Skip to content

v0.49.65

Choose a tag to compare

@github-actions github-actions released this 19 Nov 19:45
· 35 commits to master since this release

v0.49.65

This patch makes Magic Context's history summaries much cheaper on projects with many memories, on every host (issue 853).

Upgrading

Update the plugin and restart OpenCode or Pi. No database migration is needed.

Fixed

  • Summary runs sent the whole project memory with every request (issue 853, thanks @galact-byte for the detailed logs and the earlier probe): each summary request carried every active memory of the project, with no size limit. On a project with a few hundred memories that came to more than 100,000 tokens per run, while the history being summarized was often a few hundred tokens. The summarizer now gets the same budgeted set of memories that the main session already shows. In a measured run with 480 memories, a request fell from about 147,000 tokens to about 26,000.
  • Summary requests reused almost nothing from the provider's cache: the example summaries, chosen anew for each run, and the references to recent summaries came before the large memory block, so two runs in a row shared almost nothing at their start. The memory block now comes first, and the examples stay the same for a whole session. In the same measurement, the part two runs in a row share grew from 63 bytes to about 48,000.
  • A handful of short messages could start a summary run: below the force band, a run started whenever 12 messages were waiting, whatever their size. That test is gone: a run now needs enough text or tool output to be worth summarizing. Overflow recovery near the context limit, the force band and the manual commands are unchanged.
  • The logs now say why each summary run started and how large each part of its request was, as sizes only, never text.