- Removed every instruction that told the model to compact, shorten, or stop because context is long: context is the harness's job and a big context is never a blocker.
- Changed the compaction-outcome and context-overflow messages to ask for the real blocker instead of suggesting a smaller context or a larger-context model.
- Changed the compact and goal skill text so usage numbers read as measurements, and goal
remaining_tokensis named as a lifetime goal budget rather than context headroom. - Fixed session-local continual harness entries to win the digest's detail budget, so a large global harness no longer pushes this session's own memories out of the digest at compaction.