v0.9.10
Highlights
- Made post-write validation evidence-aware across syntax, browser runtime, visual review, and source review. Checks now distinguish passed, failed, and skipped or inconclusive outcomes instead of treating unavailable validation as success.
- Added rendered screenshots for static HTML and retained Canvas keyframes at 0, 4, 9, 15, and 20 seconds.
- Models without image input are explicitly told that appearance, animation timing, and terminal state were not visually reviewed.
- Converted specification-critic and requirement-fidelity continuations into internal reminders, so they are no longer attributed to users or replayed as user messages.
Reliability Fixes
- Forwarded current validation limitations to the specification critic and cleared stale limitations after canonical-path revalidation succeeds.
- Treated missing TypeScript parsers, unavailable browsers, inconclusive visual verdicts, and inconclusive source-review verdicts as limitations rather than approvals.
- Isolated source-review failures so they no longer invalidate successful syntax or runtime evidence.
- Escaped critic findings and marked them as untrusted diagnostic content before reinjection.
- Ensured browser instances close through all runtime-validation exit paths.
- Tightened automatic skill matching so broad keyword overlap and generic complete-output requests no longer activate unrelated skills.
Verification
- GitHub CI passed on Linux and Windows, including build, lint, typecheck, full tests, and CLI smoke checks.
- Local suite: 4,890 tests passed, with 41 skipped and 2 todo.
- npm package audit confirmed 24 expected files and CLI version 0.9.10.
Full Changelog: v0.9.9...v0.9.10