← Back to the screen

Design system portfolio

Design record: Faultline

The four direction lines behind this page, the generator command that produced its tokens, and the pairs it was measured against.

Incident command for one open Sev-2: INC-2291, checkout latency over its objective in eu-central. Four workflow columns, three swimlanes, ten action cards, and a clock that runs while you read

Direction

Archetype: board. The columns are the state machine: triage, investigating, mitigating, verifying. Three swimlanes cross them, one per workstream. A card moves

Data stance: tile-first for the header strip only. The strip carries four figures, and each one names the quantity it compares against. The board itself is the body of the page

Density: Atlassian default. 14px body from --ds-font-body, --ds-space-150 inside a card, --ds-space-200 around the board

The one bold thing: severity reads as length. Each card carries a bar whose width is time since detection over the incident's own elapsed clock. The oldest untouched work draws the widest bar, so the eye ranks the board before anyone reads a word

Colour names the band the age falls in (under 20 minutes, under an hour, over it), and it does not repeat the rank. Against the raised card the three chart steps measure 4.66:1, 4.51:1 and 8.91:1 in the light theme and 10.01:1, 9.78:1 and 9.30:1 in the dark one, so in dark the loudest of the three is the calmest band. Length is the encoding in both themes and colour is a second, categorical fact; the alternative is a single-hue ramp, which would paint a six-minute card red

The lane head carries the same bar one level up, and that is where the mark earns its place. A count says how much work a lane holds and cannot say whether the lane is moving, so the lane bar measures only its oldest card that has not reached mitigating: comms holds the second-oldest card on the board and reads as a quarter, because that card is already published. Three tracks of one width on one left edge, which is what the card bars cannot have across four columns

Why this product suits this system

The Atlassian design system rests on two standing assumptions: someone else may be changing this right now, and the admin has renamed everything. An incident makes both of them the operating condition rather than an edge case. The concurrency notice, the columns taken from an ordered workflow rather than from status names, and the single field-level move are that assumption on screen

Generator

atlassian_tokens.py --both          # tokens.css, byte for byte
atlassian_tokens.py --scales        # spacing, shape, typography, and motion
atlassian_tokens.py --contrast --theme light

A diff confirms tokens.css reproduces exactly from --both. That flag emits colour alone, so the page inlines the spacing, shape, type, and motion sets it needs, verbatim from --scales

Measured contrast

Card text sets --ds-text on --ds-surface-raised and measures 14.34:1 in the light theme and 9.84:1 in the dark one, against the 4.5:1 that body text needs

The age bar is a graphical object and needs 3:1. Its warning step, --ds-chart-warning-bold, measures 4.51:1 light and 9.78:1 dark, and it is the worst of the three. The success step reaches 4.66:1 and 10.01:1, the danger step 8.91:1 and 9.30:1. All three clear the floor in both themes; the ordering of those six numbers is the reason the bold thing is length and not colour

The focus ring, --ds-border-focused, measures 3.30:1 against the sunken board in the light theme, which is its worst ground, and 8.68:1 in the dark one

The unsuffixed --ds-chart-warning measures 2.47:1 on white and --ds-chart-danger 4.02:1, so the page uses neither. Against a light surface the -bold step is the one that carries a mark

Dragging, and the two routes into one move

One function, applyMove(key, toColumn, how), moves every card. The drop handler calls it, and so do the two arrows on each card, which target the neighbouring column.

No second code path exists, so a fix to one is a fix to both. WCAG 2.2 success criteria 2.5.7 and 2.1.1 ask for exactly that

At either end of the workflow the arrow disables itself and names the end it has reached. Focus follows the card, so three presses of Enter walk a card from triage to verifying with no mouse. Chromium confirms it

The second state

Another responder moved FAULT-118 while this board still shows it where it was. The notice takes the shape of a section message rather than a flag, because it wants an answer; the card carries the same mark in place; and "Show their version" resolves it through that same applyMove

Beside it, verifying holds nothing and says so once, in the column header rather than in each of its three cells. An empty cell draws nothing at rest; three dashed rectangles down the emptiest column are the board's loudest shape pointing at the one place with no work in it, and they follow the reader around as cells empty. The dashed target and its stage name appear while a card is in hand, which is the only moment a drop target is information

