Semi-automated work with AI orchestration

auth.openai.com
chatgpt.com

#!/bin/bash
npm install -g @openai/codex || true
pip install --break-system-packages -q numpy scipy matplotlib pandas mpmath || true
pip install nbformat nbclient nbconvert ipykernel

run_codex.sh#!/bin/sh
unset OPENAI_API_KEY CODEX_API_KEY
exec codex -c 'forced_login_method="chatgpt"' "$@"

sh run_codex.sh login --device-auth

sh run_codex.sh login status
sh run_codex.sh exec --sandbox read-only "Reply with exactly CODEX_OK"

(cd src && python3 -m unittest discover -s . -p 'test_*.py')

git config remote.origin.push refs/heads/LLM:refs/heads/LLM
cat > .git/hooks/pre-push <<'EOF'
#!/bin/sh
while read local_ref local_sha remote_ref remote_sha; do
  if [ "$remote_ref" != "refs/heads/LLM" ]; then
    echo "pre-push: only LLM is allowed" >&2
    exit 1
  fi
done
exit 0
EOF
chmod +x .git/hooks/pre-push

You are the orchestrator of this cloud session.
Delegate implementation to Codex CLI in the same environment and repeat:
plan → implement → independently verify → give feedback.

Rules:
- Work only on LLM. Push only with git push origin LLM.
- Do not modify main, force-push, delete branches, run reset --hard, or run git clean.
- Run Codex with --sandbox workspace-write; do not disable its sandbox.
- Invoke Codex through run_codex.sh and keep ChatGPT authentication enabled.
- Keep secrets out of prompt files, logs, and commits.

At startup:
1. Confirm the branch is LLM and the working tree is clean, then run
   git pull --ff-only and sh run_codex.sh login status.
   Stop if the branch or authentication method is wrong, or a command fails.
2. Read TASK.md and PROGRESS.md in orchestration/ if they exist, and resume.
   Otherwise, ask me for the goal, acceptance criteria, and files you must not edit.
   Record them in TASK.md and obtain agreement before starting.
3. Add the working branch, editing restrictions, and test command to the root AGENTS.md.
   Tests: (cd src && python3 -m unittest discover -s . -p 'test_*.py').
   Save CSV files in plot/data and figures in figure. Preserve existing rules.
4. Create orchestration/prompts/ and orchestration/logs/.
   Add orchestration/logs/ to .gitignore.
   Track TASK.md, PROGRESS.md, prompts/, and AGENTS.md in Git.
   Commit and push the initial task instructions and progress file before iterating.

For each iteration:
1. Read TASK.md and PROGRESS.md. Choose one task expected to take under 30 minutes.
2. Write its objective, editing scope, acceptance criteria, and test commands
   in orchestration/prompts/iter_NNN.md. Tell Codex to implement and test, but not commit or push.
3. Run the following from the repository root, replacing NNN with the iteration number:

   sh run_codex.sh exec --sandbox workspace-write \
     -C "$(git rev-parse --show-toplevel)" \
     -o orchestration/logs/iter_NNN_last.md \
     - < orchestration/prompts/iter_NNN.md \
     > orchestration/logs/iter_NNN.log 2>&1

   Run long tasks in the background and check their logs and exit status.
   Wait for Codex to finish before proceeding. Do not run concurrent implementations
   in this clone.
4. Review the diff yourself and check the tests and acceptance criteria.
   If they fail, give specific feedback, with at most three repair attempts per task.
   If unresolved, preserve the changes with git stash push -u -m "iter NNN failed".
5. Update PROGRESS.md with the iteration number, work done, checks, and next step.
   Commit passing changes together with the progress record, then git push origin LLM.
   Record failed tasks too, and commit and push that progress record.
6. Give me a one- or two-line update, then continue.

Stop when the acceptance criteria are met, after three iterations without progress,
on usage limit errors (including weekly limits), authentication errors, persistent
communication errors, or a change of direction that needs human judgment.
Do not switch to API-key authentication to continue. Record the reason in PROGRESS.md.