Standalone tools.execute calls keep failing after session abort until another chat turn

未关闭
  • 哪里不对？
#2,871 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
48/100
Issue 类型
缺陷
描述清晰度
基本清楚
活跃度
活跃
技术栈
node.js, python, typescript
领域
api, backend

调研方向

Start with the SDK's session.rpc.abort and session.rpc.tools.execute handling, then trace how cancellation state is created and cleared across direct calls and chat turns. Use the attached abort-repro.zip and its baseline, resume, and recovery runs to reproduce the failure; done means a standalone call after abort reaches the MCP server without a model turn, and standalone completion has a documented or exposed idle signal.

由索引模型根据 Issue 内容生成。

描述

After a joined extension calls session.rpc.abort({reason:'remote_abort'}), the running session.rpc.tools.execute is cancelled as expected. But the next standalone call, with a fresh toolCallId, is also rejected before it reaches the MCP server:

code: -32603
Request session.tools.execute failed with message: Tool execution was aborted.

Reproduced on Linux x86_64 with Copilot CLI 1.0.103, Node SDK 1.0.26, and Node 25.9.0. No agent turn is running when the extension starts the tool. The SDK's current main (c57638f4) has the same abort/execute contract as 1.0.26.

Repro

The attached abort-repro.zip contains a tiny joined extension, local Python MCP wait tool and launcher. Each run creates a private COPILOT_HOME. With Node, Python 3 and authenticated Copilot CLI available:

abort-repro.zip

npm install --no-save --package-lock=false @github/[email protected]
node repro.mjs baseline
node repro.mjs resume resume
node repro.mjs recovery send
  1. The extension successfully executes the local wait tool as a control.
  2. It starts a longer call and aborts on the first matching progress event.
  3. Abort returns {success:true}. The MCP server receives cancellation and finishes.
  4. Calls after 500 ms and 3.5 seconds fail with the error above, without reaching the server. Rejection takes roughly 2–7 ms.
  5. Resuming the still-active session does not restore execution. One explicit chat turn replying OK does; the next tool call reaches the same server and succeeds. The send run makes that real model turn.

After cancellation, metadata reports processing:false, abortable:false, hasActiveWork:false. This reproduced with both the CLI-bundled SDK and the published SDK package. It appears that direct calls retain an aborted cancellation scope, but I have not identified the native state responsible.

Expected / ask

The next explicit standalone invocation should have a usable cancellation scope. Could tools.execute support per-request cancellation, or could the SDK expose a supported way to clear the aborted state without running a model turn? SDK #1433 is related, but concerns cancellation reaching client handlers; here MCP cancellation already reaches the server.

Separately, successful direct execution before any abort emits no session.idle in the following one-second observation, although all activity flags are false. A completed chat turn does emit it. Please also document/provide the standalone completion signal so session indicators can return to idle when no work remains. I do not assume this lifecycle gap and the abort failure have the same cause.

主要语言
Java
星标
10.7k
派生
1.5k
30 天内没有已合并 PR

环境准备

  • 没有 Dockerfile 或 Docker Compose 文件
  • 没有 Pull Request 模板
  • 阅读贡献指南

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。

github/copilot-sdk 的其他 Issue

查看 github/copilot-sdk 的全部 Issue

相似的 Issue

更多 Java Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。