Long-horizon tasks in real containers, from payment systems to power grids. Every task has a recorded passing run.
Six tasks from the sample set. Each one is a complete Harbor task with held-out tests, a reference solution and a recorded passing run.
Find which files and byte ranges in a corrupted 7z archive can still be trusted, with 100% precision and at least 98% recall.
Repair an ISO 8583 retry detector so it tracks about 76 million live payment identities within strict memory and speed limits.
Fix an audit library so each release review uses only the evidence that existed at the time, across 60,000 measurement rows.
Reconcile four exports that disagree into one stock ledger, and split movements across bookings without double counting.
Recover the rules behind old route plans, then produce routes that pass an independent cost and timing check on new cases.
Fix a scheduling builder so one full-day safe region keeps the links between time slices, within the memory limit.
Samples use the same folders as the full dataset. Run them in Harbor with the agent harness you already use. Senior engineers write and review every task, and a fairness check removes tasks that are hard only because the spec is unclear.
The best agent we tested fails most tasks on the first try but solves most within three. That gap gives a model room to learn.
GPT-5.6 Sol with Codex, 5 runs on each of the 6 sample tasks (30 runs).
Any lab can license Terminal-Bench 4.0. We size the full set to your complexity requirements, and we can extend it with new domains or harder tasks.