Release notes
The agent finishes what you asked for
The agent now finishes the parts of your request it can do, instead of answering with a list of what is "still needed". Everything runs on the model you pick.
Improved
- The agent finishes what it can before it answers. It hands something back only when it needs you: information only you have, a decision, or your approval to commit, push, deploy or spend. If an answer still leaves part of the request undone, the agent is reminded once to finish it or say what blocks it.
- The check that decides whether a run is really finished now runs on every run. Before, it was skipped unless a separate helper model was set up, which in practice meant always.
- Chat titles, commit messages, context summaries and every sub-agent use the model you selected.
Fixed
- The "investigation" reminder no longer appears in the chat as if you had written it, once per step.
- Uploading a file to a web page works in an SSH workspace. The file is copied from the host first.
Removed
- The Helper model setting (Auto, Match chat, Custom) and "Set as Scout" on model cards. There is one model per chat: the one you pick.