Lucas Wang gn00295120

## CRITICAL_FINDING.md

      
              3 files
            
          
              0 forks
            
          
                0 comments
              
            
              0 stars
            
          
                gn00295120
                / CRITICAL_FINDING.md
            
            
              Last active
              October 24, 2025 02:39
            
              
                Twilio Audio Jittering: BEFORE vs AFTER Fix Comparison (PR #1926)
              
          
    ⚠️ CRITICAL FINDING - Test Results Contradict Recommendation

Summary

After implementing the recommended fix (320 bytes/20ms) and generating actual comparison audio, the test results show the fix performs WORSE than the current code.
Test Results

Current Code (BEFORE Fix): 400 bytes / 50ms

✅ Total flushes: 26


## compare_fix.sh
#!/bin/bash
# Simple Branch Comparison - Test main vs PR #1981
#
# Uses only Test 2 (Reasoning Token Exhaustion) - the confirmed working test
#
# Usage:
#   export GEMINI_API_KEY="your-key"
#   bash /tmp/compare_fix.sh

set -e

## GEMINI_EMPTY_CHOICES_RESEARCH.md

      
              7 files
            
          
              0 forks
            
          
                0 comments
              
            
              0 stars
            
          
                gn00295120
                / GEMINI_EMPTY_CHOICES_RESEARCH.md
            
            
              Created
              October 23, 2025 18:43
            
              
                Gemini Empty Choices Bug Tests - PR #1981 Comprehensive Test Suite
              
          
    Gemini Empty Choices Bug - Research Findings

摘要

研究發現 Gemini 確實會在多種情況下返回空的 choices/candidates 陣列，這證實了 Issue #1559 報告的問題是真實存在的，PR #1981 的防禦性修復是必要的。

已確認的觸發條件


## test_advanced_sqlite_concurrent.py
"""
Test concurrent access to AdvancedSQLiteSession to verify lock protection.

This test verifies the fix for Codex's P1 review comment:
- _get_current_turn_number(), _get_next_turn_number(), _get_next_branch_turn_number()
  and list_branches() now properly acquire self._lock when accessing shared connection.

Without proper locking, concurrent operations would cause:
- sqlite3.ProgrammingError: Recursive use of cursors not allowed
- Data inconsistency

## sqlite_stress_test_final.py
"""
Comprehensive stress test for SQLiteSession with file descriptor monitoring.

This test demonstrates the fix for two critical bugs:
1. threading.Lock() being created anew on every operation (no thread safety)
2. File descriptor leak from thread-local connections in ThreadPoolExecutor

Test setup:
- 30 concurrent writers + 10 concurrent readers
- 60 seconds duration

## medium_test.py
"""
Medium stress test - 15 workers, 30 seconds
"""
import asyncio
import threading
import time
import sys
from pathlib import Path

sys.path.insert(0, str(Path(__file__).parent / "src"))

## test_raw_sqlite_threading.py
#!/usr/bin/env python3
"""
Test SQLite file-based database threading behavior.
Verifies that thread-local connections + nullcontext() is safe for concurrent access.
"""
import asyncio
import sqlite3
import tempfile
import threading
from contextlib import nullcontext
	#!/bin/bash
	# Simple Branch Comparison - Test main vs PR #1981
	#
	# Uses only Test 2 (Reasoning Token Exhaustion) - the confirmed working test
	#
	# Usage:
	# export GEMINI_API_KEY="your-key"
	# bash /tmp/compare_fix.sh

	set -e
	"""
	Test concurrent access to AdvancedSQLiteSession to verify lock protection.

	This test verifies the fix for Codex's P1 review comment:
	- _get_current_turn_number(), _get_next_turn_number(), _get_next_branch_turn_number()
	and list_branches() now properly acquire self._lock when accessing shared connection.

	Without proper locking, concurrent operations would cause:
	- sqlite3.ProgrammingError: Recursive use of cursors not allowed
	- Data inconsistency
	"""
	Comprehensive stress test for SQLiteSession with file descriptor monitoring.

	This test demonstrates the fix for two critical bugs:
	1. threading.Lock() being created anew on every operation (no thread safety)
	2. File descriptor leak from thread-local connections in ThreadPoolExecutor

	Test setup:
	- 30 concurrent writers + 10 concurrent readers
	- 60 seconds duration
	"""
	Medium stress test - 15 workers, 30 seconds
	"""
	import asyncio
	import threading
	import time
	import sys
	from pathlib import Path

	sys.path.insert(0, str(Path(__file__).parent / "src"))
	#!/usr/bin/env python3
	"""
	Test SQLite file-based database threading behavior.
	Verifies that thread-local connections + nullcontext() is safe for concurrent access.
	"""
	import asyncio
	import sqlite3
	import tempfile
	import threading
	from contextlib import nullcontext