The agent loop stops after a tool result and never sends the reply. The answer is already in the tool output. A follow-up message such as “??” starts a new turn, and the reply comes back immediately from that same result.
What happens
- The agent runs a tool (read, search, shell, and so on).
- The tool returns the information needed to answer.
- The turn ends. No assistant message is shown.
- The user has to send another message. That new turn answers from the result that was already available.
Examples from one session
Project: a local Swift package (ChessCore), macOS, Cursor agent chat.
- Asked whether en passant repetition had tests. The test files had already been read. The turn ended. “stuck?” produced the answer.
- Asked how en passant state relates to FEN. The relevant source had already been read. The turn ended. Several pings (“do it now”, “R U THERE”) were required.
- Asked to run the rewrap and reorder skills. The scripts finished, SwiftFormat and SwiftLint were clean, and all 41 tests passed. The turn ended on the git diff. “?” / “??” produced the summary.
- Asked why BoardTests changed from a struct to an enum. SwiftFormat’s default enumNamespaces rule had already been identified. The turn ended. “??” produced the answer.
Expected
After every tool result, resume the agent until it writes a user-visible message. A project rule cannot fix this, because the model is not invoked again until the user sends another message.
Environment
- macOS (darwin 25.5)
- Cursor agent chat in the IDE
- Model in that session: Grok 4.7
- Reproduced across several turns in the same chat on 2026-10-14
