personal memory agent
Fix follow-up image lifecycle in describe.py vision analysis
The follow-up image was being closed immediately after creating async
requests, causing "Operation on closed image" errors when the batch
later tried to process them. This broke category-specific content
extraction (browsing, meeting, etc.) leaving empty content fields.
- Store follow-up images in frame_images dict keyed by frame_id
- Close image only when all follow-ups for that frame complete
- Removes stale references issue from multiple request objects
Bug was introduced in 03666da (Dec 24, 2025) when conditional close
logic was changed to unconditional.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
+10
-1
+10
-1
observe/describe.py
+10
-1
observe/describe.py
···
427
427
428
428
# Track frames by frame_id for merging follow-up results
429
429
frame_results = {} # frame_id -> result dict
430
+
frame_images = {} # frame_id -> PIL Image (for follow-up cleanup)
430
431
431
432
# Stream results as they complete, with retry logic
432
433
async for req in batch.drain_batch():
···
526
527
full_img = Image.open(io.BytesIO(req.frame_bytes))
527
528
req.pending_follow_ups = len(follow_ups)
528
529
530
+
# Store image in frame_images for cleanup (single reference)
531
+
frame_images[req.frame_id] = full_img
532
+
529
533
# Close initial image since DESCRIBE is complete
530
534
if hasattr(req, "initial_image") and req.initial_image:
531
535
req.initial_image.close()
···
570
574
f"{', '.join(cat for cat, _ in follow_ups)}"
571
575
)
572
576
573
-
full_img.close()
577
+
# Don't close full_img here - requests are async and need it open
574
578
continue # Don't output yet, wait for follow-ups
575
579
576
580
# Handle follow-up completion for parallel requests
···
615
619
616
620
# Clean up frame_results entry
617
621
del frame_results[req.frame_id]
622
+
623
+
# Close follow-up image now that all requests are done
624
+
if req.frame_id in frame_images:
625
+
frame_images[req.frame_id].close()
626
+
del frame_images[req.frame_id]
618
627
619
628
# Aggressively clear heavy fields
620
629
req.frame_bytes = None