NIM & DeepSeek put thinking
- *   in `choices[0].delta.reasoning_content`, OpenRouter in `choices[0].delta.reasoning`.
- * Without this extraction the entire thinking phase renders as a blank UI
- * (only text-delta reaches the caller) — the observed ~30s silent wait.
+ * - `reasoning-delta` / legacy `reasoning` parts (mapped by SDK providers).
+ *
+ * DUP-STREAM FIX (measured live against NVIDIA NIM nemotron-ultra, 2026-09-10):
+ * with `includeRawChunks: true` the SDK emits BOTH the mapped `text-delta`
+ * /`reasoning-delta` part AND the `raw` provider chunk carrying the SAME
+ * token. Extracting text from `raw` parts therefore duplicated every token
+ * ("Lớp 1ớp 1 (6-7 (6-7 tuổi)…"), destroying the JSON and causing the ~90%
+ * wizard prefill failure. The mapped parts are the single source of truth —
+ * `raw` parts must contribute NOTHING here.
  */
 declare function extractStreamChunk(part: any): ExtractedStreamChunk;
 declare function createStreamAbortSignal(budget: Required<StreamBudget>): StreamAbortHandle;

package/dist/index.d.ts

@@ -407,12 +407,15 @@ interface ExtractedStreamChunk {
  * Pure function so the extraction contract is unit-testable.
  *
  * Reasoning arrives through provider-specific shapes:
- * - `reasoning-delta` / legacy `reasoning` parts (mapped by SDK providers),
- * - `raw` parts (emitted when `includeRawChunks: true`) whose `rawValue` is the
- *   provider's OpenAI-compatible SSE chunk: NVIDIA NIM & DeepSeek put thinking
- *   in `choices[0].delta.reasoning_content`, OpenRouter in `choices[0].delta.reasoning`.
- * Without this extraction the entire thinking phase renders as a blank UI
- * (only text-delta reaches the caller) — the observed ~30s silent wait.
+ * - `reasoning-delta` / legacy `reasoning` parts (mapped by SDK providers).
+ *
+ * DUP-STREAM FIX (measured live against NVIDIA NIM nemotron-ultra, 2026-09-10):
+ * with `includeRawChunks: true` the SDK emits BOTH the mapped `text-delta`
+ * /`reasoning-delta` part AND the `raw` provider chunk carrying the SAME
+ * token. Extracting text from `raw` parts therefore duplicated every token
+ * ("Lớp 1ớp 1 (6-7 (6-7 tuổi)…"), destroying the JSON and causing the ~90%
+ * wizard prefill failure. The mapped parts are the single source of truth —
+ * `raw` parts must contribute NOTHING here.
  */
 declare function extractStreamChunk(part: any): ExtractedStreamChunk;
 declare function createStreamAbortSignal(budget: Required<StreamBudget>): StreamAbortHandle;

package/dist/index.mjs

@@ -13561,15 +13561,6 @@ function extractStreamChunk(part) {
     const content = part.text ?? part.delta ?? "";
     return content ? { content } : {};
   }
-  if (part.type === "raw") {
-    const raw = part.rawValue;
-    const delta = raw?.choices?.[0]?.delta;
-    const reasoning = delta?.reasoning_content ?? delta?.reasoning ?? raw?.delta?.reasoning_content ?? raw?.delta?.reasoning;
-    if (typeof reasoning === "string" && reasoning) return { thought: reasoning };
-    const text = delta?.content ?? raw?.delta?.content;
-    if (typeof text === "string" && text) return { content: text };
-    return {};
-  }
   return {};
 }
 function createStreamAbortSignal(budget) {