Problem

When the conversation context grows too large (lots of screenshots, long logs, many tool calls), the Anthropic API returns 413 request_too_large. Currently, the gateway just forwards this error to the user, making the assistant completely unresponsive until someone manually clears the session.

This happened to me on Jan 6-7, 2026 - the assistant was stuck returning 413 errors for hours until I manually cleared the context with Claude Code.

Expected Behavior

The gateway should handle this gracefully:

1. Detect approaching limit - Track token count and warn/act before hitting the 413
2. Auto-truncate on 413 - When receiving a 413, automatically remove oldest messages from context and retry
3. Preserve memory before cleanup - Write a summary to memory files before truncating so context isn't completely lost
4. Notify user - Send a message like "Context was getting too large, I've cleaned up old messages but I'm still here"

Suggested Implementation

1. In the request handler, catch 413 errors specifically

2. On 413:

   • Log the event
   • Truncate the oldest 50% of messages (or until under limit)
   • Retry the request
   • If still failing, truncate more aggressively
   • Send a notification to the user about what happened

3. Proactively:

   • Track approximate token count per session
   • When approaching 80% of model's context limit, start summarizing/truncating older messages

Impact

Without this fix, users who are away from their machine (like traveling) have no way to recover - the assistant becomes completely unresponsive.

Environment

• Model: claude-opus-4-5 (200k context)
• Heavy use of peekaboo screenshots and long tool outputs contributed to the overflow