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