Reviewing a database migration the way GitHub reviews code: a real diff of real SQL, a lock analysis of every statement attached as evidence, and one required approval standing between the change and 41.2 million rows.
Direction
Archetype. Shell. The rail is the release week, the top bar is persistent, and one migration fills the well. Someone returns to this tab every day of a release week.
Data stance. Prose-first. The review conversation is the spine. The diff, the lock table, and the checks sit under the sentences that cite them.
Density. Compact. 14px body, 12px metadata, 32px controls, and a 44px hit area on a coarse pointer through --control-minTarget-coarse rather than a taller button.
The one bold thing. State is never carried by hue. Open, merged, and closed each get an icon, a word, and a silhouette of their own: a full pill, a 3px rectangle, an outline with no fill. The rail repeats all three six times, which is what teaches the vocabulary before the reader needs it.
Generated
primer_tokens.py --both > tokens.css
primer_tokens.py --scales > scales.css
primer_tokens.py --theme light-tritanopia
--both emits colour only, so the size, spacing, type, and motion scales come from a second run of the same generator rather than from a hand-written number. Nobody edits either file. The third command is the check behind the bold thing.
What the accessible themes settled
--bgColor-open-emphasis is green #1f883d in light, orange #bc4c00 in light-colorblind, and red #cf222e in light-tritanopia. --bgColor-closed-emphasis goes from red to neutral gray in both accessible themes. --bgColor-done-emphasis stays purple #8250df throughout.
Open against merged measures 1.12:1 in light, 1.00:1 in light-colorblind, and 1.06:1 in light-tritanopia. The two fills are one tone in greyscale in every theme Primer ships, so colour could not separate them under any remapping. The icon, the word, and the silhouette carry the state; the fill only confirms it.
Contrast measured
fgColor-muted on bgColor-muted at 12px reaches 5.74:1 light and 5.94:1 dark. fgColor-onEmphasis on bgColor-open-emphasis at 14px reaches 4.52:1 light and 4.63:1 dark. That second pair is Primer's own tightest required pairing, and it is why the pill label sits at 14px medium rather than 12px.
Target size, and why the line numbers are not links
A diff row measures 19.5px at Primer's --text-codeBlock line height. Eight line-number links stacked at that pitch met neither SC 2.5.8's 24px minimum nor its spacing exemption, and moving the anchor to the whole row would not have helped: the row is the same 19.5px tall and the same 19.5px apart. Growing the pitch to 24px was the other way to pass, and it changes what the page is.
So the line number identifies the line and nothing more. The row keeps its id, so #L14 still resolves and tr:target still marks the line; the links that reach a line live in the sentences that cite it, where SC 2.5.8's inline exception applies and where a prose-first page should have put them anyway. Verified with node tools/a11y-probe.mjs github-primer: no findings, light and dark.
Copy-lint exceptions
copy-lint.ignore holds three rules with the reason for each: operator-copy, because this page is the operator's tool and "migration" is their own word; shouty-caps, because every caps run on the page is SQL; and jargon, because Request changes is the name of the control in every review tool this audience uses.
The rail
The rail carried five destinations and four of them resolved to this document. A shell whose rail does nothing is a frame around a smaller product, so the destinations are gone and the rail is the release week instead: six migrations, one of them this page.
The current one is the only link and carries aria-current="page". The other five are rows, not controls, and that is the honest treatment for a build that is one screen, and a control that does nothing is worse than a row that never claimed to. #476 and #479 in the rail are the same two the timeline cites, so the rail and the conversation agree about the week.
Below 1012px the queue is a horizontal snap strip above the page head. It reads as cards rather than as tabs, which keeps it distinct from the section nav underneath it; the previous collapse put two scrolling tab strips one above the other.
The greyscale check
.rail rendered with filter: grayscale(1) and read back: open, merged and closed stay apart with no colour at all. Open and merged are one tone, which is Primer's palette and it is measured above, so the separation is the pill against the 3px rectangle, the icon, and the word. Closed is the outline and separates on lightness as well.
The prose measure
The stance is prose-first and the spine had no measure: at 1920px a body paragraph ran 1310px, about 187 characters. .shell is capped at Primer's own --breakpoint-xxlarge (1400px) and centred, which puts the paragraph at 790px and leaves 1920 reading as 1440 does. The spec panel's own measure is capped at 34rem for the same reason.
The uplift
The page is now one pull request, #482, read as GitHub reads one: a verdict strip, a conversation with a connector line, a lock chart attached as evidence, a merge box that counts what blocks it, and a diff with syntax colour and tabular line numbers. Opened Fri 11 Sep 2026, 14:22 UTC. The page clock reads 16:05 UTC, and each reader action moves it on by one minute.
