You’re in the flow. Claude Code is working through a tricky refactor. You step away for coffee, come back, and the agent is still “thinking.” The terminal looks busy. But nothing is happening. And your API bill just grew by $20.
A freshly filed bug (#76837) on the Claude Code repository reveals a quiet money-burning behavior: when the agent hits its usage or token limit, instead of stopping and reporting the error, it keeps auto-retrying indefinitely. The only symptom is a stalled session and a climbing bill.
The Issue
The user reports that Claude Code reaches its platform-imposed usage limit (a 429 or similar token-cap response) and, rather than surfacing the error and pausing, enters an infinite retry loop. The model keeps sending the same request, getting the same rejection, and sending it again — each attempt burning API credits.
There is no user-facing indicator that this is happening. No error dialog. No warning in the TUI. The session looks active. The only way to detect it is to either spot a stalled turn (no new output for minutes) or check your API usage dashboard and see a spike in requests that returned nothing useful.
The issue is labeled for Windows and VS Code platforms, but the core mechanism — an unbounded retry loop on API rejection — is cross-platform. If Claude Code can’t get a model response, it keeps trying until your quota runs out or you manually kill the session.
Are You Affected?
Check if you’ve hit this bug:
# Check your Anthropic API usage for retry spikes
# Look for repeated 429 or 529 responses in your logs
# (path varies by setup — check your API dashboard)# If you use Claude Code CLI v2.1.207 or earlier:
# Run a task that approaches your usage limit, then watch:
tail -f ~/.claude-code/logs/*.log | grep -i "retry\|429\|rate.limit"You’re affected if:
- You use Claude Code on Windows or VS Code (labeled platforms)
- You run tasks large enough to hit Anthropic’s rate or token limits
- You’ve ever left a Claude Code session unattended and come back to a stalled agent
- Your API bill shows unexplained spikes during idle periods
The Fix
Immediate workaround: If you notice a stalled session, kill it (Ctrl+C in the terminal) and restart. Check your API dashboard for unexpected usage before resuming.
Long-term: Watch issue #76837 for an upstream fix. The expected solution is a retry budget — stop after N consecutive failures and surface the error to the user, or at minimum add a configurable max_retry_attempts setting.
Prevention:
- Set hard usage alerts in your Anthropic Console (Settings → Usage Limits) so you get notified before runaway retries cost real money.
- Keep an eye on session responsiveness. If Claude Code has been “working” silently for more than 2-3 minutes without tool calls or output, it’s likely in a retry loop.
- Run expensive or long tasks inside a
timeoutwrapper that kills the process after a reasonable duration.
Why It Happened
When Claude Code sends a request to the Anthropic API and gets back a 429 (Rate Limit) or 402 (Usage Limit) response, the core loop treats it as a transient error. The default behavior for transient errors is to retry — which is correct for network blips and temporary congestion. The gap is that there is no distinction between “try again in a second” errors and “you’ve hit your hard limit” errors.
The retry module lacks a circuit breaker. It doesn’t count successive failures, doesn’t check if the error is terminal, and doesn’t surface the failure to the user. Each retry consumes a fresh API call — meaning every attempt is billable, even though every attempt produces the same rejection.
The recently filed issue #63021 in the Hermes project shows the same class of problem being addressed there — classifying 429 quota limits as billing errors rather than transient failures. The pattern is spreading across the ecosystem.
FAQ
Q: How can I tell if I’m in a retry loop vs. just a slow model? A: A slow model produces incremental output — you’ll see tokens appear in the TUI or response stream. A retry loop is silent: no new text, no tool calls, no progress. Kill the session and check your API usage dashboard for a spike in rejected requests.
Q: Does the issue affect both CLI and VS Code versions?
A: The bug report is labeled for both platform:windows and platform:vscode. The underlying retry logic is in the core agent loop, so it could affect any platform, but these are the confirmed configurations.
Q: Can I limit API spend directly in Claude Code to prevent this? A: Not yet. Claude Code does not have a built-in spend cap. You must set limits at the API key level in the Anthropic Console. The bug is that even with API-side limits, the agent keeps retrying and burning quota rather than stopping gracefully.
Don’t overpay for AI tools. Check aiFiesta for exclusive pricing on coding agents.