Skip to content

v3.3.1

Choose a tag to compare

View all tags
@Cerlancism Cerlancism released this 08 Apr 16:59
· 40 commits to main since this release
v3.3.1
a59da13

New Features

Stream Repetition Guard (-g, --guard-repetition)

A new option that monitors streaming responses for looping output and aborts the request before it wastes too many tokens.

  • -g, --guard-repetition <threshold> - minimum number of times a pattern must repeat before the stream is aborted. Defaults to 10. Set to 0 to disable.
  • On detection, the current batch is retried automatically.
  • Applied across all translation modes: plain, array, object, and timestamp.

Agent Planning: LLM-Based Summary Fitting

Planning pass summaries that fall outside the target token range are now re-fit using the llm-summary summarise function (two-phase draft -> fit), replacing the previous direct model call for consolidation. This applies to both per-window batch summaries and accumulated consolidation steps.

  • --no-fitting - skip LLM-based fitting for both scan-window summaries and consolidation. Summaries are used as-is regardless of token range.

Full Changelog: v3.2.0...v3.3.1

Assets 2
Loading