Crash (0xC0000005) in Node utility process while running Copilot agent — identical fault address 0x6B7AF6A3 across 6 crashes (1.134.0-insider)

[vscode-crash-attachments.zip](https://github.com/user-attachments/files/30920268/vscode-crash-attachments.zip)

### Summary
VS Code Insiders **crashes repeatedly when running the Copilot agent** (agent mode). The crash is a
0xC0000005 access violation in a **Node.js utility process** (`node.CrUtilityMain`), occurring at the
**exact same address (0x6B7AF6A3) across 6 separate crashes** over 3 days — i.e. a deterministic
code defect, not random memory pressure. The faulting address is **inside V8 JIT code** (not in any
loaded module), and the process has the GitHub Copilot SDK native runtime loaded
(`extensions/copilot/node_modules/@github/copilot/sdk/prebuilds/win32-x64/runtime.node`, ~95 MB).

### Version / Environment
- VS Code: **1.134.0-insider** (user setup, x64)
- Electron: **42.8.0** (build dir `42b1420914`)
- OS: Windows 11, build **26100.8893** (x64)
- Crash reporter: Crashpad (minidumps in `%APPDATA%\Code - Insiders\Crashpad\reports`)

### Steps to reproduce
1. Open VS Code Insiders with GitHub Copilot Chat + a few MCP servers configured
(in my case: `microsoft/markitdown`, `microsoft/playwright-mcp`, `com.apify/apify-mcp-server`,
`wireshark-mcp`).
2. Start an agent-mode session in Copilot Chat (model set to a custom BYOK endpoint,
e.g. `customendpoint/.../deepseek-v4-flash`, via `chat.agentHost.byokModels.enabled`).
3. Send a message that triggers tools (e.g. `web_fetch`).
4. Within ~10–60 s the utility process crashes; the editor window dies / agent session aborts.

Crash reproduced 6 times on Aug 10–11 (13:07, 17:10 x2, 09:42 x2, 09:59) at the same address.

### Crash signature (minidump analysis)
- Exception: `0xC0000005` (ACCESS_VIOLATION) at `0x6B7AF6A3` — **identical in all 6 dumps**
- Process type: `utility`; main thread `node.CrUtilityMain` (libuv + tokio workers present)
- Faulting address is **not contained in any loaded module** → V8 JIT code region
- Distinct modules loaded: `Code - Insiders.exe`, `spdlog.node`,
`extensions/microsoft-authentication/dist/msal-node-runtime.node`, `msalruntime.dll`,
`extensions/copilot/.../runtime.node` (Copilot SDK), plus network/auth system DLLs (WAM,
Windows.Security.Authentication.Web.Core, etc.)

### Timeline from logs (`agenthost.log` / `exthost.log` / `renderer.log`, session 20260811T095750)
```

09:57:53 Agent Host starts; Copilot CLI + runtime.node loaded
09:57:58 Extension host (pid 40840) unresponsive (repeats at 09:58:05)
09:58:08 Model changed to: customendpoint/OpenCode Go/deepseek-v4-flash
09:58:09 sendMessage → MCP servers 'microsoft/markitdown' and 'microsoft/playwright-mcp' start
09:58:45 EMFILE: too many open files (extension host) ← file-handle exhaustion
09:58:46 Agent tool 'web_fetch' started; confirmation requested
09:59:10 ProxyResolver#resolveProxy DIRECT Canceled
09:59 Crash (dump written)

```

### Notes / possible aggravators
- `EMFILE: too many open files` was logged seconds before the crash — the utility/extension host
hit handle exhaustion while several stdio MCP servers were starting.
- Extension host repeatedly flagged unresponsive during agent runs.
- Crash also occurred without heavy usage (13:07 run), so the trigger appears to be agent-host
activity itself, with MCP/BYOK load increasing frequency.

### Attachments
- `vscode-crash-attachments.zip` — 4 Crashpad minidumps (incl. the 3 newest same-address dumps and
one older 2026-05-29 dump at a different address `0xBD92AA73` for comparison) + logs from the
crashing session (agenthost, exthost, renderer x2, main).

### Additional context
- This is an Insiders daily build; the faulting address is stable across the 8/10–8/11 builds.
- Please let me know if you need the full set of 7 dumps or a repro with MCP servers disabled.
```