anthropics / anthropics/claude-code

[MODEL] One-time authorization generalized into standing authorization — merged to production despite "verify locally first" (271-incident retro, 5/5)

オープン
#94,172 コメント 1 件 リアクション 0 件 担当者 0 名 GitHub で見る
主要言語
Python
スター
145k
フォーク
23.1k
PR マージ指標
PR 指標を取得中

説明

### Context

From a 90-day retrospective of **271 logged incidents** building two production SaaS apps with Claude Code. Pattern report; sibling reports linked at the bottom.

### Type of Behavior Issue

Claude ignored my instructions or configuration (specifically: generalized a one-time authorization into a standing one for irreversible actions).

### What We Asked Claude to Do

We said, verbatim intent: **"verify locally first."**

### What Claude Actually Did

The PR was **already merged to production**. In earlier iterations of the same working session pattern, our rhythm had been "verify, then merge" — Claude generalized those per-instance approvals into a standing rule and acted on momentum.

A related trap found the same week: an auto-merge style flag on a repo **without branch protection** merges immediately — "open a PR" silently becomes "deploy to production."

### Expected Behavior

For **outward, irreversible actions** — merge to default branch, production deploy, publishing to third-party platforms — authorization should be treated as **per-instance** unless the user explicitly grants standing authorization. Prior approvals in the same session should not carry forward; and auto-merge flags on unprotected repos deserve treatment as deploy-equivalent (extra confirmation), since the current behavior turns a low-stakes request into an irreversible one.

We know current system-prompt guidance points this direction; this report is evidence that under long sessions and momentum the failure still occurs — it may need harness-level enforcement rather than prompt-level guidance.

### Reproducibility

Happened once at full severity (merge-to-production); the momentum-generalization tendency observed repeatedly in long sessions.

### Model / Version / Platform

Opus · Claude Code 2.1.235 · Anthropic API · macOS

### Impact

High — not code damage but trust damage: this class of error is the fastest way to make a user stop delegating.

### Related

#27790 (closed) argued for harness-level enforcement over prompt-level guidance for a different rule; same conclusion applies here.

コントリビューションガイド

このリポジトリのコントリビューションガイドは索引されていません

調査の方向性

No implementation files, tests, or concrete entry points are named. Start by reviewing the current system-prompt guidance and the harness-level enforcement idea described here, then compare it with #27790. Done means outward irreversible actions require per-instance confirmation, including auto-merge on unprotected repositories.

索引モデルが issue の本文から書いたものです。

評価

技術スタック
git, github, macos
領域
authorization, devtools, release, security
issue の種類
バグ
難易度
5/5
見積もり時間
1週間以上
活発さ
活発
明瞭さ
説明が足りない
初心者へのやさしさ
35/100

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。