Auto-retry on 529 overloaded_error in interactive mode

Problem

When the Anthropic API returns a 529 overloaded_error at session start, Claude Code displays the raw error and the session dies immediately:

API Error: 529 {"type":"error","error":{"type":"overloaded_error","message":"Overloaded"},"request_id":"req_..."}

The user must manually restart and re-enter their prompt. There is no retry, no backoff, no recovery.

Current behavior

• 529 error → session dead, no retry
• --fallback-model exists but only works in non-interactive (-p) mode
• No settings, env vars, or hooks to configure retry behavior for API errors

Proposed behavior

• On 529 (or 503/overloaded), automatically retry with exponential backoff (e.g., 10s → 30s → 60s, up to 3 attempts)
• Show a status message like API overloaded, retrying in 10s... (attempt 1/3)
• If all retries fail, then show the error
• Optionally configurable in settings.json:

{
  "api_retry": {
    "max_attempts": 3,
    "initial_delay_seconds": 10,
    "backoff_multiplier": 3,
    "retryable_errors": [529, 503]
  }
}

Why this matters

A single transient 529 shouldn't kill a session. The user loses their prompt context and has to start over. This is especially frustrating during peak usage periods when 529s are more likely but typically resolve within seconds.

Environment

• Claude Code v2.1.78
• Model: Opus 4.6 (1M context), Claude Max plan
• Platform: WSL2 on Windows 11