Deepseek Tool Mangling

オープン
#13,442 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 8 日以内に返信

まだ誰も着手していません。

評価

難易度
4/5
見積もり時間
3〜5日
初心者へのやさしさ
35/100
issue の種類
バグ
明瞭さ
おおむね明確
活発さ
活発
技術スタック
typescript

調査の方向性

The payload names no source files or tests. Begin at the DeepSeek provider integration and the shared tool-call streaming/argument parser, then reproduce with a turn that emits at least two tool calls while comparing the raw streamed arguments against the parsed arguments. Done means multi-call turns preserve each call's fields and either apply edits correctly or fail cleanly.

索引モデルが issue の本文から書いたものです。

説明

Before submitting your bug report
Relevant environment info
- OS: Windows: 11
- Continue version: 2.0.0
- IDE version: {
Version: 1.131.0 (system setup)
Commit: e4c7e7b1d6d060162f4aa7f8225271b67ce1df75
Date: 2026-07-28T10:51:25Z
Electron: 42.7.0
ElectronBuildId: 14710161
Chromium: 148.0.7778.280
Node.js: 24.18.0
V8: 14.8.178.38-electron.0
OS: Windows_NT x64 10.0.26200
}
- Model: Deepseek
- config:
  
name: Main Config
version: 1.0.0
schema: v1
disableAutocomplete: true
models:

  - name: DeepSeek Coder
    provider: deepseek
    model: deepseek-coder
    apiKey: 
    contextLength: 128000
    shell: powershell.exe

  
  OR link to agent in Continue hub:
Description

Intermittently, when the agent emits tool calls, the arguments arrive corrupted in transit with no user input at all (no typing, no editor interaction). Required fields arrive empty (filepath, old_string, new_string), field values from sibling calls get interleaved/swapped, and the JSON is truncated mid-value. In the worst case an edit reports success but writes garbled or truncated content to disk (e.g. #pragma once → #pragma o sha; a header truncated mid-line), silently breaking source files. Failures surface inconsistently as ... argument is required or Cannot read properties of undefined (reading 'trim'). Markedly worse when multiple tool calls are batched in a single turn, pointing to a stream/serialization problem in argument generation or parsing.

To reproduce

Tool-call argument corruption
Conditions that make it reliable:

Agent is asked to perform a task that produces multiple tool calls in a single turn (e.g., "read these 3 files", or a multi-part edit touching several files).
Target files exist and the IDE editor panel is open (typing is not required).
Steps:

Give the agent a task that naturally emits ≥2 tool calls in one turn (e.g., "open A.cpp, B.cpp, and C.cpp and summarize each", or "add this same change to these 3 files").
Do not type or click in the editor while the turn streams.
Observe the tools: one or more calls fail with a missing/invalid argument — e.g. filepath argument is required, string old_string is required, string new_string is required, or Cannot read properties of undefined (reading 'trim').
In some runs, observe sibling-field contamination: an old_string/new_string containing text belonging to the other call, or filepath appearing inside changes.
In the worst case, an edit tool reports success but the file on disk is corrupted — truncated mid-line or with characters swapped (e.g. #pragma once → #pragma o sha, or a header ending abruptly at // Latest detected regions (image pixel).
Expected: each tool call receives exactly the arguments the agent emitted; edits either fully succeed with correct content or fail cleanly. Actual: arguments are empty/swapped/truncated and/or files are silently written corrupt.
Confirm it is not user-input related: keep hands completely off the keyboard/mouse for the whole turn; the corruption still occurs.

Log output

主要言語
TypeScript
スター
36.5k
フォーク
5.6k
PR マージ指標
30日以内にマージされた PR はありません

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。

continuedev/continue のほかの issue

continuedev/continue の issue をすべて見る

似ている issue

TypeScript 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。