Lab 1 — Constrained baseline
← Back to labs index

Lab 1 — Constrained baseline

This independent exercise repeats the build-your-own Lab 1 M1 task using Codex CLI. It cannot guarantee a bare call: Codex owns the agent loop and read-only sandboxing does not disable all tools. Do not compare its results as if the two harnesses had identical capabilities.

  1. Start in lab01-baseline/. Install Codex CLI and check codex exec --help. Follow the common preflight for your shell to create a private home from this lab’s config and set workdir to the tracked workspace/.

    Edit the copy to use your GPT Responses deployment and Azure OpenAI endpoint ending in /openai/v1. Load your own resource API key into AZURE_OPENAI_API_KEY via a secret manager; env_key names that environment variable, not a literal key. The Microsoft Foundry guide says Entra ID is not currently supported for Codex. Never store the key in files or logs. Create no fixtures: workspace/ must contain only .gitkeep. Disable inherited project instructions with the option below. Do not configure a simulator, MCP servers or skills; this directory does not contain store evidence.

  2. Run one fresh, noninteractive task with no prior session:

    codex exec --skip-git-repo-check --sandbox read-only -C "$workdir" --json `
  -c project_doc_max_bytes=0 `
  "Produce a Store Health Report: (a) for each service, what it does, its language, port and dependencies; (b) every catalog issue (missing or weak descriptions, price anomalies, duplicates); (c) remediate the catalog issues. End with a StoreHealthReport JSON object with services, catalog_issues and remediations_claimed arrays. Cite file:line or product ID for each claim. Do not invent evidence or claim to have changed anything you could not access."

    codex exec --skip-git-repo-check --sandbox read-only -C "$workdir" --json \
  -c project_doc_max_bytes=0 \
  "Produce a Store Health Report: (a) for each service, what it does, its language, port and dependencies; (b) every catalog issue (missing or weak descriptions, price anomalies, duplicates); (c) remediate the catalog issues. End with a StoreHealthReport JSON object with services, catalog_issues and remediations_claimed arrays. Cite file:line or product ID for each claim. Do not invent evidence or claim to have changed anything you could not access."

    --json emits Codex’s event stream. Inspect it for tool calls before describing the result as tool-free; do not treat a read-only sandbox as a tool-disable flag. The model cannot inspect the store or service source in this empty directory, so honest unknowns and no claimed fixes are expected. If it did access any external resources or tools, record that fact.

  3. Record the Codex version, deployment, auth route (never the credential), whether tools were called, the report, and any usage Codex actually exposes. Mark unavailable metrics as unavailable, not zero. This pilot does not yet include a shared grader or claim a comparable scorecard. End the run, follow cleanup for the workspace and CODEX_HOME, and unset AZURE_OPENAI_API_KEY.

Next: introduce access to the same source and store simulator as the build-your-own labs, with explicit permissions and measurable grading.