You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Better suggestions, measured with the new prompt eval (eval/) on Claude Haiku and GPT Luna:
The model now sees which files each tool touched, so suggestions name them ("add tests for slugify in
utils.py").
Follow-ups no longer ask again for something the agent said it cannot do, answer in the agent's voice, or
suggest running tests in projects without tests; they follow Spanish voseo.
Completions are shorter (1 to 8 words) and no longer include notes about the completion itself.
Both prompts include a few examples and treat the conversation as data, not instructions.
Backticks at the ends of a suggestion are kept, so code like `fetchUser` is not cut.