| Name | Modified | Size | Downloads / Week |
|---|---|---|---|
| Parent folder | |||
| manifest.json | 2026-10-01 | 1.7 kB | |
| timbal-2.7.31-linux-aarch64-gnu | 2026-10-01 | 1.1 MB | |
| timbal-2.7.31-linux-x86_64-gnu | 2026-10-01 | 1.2 MB | |
| timbal-2.7.31-macos-aarch64 | 2026-10-01 | 1.2 MB | |
| timbal-2.7.31-macos-x86_64 | 2026-10-01 | 1.2 MB | |
| timbal-2.7.31-windows-x86_64-gnu.exe | 2026-10-01 | 1.4 MB | |
| README.md | 2026-10-01 | 995 Bytes | |
| v2.7.31 source code.tar.gz | 2026-10-01 | 2.8 MB | |
| v2.7.31 source code.zip | 2026-10-01 | 3.3 MB | |
| Totals: 9 Items | 12.2 MB | 0 | |
What's Changed
- fix: scope fallback provider params per model entry (0d66a94a)
- Preserve recorded context usage for legacy conversations (ba501a8e)
- Describe compaction in terms of the context measurement (5ad61ad9)
- Pair each context measurement with the messages its request covered (87f507de)
- Salvage only the latest LLM span so a failed call does not duplicate the previous output (3e46d416)
- Cut the unsent estimate at the last model output and count the tail in every branch (5c753eae)
- Drop comments that restate the code (7773fbb2)
- Document the compaction signal: last call, unsent messages, run model window (6d74bbd4)
- Take the last LLM span directly and make the run-model test fail on a wrong model (b559952f)
- Count the messages sent after the last reported usage toward utilization (efab9320)
- Measure compaction utilization against the last LLM call and the run's model (7cf82519)
Full Changelog: https://github.com/timbal-ai/timbal/compare/v2.7.30...v2.7.31