The interaction. Apply mo-adeyemi's suggestion on line 12 and the page answers everywhere at once: the diff renumbers (+9 -3 becomes +10 -3, hunk -4,7 +4,13 becomes +4,14), a commit event joins the timeline, the replay reruns and the chart collapses from 4 m 12 s to a 0.02 s hairline, the verdict figure goes red to green, and the merge box drops from three blockers to two. Resolve the thread and request a review again and it drops to one. Revert undoes the commit. The override dialog exists because a real team needs a way past a failing budget, and it makes the reader type the PR number first.
The numbers, reconciled. The earlier page contradicted itself in several places; every figure now comes from one set. Replay: CREATE INDEX CONCURRENTLY 2 m 50 s (170 s), SET NOT NULL 4 m 12 s (252 s), DROP 0.04 s; the sum, 422.04 s, is 7 m 02 s, and the chart's x positions are those seconds over 422.04. The budget is 30 s, so the ratio is 252 / 30 = 8.4. After the suggestion the longest hold is 0.04 s. Conversation count 6, checks 4 (one failing, one warning, two passing), release week 37 with 2 of 5 applied and six rows in the rail. Times: opened 14:22, replay 14:31, review 15:04, reply 15:31, reply 15:52.
Line numbers in prose. The comments cite lines by new-file number: line 5 is lock_timeout, line 7 is the CREATE INDEX, line 12 is the constraint. The links in the prose land on the row through its id.
Dark. There is no dimmed variant in the generated tokens, so dark is the one dark theme and is checked on its own ground: the diff tints, the three label pairs and the chart bars are each measured again there. The nested-theme selector cannot re-scope inside a dark root, so the two swatches in the notes panel show the current theme and both ratios are written out beside them.
Keyboard. ? lists them. J and K walk the conversation, R re-runs the checks, A applies the suggestion, / focuses the rail filter. The key hints on buttons are hidden on a coarse pointer.
Deliberately left. direction_check reports accent-spread on the page, as it did before the uplift. The colours it counts are Primer's open, merged, closed and label states, which are the product's vocabulary rather than competing accents. The primary button is the only green control.
Round two: the first screen
The verdict's first cell is now the replay drawn as one run, full width, 88px tall (64px on a phone). Three segments on one time axis, each on a functional role: the concurrent index build in success (2 m 50 s, 0 to 40.28%), the first 30 s of the exclusive lock in attention (the budget, 7.11%), the rest in danger (3 m 42 s over, to 99.99%), with the 0.04 s DROP as a tick at the end. The 4 m 12 s figure stands on a stem that drops onto the bar where the lock begins, so it is read off the thing it measures. Blocking apply and Apply window are a quiet strip beneath. Apply the suggestion and the segments collapse to a ghost of the old hold, the figure ticks to 0.04 s and the stem goes green; the chart below the conversation still carries the same numbers and the text equivalent.
Round four (2026-10-29)
The lock bar had three segments built three ways: a pale green fill, a yellow budget window, and a saturated red slab that read as an error banner in light and left the green and yellow as dark blanks in dark. All three are now one construction, a muted fill in the segment's own colour under a 4px cap in its emphasis colour, so the bar reads the same in both themes and the cap, not the fill, is the drawn mark. Every segment names itself: CREATE INDEX 2 m 50 s, BUDGET 30 s (inside the bar where the track is 760px or wider, and in the coloured note under it at every width) and SET NOT NULL ACCESS EXCLUSIVE, which now starts at its own segment's edge. The figures are unchanged: the lock is the 30 s budget plus 3 m 42 s over, 4 m 12 s, after a 2 m 50 s index build, inside a 7 m 02 s replay.
At 390 the release-week list and the section tabs scroll sideways by design, but the cut-off words read as a layout fault. Each row now fades at the edge it continues past (data-more, set from scroll position and size), so "Widen tracking_code to 64" and "Lock analysis" read as scrollable. The reflow was considered and left: Primer's own UnderlineNav scrolls or collapses to a menu, and four icon tabs do not fit in 358px.
Round eight (2026-10-29)
- Reviewers column vanishes on scroll (1440 light). Real. The left rail was sticky and the right
.panewas not, so past the first screen the details scrolled away and left an empty third..paneis now sticky under the tabs (top: tabs-h + space-lg) from 1012px wide, but only when the window is 820px or taller, because the column measures about 740px and a sticky box taller than the window would be cut off instead of scrolled. - Tab row cut at "Lock an..." (390). The row already scrolled with an edge fade, but the fade fired before the fonts loaded and so was stale. At 543px and under the four tabs now drop their icons and tighten their padding, so all four fit at 390. Below about 380 they still scroll, with the fade, and the cue is re-measured when the fonts are ready.
