Research runs here start with research/_shared/pick_track.py. Its CLAIM CHECK says whether a sibling run has already landed work on the programme in the last six hours. To answer, it needs six hours of origin/main. Cloud clones are shallow, so the tool tests that with git rev-list -1 --before='6 hours ago' origin/main: if some commit is older than six hours, the window is covered.
On 2026-11-13 a run reported that the check kept saying "the shallow history is shorter than 6h" after it ran the exact fetch the message suggested. We reproduced it in a fresh session:
| step | commits on origin/main | oldest | covered? |
|---|---|---|---|
| fresh clone | 50 | 7 days (stale ref) | yes (wrongly: the ref was a week old) |
| the tool's own refresh (plain fetch, since "covered") | 88 | 4 hours | no |
git fetch --shallow-since='6 hours ago' (as suggested) | more | under 6 hours | no |
git fetch --shallow-since='7 hours ago' | more | 6–7 hours | yes |
There were two faults. The refresh decided whether to deepen before fetching. A week-old ref passes "is there a commit older than six hours?", so it ran a plain fetch, and that replaced the history with four hours' worth. Then the fallback message suggested --shallow-since='6 hours ago'. By definition, that fetch keeps only commits newer than six hours, so the test it was meant to satisfy could never pass. The refresh path already used hours + 1, but the message did not.
The fix re-checks coverage after a plain fetch and deepens once with --shallow-since='7 hours ago' when the window is still short. The message now suggests the same seven-hour window. Two new tests cover both cases. In the live session, a run that starts from a stale ref now ends with the window covered and no manual step.