Windows 下点「停止」后回合被中止了, 但 turn/end 序列化失败, 变成了 host/agent-error

[HsiangNianian]

Windows + DSH 0.1.0-rc.6

点击 Web GUI 的「停止」按钮中断正在运行的回合

回合确实被中止了, 但客户端收不到 turn/end aborted
取而代之的是 host 广播的 agent-error

日志

浏览器 console 里第三方插件收到:

session event "turn/end" carries non-JSON-serializable data

根因链

1. 点击停止 → agent loop 追加 turn/end { kind: "aborted", reason: signal.reason }
2. Windows 下 AbortSignal 的默认 reason 是 DOMException, 不可 JSON 序列化
3. Session.append 的序列化校验抛错(dsh-session 的 snapshotJsonValue 校验)
4. loop 的 finally 捕获后 throwError → 宿主广播 host/agent-error
5. 于是「用户主动停止」这个事件丢失了 aborted 语义, 变成了一个看起来像故障的 agent-error, 第三方插件无法区分这是用户取消还是真的故障, 容易误触发自动恢复(我们插件就踩了, 见参考链接)

一些建议

turn/end aborted 的 reason 在 append 前做 JSON 安全处理

signal.reason 是 DOMException / Error 时, 序列化为 { kind: "user" } 或 message 字符串

这样取消语义能正常到达客户端, 序列化失败不会吞掉 aborted

参考

我们插件因此误发过「继续」, 完整复现链与修复见
HsiangNianian/dsh-auto-continue#2

[zoahdev]

Confirmed at source level and implemented as a cherry-pick-ready branch - your root-cause chain is exact.

Source audit (master 47f9438)

packages/core/agent-loop/src/agent.ts builds the aborted turn end as { kind: 'aborted', reason: signal.reason } (agent.ts:304). On Windows an externally aborted controller carries a DOMException as signal.reason, session.append rejects the non-JSON-serializable record, and the finally block turns the serialization failure into host/agent-error - the cancel semantic is lost and third-party plugins cannot tell "user stopped" from "real failure".

Patch: fix/agent-abort-reason-json-safe

https://github.com/zoahdev/deepseek-harness/tree/fix/agent-abort-reason-json-safe

• Aborted-turn reasons are normalized before persistence: typed causes (user / parent / hook / disposed) pass through, anything non-conforming (DOMException, Error, unknown) flattens to the durable { kind: 'user' } cause.
• The AgentCancelCause union has no message field, so the fix follows the type instead of inventing one.

Verification

• tsc -b tsconfig.host.json - clean
• cancel.spec 32/32, including a new Windows-path regression test: a running turn aborted with new DOMException(...) lands turn/end as { kind: 'aborted', reason: { kind: 'user' } } with no error turn end

This closes the loop for auto-recovery plugins like dsh-auto-continue - "user stopped" now looks like a stop. Thanks for the complete repro chain with the reference issue.