Lab 2A — Tool loop
← Back to labs index

Lab 2A — Tool loop

Capability: Codex already owns the tool loop and shell tool; you cannot replace its call/result pairing with the build-your-own loop. Compare an empty-workspace run to a run with a synthetic service manifest.

Install Codex CLI. Set CODEX_HOME to a private temporary directory, copy this lab’s config.toml.example to $CODEX_HOME/config.toml, and replace 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 credentials in files. Workspace: start in lab02-tool-loop/ and follow the common preflight. It sets workdir to this lab’s workspace/ and creates a private home from this lab’s config. Create the fixture (besides .gitkeep, this is the only input):

"service: order-service`nlanguage: Go`nport: 3000`ndepends_on: RabbitMQ" |
  Set-Content (Join-Path $workdir 'service.txt') -Encoding utf8

printf 'service: order-service\nlanguage: Go\nport: 3000\ndepends_on: RabbitMQ\n' > "$workdir/service.txt"

Run in read-only mode (not unrestricted shell execution):

codex exec --skip-git-repo-check -C "$workdir" --sandbox read-only --json `
  "Report the service name, language, port and dependency from service.txt; cite the line numbers. Do not modify files."

codex exec --skip-git-repo-check -C "$workdir" --sandbox read-only --json \
  "Report the service name, language, port and dependency from service.txt; cite the line numbers. Do not modify files."

Inspect the JSON events: find a tool invocation, its result and the final answer; check the cited line against service.txt. Run again with -C pointing to this lab’s empty workspace-b/ (only .gitkeep) and -c project_doc_max_bytes=0, and note the difference in grounded answers. For PowerShell, use -C (Join-Path $labdir 'workspace-b'); for bash, use -C "$labdir/workspace-b". Report whether tools were actually called. Do not use danger-full-access or disable approvals: the build-your-own Lab 2A’s deliberately unrestricted mode is not a safe requirement for this path. Record any inaccessible internal loop details as unavailable.

Follow reset and cleanup after both runs.