[BUG] Bun 1.4.1 segfault at 0xFFFFFFFFFFFFFFFF in JSC GC mark phase, ~90 min into resumed session (CLI 2.1.260 native, Windows 11 26200, agent-teams messaging active)

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report
• [x] I am using the latest version of Claude Code

What's Wrong?

A long-lived session resumed with claude --resume <id> (running for ~90 min: Elapsed: 5395118ms, RSS: 0.42 GB | Peak: 0.50 GB | Commit: 0.83 GB | Faults: 1204752) crashed hard with a Bun segmentation fault while the model was in an extended-thinking turn (effort: high). Right before the crash the session had sent a message to another local Claude Code session through the experimental agent-teams feature (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1) and had run one shell command (an ssh to a remote host). The crash killed the process; the in-progress turn was lost.

The decoded crash report (bun.report) places the fault in the JavaScriptCore garbage collector mark phase, on a dedicated GC thread:

WTF::Dependency::loadAndFence (Atomics.h:446) → JSC::MarkedBlock::aboutToMark (MarkedBlock.h:601) → JSC::Heap::testAndSetMarked (HeapInlines.h:75) → JSC::SlotVisitor::markAuxiliary (SlotVisitor.cpp:302) → visit of JSObjectWithButterfly / JSScope / JSLexicalEnvironment → JSC::Heap::runBeginPhase (Heap.cpp:1647) → WTF::Thread::entryPoint (Threading.cpp:309)

What Should Happen?

The turn completes; the process does not crash. claude --resume <id> afterwards restores the transcript correctly, so the only loss is the interrupted turn.

Error Messages/Logs

============================================================
Bun v1.4.1 (83c373b27) Windows x64
Windows v10.26200
CPU: sse42 avx avx2
Args: "C:\Users\<user>\.local\bin\claude.exe" "--resume" "<session-id>"
Features: Bun.stderr(2) Bun.stdin(2) Bun.stdout(2) WebSocket(2) abort_signal(5) fetch(177) jsc spawn(1) standalone_executable process_dlopen yaml_parse(393) claude_code
Builtins: "bun:ffi" "bun:main" "node:assert" "node:async_hooks" "node:buffer" "node:child_process" ... "node:zlib" "ws"

Elapsed: 5395118ms | User: 75921ms | Sys: 32421ms
RSS: 0.42 GB | Peak: 0.50 GB | Commit: 0.83 GB | Faults: 1204752 | Machine: 68.39 GB

panic(thread 8364): Segmentation fault at address 0xFFFFFFFFFFFFFFFF
oh no: Bun has crashed. This indicates a bug in Bun, not your code.

To send a redacted crash report to Bun's team,
please file a GitHub issue using the link below:

 https://bun.report/1.4.1/w_183c373bmgkgkQ+hogCkttsoCgph46Bw33zmCw34r+BghzwjCsostoCwwvsoC231soCit7q8BiCk5k1zC+w84vCmyq+8DCYKERNEL32.DLLurmLCSntdll.dll4ulrBA2DD

Steps to Reproduce

Not deterministic; observed once so far. Conditions at the time of the crash:

1. Start Claude Code 2.1.260 on Windows 11 (standalone claude.exe, installed via the native installer in ~/.local/bin) with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 and defaultMode: auto in settings.
2. Resume an existing session with claude --resume <id> and keep it running for over an hour with several tool calls (Bash/ssh) and MCP servers loaded.
3. Have a second local Claude Code session running; from the first session, send it a message via the agent-teams feature ("sent to <session-name> — another Claude session on this machine").
4. Let the model continue with a high-effort extended-thinking turn.
5. Bun crashes with the segfault above during the GC mark phase.

Claude Model

Other (Fable 5.1, claude-fable-5-1, effort high)

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.260 (Claude Code)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Windows Terminal

Additional Information

• Shell inside the pane: Windows PowerShell 5.1 (powershell.exe), running under a terminal multiplexer (herdr 0.8.2, ConPTY-based). Other Claude Code sessions in sibling panes were unaffected.
• No Windows Error Reporting event was written for claude.exe: Bun's panic handler exits the process itself, so this class of crash is invisible in the Windows Application event log.
• The bun.report link above is the redacted stack; happy to provide more details on request.
• Related reports in the same 0xFFFFFFFFFFFFFFFF family, on earlier Bun versions: #78436 (Bun 1.4.0, closed), #28789 (closed), #22632 (closed), and #87616 (Bun 1.4.0, open, different address). This one is on Bun 1.4.1 / CLI 2.1.260 with a decoded stack inside the JSC GC mark phase, which is why I'm filing it separately rather than commenting on a closed issue.