Distill v2.0.30
Utility model saves more expensive tokens
- New
ask_stored_outputtool: ask the utility model about a stored or truncated tool output and get verbatim excerpts, instead of re-reading the whole file. - Large exact-output commands (
cat,sed -n,git show), documents and whole-file reads now go through extractive utility selection; the original is stored. - Stack traces, test output and HTML are reduced when the original is stored first.
- Subagent results are compressed by the utility model; resume instructions are always kept.
Fixes
- Output-reservation context overflows no longer end the turn on any model: the safety margin scales with the input, and a provider rejection lowers a learned per-route output cap and resubmits without compacting.