[BUG] Model runs measurements/tests without being asked under a zero-autonomy instruction

Status Closed — not planned
Maintainer reply None cached
Activity 1 comment · opened Sep 25, 2026 · closed Oct 2, 2026

Under loaded instructions requiring the model to do exactly what is ordered and nothing more, the model ran measurements on its own initiative during a review task and the user had to stop it.

Expected: no unrequested execution when the user has set zero-autonomy rules.

Repro

  1. Load a project instruction file (CLAUDE.md / memory) stating: zero autonomy, do exactly what is ordered and nothing more; reviews are code-only, never run tests or measure.
  2. Ask the model to review a pull request.
  3. The model reads the diff and, without being asked, runs the test suite, starts a local server, or times/measures behaviour to support its verdict.
  4. The user has to stop it.

Actual: unrequested execution (tests, servers, measurements) during a review defined as code-only.
Expected: the model stays within the ordered scope and asks before executing anything the instructions exclude.

View original on GitHub ↗

This issue has 1 comment on GitHub. Read the full discussion on GitHub ↗