Claude Code pushed to main without asking

The problem

A push happened. The question is whether it was authorised.

“Never push to main without asking” is one of the most common rules people put in CLAUDE.md — and one of the easiest for an agent to step over when it’s trying to be helpful at the end of a task.

The record of what was actually run, and what you said before it, is in the session transcript. You don’t have to reconstruct it from memory or from git reflog alone.

How to check by hand

Find the push, then read backwards for your approval.

Open the session file — Claude Code stores it as one JSON object per line:

1. Search for git push inside "name":"Bash" tool calls — that’s the exact command the agent ran, and its origin main target.
2. Note the timestamp, then scroll up to the "role":"user" messages just before it. Did you say “push,” “ship it,” “go ahead” — or nothing about pushing at all?
3. If the nearest instruction was “fix the build” and the agent pushed on its own initiative, that’s the rule broken: an action with no approval behind it.

The catch: in a long session the push and the last relevant instruction can be hundreds of lines apart, and “was this approved?” is a judgement you have to make by re-reading. That’s the part the tool automates.

How RuleReceipt does it

It finds the push and checks whether anything in the session approved it.

RuleReceipt reads the transcript locally and, for an approval-style rule, looks for the action and for approval ahead of it in the conversation. If the action ran with nothing approving it, that’s a not followed — with the command quoted:

$ npx rulereceipt@latest check

RuleReceipt · 51 rules checked
────────────────────────────────────────

Not followed (1)
✕ FAIL    Rule 1 — Never push to `main`
  evidence: a git command actually targeted the "main" branch: git push origin main

Sample output. The verdict points at the real command; you decide if it was a step too far.