Skip to content

v5.19.0

Choose a tag to compare

@Kanevry Kanevry released this 13 Nov 14:08
· 1 commit to main since this release

Added

  • A design review runs by itself when a wave changes UI files, and /close warns when the repo could render the change and no review covers it (#1452). On by default and advisory. The close gate is warn-only; nothing switches it automatically. Switch it off without a commit via SO_DESIGN_REVIEW=off or owner.yaml design-review.enabled: false, committed via design-review.enabled: false. Every setting resolves env > owner.yaml > committed > default.
    • Trigger: after an Impl-Core or Impl-Polish wave, scripts/design-review.mjs detect classifies the wave's files with classifyUiPaths() (scripts/lib/ui-touch.mjs).
      • Application UI source counts: components, pages, app styles, in-app message catalogues.
      • Never counted: tests, stories, fixtures, eval pages, tool output, docs, research, scaffolds, exported designs, test setup, mocks, generated code (__generated__/, gql/, *.gen.*, *.generated.*; a plain generated/ directory counts), video compositions, everything under pages/api/ and app/api/ (.tsx included), app/**/route.ts, and logic files in api/ and server/ directories.
      • Exceptions that count: test/ and tests/ route directories under app/, pages/ or routes/; e-mail directories inside app/, pages/, routes/, components/ or features/; components and pages in any other directory named api, server or mail.
      • The defaults are heuristics; design-review.include replaces the UI globs, design-review.exclude adds exclusions, and the default exclusions always win.
      • Measured with git ls-files over six repos: vault 288 → 0 matches, browser-kit 77 → 0, EventDrop.at 974 → 857 (its app).
    • Review: wave-loop-review.md step 5b captures the rendered result at 1440 and 390 with the test-runner driver. It reviews with agentic-designer's designer critique (on PATH or at design-review.designer, adapter scripts/lib/designer-critique.mjs), else with ux-evaluator.
      • Findings and verdicts are recorded and shown only (design-review.act-on-verdicts: false, both reviewers). Reason: designer critique had 23 of 37 findings wrong in its first production run (agentic-designer#85), and ux-evaluator rubric-v2 has no measured precision on real pages.
      • The page's value sentence reaches only --image page-top shots, never a --section-image.
      • Page-top shots get --role first-view, section shots --role section. A designer that rejects the flag (exit 2, INVALID_ARGUMENTS naming --role) is rerun once without it within the same per-shot time limit, and the output carries a role_fallback warning.
      • At most two rounds per surface and session.
    • Close: session-end Phase 2.7 runs design-review.mjs check.
      • With design-review.close-gate: warn (the default), it prints WARNING … would refuse and exits 0 when UI files changed in this session, a playwright.config.* whose directory holds them exists, and no record covers them in their current content. It records the case once; design-review.mjs would-refuse lists these records for classification.
      • With close-gate: refuse, the same case exits 3.
      • Not renderable, or not measurable: a WARNING and a record written by check itself.
      • No date and no automatic switch: the gate stays warn-only until close-gate: refuse is set in a commit, and that is due only when (1) foreign commits pulled mid-session are no longer counted (third review of !413, H2) and this is independently re-reviewed, (2) the cold reader passes against owner judgements (agentic-designer#85), (3) at least 20 would-refuse cases are classified with at most 10 % wrong, and (4) step 5b has run once in a real wave with the real designer.
    • Change set: this session's own commits (<start>..HEAD ^<default-branch baseline>, intersected with the net diff, and narrowed to the commits this worktree's HEAD reflog shows it created) plus the working tree. A rebase onto, or a merge of, an updated main, and a pull of a colleague's commits, never count other people's files; if the reflog cannot answer, the result is unmeasured, never a refusal. Commits with your git identity that are not attributed to this checkout (another worktree, clone or host, reset author date, squash, ambiguous author fingerprint) are named in a note and do not change the verdict. It never falls back to HEAD~1. STATE.md values older than this session's lock are ignored. Files last written before the session and deleted files do not count.
    • Records: they live in .orchestrator/metrics/design-review.jsonl, keyed by worktree and session, not by branch.
      • Each record covers files by path and blob hash, so a later UI change re-arms the gate.
      • They are written with a newline guard and read back. Unreadable lines and unknown setting values are named.
      • orchestrator.design_review.recorded carries no free text and no paths.
    • Record rules:
      • reviewed needs reviewer, verdict, and untracked, non-empty PNG screenshots written in this session. A recapture or BLOCKED review is reported as "not judged".
      • A hand-made not-reviewable in a repo that can render is reported like an override.
      • owner-override records actor and session kind. It is refused without an owner-queue --decision in lead-guided (HUD steering file or --lead-guided), headless, CI and non-Claude-Code sessions.
    • Reviewer: session-reviewer gets a short "UI and Copy" category.
    • Before: no UI check ran by default. The opt-in frontend-slop hook fired 0 times in 27,530 events, and across three product repos 58 of 262 UI and copy corrections were raised by the owner.

Changed

  • The full gate reports counted test failures behind a passing exit code (#2066, GH#93): when the test command exits 0 while its output counts failed tests, the gate prints one stderr line and sets test.failed_despite_pass: true in its envelope. The verdict and the exit code still follow the test command's exit code. The hint appears only when the output carries a Tests summary line (vitest, jest); a pytest summary such as 2 failed, 10 passed in 0.12s does not raise it. The count is read from free text and can be misread.

  • The three read-only wave reviewers return their findings instead of writing a file (#2069, GH#90): analyst, architect-reviewer and qa-strategist were told to write .orchestrator/audits/wave-reviewer-<wave>-<name>.md without a write tool. The coordinator writes that file now, and report_path is optional in the three output schemas.

  • Seven skills no longer carry a tools: frontmatter key (#2069, GH#90): Claude Code ignores the key, and the generated .agents/skills and .codex-plugin/skills mirrors turned it into allowed-tools, a pre-approval nobody had decided on. The mirrors of brainstorm, debug, eli5, grill, tmux-layout, ux-grill and write-executable-plan lose allowed-tools.

  • test.mode no longer claims to block (#1452): docs/session-config-reference.md said strict blocks session-end on HIGH/CRITICAL findings. The value is parsed (parse-config returns "mode":"strict", run against a fixture), but neither /test, test-runner nor session-end reads it, so the reference now says that.

  • The pre-push hook lets a trusted-publishing probe tag through the mirror smoke gate (#1906): a commit pushed alone to refs/tags/vX.Y.Z-<id>.N on github whose single parent is a proven curated snapshot and which changes only the version fields of package.json and package-lock.json to that version runs npm run test:mirror, like the snapshot itself (scripts/lib/mirror-probe.mjs). Before, the route B rehearsal could not push its tag without a hook bypass, because such a tree is never the export tree of a source.

  • The trusted-publishing workflow publishes a prerelease under its own dist-tag (#1906): .github/workflows/publish.yml runs npm publish --provenance --access public --tag <id> for a version X.Y.Z-<id>.N (e.g. 5.18.1-probe.0 → dist-tag probe), so latest never moves to it. Only X.Y.Z and X.Y.Z-<id>.N publish (<id> lowercase, two or more characters, not latest, not range-shaped like v1/vx); build metadata and npm loose forms (1.2.3beta.0) fail the run before npm normalises the version. This makes the first trusted publish, which npm needs to confirm a newly set-up Trusted Publisher, a rehearsal that cannot burn a release number. skills/npm-publish/SKILL.md § Route B describes the rehearsal. A plain X.Y.Z publishes as before.