Research2026-08-04

An independent researcher recorded exactly what OpenAI's Codex command-line coding tool sends to the model, using a local server that captured requests instead of calling a real provider. A 16-character prompt produced a request of about 43,000 bytes, almost all of it built-in instructions, tool definitions and environment details rather than the prompt itself. Files that were never read, including ignored files and a fake credentials file, did not leave the machine, but explicit reads bypassed .gitignore. Token counts were local estimates, not billing figures.

What changed

The contents of what the Codex coding assistant sends on each turn were not publicly measured in detail.

What it unlocks

Seeing which local files, instructions and command output actually leave the machine, and estimating how quickly a session's request grows.

  • 16-character prompt produced a 42,980-byte request
  • roughly 9,435 estimated tokens in the first request, of which about 25 were the prompt
  • 1,000 synthetic instruction markers added 24,197 bytes
  • a 10,000-line command result pushed one request to 88,480 bytes

What you need to act on it

  • Codex CLI configured to point at a local recording server

Send this to someone who needs it

Shares the story and its sources. Nothing about you.

What does this mean for your job?

This is the story as everyone gets it. Once a week we send you the version written for your role — what changed, why it matters for the work you actually do, and one thing to try. Free while we tune it.