## Summary
- **Phase 4 — Error Paths** (`test_4_error_paths.py`): 28 cases across 8
categories that send intentionally invalid SQL and assert 4xx not 5xx.
Self-contained, runs in ~24s without data dependency.
- **Phase 5 — Concurrency Stress** (`test_5_concurrency.py`): 26 queries
(2 per category) run simultaneously at 5/10/15 worker levels via
ThreadPoolExecutor. Each thread gets its own client. Catches deadlocks
and thread-safety bugs.
- **EXPLAIN ANALYZE** (`test_explain_analyze.py`): Verifies
EXPLAIN/EXPLAIN ANALYZE return 200 with plan output and no
`partial_error`. Also validates EXPLAIN ANALYZE on cross-schema
dot-notation queries returns 4xx (not 500 crash).
- **Fix**: `load_all_queries()` now skips JSON files without a
`"queries"` key.
## Test plan
- [x] Error path tests pass on pentest (28/28, all return 400)
- [ ] CI run: `rye run pytest tests/query_agent/test_4_error_paths.py
-v`
- [ ] CI run: `rye run pytest tests/query_agent/test_5_concurrency.py
-v`
- [ ] CI run: `rye run pytest tests/query_agent/test_explain_analyze.py
-v`
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
