[BUG] API returns 500 Internal Server Error when context is near capacity (~92%)

Description

Claude Code returns 500 Internal Server Error when context usage is high (~92%), even though there are still ~16k tokens available. The API should either work normally or return a graceful error message, not an unhandled 500.

Error Message

API Error: 500 {"type":"error","error":{"type":"api_error","message":"Internal server error"},"request_id":null}

Context at Time of Error

• Total context: 184k / 200k tokens (92%)
• Free space: 16k tokens (8%)
• Model: claude-opus-4-5-20251101

Context Usage: 184k/200k tokens (92%)
⛁ System prompt: 5.2k tokens (2.6%)
⛁ System tools: 16.0k tokens (8.0%)
⛁ Custom agents: 370 tokens (0.2%)
⛁ Memory files: 6.1k tokens (3.1%)
⛁ Messages: 156.7k tokens (78.4%)
⛶ Free space: 16k (7.8%)

Steps to Reproduce

1. Use Claude Code until context reaches ~90%+
2. Attempt to send a message or perform an edit
3. API returns 500 Internal Server Error

Expected Behavior

• Either the operation should succeed (16k tokens is still significant space)
• Or return a graceful error like "Context too large, please run /compact"

Actual Behavior

• Unhandled 500 Internal Server Error
• No useful error message
• User has no idea what went wrong

Impact

• Users paying for 200k context window effectively can't use the last ~10-15%
• Poor UX with cryptic error message
• Forces users to compact earlier than necessary

Environment

• Claude Code version: 2.0.75
• Model: claude-opus-4-5-20251101
• Plan: Max 20X
• Auto-compact: Disabled