openaiprovider: Chat tool continuation drops reasoning_content from the originating assistant message

オープン
#1,409 コメント 1 件 リアクション 0 件 担当者 0 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

@nicksin-leroux がすでに取り組んでいます。

2026年11月20日 から。

  • #1410 @nicksin-leroux による — オープン

評価

難易度
3/5
見積もり時間
1〜2日
初心者へのやさしさ
74/100
issue の種類
バグ
明瞭さ
明確に書かれている
活発さ
活発
技術スタック
go

調査の方向性

Start in the Go Chat Completions provider path used by openaiprovider.NewChatCompletionsAgent and trace how the assistant response is added to the automatic tool loop. Add a regression test with the fake HTTP transport and local lookup tool described in the issue. Done means the immediate follow-up includes the supplied reasoning_content, omits it when absent, and does not carry it into unrelated later turns.

索引モデルが issue の本文から書いたものです。

説明

Problem

When a Chat Completions endpoint returns reasoning_content alongside a tool call, the Go Chat provider omits that field from the assistant message on the next request. The tool runs and its result is paired with the correct call. However, an endpoint that requires the original field can reject the follow-up instead of completing the answer. We’ve observed this requirement with Kimi.

Reproduced on both v0.3.1 and main at 045c7be52783e6ae383c816cf9c64310f8d25240, using openaiprovider.NewChatCompletionsAgent and the standard automatic tool loop with streaming disabled.

How to reproduce

Use a fake HTTP transport and one local lookup tool. No API key or live endpoint is needed.

  1. Have the transport return a completion with finish_reason: "tool_calls" and this assistant message:
{
  "role": "assistant",
  "content": "I'll look that up.",
  "reasoning_content": "synthetic-marker",
  "tool_calls": [{
    "id": "call_1",
    "type": "function",
    "function": {"name": "lookup", "arguments": "{}"}
  }]
}
  1. Let the automatic loop execute lookup. Inspect the messages array in the next HTTP request.

Actual: The assistant message still contains call_1, and the tool result references call_1, but reasoning_content is missing.

Expected: For endpoints that require it, the assistant message containing call_1 should retain the exact value "synthetic-marker" on that immediate follow-up. If the endpoint didn’t supply the field, it should remain absent. The value shouldn’t appear in visible output, logs or unrelated later turns.

Related work and proposed contribution

PR #630 preserves reasoning that has already reached the automatic tool loop. This case occurs one step earlier: the Chat provider doesn’t carry the endpoint’s reasoning_content field into that loop.

We’ll open a PR linked to this issue with a proposed fix and a regression test. The test will check that reasoning_content is returned with the matching tool result when supplied, stays absent when omitted, and isn’t carried into unrelated later turns.

主要言語
Go
スター
683
フォーク
70
平均マージ
17時間 50分
マージ済み PR(30日)
64

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

microsoft/agent-framework-go のほかの issue

microsoft/agent-framework-go の issue をすべて見る

似ている issue

Go の issue をもっと見る

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。