[BUG] darwin-x64 ships AVX2-requiring Bun while linux-x64 ships baseline — SIGILL on pre-Haswell Intel Macs; fix is aligning the build variant

Open 💬 0 comments Opened Jul 8, 2026 by f-hoedl

Summary

Since v2.1.113, @anthropic-ai/claude-code-darwin-x64 crashes with SIGILL (Illegal instruction: 4, exit code 132) on Intel Macs without AVX2 (Ivy Bridge and older, e.g. 2012 Mac mini / i5-3210M). The linux-x64 package does not have this problem — because it is compiled with Bun's baseline variant, which only requires SSE4.2. The darwin-x64 package is not.

This looks like a one-line build-configuration inconsistency, not a hard platform limitation.

Evidence

Extracted from the published npm tarballs of 2.1.204 (strings <binary> | grep "Bun v"):

| Package | Embedded Bun signature | Runs on non-AVX2 x64 |
|---|---|---|
| claude-code-linux-x64@2.1.204 | Bun v1.4.0 (63bb0ca0d) Linux x64 (baseline) | ✅ yes (verified, see below) |
| claude-code-darwin-x64@2.1.204 | Bun v1.4.0 (63bb0ca0d) macOS x64 | ❌ SIGILL |

Same Bun version, same commit — only the variant differs.

Cross-check on identical hardware: on a 2012 Mac mini (i5-3210M, no AVX2, macOS 13.7.8):

  • claude darwin-x64: every version 2.1.113 → 2.1.204 exits 132 (SIGILL). 2.1.112 (last Node-based npm build) works.
  • The linux-x64 2.1.204 binary runs correctly on the very same CPU inside a Lima VM (Ubuntu, vz) — confirming the silicon is capable and only the darwin build variant is the blocker.

Repro

  1. Intel Mac without AVX2 (any pre-Haswell, e.g. i5-3210M; sysctl machdep.cpu.leaf7_features shows no AVX2)
  2. npm install -g @anthropic-ai/claude-code@2.1.204
  3. claude --version → killed by SIGILL, exit code 132

Expected

Ship darwin-x64 compiled against bun-darwin-x64-baseline, exactly as linux-x64 (and per earlier reports win32-x64) already do. Alternatively publish a claude-code-darwin-x64-baseline optional package and select it in install.cjs when CPUID lacks AVX2 — the platform map there already handles finer-grained variants (musl, android).

Prior reports

#50466 (this exact regression, closed as duplicate), #21018, #24456, #19967, #19904 — all closed as duplicates without a resolution for macOS. The linux side was evidently fixed by switching to baseline; darwin was left behind.

Environment

  • Claude Code: 2.1.113–2.1.204 (all affected); 2.1.112 OK
  • Hardware: Mac mini 2012, Intel i5-3210M (Ivy Bridge — AVX yes, AVX2 no)
  • OS: macOS 13.7.8 (22H730)
  • Install: npm global (native installer ships the same binary)

View original on GitHub ↗