Three cases
gstack 1.91.185.0 repairs three flaky paid eval cases
gstack's newest release repairs three paid eval cases that kept failing at random: two for plan design review, one with a ten-minute stall, and one for the ship docs sync.
Rating
“4/10 on design” now counts as a design rating
Opus 5.5 rated a plan four out of ten on design, and the focus check failed it for leaving out the words design completeness. Any sentence that scores the design out of ten now counts.
Timeout
The design floor case stops stalling until its 610 s timeout
The design floor case sat until its ten-minute timeout at two menus: a Create menu whose title had scrolled away, and a focus menu worded differently. Both are now recognized by their structure, and answered.
Docs sync
The docs sync rule moves to where the agent reads it
Twice, the agent tacked a help flag onto its first line of reads, which the fixture forbids. The own-call rule now sits in the sentence that introduces those commands, not after them.
Gate
No threshold, budget, timeout or trial count changed
CHANGELOG.md1.91.185.0 · Gate integrityTwo checks now accept more: the design rating and the focus menu. No threshold, budget, timeout or trial count changed. Do you run /plan-design-review before you build a UI?
