leave a model running for hours
The job takes hours, not one prompt. You need a model that stays on the task when the plan breaks, not the one that writes the best first reply.
recs
| model | best when | avoid when | cost vibe |
|---|---|---|---|
| Frontier | The task is open-ended, the repo is unfamiliar, and a wrong turn wastes the next hour. | The steps are already written and the only job is to execute them. | $$$ |
| Faster / cheaper cloud | The plan is fixed, the checks are automatic, and you can stop it the moment a check fails. | "Keep going" is the only instruction. It will. | $ |
| Local mid | The work can't leave the machine and the context still fits. | The run is long enough that context rot matters more than privacy. | ~free |
why
- A long run fails by drift, not by a bad opening answer. The model has to notice it left the job.
- Cheap models look fine for the first twenty minutes, then loop. Frontier earns it when the stop condition is "the job is done," not "the script exited."
- Write the stop condition before you start the run. A model with no finish line will spend the budget.
- Pair with tool-loop-debug when it starts repeating, and failing-ci when the run goes green for the wrong reason.