Skip to content

Releases: FelipeFariasAlfaro/glia

v0.6.2

Choose a tag to compare

@FelipeFariasAlfaro FelipeFariasAlfaro released this 19 Oct 01:17

Added

  • glia selftest command that verifies a build end to end: SQLite, the embedding backend import, an init/scan/learn/recall/doctor round trip in a temporary workspace, and a real MCP stdio handshake against the same executable. --with-model additionally loads the configured embedding model and checks that it returns 1024-dimensional vectors.
  • Reusable Build binaries workflow that can be run manually to validate all six binaries before tagging a release.

Fixed

  • Fixed the macOS binaries, which crashed at startup with AttributeError: module 'pkg_resources' has no attribute 'NullProvider'. The binaries no longer bundle pkg_resources or setuptools, so PyInstaller's pkg_resources runtime hook is never included, and the benchmark tooling inside onnxruntime that imported it is excluded.
  • Fixed glia init reporting "already initialized" when .glia/ existed but contained no memory yet, which broke the recommended layout of placing the binary at .glia/glia before initializing.
  • The full Linux binary now bundles CPU-only PyTorch. The CUDA build produced a 2.8 GiB executable, beyond GitHub's 2 GiB release-asset limit.
  • Binaries no longer crash on emoji output in consoles that do not use UTF-8, such as Windows cp1252 pipes.

Changed

  • Updated PyInstaller to 6.22.3 and pinned pyinstaller-hooks-contrib.
  • The light binary is smaller (about 71 MB on Linux, previously 114 MB) because PyTorch, ONNX tooling, and build tools are no longer collected.
  • Every binary is now verified with glia selftest, including real embedding inference, before release. PyPI publication waits for all six binaries to pass, so a binary failure no longer consumes a PyPI version.

Validation

  • Full test suite: 46 passed on Python 3.11 (the first full run with a working SQLite build).
  • Linux light and full binaries were built locally and passed glia selftest --with-model outside the build environment, with no Python on the PATH: real ONNX and PyTorch inference returned 1024-dimensional vectors, and the MCP handshake listed 9 tools.
  • The previous macOS failure was reproduced locally by leaving a stale pkg_resources package in the build environment; the new specs build cleanly in that state.

0.6.1

Choose a tag to compare

@FelipeFariasAlfaro FelipeFariasAlfaro released this 18 Oct 22:52

Full Changelog: v0.6.0...v0.6.1

Versión 0.3.0

Choose a tag to compare

@FelipeFariasAlfaro FelipeFariasAlfaro released this 24 May 23:04

Se añade una mejora al proceso watch para que funciona como proceso en segundo plano identificando los cambios en el codigo y registrandolos

Versión 0.2.0

Choose a tag to compare

@FelipeFariasAlfaro FelipeFariasAlfaro released this 24 May 18:18
v0.2.0

Release v0.2.0 - multi-provider support, glia.env config