[BUG] Why is the 13K buffer (`nx1 = 13000`) needed in the autocompact threshold calculation?

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

In the npm package @anthropic-ai/claude-code, the file cli.js (or cli-new.js in the source) contains:

Definition:

var nx1 = 13000,
  tyI = 20000,
  eyI = 20000;

Usage:

function AkI() {
  let A = s4A(),     // Available input tokens
    B = A - nx1,     // Auto-compact threshold = Available - 13000
    Q = process.env.CLAUDE_AUTOCOMPACT_PCT_OVERRIDE;
  // ... environment variable override logic ...
  return B
}

The nx1 = 13000 constant is used to calculate the auto-compact threshold by subtracting 13K from the available input tokens.

🔍 Current Understanding

Based on code analysis, the 13K buffer appears to be designed to:

• Reserve space for the compression prompt (~2,000 tokens)
• Account for system overhead (~720 tokens)
• Provide a safety margin for token counting errors

However, there is no comment or documentation explaining:

• Why exactly 13,000 tokens?
• What is the calculation basis?
• Why not 3K, 5K, or 10K?

What Should Happen?

🧪 Experimental Findings

I conducted experiments modifying the nx1 value:

1. Changed to 3K (nx1 = 3000):

• ✅ System runs normally
• ✅ Compression works correctly
• ✅ Model receives more context

1. Changed to 0 (nx1 = 0):

• ✅ System runs normally
• ✅ Compression works correctly
• ✅ Model receives even more context

❓ Questions

1. Why is 13K specifically chosen?

• Is this an empirical value from testing?
• Is there a specific calculation or formula?
• What edge cases does 13K protect against that 3K or 0 cannot?

1. Is 13K necessary?

• My experiments show that 3K and even 0 work fine
• The actual compression prompt only needs ~2,720 tokens
• Why reserve 13K when only ~2.7K is actually used?

1. What are the risks of reducing it?

• Are there specific scenarios where 13K is critical?
• Does it depend on model type, context size, or other factors?
• What happens in edge cases with very large contexts?

1. Documentation request:

• Could you add comments explaining the rationale for 13K?
• Or provide documentation on how this value was determined?

This has been puzzling me for a while.

Error Messages/Logs

Steps to Reproduce

1. change 13k -> 3k or sth else
2. work until auto compression

Claude Model

None

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.0.28

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

VS Code integrated terminal

Additional Information

_No response_