Repository navigation
[Bug] Desktop host killed by PartitionAlloc OOM (exit 0xE0000008): whole app exits, and the crash popup shows only the last 64 KB of stderr #10492
Replies: 2 comments
|
Additional Windows reproduction on 0.2.1-rc.3, with a directly correlated Resource-Exhaustion-Detector event. This supplements #10492 rather than opening a duplicate. Observed evidence (2026-10-26, UTC+08)
The stderr attached to the exit also mentions incompatible third-party bundles. Those lines initially distracted our diagnosis; the Windows event and exact exit code identify allocation failure, not proof that those plugins caused this termination. Recovery impact, separately verified
Requests to maintainers
We have not intentionally exhausted memory or rerun a destructive reset on the user's live profile. A controlled reproduction should use a disposable profile and no real credentials. |
|
Supplementary data point: What happened
Environment: DSH Desktop 0.2.1-rc.5 (nightly, build commit Load at crash time (context, cause not established)
Probe:
|
| input | exit code | stderr tail |
|---|---|---|
console.log('hi') |
0 | — |
throw new Error('boom') |
1 | JS stack |
process.abort() |
134 | abort message |
process.crash() |
4294930435 (0xFFFF7003) | …crashpad_client_win.cc:867] not connected |
The last row matches the crash report byte-for-byte. Crashpad defines this value explicitly in util/win/termination_codes.h:
//! \brief A dump was requested for a client that was never registered with
//! the crash handler.
kTerminationCodeNotConnectedToHandler = 0xffff7003,So 0xFFFF7003 does not identify the fault: it is the code the process terminates itself with when a dump is requested while crashpad was never connected — which is the normal situation for the desktop host, spawned by the shell with ELECTRON_RUN_AS_NODE=1. Every native fault in the host is therefore guaranteed to surface as this exact code, with no dump. That is consistent with the four 0xFFFF7003 events in this thread and with #10559, and it explains why they all look identical.
Requests
- Connect the host to a crashpad handler (or provide an opt-in/native-dump path for
ELECTRON_RUN_AS_NODEchildren). Without it, this entire class of host bugs can never be localized. - Decode the termination code in the report/dialog:
4294930435 (0xFFFF7003) — native crash; crashpad not connected; no dump, and always keep the first line of the error, not only the tail. - Keep routine plugin info logs off host stderr (or rate-limit them); today they can consume the entire retained stderr tail.
- Optional: record the last active session/tool in the report, as suggested above.
Only one occurrence so far on rc.5, no reproduction yet. I can attach the full crash-*.log (~5 KB, mostly catalog noise) and provide a native dump if a supported capture path exists.
Uh oh!
There was an error while loading. Please reload this page.
中文摘要:9 天内 6 次 Desktop host 被内存分配失败硬终止(退出码
0xE0000008),一次崩溃带走全部会话;崩溃时 Windows 系统日志 1003「虚拟内存不足」比崩溃早 4 秒出现(提交量 41.5/上限 47.9 GB)。另外 4 次退出码是0xFFFF7003,日志里都伴随crashpad_client_win.cc:867 not connected。两个请求:①崩溃弹窗不要只堆原始退出码 + 最后 64 KB stderr(这次后面跟的是windows-computer-use MCP server 0.5.0 ready,很容易被误读成插件问题);②host 崩溃后能否自动重启/恢复会话,或按会话拆 host 以缩小爆炸半径。Environment
10.0.26100host build,10.0.26200platform), 32 GB RAMAutomaticManagedPagefile=False,C:\pagefile.sys 16384 32768→ commit limit 47.9 GBDeepSeek Harnessshell + host + MCP servers), combined private memory ≈ 2.4 GB%APPDATA%\@deepseek-ai\dsh-desktop\logs\crash-*.logSymptom
6 host crashes in 9 days, all at
phase: running(not during boot):0xFFFF70030xE00000080xFFFF70030xFFFF70030xFFFF70030xE0000008(3758096392)Popup text (verbatim):
Evidence
1.
0xE0000008=partition_alloc::win::kOomExceptionCode— Chromium/V8 allocation failure → immediate fail-fast inside the process:2. Windows itself reported commit exhaustion seconds before the crash (System log, Event ID 1003 "Your computer is low on virtual memory"):
2026-10-13 22:55:2010032026-10-13 23:00:4810032026-10-13 23:07:0410032026-10-13 23:12:171003 → crash at 23:12:21 (4 s later)3. The other 4 crashes (
0xFFFF7003) all contain this line, and the two0xE0000008ones do not:I could not determine which of the two publicly documented meanings applies to
0xFFFF7003in this context (CrashpadkTerminationCodeNotConnectedToHandlervs. another V8/Node heap-exhaustion termination path), so I am reporting the correlation only — not a conclusion.4. No conventional crash artifacts: nothing in
%LOCALAPPDATA%\CrashDumpsfornode/DSH, and no related Application Error / WER entries in 30 days (the only WER item nearby isdwm.exe RADAR_PRE_LEAK_64). Consistent with an in-processRaiseExceptionfail-fast rather than an access violation.5. Collateral damage: the whole app exits — every running session dies at once, and the in-flight
session.v4.jsonl.zstdends mid-frame.Two requests
(A) Make the crash dialog diagnosable at a glance.
Today:
host exited with <raw NTSTATUS in decimal>: <last 64 KB of stderr>.3758096392→ "out of memory").windows-computer-use MCP server 0.5.0 ready,[modlens] … not an adapter …). I initially misdiagnosed this as a plugin bug.Suggested:
(B) Crash resilience & blast-radius.
Workaround I applied locally (for other users hitting this)
Raised the pagefile to
24576 49152MB (commit limit becomes ≈ 56.5 GB, max ≈ 80 GB) and reduced parallel heavy sessions. Reboot pending. Not a fix — just headroom.If useful, I can share the two read-only PowerShell helpers I wrote while diagnosing this: a crash-forensics reporter (exit-code semantics + signature scan + Event 1003 correlation + commit/pagefile/process snapshot) and a background memory sampler, so that "how much commit charge was there at the moment of the crash" is answerable next time.
Reported from a real long-running setup (multi-session agent work, 30+ days uptime of the discipline repo); happy to provide any additional log excerpt you need — I have all 6 crash logs locally.
All reactions