Reframe the summarization prompt as creating a handoff for a different assistant
(rather than self-summarization) to produce more thorough, explicit summaries.
Add a pre-turn compaction check so context is trimmed before building the LLM
request, not just after. Extract recent user messages from discarded turns during
compaction so the model retains the user's actual words alongside tool call/result
pairs.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
f459aa78d7