Native V8 array-length crash during active tool-heavy turns and session resume
- 主要语言
- Shell
- 星标
- 11.2k
- 派生
- 1.9k
- 平均合并
- 14 小时 16 分钟
- 30 天内合并 PR
- 6
描述
### Describe the bug
The packaged Linux x64 native binary aborts inside V8 during active, tool-heavy turns as well as while resuming conversations.
The original report attributed this to several concurrent resumes. Further testing disproved that: Herdr automatic restoration was disabled, failed panes were cleared, and an already-running conversation crashed mid-turn while displaying:
```text
Reviewing brief clarity · 29.4 KiB
# Fatal error in , line 0
# Fatal JavaScript invalid array length 0
```
A second active conversation failed similarly while displaying `Shipping Lakebase refactor · 53.7 KiB`. Both ended with SIGILL. These were normal active turns, not startup or bulk restore.
The fatal string maps to V8's `Accessors::ArrayLengthSetter`, where `JSArray::SetLength(..., 0)` returned an exception and V8 called `FATAL`:
https://github.com/v8/v8/blob/main/src/builtins/accessors.cc#L210-L214
### Affected version
GitHub Copilot CLI 1.0.70
### Steps to reproduce the behavior
A minimal deterministic reproduction is not yet available. The recurring observed sequence is:
1. Start or resume a Copilot conversation.
2. Run a long, tool-heavy workflow with many subagents.
3. Start another turn.
4. While Copilot is actively streaming/reviewing the response, the native process aborts with `Fatal JavaScript invalid array length 0` and SIGILL.
Two affected conversation histories were approximately:
- 14.1 MB / 2,482 events / 929 completed tool calls / 23 completed subagents
- 21.3 MB / 3,959 events / approximately 1,500 completed tool calls / 24 started subagents
The final persisted event in each is `assistant.turn_start` without a matching `assistant.turn_end`, confirming the failure occurred mid-turn.
### Expected behavior
Long or tool-heavy conversations should continue normally. If an internal array operation fails, the CLI should surface a recoverable JavaScript error rather than aborting the native process.
### Additional context
- OS: Ubuntu 24.04 under WSL2, Linux `6.6.114.1-microsoft-standard-WSL2`, x86_64
- Terminal: Windows Terminal -> Herdr PTYs
- Shell: zsh
- Memory after the latest crash: 31 GiB total, 28 GiB available, no swap in use
- Copilot's embedded runtime reports Node.js `v24.16.0`
- Kernel logs show repeated SIGILL failures at the same native address
- `app.js` and `index.js` checksums match between the npm platform package and Copilot's update cache
- The npm loader's subsequent `no platform package found` message is misleading; the platform package launched successfully and its child process terminated by signal
- Reinstalling npm replaces 1.0.70 with the same 1.0.70 and does not resolve the failure
- Herdr's `resume_agents_on_restore` setting is not required to trigger this
- Running the packaged JavaScript entrypoint under system Node.js `v24.11.1` passes startup/version and interactive Herdr smoke tests. I am testing this as a temporary workaround, but it has not yet had the same long-duration workload.
贡献指南
调研方向
先从打包的 JavaScript 入口 app.js 和 index.js 开始,然后比较它们在已记录的系统 Node.js workaround 下的行为。如果可能,复现一次漫长且大量使用工具的 active-turn workflow,并捕获原生故障。完成的标准是 workload 不再因 V8 invalid-array-length SIGILL 而中止,或者 CLI 改为显示一个可恢复的错误。
由索引模型根据 Issue 内容生成。
评估
- 技术栈
- javascript, node.js
- 领域
- cli
- Issue 类型
- 缺陷
- 难度
- 5/5
- 预计耗时
- 一周以上
- 活跃度
- 冷清
- 描述清晰度
- 需要澄清
- 新手友好度
- 25/100