- Diff code clipped (390). The code already had its own horizontal scroll (
.file__scroll) with no hint of it. All three scrollers now carry the samedata-moreedge fade the tabs use. - Release-week rail clips items (390). The 13rem cards put half of a second card's words under the fade. At 543px and under each card is the width of the row minus 4rem, with mandatory snap, so one card reads whole and the next shows as a card edge to swipe to. Declined: stacking all six, which would push the lock verdict, the page's one moment, off the first screen.
Round nine (2026-11-09)
Read below the fold, with the states driven, at 360 and 390 on a touch context, and in dark.
- The window was a Sunday. The apply window read "Sat 13 Sep" in four places, and 13 September 2026 is a Sunday: the hour count (9 h 55 m from 16:05 on Fri 11 Sep) was already a Saturday. It reads Sat 12 Sep everywhere.
- The replay total outlived the fix. After the suggestion applied, the header still said "7 m 02 s end to end" while the check under it said 2 m 50 s. The header now follows the replay and says what it was.
- A decimal point that does not read. In Mona Sans at 14px the point in 8.4 closes to nothing and the line reads 84 times. The verdict now says "More than 8 times the limit" and "Under 1% of the limit", both true of 252 / 30 and 0.04 / 30.
- Linked issue titles in sans. The two timeline events (#476, #479) set their whole title in monospace, which wrapped badly on a phone. The title is semibold sans with the identifier in code and the number muted, as GitHub sets it.
- Sideways code is keyboard reachable. The three code scrollers take focus, a role and a name only while they overflow, and draw their ring inside so the edge fade does not clip it.
- Phone. Fields are 16px on a coarse pointer so iOS does not zoom on focus; the top-bar links and the diff link grow a 44px hit area the way the buttons do; the four tabs fit at 360 without scrolling; the red segment of the lock bar names itself; the shortcuts button is not shown on a phone, which gave the bar a third row at 360.
- Wide. The top bar's content aligns to the 1400px column at 1920 instead of hugging the window edge.
- Filtered to empty. The rail's empty state offers "Clear filter", and focus returns to the field.
- Tab highlight. The section nav marks the section the reader is in from 96px down, not 140px, so Lock analysis no longer stays lit over the next comment.
- File path breaks at slashes and underscores, not mid-word.
Left, severity 0 to 1: the sidebar sits after the diff on a phone (the verdict is the page's first screen, so the pane does not move above it); decimals in prose elsewhere ("41.2 million") keep Mona Sans's small point.
Round ten (2026-11-09)
A second reader, cold, held the merge box against GitHub's own. Round nine scored it 36/40; the fresh read scored it lower, on the one act the page exists for.
- Applying did nothing (severity 3). The Apply button enabled when blockers cleared and had no handler, and the blocker count could never reach zero: the review row counted as a blocker for ever, so Apply could not be reached either. Both are fixed. After the suggestion, a resolved conversation and a re-request, mo-adeyemi approves the split version (the review row, the sidebar, the timeline and the count all answer) and Apply opens.
- A commit with one press and no way back (severity 3). Apply now opens a confirm dialog that says first when it runs (Sat 12 Sep, 02:00 to 04:00 UTC, not now), then the longest lock and its mode, then what is still open (the freeze warning, which lifts on Monday). If the lock budget was overridden, the dialog names 4 m 12 s of ACCESS EXCLUSIVE and the confirm button is the danger one. Confirming queues the migration: the box reads Queued, the verdict's two figures follow, a timeline event is added, and Remove from queue sits in the box. Focus goes to the box heading and the reason and the undo scroll into view.
- The merge box said nothing about the window before the press. The line beside Apply now states the window, the freeze it sits inside and the longest lock, in that order.
- Withdrawing the override reset the whole review (severity 2). Withdraw and Revert both called one reset that also deleted a comment the reader had typed. Withdraw now removes only the override and logs that it did; Revert keeps typed comments and never runs the clock backwards.
- A queued review can be undone from behind it (severity 2). While queued, Unresolve, Re-run, Apply suggestion, Revert and Withdraw are disabled, and the box says to remove it from the queue first.
- Conversation count. The tab read 6 until the script ran and then 5. It reads 5, the five comments.
