Skip to content

CODEX.md

Latest commit

 

History

History
78 lines (54 loc) · 2.5 KB

CODEX.md

File metadata and controls

78 lines (54 loc) · 2.5 KB
Raw

Running This With Codex

This repo is built for people who already pay for ChatGPT and want to run the loop from that plan instead of wiring an API key on day one.

npm install -g @openai/codex
codex login

Then:

scripts/run_codex_step.sh improve_scoring

The sandbox flag is not optional

improve_scoring has to edit config/scoring.yaml and run the eval. Codex defaults to a read-only sandbox, which means without the right flag the step runs, prints something reasonable, and changes nothing.

scripts/run_codex_step.sh therefore always passes:

codex exec -m "$model" --sandbox workspace-write "$(cat "$prompt")"

If you invoke Codex by hand, pass it yourself. A step that appears to succeed while editing nothing is the most annoying failure mode in this repo, because the gate will keep reporting the same score and it looks like the improver has nothing to say.

Why files instead of a database

Codex is good at reading local context, editing files, running commands, and explaining what it changed. That means the loop needs no infrastructure:

  • YAML for the rules a human argues with
  • JSONL for memory that only ever appends
  • shell scripts for repeatable runs
  • an exit code for the gate

Everything is inspectable with cat and reviewable in a diff. That is the property that makes the pull request readable.

AGENTS.md

Codex reads AGENTS.md from the working directory before it acts. That file is the law: what it may edit, what it must cite, and when it must stop. The scripts cd to the repo root so the law is always in scope.

Keep it short. If it needs a table of contents it is too long to be enforced.

Choosing a model

CODEX_MODEL=<model> scripts/run_codex_step.sh improve_scoring

Judgment work wants a strong model. The improver reads outcome reasons and argues about weights, so this is not the place to economise.

Recommended loop

  1. Run python3 scripts/propose_improvement.py offline until the mechanics are boring.
  2. Replace the sample outcomes with your own.
  3. Rewrite evals/fixtures.yaml around accounts you wish had been routed differently.
  4. Run scripts/run_codex_step.sh improve_scoring by hand, twice, and read both diffs.
  5. Only then put scripts/weekly_tune.sh on a schedule.

Do not automate the merge. Ever.

Docs