claude-code - 💡(How to fix) Fix Stillborn sessions: 16-28 generic `internal_error` events at startup, no API call made (2.1.132 + 2.1.133)

Summary

In Claude Code 2.1.132 and 2.1.133, I observe a small fraction of session boots where the process emits a burst of 16-28 internal_error events within ~600ms of startup, then exits silently without making any API call. The session is "stillborn" — no user prompt is ever accepted.

I'd like to:

1. Surface this pattern as a known issue if it's already tracked.
2. Provide enough diagnostic detail to help if it isn't.

Pattern observed

Over a 3-day window of normal use I captured two occurrences in the OpenTelemetry log stream:

Both sessions also emitted 4 mcp_server_connection events with status: connected (so MCP servers came up fine), then the error burst fires, then the process exits.

Common signal across error events

Every error event in both sessions shares:

• event.name: "internal_error"
• body: "claude_code.internal_error"
• error_name: "Error" (the generic Error constructor, no further class)
• severity: "" (empty)
• trace_id: "" (empty)
• span_id: "" (empty)
• request_id: null
• No model, no tool_name, no mcp_server_name

So no stack trace, no error class, no request context — just a count of unhandled errors. I don't have enough surface to file a specific root cause; whatever is throwing isn't being attributed.

The errors arrive batched in groups of ~6 with event.sequence 0-5 repeating, suggesting some kind of retry loop or fan-out pattern that fails 6 ways.

Reproduction

I cannot reliably reproduce. Both sessions started normally:

• cd ~/workspace/somewhere && claude (no resume, no special args)
• KDE on Ubuntu 24.04, XDG_CURRENT_DESKTOP=KDE
• konsole terminal
• Standard set of MCP servers (Atlassian, Microsoft Learn, plugin:slack:slack, plus one user-scope MCP)
• ~5 plugins enabled (slack, code-review, superpowers, multi-agent-code-review, session-report, claude-hud)
• ~6 user-scope hook scripts

The two failures are 3 days apart, suggesting a race condition rather than a deterministic config problem. Most session boots succeed cleanly.

Diagnostic ask

If there's a way to surface more detail in the internal_error event payload (stack trace, originating module, error class name) that would help me file a more specific issue if this recurs. Right now the OpenTelemetry layer captures the event but not the cause.

Environment

• Claude Code 2.1.132 and 2.1.133 (same pattern on both)
• Ubuntu 24.04, brew install via /home/linuxbrew/.linuxbrew/bin/claude
• Node v24.13.1
• OTel telemetry enabled (CLAUDE_CODE_ENABLE_TELEMETRY=1, OTEL_LOGS_EXPORTER=otlp)