- Copy-lint comments.
operator-copyno longer fires and is dropped from the ignore file.shouty-capsalso fires on the avatar initials PN MA LH, and the lock bar's BUDGET label (not SQL) is now Budget.jargonfires only on Re-request review; the comment says so.
Inventions, each minimal: (1) the reviewer approves about a minute of page clock after a re-request, only once the suggestion is applied, because his own comment says he will approve the split version; before that the row says what he is waiting on. (2) Apply queues for the window instead of running now, because the page already shows an apply window and a freeze check; the page never simulates the window opening, so no state is merged. (3) Remove from queue exists until then.
Left, severity 1: the reviewer approves minutes after the request, not "tonight" as his comment says; the timeline events wrap their sentence on a 360 phone as the existing ones do.
Round eleven (2026-11-10)
- Five rail rows looked like navigation and did nothing (severity 2). #484, #483, #479, #478 and #476 were inert spans in the same row shape as the live #482, inside a
nav. Reproduced at 1440 light: the five reportSPANwith no href, Tab skips them, a click moves nothing. The brief gives each of them a title and a state and nothing else, so the page cannot honestly open one without inventing its diff and review. They stay a plain list: Primer muted text and no hover, which they already had, plus a line under the filter that says so ("Only #482 opens here. The rest of the week is listed for context."), tied to the list witharia-describedby. The container is now anasidewith the same id (the skip link still lands on it), because anavlandmark holding one link and five labels is a promise of navigation. The filter still matches all six.
Inventions: none beyond that one sentence of rail copy. Left, severity 1: the dialog is still a centred card on a phone, "production-eu" can still break at its hyphen, and the Open pill does not change to a Queued mark after queueing; this round was given the rail only.
Round thirteen (2026-11-10)
The reviewer would ship it "nearly": the Diff tab that is only a scroll anchor, with the five inert rail rows, is where the page claims navigation it does not have.
- The Diff tab was a scroll anchor dressed as a tab (reproduced, 1440 light). The four section tabs are links; clicking Diff moved the page 3,181px down and kept it lit, and the same was true of Lock analysis and Checks. An underlined tab row says "this view replaces that one", and nothing was replaced.
- Five rail rows (reproduced, 1440 light). #484, #483, #479, #478 and #476 are SPANs, cursor auto, outside the tab order, in the same row shape as the live #482, interleaved with it. Round eleven captioned them; the caption sat above a list that still read as a list of pull requests.
Decision: an in-page link row, not real tab panels. This is one document in which every act answers in every section: applying the suggestion rewrites the diff, the replay, the verdict and the merge box. Panels would put Apply suggestion and the diff it changes on different tabs, and the brief sets the diff, the lock table and the checks under the sentences that cite them. GitHub itself makes the diff a separate page and a design lead would only split here if the page were split. Primer's UnderlineNav is a nav of links, so the row stays one, and it now says what it is.
- The section row. A caption above it reads "Jump to a section of this page"; the nav is named by it (
aria-labelledby), and the section the reader is in isaria-current="location"instead of the tab-like"true". Behaviour, scroll spy and the underline are unchanged, so nothing else moved. The caption is non-sticky, so it costs one 20px line once, at every width. - The rail. The one live row, #482, stands alone under the label. A rule and the caption "Also in week 37. Listed for context, they do not open here." separate it from the five others, which keep Primer muted text, cursor auto, no hover, no focus stop (checked: hover leaves the background transparent, Tab from the filter lands on #482 and then on Conversation). Not made links: the brief gives each of the five a title and a state and nothing else, and #476 and #479 appear in the timeline only as mentions, so a row opening them would scroll to a sentence, not to the pull request the row names. The filter still matches all six; the caption and the context list hide when none of the five match. On a phone the live row is one full-width row and the five stay the sideways strip.
Inventions: none beyond two strings of chrome copy ("Jump to a section of this page", the reworded rail caption). Order of the rail changes (#482 first, then #484 down to #476) because the live row cannot be both in the middle of a list of context and apart from it. Left, severity 1: the section row still looks like Primer UnderlineNav, because it is one.
- The Open pill after queueing (the round twelve sev 1). Queueing leaves the pull request open and waiting, so the title pill and the rail row now show a fourth mark: Queued, a clock icon, a dashed outline in the attention colours (a muted attention fill, attention text, 3px corners). It joins the three-mark vocabulary by silhouette as well as colour: pill, filled rectangle, solid outline, dashed outline. Remove from queue puts the Open pill back in both places; the rail filter also matches "queued" while it holds. Computed from the rendered pairs: 4.52:1 light (rgb 154 103 0 on 255 248 197, the tightest, just over 4.5), about 6.3:1 dark. Invention: the clock icon is the page's own (it already marks the merge box when queued). I could not check GitHub's merge-queue icon against a source here, so this is the attention-family reading of it, not a copy.
