Kimi Code的VSCode插件，会不断调用TaskOutput，提示词要求停止也无效。

[OpenBST]

What version of Kimi Code is running?

kimi code for vs code 0.6.7

Which open platform/subscription were you using?

Allegro

Which model were you using?

k3

What platform is your computer?

windows 11(intel)

What issue are you seeing?

执行后台构建等长耗时任务时，会疯狂轮询，即使向agent要求了停止轮询也没用

vscode版本信息：
版本: 1.132.0 (user setup)
提交: df53daabb18cd157bdb08c7f01c34df936cf12f4
日期: 2026-08-04T15:30:20Z
Electron: 42.7.1
ElectronBuildId: 14760073
Chromium: 148.0.7778.280
Node.js: 24.18.0
V8: 14.8.178.38-electron.0
OS: Windows_NT x64 10.0.26200

What steps can reproduce the bug?

...

What is the expected behavior?

No response

Additional information

No response

[SnowingFox]

Diagnostic from a source-level trace (no code change — there is no minimal, unit-testable defect to fix; leaving a verified root-cause trace for triage).

1. The hypothesis of a VSCode-extension polling/re-emit loop is not supported: TaskOutput has zero matches under apps/vscode, and the extension is event-driven (SessionRuntime.onSdkEvent) with no polling loop.
2. The repeated TaskOutput calls come from the model invoking the engine TaskOutput agent tool (packages/agent-core-v2/src/agent/tools/task/task-output/taskOutputTool.ts), which returns a 32 KiB not_ready snapshot until the task is terminal.
3. The tool description already forbids polling ("Prefer relying on automatic completion notifications… Do not use TaskOutput to wait"), and the engine already delivers automatic completion notifications (taskService.ts), so this is model-behavior, not an extension bug.
4. "Stop request ineffective": while streaming, the webview enqueues messages (chat.store.ts sendMessage) and queued messages are only sent after the turn ends — so a natural-language "stop" waits for the polling turn. The documented immediate interrupt is the Stop button (abortChat → SessionRuntime.cancel()).

If the desired UX is that an in-stream user message also interrupts the current tool-invocation turn (rather than queueing), that would be a product decision; happy to help with a targeted change if maintainers want it.