Description

Intermittent error message surfaces in Telegram chat that reads:

⚠️ API provider returned a billing error — your API key has run out of credits or has an insufficient balance. Check your provider's billing dashboard and top up or switch to a different API key.

The error is misleading — no API billing limits have actually been reached. More critically, when this error appears, the agent's actual response is sometimes not delivered, requiring the user to re-send their message to get a reply.

Environment

• OpenClaw version: 2026.2.9 (33c75cb)
• Channel: Telegram
• Auth: Claude Max 20x (OAuth token)
• Other APIs: Grok (xAI) for web search, Gemini for memory embeddings
• Context at time of error: 100% (200k/200k)

Reproduction

• Occurs intermittently during active conversation
• No specific trigger identified — happens on normal text responses (no tool calls)
• Not correlated with web_search, memory_search, or any specific tool
• More frequent during long/heavy sessions near context limit

What we've ruled out

• Anthropic weekly allowance — confirmed ~90% remaining
• Grok/xAI credits — confirmed $23+ balance remaining
• Gemini API — billed through Google Workspace, no limits hit

Key issues

1. Error message is misleading — says "billing" but no billing limit is actually reached
2. Swallows the agent's response — user sees error instead of the actual reply, has to re-ping
3. Not logged anywhere — checked journalctl -u openclaw, session transcripts, and ~/.openclaw/logs/ — no trace of the error
4. Intermittent — not every message, making it hard to diagnose

Expected behavior

• If a transient API error occurs, retry before surfacing to user
• If surfaced, don't swallow the actual response
• Log the underlying error (HTTP status code, provider, endpoint) for diagnosis

Suspected cause

Possibly transient Anthropic rate limiting (per-minute, not weekly allowance) being caught and displayed as a billing error. May be related to high context usage (100%) triggering compaction-related API calls.