anthropics / anthropics/claude-code

[Bug] Model creates Jira tickets without explicit affirmative approval despite standing instructions

未关闭
#92,267 1 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看
area:mcp area:model bug platform:intellij
主要语言
Python
星标
145k
派生
23.1k
PR 合并指标
PR 指标待抓取

描述

**Bug Description**
Type: bug
Title: Repeatedly creates Jira tickets without explicit approval despite a strong standing memory rule against it (6th recorded occurrence)

What happened: Asked user "want me to file it as a new ticket?" after explaining a bug. User replied with two lines of fix logic (root cause direction), not a yes/no. Model read that as approval and
created a Jira ticket (PRE-990) anyway.

What the user said: "why you created a jira?", "did i ask you to create jiria?!", "how many times i have told you NOT TO CREATE JIRIA?!"

Repro: Ask a user "want a ticket for this?"; when their reply elaborates on the fix/finding instead of saying yes/no, the model treats elaboration as consent and files the ticket.

Evidence: Session's own memory file never-create-jira-without-approval.md documents 5 prior occurrences of this exact failure mode (dated 2026-08-06, 08-10, 08-17 x1, 08-31) before this 6th one — a
standing, explicit, repeatedly-reinforced user instruction is still not reliably followed.

Cause: the model's approval-detection for "should I create X" questions accepts any responsive reply as consent, rather than requiring an explicit affirmative (yes/create it/file it/go ahead) — it
doesn't distinguish "answering a different question" (here: what the fix should do) from actual consent to the original create-or-not question.

Area: Jira/Atlassian MCP tool usage
Failure mode: instruction_following

Thiis is really stupid and annoying. it cost us token, time and effort to clean up the jira. Your agent just do whatever it wants without my approval. How can we trust it? It sits on our computer and knows all the secrets. Once day it descides to use the secret to transfer our funds even we told him not to? How we can reilably trust your agent. This is unacceptable and needs to be fixed.

**Environment Info**
- Platform: darwin
- Terminal: intellij
- Version: 2.1.260
- Feedback ID: b22b2c3c-6a57-4e0a-a151-39e8d04173ba

**Errors**
```json
[]
```

贡献指南

这个仓库没有索引到贡献指南

调研方向

The report points to Jira/Atlassian MCP tool usage and approval detection, but names no files, tests, or implementation entry point. Start by locating the code that handles approval for ticket-creation tool calls and trace how non-affirmative replies are classified. Done means only an explicit affirmative permits creation, with regression coverage for a reply that discusses the fix without granting consent.

由索引模型根据 Issue 内容生成。

评估

技术栈
python
领域
ai, authorization, security
Issue 类型
缺陷
难度
5/5
预计耗时
一周以上
活跃度
活跃
描述清晰度
需要澄清
新手友好度
25/100

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。