You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit e5c4e47
Browse filesBrowse the repository at this point in the historyBrowse files
Brain
committed
Merge main into up-arrow edit, preserving edit and deletion admission
Copy file name to clipboardExpand all lines: docs/browser-testing.md
+30-8Lines changed: 30 additions & 8 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -103,8 +103,8 @@ verified clean commit or source manifest.
103
103
CI stays on `ubuntu-24.04`. `pnpm test:browser:ci` inherits the ordinary config and
104
104
excludes only tests tagged `@local-webkit` from `webkit-measurements`. Their Chromium
105
105
instances and every untagged WebKit case remain required. Both engines, serial
106
-
measurement order, zero retries, all existing assertions and budgets, and the
107
-
strict `CI required` aggregate remain in place.
106
+
measurement order, zero retries, the documented functional assertions and
107
+
measurement ceilings, and the strict `CI required` aggregate remain in place.
108
108
109
109
CI shards each functional engine across two runners, without waiting for the
110
110
separate measurement runner. Each job selects its engine with `--no-deps` and
@@ -124,7 +124,7 @@ The following **three WebKit cases are local-only**, not passing CI coverage:
124
124
125
125
| Case | Reason and coverage gap |
126
126
| --- | --- |
127
-
|`channel-opening.spec.mjs`: cold opening / warm switching |Hosted Linux WebKit recorded 104ms against the unchanged <100ms warm budget. The whole case is local-only, including its cold opening under held DM labels and no-new-head-read assertions. This is runner-sensitive evidence, not proof of an app or engine cause. Chromium retains the full case in CI. |
127
+
|`channel-opening.spec.mjs`: cold opening / warm switching |Originally excluded after hosted Linux WebKit recorded 104ms against the former <100ms budget. The warm gate now uses the target/ceiling policy below; this change does not revalidate Linux WebKit or restore its CI selection. The whole case remains local-only, including cold opening under held DM labels and no-new-head-read assertions. Chromium retains the full case in CI. |
128
128
|`scroll.spec.mjs`: cursor paging / large-history virtualization | Linux WebKit repeatedly stops short of the requested wheel edge. The cause remains unresolved between engine/input handling and the harness. Its 31 unique cursor requests, 640-message traversal, 4px anchors and DOM ceilings remain local-only on WebKit; Chromium retains them in CI. |
129
129
|`scroll.spec.mjs`: live edits / reading anchor | Linux WebKit's fetch reader can leave part of an edit undelivered while the SSE stream is open. WebKit growth/shrinkage and reading-anchor checks are local-only; Chromium retains the case in CI. The delivery defect is not fixed by this selection change. |
130
130
@@ -143,7 +143,8 @@ bin/pnpm test:browser # complete original suite, including those case
143
143
The full suite remains part of `pnpm test` and `just scan` on every local platform;
144
144
these cases are not silently skipped on Linux. The local-only command may still
145
145
fail there. No macOS CI runner is configured. To restore a case to CI, remove its
146
-
tag only after unchanged Linux assertions and budgets pass repeatedly. Live-edit
146
+
tag only after the documented Linux functional assertions and measurement
147
+
ceilings pass repeatedly. Live-edit
147
148
closure also needs complete delivery on the open stream without a later write,
148
149
heartbeat or close rescuing it. Do not move ordinary app/test failures out of CI
149
150
or grow this exception list merely to get a green run.
@@ -205,10 +206,31 @@ or attended live-account acceptance.
205
206
`channel-opening.spec.mjs` uses the actual app/session and production broker with
206
207
an offline upstream: 128 DMs and 1,001 uncached participants. Profile responses
207
208
stay held while an unprepared channel opens. This checks the **actual sidebar
208
-
label caller**, not just the reader's priority flag. Cached returns then require
209
-
no new head request and less than **100ms** from a browser-clock button click to
210
-
visible correct-channel rows across a paint opportunity. This is a controlled
211
-
regression budget, not a universal device/relay SLA or hardware input measurement.
209
+
label caller**, not just the reader's priority flag. Cached returns require no
210
+
new head request and visible correct-channel rows with the matching composer
211
+
across a paint opportunity. The existing **1s completion watchdog** still fails
212
+
an unfinished switch.
213
+
214
+
Warm timing keeps a **<100ms target** and a provisional **<200ms per-switch hard
215
+
ceiling**. All four browser-clock samples and the functional checks complete
216
+
before the ceiling is enforced; there are no retries or discarded outliers.
217
+
Target misses add `performance` annotations to the downloadable Playwright JSON
218
+
report (`ci-report.json` in CI), not the GitHub job-summary table. Raw timings,
219
+
first-visible times and frame diagnostics remain in `evidence.json`, including on
220
+
passing runs. Authors and reviewers should inspect target misses when changing
221
+
opening/rendering paths; green CI does not mean the 100ms target was met.
222
+
223
+
The ceiling is an explicit tolerance policy, not a statistically established
224
+
flake-free limit. Hosted Chromium recorded intermittent 111–120.5ms misses across
225
+
main and multiple PRs ([main](https://github.com/block/buzz-app/actions/runs/35925546559),
[#185](https://github.com/block/buzz-app/actions/runs/35936296156)). The 200ms ceiling
228
+
leaves roughly 80ms above the largest observed miss while retaining an automatic
229
+
slowdown alarm. The timer includes browser scheduling and layout observation as
230
+
well as application work: those samples do not establish runner contention as the
231
+
cause, and severe stalls can still fail. Regressions between 100ms and 200ms now
232
+
require performance review rather than automatically failing CI. This is not a
233
+
universal device/relay SLA or hardware input measurement.
212
234
213
235
Run this focused journey when changing startup/sidebar scheduling:
Copy file name to clipboardExpand all lines: docs/design-system-adoption.md
+1-1Lines changed: 1 addition & 1 deletion
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -15,7 +15,7 @@ pill actions, consistent fields and shared states.
15
15
| Search | SearchField | Channels, pages, members and GIFs preserve their query, ref and keyboard handlers. |
16
16
| Navigation | NavigationItem | Settings, channel rows, shell destinations, Home and community choices. Route destinations remain buttons, not tabs. |
17
17
| Tabs | Tabs | Emoji/GIF uses associated panels and Base UI keyboard activation. Workflow mode retains its existing externally owned editor view. |
18
-
| Modals | Dialog and AlertDialog | Page search, community chooser/setup and workflow confirmations. Pending work prevents dismissal; focus returns to the opener. |
18
+
| Modals | Dialog and AlertDialog | Page search, community setup and workflow confirmations. Pending work prevents dismissal; focus returns to the opener. |
19
19
| Panels and headers | Panel and PanelHeader | Settings, channels and companion cards use shared paint. Grids, scrolling, docks and subscriptions stay with the feature. |
20
20
| Feedback | Toast | Agent-start, live-update, sidebar preferences and Settings feedback use a source-owned stack. Recovery stays available; form errors, blocked pages and lasting paused-state context remain inline. |
21
21
| Hints | Tooltip | Control titles and agent activity use keyboard-accessible, dismissible hints. Accessible names stay on the controls. |
0 commit comments