[BUG] Claude Code CLI freezes and becomes unresponsive (SIGKILL required)

Summary

Claude Code CLI becomes completely unresponsive during active sessions after approximately 10 minutes of operation, requiring SIGKILL to terminate. Session resumption via claude --resume also hangs indefinitely on "Resuming conversation."

Environment

• OS: Fedora Linux 43
• Kernel: Linux 6.12.28 #1 SMP Tue May 20 15:19:05 UTC 2025 aarch64 GNU/Linux
• Architecture: aarch64 (ARM64)
• Claude Code version: 2.1.37
• Claude Code launch flags: --dangerously-skip-permissions

Steps to Reproduce

1. Launch Claude Code CLI with claude --dangerously-skip-permissions.
2. Begin a task and allow Claude Code to work for approximately 10 minutes.
3. Observe that the CLI becomes completely frozen — no output, no response to keyboard input.
4. Attempt to send signals via keyboard (Ctrl+C, Ctrl+Z) — no response.
5. Manually send all signals via kill (SIGINT, SIGTERM, SIGHUP, SIGTSTP, etc.) — none have any effect.
6. Only SIGKILL (kill -9 <pid>) successfully terminates the process.
7. Relaunch with claude --dangerously-skip-permissions --resume.
8. Select the previous session when prompted for session resumption.
9. CLI displays "Resuming conversation" and hangs indefinitely — again only responsive to SIGKILL.

Expected Behavior

• Claude Code should remain responsive throughout the session.
• Standard signals (SIGINT, SIGTERM) should be handled gracefully.
• Session resumption should load the prior conversation and return to an interactive prompt.

Actual Behavior

• The CLI freezes completely after ~10 minutes of active work.
• All signals are ignored (including all signals sent manually via kill) except SIGKILL, suggesting the Node.js event loop is blocked or the process is stuck in an uninterruptible state.
• Resuming the session hangs on "Resuming conversation" and never recovers, suggesting the serialized session data may itself trigger the hang (possibly a corrupted or oversized conversation state).

Additional Context

• First noticed: Approximately 1–2 weeks ago (late January / early February 2026).
• Reproducibility: Consistent — occurs reliably within ~10 minutes of active use.
• Platform note: Running inside an Apple container (full VM, not Docker) on aarch64 with kernel 6.12.28.
• The fact that --resume also hangs on the same session suggests the issue may be tied to conversation state (e.g., size, specific tool output, or a serialization/deserialization bug) rather than purely a runtime timeout or resource leak.

Possible Investigation Areas

• Event loop blocking in the Node.js runtime (long-running synchronous operation or unresolved promise).
• Conversation state serialization — the saved session may contain data that causes a hang on deserialization or replay.
• Signal handler registration — SIGINT/SIGTERM handlers may not be properly registered or may be swallowed.
• aarch64-specific issues — potential platform-specific bug in a native dependency.
• Memory or file descriptor leak that accumulates over the ~10-minute window.

'