Cline's compaction trigger was estimating tokens at 3 characters each and never firing on dense content

GitHub·high signalXBlueskyLinkedInCopy link

Cline v4.1.19 fixes long tasks running out of context instead of compacting. The trigger compared a character-based estimate of the conversation at roughly 3 characters per token against the model's context limit, so content that tokenizes far denser (disassembly, image dumps, minified sources) filled the real window while the estimate stayed under threshold, and the turn was squeezed down to a handful of output tokens. The trigger now also uses the token count the provider itself reports and scales retained history by how far off the estimate proved, and the summarizer's output budget doubled so a reasoning model can no longer spend the entire budget thinking and return no summary.