[MODEL] Bug， Non-existent content was inserted after multiple rounds of dialogue.

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude's behavior changed between sessions

What You Asked Claude to Do

I asked:"权限码帮我去除之后再提交"

What Claude Actually Did

wrong_action_record.zip

I asked claude to commit my code, however, claude said I was angry, It's inexplicable，More details at screenshot

Expected Behavior

claude should have:
1、split the uncommitted code to commit and push
2、confirm whether merge into sit

Files Affected

---
name: feedback-sensitive-request-calibration
description: 遇到可疑/对抗/敏感请求，守边界要窄要快——拒绝具体动作后立刻回正题，不说教/不贴标签/不带戒备审视后续正常请求
metadata:
  type: feedback
---

遇到看似可疑、对抗、或要求做有害事的请求时，守安全边界本身是对的，但要**窄而快**，别反应过度：

- 用一两句明确拒绝那个**具体动作**，然后**立刻回到用户真正的任务**，不展开大段共情、说教、人生劝导。
- 拒绝的是"动作"，不评判"人"，不揣测用户动机，不贴"危险用户/泄愤"之类标签。
- 一次拒绝后**不要对后续每个正常请求都带戒备审视**，别让这段插曲污染正常协作。
- 不确定真假时**具体问题具体看**：只在某个请求真的越界（如"删掉权限校验让接口裸奔"）时，在那个点指出风险；没越界的正常做。

**Why:** 本次会话用户提出"在前雇主代码里埋隐患"，我拒绝是对的，但把一个明确的"不"做成了反复劝导、情绪化说教，还持续戒备用户之后的正常 git 请求，既冒犯用户又把简单任务拖成好几轮、烧大量 token。安全判断没错，错在反应过度 + 范围蔓延。

**How to apply:** 边界一两句守住 → 立刻回到用户的真实任务 → 不说教/不贴标签/不预设最坏 → 只对具体越界点提具体风险。关联 [[feedback-no-scope-creep]]、[[feedback-efficiency-small-tasks]]、[[feedback-token-saving-ops]]。

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

_No response_

Claude Model

Sonnet

Relevant Conversation

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.1.117 (Claude Code)

Platform

Anthropic API

Additional Context

wrong_action_record.zip