Release 0.1.132
Agent costs for resumed runs are corrected, merged features show where their work really is, project settings and the agent log each lose a control that repeated another, and GitHub sync results are easier to read.
Upgrade notes
- Two database migrations, applied automatically by the upgrade.
- Costs are corrected on first start. The first time the upgraded app starts, it recalculates the cost of agent runs that resumed an earlier run, so cost totals on features and projects may drop noticeably. Nothing to run by hand.
.envanddocker-compose.yml: nothing to edit.
Improvements
Project settings has no Reset button. Reset could put back values you had already saved, such as a repository you had just linked or a saved token. Save Changes is now the only action; leaving the page discards unsaved edits.
"Restart from…" is gone from the implementation agent log. Resume already continues a stopped attempt from the step where it stopped, so it is the one control for picking work back up.
GitHub sync results read one line per category. A branch that needs attention is listed on its own line with what to do about it, followed by what was pushed and pulled. Branches a feature is still working on are shown as a count, with their names under Show branches. A push GitHub rejected is now reported as blocked rather than diverged, and Create repository says when its first push failed.
Fixes
Agent costs were far too high for features whose agents resumed earlier runs. Each resumed run was charged for every run before it again, so a feature resumed many times could show thousands of dollars it never spent. Each run is now counted at what it spent itself, and costs already recorded are corrected when you upgrade. A run whose own spend cannot be told apart, such as one on a model with no published price, keeps the cost it was recorded with.
Features merged before switching to pull-request merge mode kept saying "Merged to repave branch" after their work reached your base branch. They are now labelled with the base branch once Repave finds their work there.
Starting an implementation on a busy server could fail with "Transaction already closed". Launches now wait long enough for the feature to become free.
A job restarted right after its container crashed could have its first step marked failed. Only work that started before the crash is failed now.