How to lengthen Claude Code's retry delay on 529 overloaded errors

When the API rejects a request with a 529 overloaded error, Claude Code retries automatically with exponential backoff. You can stretch that backoff with environment variables.

CLAUDE_CODE_OVERLOADED_RETRY_BASE_DELAY_MS sets the starting delay. The default is 500 milliseconds, and it accepts whole numbers from 500 to 32000; any other value is treated as unset.

v2.1.296 also added CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS for a longer maximum delay. The changelog doesn't give its range, and it isn't in the environment variable docs yet.

// ~/.claude/settings.json
{
  "env": {
    "CLAUDE_CODE_OVERLOADED_RETRY_BASE_DELAY_MS": "4000"
  }
}

If retries run out quickly while the API is busy, raise the starting delay to spread them over a longer window. On anything older than v2.1.296, run claude update first.