Deviations

A theme toggle ships. Inside Jira the host owns the theme and an app reads it, and the system says not to ship one. This page stands alone, and the contract asks for both themes with an in-page toggle that wins in both directions

No @atlaskit packages, since the page runs no build step. The page hand-builds the lozenge, avatar, avatar group, button, and section message against the token contract. The foreign-component-library and native-control rules in ads_lint read .tsx and .jsx, and have nothing to say about a <button> here

Atlassian Sans reaches no Google Fonts URL, so --ds-font-family-body falls back to its own stack. The page loads no web font

A stage switcher replaces the columns under 48rem. Four columns measure 1076px inside a 390px scroller: two and a half screens sideways, with the stage name scrolled off the top for most of the interaction. The switcher keeps all four counts on screen, the board shows one stage, the lanes and their bars stay, and a card pushed out of the stage you are standing in is what the status line is for. Nothing scrolls sideways at 390px, inside the board or outside it

contain: paint bounds the board scroller. Without it the column widths reached the document's scroll width and the whole page scrolled sideways at 375px, even while the scroller itself clipped correctly. A browser measurement found it. The board now scrolls inside its own box at every width, and the page never does

Uplift, 2026-10-29

The page was a board under a header strip. It is now the screen a Jira or Confluence user would recognise: a top navigation, a side navigation with the responders, a chart, the board, and an issue column. What changed, and why:

Shell. A grid of top nav, side nav, page and issue column, with elevation by surface: raised cards over sunken cells, overlay surfaces for the panel, the flag and the menu. Under 64rem a three-item bottom bar replaces the side navigation, because a labelled sidebar costs a phone 240px it does not have

One chart, drawn by hand. Checkout p99 over a deterministic series, with three marks (impact begins, detected, traffic shed), the 1.2 s objective and a keyboard scrubber (role=slider, aria-valuetext per sample). The figure in the heading, the final sample, the 4.0x ratio and the sentence under the chart read from the same series. Every time on the page is an offset from one anchor, so elapsed, detected, next update and the activity log cannot disagree

WIP limits. Each column states its limit and its count. A move past a limit is flagged and not refused, because in an incident the work has to land somewhere. The column head turns to "Over by n", and the flag says what the move cost. Over-limit is a status, so it is the only lozenge besides severity and state

Issue panel. A card opens an issue column: the four-stage track (a second keyboard route into applyMove), assignee menu (listbox, aria-activedescendant), inline edit of the title and description, the description's own log, a comment composer and an activity list. Docked at 90rem and up, a modal drawer from 48 to 90rem (inert behind it, focus trapped, Escape closes), a bottom sheet under 48rem. Between 90 and 100rem the side navigation folds to an icon rail while the panel is open, so the board keeps its width

Product thinking: undo. Every move raises a flag, announced in the status region, with Undo. Undo restores column, stage clock, log and the concurrency notice through the same applyMove, so drag, arrows, track, notice and undo are one path (WCAG 2.5.7, 2.1.1)

Decided against. A drop refused at a limit (hides the consequence); colour-only severity (see above); a toast stack (one flag, because two cannot both be read)

Kept from the gates. .card .keybtn carries a 24px target; the skip link is clipped, not translated, so the focus probes see it in place; .app no longer clips overflow, so the top nav sticks to the viewport. A boson focus-obscured-risk on the skip link is left: it is a positional reading (the link is fixed at z-index 40, above the bar at 8), and every colour finding from boson's one-file copy is an artefact of tokens.css not travelling with it

Round two, first screen, 2026-10-29

A critic read the first 900px at 1440 as a sparse sidebar, small text, a board cut at the fold and one pink fill. What changed:

The banner. --ds-background-danger-bold is spent once, as Atlassian's own banner under the top navigation: severity, state and the two clocks (elapsed, next update). The title beneath it is Heading XXL (Heading XL under 64rem). The lozenges inside it are chips of the page surface, so each keeps its own status colour on the bold ground

The chart has two chart hues, by phase. --ds-chart-danger-bold while it worsens up to the shed, --ds-chart-information-bold while it holds after it. The fills are gradients of the plain --ds-chart-danger and --ds-chart-information steps, and the notes and the two direct labels take the matching text tokens. Information is used after the shed, and not warning or success, because the curve is still 4.0x over its objective: it is held, not better, and the card-age bars keep the success, warning and danger steps for age

