Your first real task

Use the disposable workspace from Initial configuration, after the ChatGPT file-read check succeeds. This exercise runs one harmless command that prints a marker. It does not modify project files.

Grant command execution deliberately

Stop the demo's foreground serve with Ctrl+C. In that demo's config.toml, set execution_enabled = true under [machines.local]. Leave workspace_write_enabled = false; typed file writes are not needed here.

Restart with the same paths:

openmachine --config "$DEMO/config.toml" --state-dir "$DEMO/state" serve

This is a native-account command grant, not a sandbox restricted to the project directory. Only proceed on an account and workspace appropriate for testing. The client may separately ask you to approve the command.

Ask for the command and its original result

Use OpenMachine on machine local, workspace project. Inspect the current
workspace and relevant execution Skill. I authorize one command in that exact
workspace: python3 -c 'print("openmachine-ready")'. First resolve python3 without
running it. Save one execution ID, run the command with a 30-second timeout,
and read that original execution until its result is known. Do not write files,
install anything or change configuration. Report the ID, actual output, status,
exit code and cleanup result. Do not rerun it after a lost reply.

The command's tool arguments use an argv array, not a shell string:

["python3", "-c", "print(\"openmachine-ready\")"]

The assistant supplies the observed workspace cwd, current HEAD where present and a fresh execution ID. It must use the actual discovered interpreter when the selected runtime PATH does not resolve python3.

Check success

The original receipt must show status=succeeded, exit code 0, output openmachine-ready followed by a newline, and verified cleanup. Save the execution ID. Merely seeing an accepted start or a wrapper exit is insufficient.

You can then ask for a read-only retrieval using that same ID. It should return the original result, not launch a new command. A client timeout does not prove that the command failed or never ran.

Try a durable multi-stage task next

After separately enabling typed workspace writes, this is a bounded example:

In the same disposable project, create one durable OpenMachine task to create
hello.txt with exactly 'hello from OpenMachine' plus a newline, then verify its
actual contents and SHA256. Do not overwrite an existing hello.txt. Read the
applicable Skills and bind the task to the observed workspace. Keep the original
task, change and execution IDs; finish only after verification. If interrupted,
report the existing task ID rather than creating a replacement.

This second example needs its own explicit write grant and task contract. Tasks explains the lifecycle and recovery. The first one-command check does not require a fabricated multi-stage contract.

After testing, disable the demo grants or stop its server. Do not delete a state directory or environment while original commands still depend on it.