Test one live iteration

Use a temporary project so the test does not change the user's backlog. The test starts a real agent and can consume its usage allowance. A local stub checks the runner, but does not establish live agent behavior.

Workflow

1. Resolve scripts/run-live-test.sh relative to this skill's directory. To test a particular checkout, set LOOPER_BIN to its absolute looper binary path, such as build/looper. The runner otherwise prefers its checkout over PATH.
2. Run the script. It creates PROJECT.md and a two-task backlog, then initializes Git, then executes Looper with --once --jobs 1 and MAX_ITERATIONS=1. The integration check requires a nonempty README.md.
3. Inspect the task entry, result, and final report. The expected selection is T2, ahead of T10 at equal priority.
4. Verify the actual outcome: T2 is done, T10 remains todo, and a nonempty README.md exists. The runner checks these conditions. Exit code 2 is expected because one task remains.
5. Report the executable, agent, selected task, checks, outcome, and temporary paths. Keep the artifacts for inspection. Do not report a blocked task or an exit code alone as a successful test.

If the runner is unavailable

Create the same fixture in a temporary directory and run the intended Looper executable with CODEX_JSON_LOG=0, LOOP_DELAY_SECONDS=0, and MAX_ITERATIONS=1. Check the same outcomes before reporting success.