Skip to content

v0.18.0

Choose a tag to compare

@a-lottes a-lottes released this 13 Nov 09:21
· 11 commits to main since this release
9dcb325

Added

  • Light path for small changes (light-path, #128): at the spec gate /story-time runs six sizing checks and, only if all hold, offers a light path. The user chooses. The spec stays; /spark writes a short plan for approval, then build, one review, no /demo-day, then /go-live. Any open decision, a sixth task, a file outside the plan or an open Major after review returns the feature to the full loop. Switch off with Light path: off in constitution §11.
  • Visual QA review (visual-qa-review, #124, issue #1): QA records one Visual Review row per screenshot of each UI state at wide desktop and phone width; a deviation from a binding design direction is at least Major; the ux lens checks start-up robustness; the Designer's craft check names scale and use of space.
  • Constitution §11 Models & Limits (#127): optional per-role model overrides and limits.

Changed

  • Default model is sonnet for every agent (#127). Up to 0.17.0 the Engineering Manager and the Reviewer were pinned to opus. This lowers cost and may change review depth; name reviewer: opus in §11 to keep the old behaviour.
  • Reviewer rounds are capped at two (#127), one on the light path. What is still open after the cap goes to the user.
  • The closing table is built from the trail (#127): /spark reads the Status rows on disk instead of the session's own account of what passed.
  • Companion marker in tools/README.md moves to 0.18.0, because aspark-guard 0.3.0 adds a rule (code-requires-approved-plan) and aspark-graph 0.8.0 adds impact --brief. Users who have not declared both companions installed are offered them once more.

Fixed

  • Nothing claimed as fixed.

Known limitations

  • #127 and #128 have no review or QA of their own; their evidence is dry runs, one session per case, with no rate observed.
  • The light path's token saving is not measured, and no full light loop through to a release has run yet.
  • The Factory does not read the spec's Path row yet; a Factory line runs every ticket through the full loop.
  • This is prompt material: a best-effort check, not a guarantee.