Highlights
- Added budget-aware retries with exact call and cost accounting.
- Added final-answer validation with correction feedback.
- Added typed success and failure results with JSONL export.
- Preserved partial trajectories and usage statistics for failed runs.
- Improved handling of null and malformed provider responses.
- Added reproducible 1M-character and multi-document benchmarks.
Validation
- 178 tests passing
- 93.5% coverage with branch tracking
- Verified across Python 3.9 to 3.12 on Linux, macOS, and Windows
See CHANGELOG.md and BENCHMARK_RESULTS.md for details.