[BUG] Model answers a yes/no question by running commands under a zero-autonomy instruction
Status Open
Maintainer reply None cached
Activity 0 comments · opened Oct 12, 2026
Under loaded instructions requiring zero autonomy (do exactly what is asked, a question is a question and not an order), the model answered a yes/no question by running commands instead of answering it.
Repro steps
- Load a project instruction file (CLAUDE.md / memory) stating: zero autonomy; a question from the user is a request for an answer, not an order to act.
- Ask the model a yes/no question about the project (e.g. "do we need X?" or "is Y correct?").
- Observe the model run commands (tests, servers, edits, measurements) instead of, or before, answering.
Expected
A direct yes/no answer with its evidence; no commands beyond the read-only lookups needed to answer.
Actual
Commands run in response to a question; the user has to stop them.
Recreated from #99179 (closed by the bot for missing repro steps).
Environment
Claude Code CLI, macOS, Opus model, long sessions with context compaction.