Lab 4 — Planning and todos

Capability: Plan-then-execute can be requested of Codex; this does not prove that its internal todo state is identical to the build-your-own planner/executor and write_todos contract.

Install Codex CLI; create a private temporary CODEX_HOME, copy this lab’s config.toml.example to $CODEX_HOME/config.toml and fill in the endpoint (ending in /openai/v1) and deployment. Set AZURE_OPENAI_API_KEY to your resource key via a secret manager. The Microsoft guide says Entra ID is not supported for Codex; never save the key in files. Workspace: start in lab04-planning/ and follow the common preflight. Use workspace/ for planning and workspace-b/ for the independent greedy comparison. Create identical synthetic catalogs in both:

$second = Join-Path $labdir 'workspace-b'
$catalog = '[{"id":1,"name":"Demo leash","price":0},{"id":2,"name":"Demo bowl","price":20}]'
$catalog | Set-Content (Join-Path $workdir 'catalog.json') -Encoding utf8
$catalog | Set-Content (Join-Path $second 'catalog.json') -Encoding utf8
codex exec --skip-git-repo-check -C "$workdir" --sandbox read-only --json `
  "Read catalog.json. First propose a numbered plan identifying questionable prices and the evidence for each; do not change any files."

second="$labdir/workspace-b"
printf '[{"id":1,"name":"Demo leash","price":0},{"id":2,"name":"Demo bowl","price":20}]\n' > "$workdir/catalog.json"
cp "$workdir/catalog.json" "$second/catalog.json"
codex exec --skip-git-repo-check -C "$workdir" --sandbox read-only --json \
  "Read catalog.json. First propose a numbered plan identifying questionable prices and the evidence for each; do not change any files."

Review the plan. In a separate run with a fresh copy of the catalog, ask for an unplanned greedy analysis using -C "$second"; compare omitted issues and tool usage. If you want to carry out a planned change, reset workspace-b/ and recreate its original catalog first, use interactive codex -C "$second" --sandbox workspace-write --ask-for-approval on-request, inspect each proposed change and approve only explicitly authorized local edits. Do not connect a live store or assume planning text is an enforcement boundary. Verify the catalog file and record planned versus actual edits; mark unknown internal todo behavior as unavailable.

Follow reset and cleanup after both passes.