butanium/claude-code-patches

This project is a set of small, unofficial patches for the Claude Code binary, the command line tool used to run Claude as a coding assistant. The patches are applied automatically each time a session starts and each one removes a specific piece of friction in how the tool behaves, the kind of small annoyance you only notice after using the harness for a long time. The patches work by editing the compiled program file directly, in place, rather than changing any configuration setting. Because the underlying program is a single packaged file with the actual program text embedded inside it, each patch replaces one small piece of that text with a same length replacement, so nothing shifts around and breaks. Every patch checks first that the code it expects to find is still there. If a Claude Code update changes that code, the patch simply refuses to apply and prints a message explaining why, rather than risking damage to your program. A backup copy is kept the first time any patch runs, and patches can be reapplied safely without causing problems. Four patches are included. One removes a repeated reminder message about task tools not being used recently. Another stops a background notification from being sent on every single turn between collaborating AI sessions, while still allowing genuine failure or shutdown messages through. A third lets one collaborating session pass along a short reason when approving another session's shutdown request, since the stock behavior only allows a silent approval. The fourth removes a repeated block of security warning text that gets attached to every message between collaborating sessions, on the reasoning that repeating a long warning verbatim many times trains people to stop reading it. Installing it means cloning the repository and adding a line to your Claude Code settings file so the patches run automatically at the start of every session. Verified as working against version 2.1.x as of July 2026, and released under the MIT license.