Repository navigation
Replies: 1 comment
|
A use case for a per-project switch: a 6 GB Azure DevOps repo with 9,706 checkpoint refs. Captures there time out (the server logs that the baseline was skipped after repeated capture timeouts), yet every finished turn still runs about 9 checkpoint git processes on a server where each spawn is expensive (#31422). I'd turn checkpoints off for that repo and keep them for the smaller ones. Drafted with Claude Opus 5.5 (Claude Code) on behalf of @lingolooper-simon. |
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
T3 Code writes a checkpoint commit on every turn and keeps each one under
refs/t3/orchestration-v2/checkpoints/. Nothing in the settings turns this off. I readpackages/contracts/src/settings.ts, which defines 492 top-level keys, and none of them mentions checkpoints.packages/shared/src/serverSettings.tshas none either.apps/server/src/orchestration-v2/CheckpointCaptureService.tsdecides whether to capture by reading the run status. It never reads configuration. Three run states stop a capture, those being completed, interrupted and cancelled, along with runs that already rolled back. A user has no setting, no environment variable and no config file to change.Nothing prunes the refs once they exist, so they accumulate for as long as the repository lives, and no limit caps how large a capture can get. Issue #38102 reports 1,828 refs over 18 days and a
.gitof 23 GB against 1.4 GB for the same repo on a machine without checkpoints, which filled a 96 GB disk and started failing turns. Issue #26227 reports capture attempts that hit a 30 second git timeout, never finish, and leave thousands oftmp_objdir-bulk-fsyncdirectories behind. Pull request #34973 would remove a thread's refs when the thread is deleted, and it is still unmerged.I am asking for a setting next to the background activity controls.
checkpoints: per-turnkeeps the current behaviour, capturing every turn and offering revert as it works today.checkpoints: offcaptures nothing and disables the revert action for threads created in that mode.checkpoints: manualcaptures only when the user asks.A retention limit that prunes by age or by refs per thread would help as well, though the opt-out is the part I care about.
I understand what
offcosts. The revert model depends on per-turn capture, so turning it off removes the ability to roll a thread back rather than adjusting a preference. I also know that capturing untracked files which are not ignored is intended, per #26487. I am not asking for that to change.All reactions