[BUG] Scratchpad/temp files accumulate without cleanup (Claude Code and Cowork) until the host OS freezes and requires a hard power-off

### Preflight Checklist

- [x] I have searched [existing issues](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20state%3Aopen%20label%3Abug) and this hasn't been reported yet
- [x] This is a single bug report (please file separate reports for different bugs)
- [x] I am using the latest version of Claude Code

### What's Wrong?

Scratchpad/temp files under %LOCALAPPDATA%\Temp accumulate without ever being cleaned up. Once the directory grows large enough, the entire host OS freezes — no mouse, no keyboard, no Ctrl+Alt+Del — and only a hard power-off recovers it. Worse, relaunching Claude afterwards while the scratchpad is still populated freezes the machine again within seconds, creating a reboot loop the user cannot escape from inside the app. 8 unexpected shutdowns in 14 days, all during active Claude sessions. Deleting the scratchpad folder manually stopped it completely.

### What Should Happen?

Scratchpad artifacts should be cleaned up when a session ends, and on startup after an abnormal termination. Per-document intermediates (.docx builds, image rendering) should be removed as soon as each build completes rather than left for end-of-session. The scratchpad should be bounded by file count and size, and startup must never block on scanning it — that blocking scan is what makes the failure unrecoverable.

### Error Messages/Logs

```shell
Kernel-Power Event 41 + EventLog 6008 (unexpected shutdown) x8 in 14 days:
2026-08-11 22:07 | 2026-08-13 06:27
2026-08-22 04:47, 05:25, 06:32, 06:35 <- reboot loop
2026-08-24 23:57 (stale value, see note) | 2026-08-25 00:26

LiveKernelEvent 141 (VIDEO_ENGINE_TIMEOUT_DETECTED), WATCHDOG dumps at
00:25:45, 00:26:04, 00:29:09 -> C:\Windows\LiveKernelReports\WATCHDOG\

No WHEA-Logger entries. Both SSDs SMART Healthy, PredictFailure False.

Application log: MoAppHang for Claude_1.34493.0.0_x64__pzs8sxrjxfjjc
at 2026-08-21 12:42 (app itself hung; possibly related)

NOTE: Event 6008 reports the 08-24 freeze at 23:57:35, but file writes
continued until 00:18 — impossible during a freeze. 6008 is a stale
heartbeat; the real freeze was just after 00:18.
```

### Steps to Reproduce

Run a long Claude Code or Cowork session on a software project, including .docx generation and image rendering.
Let the scratchpad under %LOCALAPPDATA%\Temp grow without clearing it.
Host OS freezes completely; hard power-off is the only recovery.
Reboot and launch Claude again without clearing the scratchpad → freezes again within seconds.
Delete the scratchpad folder from outside the app → problem disappears entirely.

Observed trigger in the most recent case: three .docx files (43 KB / 338 KB / 43 KB) generated in 13 minutes, preceded by an SVG→PNG diagram render. Freeze occurred immediately after the last write, not during code editing.

### Claude Model

Opus

### Is this a regression?

I don't know

### Last Working Version

_No response_

### Claude Code Version

1.34493.1

### Platform

Anthropic API

### Operating System

Windows

### Terminal/Shell

PowerShell

### Additional Information

Reproduces in both products, which points at shared scratchpad handling rather than a product-specific bug.

Ruled out: RAM/CPU/PCIe (zero WHEA entries), disk failure (SMART clean on both SSDs), disk-full (91 GB free), VSS exhaustion (0.98 GB of a 7.71 GB cap), and runaway files inside the project folder (project tree clean — its file count is entirely .venv + node_modules). The GPU is a legacy AMD card on a frozen driver, which is a fragility factor, but the machine ran years on that exact config and the freezes correlate strictly with Claude sessions.

Independent signal: Glary Utilities became unresponsive attempting to delete the scratchpad folder and never finished — it had to be removed with shift+delete.

Honest limitation: no exact file count was captured before deletion. NTFS MFT high-water analysis afterwards showed only a ~110,000-record gap, which does not support a "millions of files" claim — the scale was likely tens of thousands. This report doesn't rest on a number; it rests on the reproduction sequence. Project and file names withheld (unpublished project); exact names available privately to a maintainer on request. A monitor now logs scratchpad size every 2 minutes, so I can supply hard numbers if it recurs.

Start by reproducing a long Claude Code or Cowork session on Windows while monitoring %LOCALAPPDATA%\Temp, then inspect what happens after .docx and SVG-to-PNG work completes and after an abnormal termination. Compare behavior when the scratchpad is populated versus manually deleted. Done means completed intermediates and session artifacts are cleaned up, startup does not block on the scratchpad, and file count and size remain bounded.

由索引模型根据 Issue 内容生成。