bloby-bot 0.47.3 → 0.47.4

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "bloby-bot",
-  "version": "0.47.3",
+  "version": "0.47.4",
   "releaseNotes": [
     "1. # voice note (PTT bubble)",
     "2. # audio file + caption",

package/supervisor/harnesses/pi/providers/stream-google.ts

@@ -8,6 +8,7 @@
  * Endpoint: POST {baseUrl}/models/{modelId}:streamGenerateContent?alt=sse&key={apiKey}
  * Stream:   SSE — each `data: {...}` is one candidate update.
  */
+import { log } from '../../../../shared/logger.js';
 import type {
   PiStreamRequest,
   PiStreamEvent,
@@ -149,9 +150,18 @@ export async function* streamGoogle(req: PiStreamRequest): AsyncIterable<PiStrea
   let lastFinish: string | undefined;
   let promptBlockReason: string | undefined;
   let usage: { inputTokens?: number; outputTokens?: number } | undefined;
+  // Debug counters — drop once this stabilises.
+  let chunkCount = 0;
+  let thoughtPartCount = 0;
+  let emptyTextPartCount = 0;
+  let firstChunkSummary = '';
 
   try {
     for await (const chunk of parseSse(res)) {
+      chunkCount++;
+      if (chunkCount === 1) {
+        try { firstChunkSummary = JSON.stringify(chunk).slice(0, 600); } catch {}
+      }
       // The whole prompt can be rejected before we even get a candidate.
       if (chunk?.promptFeedback?.blockReason) {
         promptBlockReason = chunk.promptFeedback.blockReason;
@@ -161,10 +171,12 @@ export async function* streamGoogle(req: PiStreamRequest): AsyncIterable<PiStrea
       for (const part of parts) {
         // Thinking models emit reasoning parts with `thought: true`. They
         // shouldn't be shown to the user as part of the visible answer.
-        if (part?.thought) continue;
+        if (part?.thought) { thoughtPartCount++; continue; }
         if (typeof part?.text === 'string' && part.text.length > 0) {
           accumulated += part.text;
           yield { type: 'text_delta', delta: part.text };
+        } else {
+          emptyTextPartCount++;
         }
       }
       if (candidate?.finishReason) lastFinish = candidate.finishReason;
@@ -185,6 +197,18 @@ export async function* streamGoogle(req: PiStreamRequest): AsyncIterable<PiStrea
     return;
   }
 
+  log.info(
+    `[pi/google] stream done — chunks=${chunkCount} text=${accumulated.length} ` +
+    `thoughtParts=${thoughtPartCount} emptyTextParts=${emptyTextPartCount} ` +
+    `finishReason=${lastFinish || 'none'} ` +
+    `promptTok=${usage?.inputTokens ?? '?'} outTok=${usage?.outputTokens ?? '?'}`,
+  );
+  if (chunkCount > 0 && !accumulated) {
+    log.info(`[pi/google] first chunk (truncated): ${firstChunkSummary}`);
+  } else if (chunkCount === 0) {
+    log.warn(`[pi/google] SSE stream parsed zero chunks — check response shape (status=${res.status} content-type=${res.headers.get('content-type') || ''})`);
+  }
+
   // Prompt-level block: nothing came back at all.
   if (promptBlockReason) {
     yield { type: 'error', error: `Gemini blocked the prompt (${promptBlockReason}).` };
@@ -199,7 +223,10 @@ export async function* streamGoogle(req: PiStreamRequest): AsyncIterable<PiStrea
     const reason = lastFinish && lastFinish !== 'STOP' && lastFinish !== 'FINISH_REASON_STOP'
       ? lastFinish
       : undefined;
-    yield { type: 'error', error: finishReasonMessage(reason) };
+    const hint = thoughtPartCount > 0 && !lastFinish
+      ? ' (model emitted thinking but never the final answer — try a non-thinking model like gemini-2.5-flash, or raise maxOutputTokens)'
+      : '';
+    yield { type: 'error', error: finishReasonMessage(reason) + hint };
     yield { type: 'done', stopReason: mapStopReason(lastFinish), usage };
     return;
   }