Claude Code 2.1.113+ crashes on startup (SIGILL in bundled Bun 1.3.14) on Intel Emerald Rapids / Ubuntu 24.04

[pedram-tahoe]

Summary

Every Claude Code release from 2.1.113 onward (the switch from a Node-based cli.js to a Bun single-executable bin/claude.exe) crashes on startup with a SIGILL inside the bundled Bun v1.3.14 runtime. The crash fires in ~4ms, before any Claude Code code runs — --version and --help both die.

Reverting to 2.1.112 (the last Node-based release) works perfectly on the same machine.

Reproduction

curl -fsSL https://claude.ai/install.sh | bash
# OR
npm install -g @anthropic-ai/claude-code
claude --version

Output:

============================================================
Bun v1.3.14 (521eedd6) Linux x64 (baseline)
Linux Kernel v6.8.0 | glibc v2.39
CPU: sse42 popcnt avx avx2 avx512
Args: "claude" "--version"
Features: jsc standalone_executable claude_code
Elapsed: 4ms | User: 0ms | Sys: 5ms
RSS: 24.44MB | Peak: 12.58MB | Commit: 24.44MB | Faults: 0

panic(main thread): Segmentation fault at address 0x0
oh no: Bun has crashed. This indicates a bug in Bun, not your code.

https://bun.report/1.3.14/B_1521eeddkggggEggggC+ypRq9yxpC09vypCovs4wCy5/qpC6z0qpCy8wK0oxK6iyqpCA2AA

Illegal instruction (core dumped)

Environment

Field	Value
Claude Code	2.1.113 through 2.1.161 (all reproduce)
Bundled Bun	v1.3.14 (521eedd6) Linux x64 baseline
OS	Ubuntu 24.04.1 LTS (Noble)
Kernel	6.8.0-1044-oracle (Oracle-flavored, OCI-backed K8s host)
glibc	2.39-0ubuntu8.4
CPU	Intel Xeon Platinum 8592+ (Emerald Rapids, 2× 64-core / 256 thread)
CPU flags	full avx, avx2, avx512f/dq/bw/vl/vnni/bf16/fp16, avx_vnni, amx_tile, amx_int8, amx_bf16
Container	Kubernetes pod (RunAI), seccomp mode 2, 1 filter installed

Notably, the CPU has all baseline Bun requirements (and far more) — this isn't a missing-instruction issue at the hardware level.

Diagnostics already performed

• strace shows the crash happens right after JSC sets up its 128 GiB JSGigacage anonymous mmap and a 2 GB PROT_NONE mprotect. The final syscall before SIGILL is the gigacage mprotect. No -EPERM or -EACCES from seccomp in the trace.
• The SIGILL fires at address 0x2b05212 inside the binary's text segment, signaled as ILL_ILLOPN (illegal opcode), then Bun's handler triggers a secondary SIGSEGV at 0x0 that produces the panic line.
• Env-var workarounds that do not help (crash happens before they're consulted):
  BUN_JSC_useJIT=0, BUN_DISABLE_JIT=1, BUN_JSC_useDFGJIT=0, BUN_JSC_useConcurrentJIT=0, JSC_useJIT=0, JSC_useGigacage=0 (rejected as "invalid option"), JSC_useWebAssembly=0, WEBKIT_JSC_useJIT=0.
• npm package and the curl-installed standalone binary both reproduce identically (same bundled binary).
• The crash is fully deterministic — happens on every invocation including --version.

Workaround

npm install -g @anthropic-ai/claude-code@2.1.112
# also disable the auto-updater so it doesn't roll forward into the broken Bun build:
echo '{"autoUpdaterStatus":"disabled"}' > ~/.claude/settings.json
export DISABLE_AUTOUPDATER=1

Bun crash report URL

For forwarding to the Bun team if helpful:

https://bun.report/1.3.14/B_1521eeddkggggEggggC+ypRq9yxpC09vypCovs4wCy5/qpC6z0qpCy8wK0oxK6iyqpCA2AA

Asks

1. Consider bumping the bundled Bun version (or building Bun with a different baseline target) in the next Claude Code release — there are multiple Bun 1.3.x reports of startup SIGILL on Sapphire / Emerald Rapids Xeon hosts.
2. In the meantime, would it be possible to keep publishing a Node-based fallback (a la cli-wrapper.cjs actually running the JS bundle in-process) for environments where the Bun SEA crashes? The wrapper currently only re-spawns the native binary, which provides no escape hatch.
3. Document the 2.1.112 rollback in a known-issues note so other users on Emerald Rapids / Sapphire Rapids K8s hosts can find the workaround.

[serkanz]

#51464

[BikesAndBBQ]

Confirming the same crash (same Bun build, bun.report/1.3.14/B_1521eedd…) on a very different CPU, with a couple of details that may help narrow it.

Environment

• CPU: Intel Core i5-5300U (Broadwell, 2015) — flags: sse4_2 popcnt avx avx2 (no AVX-512).
• Containerized (k3s), Linux kernel 6.17, x86_64, npm install -g.

Two differences from the original report:

1. claude --version does not crash here (exits 0) — only an actual run crashes, ~2s in:

   [926] embedder failed to suspend thread 0x… for TLC 0x…
   Bun v1.3.14 (521eedd6) Linux x64 (baseline)
   Linux Kernel v6.17.0 | musl
   CPU: sse42 popcnt avx avx2
   Args: "claude" "-p" "--output-format" "json"
   Elapsed: 2152ms | ...
   panic: Segmentation fault at address 0x0
   oh no: Bun has crashed. This indicates a bug in Bun, not your code.
   Illegal instruction (core dumped)
   

   Deterministic: printf hi | claude -p --output-format json → crash every time.

2. Crashes on both musl and glibc — identical failure on node:22-alpine (claude-code-linux-x64-musl) and node:22-slim (glibc 2.36, claude-code-linux-x64). Not libc-specific.

Same version boundary you found: 2.1.112 (last Node-based) works; the Bun builds (2.1.113+) crash. Pinning to 2.1.112 is our workaround.

Worth noting: your Emerald Rapids (newest, AVX-512) and this Broadwell (2015, AVX2-only) both hit the identical Bun build (521eedd6) — so the affected CPU range is very wide — and on the older chip the SIGILL only fires at runtime, not at --version. That argues against a simple "missing-ISA gate at startup."

[github-actions]

Closing for now — inactive for too long. Please open a new issue if this is still relevant.