ai: Thinking Shows Twice When a reasoning.text Detail Repeats Streamed reasoning_content

未关闭 适合新手
#64,723 1 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
2/5
预计耗时
1-3 小时
新手友好度
78/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
活跃
技术栈
typescript
领域
backend, testing

调研方向

从 packages/ai/src/protocols/openai-chat.ts 开始,重点查看 detailText 和 reasoningTextObserved,然后检查 packages/ai/test/provider/openai-chat.test.ts 中的 OpenAI chat provider 测试。重现 reasoning_content 后跟随匹配的 reasoning.text 详情的流。完成标准是 Thought 块中只包含一次该文本,并且回归测试通过。

由索引模型根据 Issue 内容生成。

描述

area/providers platform/macos status/repro-simulated type/bug version/v2
Description

Claude's thinking shows twice in the Thought block when the model is reached through an OpenAI-compatible gateway that streams reasoning_content and then sends the signed thinking block as one reasoning.text entry in reasoning_details. I see it in OpenChamber. The doubled text is also in the stored session, so it does not depend on the interface.

The next request sends the doubled text back as reasoning_content. The signed entry itself is sent back correctly.

The cause is in packages/ai/src/protocols/openai-chat.ts. detailText returns the text of every reasoning.text entry, and OpenCode appends that text as another reasoning delta after the streamed chunks. reasoningTextObserved already stops this for Kimi's summary entries but has no effect on reasoning.text.

The code is the same in v2.2.7, v2.3.2 and dev (ae88eb6a7).

In my local database, 730 of 918 reasoning blocks from Claude Sonnet and 188 of 219 from Claude Haiku are exactly doubled. The GPT model I use (321 blocks) shows no doubling.

I have a two-line fix with a regression test, and it passes the related existing provider tests. I can open a PR if a maintainer wants one.

Plugins

Two local plugins (memory and environment setup). Neither mentions reasoning or thinking.

OpenCode version

2.2.7 (I also read the source of 2.3.2 and dev)

Steps to reproduce
  1. Use a Claude Sonnet or Haiku model through a gateway that streams the chunks below on /v1/chat/completions. OpenCodex does this. OpenRouter sends both fields in one frame and is not affected.

    {"delta":{"reasoning_content":"Think "}}
    {"delta":{"reasoning_content":"hard."}}
    {"delta":{"reasoning_details":[{"type":"reasoning.text","text":"Think hard.","signature":"...","format":"anthropic-claude-v1","index":0}]}}
    {"delta":{"content":"Done"},"finish_reason":"stop"}
    
  2. Ask a question that makes the model think.

  3. Open the Thought block. It reads "Think hard.Think hard." and should read "Think hard."

The same stream as a test in packages/ai/test/provider/openai-chat.test.ts fails with Received: "Think hard.Think hard.".

Screenshot

No response

Operating System

macOS 27.2

Terminal

OpenChamber (web UI)

主要语言
TypeScript
星标
227k
派生
30.7k
平均合并
5 小时 14 分钟
30 天内合并 PR
486

环境准备

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

anomalyco/opencode 的其他 Issue

查看 anomalyco/opencode 的全部 Issue

相似的 Issue

更多 TypeScript Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。