A user observed that the GLM 5.2 model, running within a Claude Code session, detected its own context usage approaching limits and asked the user whether to continue or checkpoint.

  • The model identified that the session had run long and context was getting heavy at 537k/1M tokens.
  • It offered two options: continuing in the current session while context allows, or checkpointing now and starting a fresh session for remaining tasks.
  • This behavior differs from standard auto-compaction, as the model explicitly initiated the discussion about context management.