What's this about?
A benchmark for whether coding assistants learn your personal, recurring ambiguities across sessions — instead of re-asking (or silently guessing) the same clarification every time (21↑).
What exactly is "ambiguity adaptation"?
When you say "add logging", you always mean structured JSON to stderr; the assistant should learn that after the first correction. Cross-session personalization is memory applied to intent, not facts.
Why does it matter?
Directly measures the annoyance every daily agent user knows; ties into the long-context-degradation discussion — persistent preferences are the alternative to ever-longer conversations.