Agent Skills 0.6.17 adds an accessibility-engineering skill for auditing and fixing WCAG 2.2 AA issues, fixes a rollback workflow that could never log in, and lets you regrade eval runs without paying for them again.
New skill
- accessibility-engineering (#261, @Chukwuebuka-2003, reviewed by @nucliweb). Audits an existing interface with pinned axe-core and pa11y, triages the findings, and walks through the manual keyboard and screen reader checks the tools can't do. It ships an
accessibility-auditorpersona, an/a11ycommand, and an audit script that fails a partial run instead of reporting zero violations. Building new UI stays with frontend-ui-engineering. The shared accessibility checklist now targets WCAG 2.2 AA and matches the skill.
Fixes
- Rollback can log in (#799, @divya0795). The ci-cd-and-automation rollback workflow ran
vercel rollbackon a runner with no Vercel login, so it failed exactly when needed. It now passes the token, like the preview step already did. - Review evidence from the right tree (#795, @TanmayShah29, reported by @Scroller251-cloud in #788). code-review-and-quality checks that the checkout is clean and on the reviewed ref before running tests or probes, since a working copy can hold another session's edits.
- Module specs under
spec/(#797, @SajjadPSavoji, #781). The capability map and per-module specs no longer pile up at the project root./buildalready readsspec/. - A Gemini subagent example that loads (#798, @divya0795). The example in the per-agent configuration guide now has the required
nameanddescriptionand the realgrep_searchtool.
Evals
--recordsaves a behavioral run's transcript, skill text and staged files, and--regradegrades it again against an updated rubric without another executor run (#667, @mavericksea-ai, #654).--modelpins the model for either.
Thanks
Thanks to everyone who contributed and reviewed this release: @Chukwuebuka-2003, @nucliweb, @divya0795, @TanmayShah29, @Scroller251-cloud, @SajjadPSavoji, @hinson0 and @mavericksea-ai.
Full Changelog: 0.6.16...0.6.17