Repository navigation
xbbg 1.6.1
Changed
- The Python LangGraph adapter requires xbbg 1.6.0 or newer.
py-xbbg-langgraphnow declaresxbbg>=1.6.0,<2, the first release with theadelay_status()function andon_delayedrequest parameter behind itsxbbg_delay_statustool andon_delayedarguments; xbbg 1.5 has neither. - DuckDB tests run on Python 3.15. duckdb 2.0 publishes stable-ABI wheels that install on CPython 3.15, so the
testextra no longer leaves duckdb out there, the 3.15 CI jobs run the DuckDB tests, andxbbg[duckdb]on 3.15 installs duckdb 2.0 from a wheel instead of needing a source build. Theuv.lockpin moves from duckdb 1.5.6 to 2.0.0. - BREAKING (Rust): typed field hints are parsed once during request preparation.
ArrowType::parseis fallible, response-state constructors consumeHashMap<String, ArrowType>hints, andTypedBuilder::append_valuereturns a conversion-miss flag instead of discarding it.
Fixed
- Stub generation no longer depends on the default text encoding, and no longer breaks TOML config files.
generate_stubs(),generate_ta_stubs(), andconfigure_ide_stubs()read and wrote files in the platform's default encoding. Python 3.15 made that UTF-8 (PEP 686), but on Windows with Python 3.11–3.14 it is the ANSI code page, so stub modules generated from schema descriptions with characters such as’failed to import or to write, and editing.vscode/settings.json,pyrightconfig.json,pyproject.toml, orty.tomlcould garble existing non-ASCII entries, raise, or silently skip the file. These files are now always read and written as UTF-8. The stubs path inserted intopyproject.tomlorty.toml, and shown in the manual setup instructions, is now quoted: Windows backslashes were written unescaped, so on Windows the file stopped parsing as TOML on every Python version. - Text-delivered datetime fields no longer disappear after metadata resolution (#363). Generic
Datetimemetadata with a Character/String field kind resolves to string, including the existing textual aliases. The field-cache schema is bumped so persisted incorrect timestamp hints are discarded. - Field type errors are observable (#364). Unknown override types raise validation errors instead of silently selecting string. Failed non-null conversions retain the declared schema and null result but emit one value-free request warning and record per-field counts in
xbbg.type_misses. Streams carry batch deltas, shard aggregation preserves totals, native table composition sums supplied batch diagnostics, and Python native/pandas conversion paths retain the metadata. Public field resolver APIs normalize native validation errors to the publicBlpValidationError. - Related conversions no longer wrap or overflow silently. Integer narrowing and floating-point integer bounds are checked; date-to-timestamp conversion rejects overflow, pre-epoch timestamp-to-date conversion floors correctly, and time-of-day values outside the valid day range are rejected and counted.
- Python LangGraph compatibility on Python 3.14. The adapter bounds LangSmith below 0.9 on Python 3.14: newer releases call a deprecated asyncio API during import, failing applications and tests that treat deprecation warnings as errors. Python 3.11–3.13 resolution is unchanged.
Full Changelog: v1.6.0...v1.6.1