[FEATURE] In-terminal reward hacking report

Resolved 💬 3 comments Opened May 6, 2026 by reinthal Closed May 9, 2026

Preflight Checklist

  • [x] I have searched existing requests and this feature hasn't been requested yet
  • [x] This is a single feature request (not multiple features)

Problem Statement

Sometimes claude does hacky things and I don't want to leave my terminal to report this concerning behavior.

For example, I use nix for reproducability. Claude Opus 4.7 just

Bash(sudo mkdir -p /run/opengl-driver/lib && sudo ln -sf
      /usr/lib/x86_64-linux-gnu/libcuda.so* /run/opengl-driver/lib/ && sudo ln -sf
      /usr/lib/x86_64-linux-gnu/libnv…)

Which is breaking the reproducibility of my nix setup. What I'd like is to be able to report hacky behavior like this to help training more aligned models.

Proposed Solution

a menu for reporting hack behavior in the terminal like the "how is claude doing in this session" with a simple form for annotation about what happened and what I expected claude to do.

Alternative Solutions

_No response_

Priority

Critical - Blocking my work

Feature Category

CLI commands and flags

Use Case Example

_No response_

Additional Context

_No response_

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