PRODUCTION SCENARIO
Your autonomous coding agent needs several context windows to port a 300-test service. The first window has already written a tests.json tracker and an init.sh setup script. Each later window is handed a compacted summary of the one before it, and those windows keep re-implementing modules that already passed, or deleting tests that were failing.
How should the later windows be started?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Start each later window fresh and let it recover state from tests.json, init.sh, and the git log
When a context window is cleared, a brand new window beats a compacted summary, because the latest models recover state effectively from the local filesystem and git: the tracker file, the setup script, and the commit history. Removing or editing tests so that progress looks better is never acceptable.