What
apps/framework/harness/run-eval.ts:
const FORCE = !args.has('--skip-existing');So the default for every run is to re-run and overwrite cells that already have results, and preserving them requires opting in.
Why that is the wrong way round here
A 24-cell Outpost run costs $35–40 and a couple of hours, and a partial run is the normal way to recover — three runs were interrupted on 21–22 August by a wiped config, exhausted credits and a stopped Docker daemon. Each time the obvious command re-ran work already paid for.
It also misled me directly: I told the user the harness would resume automatically and save three completed cells. It did the opposite.
Done when
Either the default preserves existing results and re-running is the opt-in, or the polarity is documented with the reason it is that way. There may be a good reason — a stale result silently surviving a code change is its own hazard — but if so it should be written down, since the current behaviour reads as an accident.
What
apps/framework/harness/run-eval.ts:So the default for every run is to re-run and overwrite cells that already have results, and preserving them requires opting in.
Why that is the wrong way round here
A 24-cell Outpost run costs $35–40 and a couple of hours, and a partial run is the normal way to recover — three runs were interrupted on 21–22 August by a wiped config, exhausted credits and a stopped Docker daemon. Each time the obvious command re-ran work already paid for.
It also misled me directly: I told the user the harness would resume automatically and save three completed cells. It did the opposite.
Done when
Either the default preserves existing results and re-running is the opt-in, or the polarity is documented with the reason it is that way. There may be a good reason — a stale result silently surviving a code change is its own hazard — but if so it should be written down, since the current behaviour reads as an accident.