Round fifteen (2026-11-10)
The round fourteen reviewer would ship it "yes, with a note that the rail's inert rail rows and anchor tabs need captioning or removing". Round thirteen captioned both and they were still listed, so this round removes the look of each.
- Inert rail rows (reproduced, 1440 light). Before: five
SPANrows (#484 to #476) in the same row shape as the live #482. Chosen: remove them, keep #482. The rail is the label, the #482 row, and one plain muted line, "5 other migrations this week", with no row, no state mark, no rule. Nothing in it now looks like a destination that is not one. The count badge on the label (6) went with them. The rail filter went too: a filter over one row filters nothing, and the/shortcut and its two lines in the shortcuts dialog went with it (Esc now reads "Close a dialog"). The Closed and Merged marks still appear in the timeline, so the three-mark vocabulary is intact. - Section row read as UnderlineNav tabs (reproduced, 1440 light). Before: four links with an icon each, a 2px bar under the current one, inside a sticky row. Now a plain list of in-page links: a leading label "On this page" (visually hidden on a phone, still the nav's name), link-coloured text, no icons, no bar; the section the reader is in is dark and semibold,
aria-current="location"as before. The row still sticks, because the diff header and the details pane sit under it. Scroll spy unchanged. The 360 and 390 widths fit with no sideways scroll (sw equals cw). - Other severity 1 (the reviewer's one new note). The Change window check warns that the apply window sits inside the EU fleet freeze, while the merge box read "Ready to apply in the window" over "Checks passed". The check heading now reads "Checks passed with a warning" (counted, so "with 2 warnings" if a second ever shows) and the box heading reads "Ready to apply in the window, inside the EU fleet freeze". Both facts are the page's own. Queued heading unchanged.
Inventions: the string "5 other migrations this week" (the reviewer's wording) and "On this page". No product facts. Left, severity 1: the phone has no sticky merge bar and the diff sits below the comment box (reviewer scores 3 on both, no defect named).
Round seventeen (2026-11-10)
The round sixteen reviewer scored 37/40 and held three things: focus drops to the page body after three acts, the diff is cut at 360 with only a fade as a cue, and the phone is the desktop restacked (8,343px at 390, 8,694px at 360) with the diff below the comment box.
- Focus after Apply suggestion, Re-request review and Re-run all (reproduced at 1440, 390, 360, 1920:
document.activeElementwas BODY after each). The pressed button is disabled or hidden by the act, so the browser dropped focus. Each act now moves it to its own result: Apply suggestion to the "Applied in c41d7ae" line where the button was; Re-request review to the review row title (now "Review requested"); Re-run all to the checks row title ("Checks are running", then the result); a single check re-run to that check row. Thershortcut does the same as the button. Those five elements carrytabindex="-1", so they take focus and never join the Tab order. Nothing moves when a re-run is refused (queued or already running). - Code cut at 360 (reproduced: the diff scrolled 590px inside a 334px region, the right edge faded, the SQL index definition unreadable without a drag). On a phone (up to 767px) code lines now wrap, as GitHub's mobile diff does, with
pre-wrapandoverflow-wrap: anywhere; the two gutters give back width (2.25rem, 4px padding). The diff and both suggestion snippets no longer overflow at 360 or 390 (scrollWidth equals clientWidth), so nothing is cut and the sideways-scroll region, which the script only exposes while a region overflows, is not needed. Wider widths are unchanged and still scroll inside their labelled, focusable region. - Phone reading order. From 767px down, the merge box and then the diff move in the DOM ahead of the conversation (verdict, checks and merge box, diff, conversation, comment form, details). The nodes move with
matchMediaand move back on a wider window, so Tab and a screen reader follow what is on screen; the order at 1440 is as before. The "On this page" links follow the same order on a phone (Checks, Diff, Conversation, Lock analysis), and the section highlight is now the section nearest above the edge instead of a fixed three-step test, so it is right in either order. Conversation now links to#timeline(the first child of the old target, so the same place on desktop). Measured at 390: merge box at 836px (was 3,995), diff at 1,792px (was 5,214). Page height, first paint: 390 8,343 to 8,538px; 360 8,694 to 9,006px. It grew, because wrapped code takes more lines; the decision and the diff are reachable, the page is not shorter.
Inventions: none. Left, severity 1: the release rail and top bar still sit above the title on a phone (about 370px before the verdict), and the phone has no sticky merge bar.