claude-code - 💡(How to fix) Fix Transient 529 Overloaded API errors abort long-running tasks with no auto-recovery [2 comments, 3 participants]

Description

When the Anthropic API returns a transient 529 Overloaded response mid-session, Claude Code surfaces the error and halts the in-flight task. There is no automatic retry, backoff, or resume behavior — the user must manually re-issue the prompt, which is especially disruptive for long-running, multi-step workflows (e.g. /goal-style tasks that span many tool calls).

Observed error

API Error: 529 Overloaded. This is a server-side issue, usually temporary — try again in a moment. If it persists, check status.claude.com.

Steps to reproduce

1. Start a long-running task in Claude Code that involves many sequential tool calls (e.g. a /goal-style multi-step workflow, large refactor, or extended agent loop).
2. Wait for the API to return a transient 529 Overloaded (more likely during peak load).
3. Observe that Claude Code aborts the current turn and returns control to the prompt with no further action.

Expected behavior

Transient 5xx responses (particularly 529) should be treated as retryable. Claude Code should:

• Retry with exponential backoff (e.g. 1s / 2s / 4s / 8s, capped) for a bounded number of attempts.
• Preserve task state across the retry so the in-progress goal/loop continues from where it left off rather than discarding context.
• Only surface the error to the user if retries are exhausted.

Actual behavior

A single 529 immediately terminates the current turn. Any in-progress multi-step task (/goal, agent loop, long tool-call chain) is lost and must be manually restarted, often re-doing significant work.

Impact

• Long-running tasks are effectively unreliable during periods of elevated API load.
• Work performed before the error (tool calls, partial edits, gathered context) is not resumed automatically.
• Users running unattended/autonomous workflows return to find them halted partway through.

Suggested improvements

1. Add automatic retry with exponential backoff for 429, 503, and 529 responses.
2. Make the retry policy configurable (max attempts, base delay) via settings.
3. For multi-step / autonomous tasks, expose a resume mechanism so the loop can continue after a recovered transient error.
4. Optionally, log retry attempts so users can see that the harness recovered rather than silently stalling.

Environment

• Claude Code CLI (latest)
• macOS
• Errors observed during normal interactive use; no unusual configuration.