/eval-plan
gstack’s new /eval-plan shows what a paid eval will cost before you approve it
gstack's new /eval-plan preregisters a paid eval for your product. A dry run on a free stub model checks prices and prompts. A pilot of about three runs projects the cost before you approve.
Prices
/eval-plan now prices its two default models out of the box
claude-opus-5-5$4 / $20 per MTokgpt-6.1-sol$2 / $10 per MTokIts default models, Claude Opus 5.5 and GPT 6.1 Sol, used to stop on a missing price. Both now have verified prices, so the plan runs. A model without a verified price still fails closed.
Lanes
gstack-lane-check names the open PR that touches your planned files
gstack-evidence verifyverified | unverifiedgstack-lane-check compares your planned files with in-flight branch and PR, and names each overlapping PR. Evidence verify reruns a bundle's probes and marks each claim verified or unverified.
Owner brief
Autoplan ends on a one-page brief for the owner
close --plan · each draft direction becomes your decision and its dateAutoplan now ends on a one-page brief. Urgent items outside the plan come first, then a summary, then numbered decisions, seven per page. Closing the plan writes your decisions in.
/review
/review now reviews outside contributors’ PRs as evidence
eval-plan/SKILL.mdthe /eval-plan guide/review now reviews an outside contributor's PR as evidence, and credits them when their diagnosis or test carries over. It flags banned terms in every diff, and the plan review's reality check names the file it reads. The /eval-plan guide has more.
