fix(storybook): freeze the clock per render and skip baseline-only pushes

MockDate.set was gated on inStorybookTestRunner() at preview module
scope, but the test runner appends its UA marker via an injected script
after the page loads, so the check read false and the clock was never
frozen - rendered wall-clock times drifted between runs (caught by
Visual Review on the autoresearch dashboard stories; the drift sat just
under jest-image-snapshot's 1% threshold, which is why local verify
passes missed it). Freeze the clock in the outermost decorator, where
the marker is visible, and pin the autoresearch fixture epoch since
module-scope fixtures evaluate before any decorator runs.

Also skip the visual-regression job when a push is the VR bot
committing an approved baseline: paths-filter diffs the whole PR, so
the existing filter exclusion can never catch it; compare the push
delta instead, as posthog/posthog does.

Generated-By: PostHog Code
Task-Id: c8fa7392-2afe-4919-b8be-e6e3b56e1419

Author: pauldambra

.github/workflows/code-storybook.yml

@@ -28,6 +28,7 @@ jobs:
       pull-requests: read
     outputs:
       code: ${{ steps.filter.outputs.code }}
+      baseline_only_push: ${{ steps.baseline_push.outputs.result }}
     steps:
       - name: Detect relevant changes
         id: filter
@@ -36,18 +37,56 @@ jobs:
           predicate-quantifier: every
           filters: |
             # Anything that can change a rendered story. The baseline manifest
-            # is excluded so the VR backend's approval commit (which touches
-            # only snapshots.yml) doesn't re-run the suite for nothing.
+            # is excluded, but note paths-filter diffs the whole PR against
+            # its base, so once a PR has code changes this filter stays true
+            # for every later push — the baseline_push step below is what
+            # skips the VR backend's own approval commits.
             code:
               - "{apps/code/**,packages/**,pnpm-lock.yaml,.github/workflows/code-storybook.yml}"
               - "!apps/code/snapshots.yml"
               - "!**/*.md"
 
+      # Visual Review commits the approved baseline back to the PR branch as
+      # a bot; re-running the suite for that push is pure waste (only the
+      # manifest changed). Compare the push delta (before...after), not the
+      # PR diff, so it only matches the bot's own baseline commit.
+      - name: Detect Visual Review baseline-only push
+        id: baseline_push
+        if: github.event_name == 'pull_request' && github.event.action == 'synchronize'
+        env:
+          GH_TOKEN: ${{ github.token }}
+          REPO: ${{ github.repository }}
+          ACTOR: ${{ github.actor }}
+          BEFORE: ${{ github.event.before }}
+          AFTER: ${{ github.event.after }}
+        run: |
+          set -euo pipefail
+          result=false
+          case "$ACTOR" in
+            *'[bot]' | posthog-bot) is_bot=true ;;
+            *) is_bot=false ;;
+          esac
+          zero=0000000000000000000000000000000000000000
+          if [ "$is_bot" = true ] && [ -n "$BEFORE" ] && [ "$BEFORE" != "$zero" ] && [ -n "$AFTER" ]; then
+            # compare uses merge-base(before, after), which for an appended
+            # commit is exactly the push delta; a force-push/rebase widens it
+            # and won't match.
+            files=$(gh api "repos/${REPO}/compare/${BEFORE}...${AFTER}" --jq '.files[].filename' || true)
+            if [ -n "$files" ] && ! grep -qv '^apps/code/snapshots\.yml$' <<<"$files"; then
+              result=true
+            fi
+          fi
+          echo "result=$result" >>"$GITHUB_OUTPUT"
+          echo "Visual Review baseline-only push: $result"
+
   visual-regression:
     needs: changes
     # Fail closed: if change detection failed (or this is a main push, where
-    # `changes` is skipped), run instead of skipping.
-    if: ${{ !cancelled() && (needs.changes.result != 'success' || needs.changes.outputs.code == 'true') }}
+    # `changes` is skipped), run instead of skipping — unless the push was
+    # just the VR bot committing an approved baseline.
+    if: >-
+      ${{ !cancelled() && (needs.changes.result != 'success' || needs.changes.outputs.code == 'true') &&
+      needs.changes.outputs.baseline_only_push != 'true' }}
     runs-on: ubuntu-latest
     timeout-minutes: 30
     permissions:

apps/code/.storybook/preview.tsx

@@ -10,14 +10,6 @@ function inStorybookTestRunner(): boolean {
   return navigator.userAgent.includes("StorybookTestRunner");
 }
 
-// Visual regression snapshots need identical pixels on every render: freeze
-// the clock (elapsed-time counters, relative timestamps) and re-seed
-// Math.random before each story (random status verbs), including on the
-// re-renders the test runner triggers for theme flips and retries.
-if (inStorybookTestRunner()) {
-  MockDate.set("2026-07-01T10:30:00Z");
-}
-
 function seededMathRandom(): void {
   let state = 0x9e3779b9;
   Math.random = () => {
@@ -63,8 +55,18 @@ const preview: Preview = {
       );
     },
     // Last in the array = outermost, so it runs before every story render.
+    // Visual regression snapshots need identical pixels on every render:
+    // freeze the clock (elapsed-time counters, rendered timestamps) and
+    // re-seed Math.random (random status verbs), including on the re-renders
+    // the test runner triggers for theme flips and retries. This must live in
+    // a decorator, not module scope: the test runner appends its UA marker
+    // via an injected script AFTER the page (and this module) loads, so a
+    // module-scope inStorybookTestRunner() check reads false. Story-module
+    // fixtures evaluated at import time still see the real clock — use fixed
+    // dates there.
     (Story) => {
       if (inStorybookTestRunner()) {
+        MockDate.set("2026-07-01T10:30:00Z");
         seededMathRandom();
       }
       return <Story />;

packages/ui/src/features/autoresearch/AutoresearchFullDashboard.stories.tsx

@@ -35,7 +35,10 @@ interface FullDashboardStoryProps {
   approach: string;
 }
 
-const STORY_NOW = Date.now();
+// Fixed date: module-scope fixtures evaluate at import time, before the
+// visual-regression clock freeze, so wall-clock times here would make the
+// rendered Time column drift between snapshot runs.
+const STORY_NOW = new Date("2026-07-01T10:30:00Z").getTime();
 const STARTED_AT = STORY_NOW - 26 * 60_000;
 
 function FullDashboardStory(props: FullDashboardStoryProps) {