Frequent API 500 errors interrupt autonomous workflows

Environment

• Claude Code version: Latest (as of 2025-02-25)
• OS: macOS Darwin 24.5.0
• Shell: zsh
• Model: claude-opus-4-6

Problem

During autonomous multi-step workflows (implementing features, running tests, committing, creating PRs), Claude Code sessions are frequently interrupted by API 500 errors from Anthropic's servers:

API Error: 500 {"type":"error","error":{"type":"api_error","message":"Internal server error"},"request_id":"req_011CYVC2GJkEsBx7qPPnmbUW"}

These errors halt execution entirely and require the user to manually re-trigger the conversation. In a workflow that should run autonomously (code → test → commit → PR → merge → release), this means the user has to babysit the session and press enter every few seconds when the errors occur.

Expected Behavior

Claude Code should automatically retry on transient 500 errors with exponential backoff, rather than stopping and requiring user intervention. The session context is preserved, so retrying should be safe and transparent.

Impact

• Breaks autonomous workflow execution
• Requires constant user monitoring for what should be hands-off operations
• Particularly painful during long multi-step release cycles (implement → test → commit → push → PR → merge → tag → release → docs)

Reproduction

Occurs intermittently during any session, but more noticeable during longer autonomous workflows with many sequential tool calls.