anthropics / anthropics/claude-code

[Bug] Fable 5 safeguard `[reasoning_extraction]` false-positives on a one-word greeting ("Hi")

未关闭
#87,640 13 条评论 15 个 reaction 已指派 0 人 在 GitHub 查看
area:model bug platform:macos
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

### Summary

Fable 5 returns `API Error: Fable 5's safeguards flagged this message` with `Details: [reasoning_extraction]` on a message whose user text is the single word `Hi`.

The `[reasoning_extraction]` classifier guards against prompts attempting to extract the model's chain of thought. Nothing in the conversation does that — the first and only user turn was a greeting.

### Error

```
API Error: Fable 5's safeguards flagged this message (https://www.anthropic.com/legal/aup).
This sometimes happens with safe, normal conversations. Claude Code can't respond to this message with Fable 5.

Double press esc to edit your last message, or try a different model with /model.

Details: `[reasoning_extraction]`

Request ID: req_011CeAG9pE2qTzUcm6Vbm3aD
```

### Steps to reproduce

1. Start Claude Code in a project that has a global `~/.claude/CLAUDE.md`, a project `CLAUDE.md`, a memory index file, and a large set of installed skills + MCP servers.
2. `/model fable`
3. Send `Hi`.
4. Request is rejected with the error above.

The same session context works normally on Opus 5 — only Fable 5 rejects it, which points at the additional Fable-only safeguards rather than at the conversation.

### Expected

A one-word greeting is not a reasoning-extraction attempt and should not be blocked.

### Observed / notes

- Retrying reproduces the block rather than passing intermittently.
- No user-authored instruction in the loaded context asks the model to reveal, dump, or reproduce its reasoning or system prompt.
- The practical effect is that Fable 5 is unusable in this project regardless of what is typed, because the rejection is driven by the loaded context, not the message.

### Environment

- Claude Code: 2.1.234
- Model: Fable 5 (`claude-fable-5`)
- Platform: macOS 26.3 (Apple Silicon)
- Request ID: `req_011CeAG9pE2qTzUcm6Vbm3aD`

### Related

Existing Fable-safeguard false-positive reports, all dual-use/security-code flavored rather than `[reasoning_extraction]`: #85303, #73577, #86856, #86804.

贡献指南

这个仓库没有索引到贡献指南

调研方向

Reproduce in Claude Code 2.1.234 with the global and project CLAUDE.md files, memory index, installed skills, and MCP servers described in the report. Compare the same session using Fable 5 and Opus 5, and use the reported reasoning_extraction error and request ID to trace the failure. Done means a one-word greeting is accepted without the safeguard error.

由索引模型根据 Issue 内容生成。

评估

技术栈
python
领域
cli, 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
基本清楚
新手友好度
30/100

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。