The side navigation is a rail from 64rem, 3.5rem wide, until "Show navigation labels" widens it to 15rem. The 240px of labels went to the board. Names stay in the accessible tree and in each avatar's title, and the Board count becomes a badge on its icon

The board is in the fold at 1440x900: the section message is one row (text left, actions right), the chart's header is one line, and the first lane's three cards sit above 900px. Two stat lozenges carry what the figures prove: 6m 41s is inside the 10m objective, and 12,480 is 1,340 above 30 minutes ago

Round four

Round nine, 2026-11-09

A below-the-fold, driven-state, phone and dark pass. What a browser found and what changed:

The warning glyph vanished in dark. The section message used the inverse warning icon token, which is built for a bold yellow ground and is near-black. On the subtle warning ground it read as a smudge on dark. It now takes --ds-icon-warning, orange in light and yellow in dark.

The phone stage switcher was top-aligned. The four tiles baseline-aligned their text to the top of a 40px box, so the selected tile looked empty below its label. Centred, and 44px tall.

The chart's held-phase label crossed the shed rule at the width the docked issue panel leaves the chart (1440px with a card open). The label now measures itself after it is drawn, moves to the right of the rule if it would cross it, and shortens to "Holding" if it still does not fit.

Units split across lines. "1.2 s" broke between the number and the unit in the chart caption on a phone, and a limit of "3" was left alone on a line in the section message. The seconds unit now uses a narrow no-break space, and the limit and card counts are bound to their words.

The rail's count badge touched the hairline. It sat flush against the rail's edge and over the icon. Smaller, nudged in, centred in its own box, clear of the icon's top.

The comment composer's avatar sat on its own row above the box. It now sits beside the box, as in Jira.

Left: boson reports focus-not-visible on most controls and focus-obscured-risk on the skip link. A real Tab through the page shows a 2px solid ring on every one of them and the skip link in place above the bar, so these are the tool's synthetic focus, not a finding. Taste, not changed: the design-notes column stops 120px short of the board's right edge at 1920.

Round ten, 2026-11-09

A second, harsher read, driven at 1440, 1920, 390 and 360 (touch, 3x), both schemes. Scored 33/40 before, 36/40 after.

Show their version and Keep my view had no way back (severity 3). Accepting the other responder's move put Mitigating over its limit with no flag and no Undo, and dismissing the notice removed it for good. Both now raise the flag with Undo (and Ctrl+Z). Undoing an accept returns the card, the stage clock, the log, the notice and, on a phone, the stage you were on. Undoing a dismiss brings the notice back and focuses its primary button.

A move can leave the stage you are looking at (severity 2). On a phone the card goes to another stage and the only trace is a figure ticking in the switcher. The flag now has "Show <stage>" when the card has left the stage on screen; it switches the stage and focuses the card.

Nothing explained the bars or the limits (severity 2). A disclosure beside the board heading, "How to read the bars and limits", opens one panel: the card bar (age over time since detection, the three bands drawn as swatches in the same tokens), the lane bar, and what a stage limit means. Escape closes it.

A comment could not be taken back (severity 2). Your own comments carry Delete; the flag offers Undo, which restores it in place.

The over-limit note showed on the laptop board (severity 2). The phone's one-line note had no rule hiding it above 48rem, so it printed a stray sentence above the column heads after any over-limit move. Hidden outside the phone layout.

The flag on a phone takes two rows (sentence and dismiss, then the actions) so a third button never squeezes the sentence into a column. Flag buttons and the help button are 44px on a coarse pointer.

"Incidents" in the breadcrumb linked to the top of the same page; it is plain text now, since there is no list to go to.

Inventions (none are product facts from the brief): the Delete action on your own comments; the wording of the three help entries; "Show <stage>" on the flag; that Keep my view is itself undoable.

Left: the time to detect reads 6m 41s while the two clock times, shown to the minute, differ by 7 minutes (severity 1, rounding of the display). The chart's scrubber has no visible hint that it can be dragged (severity 1). No copy-lint.ignore exists for this page.

Round eleven (2026-11-10)

An independent reviewer scored 32/40 and named four defects. All four reproduced before any edit (390 touch, 1440x900, 390 dark).

Touch targets (severity 2). On a coarse pointer the card move arrows were 28x28 and "Open FAULT-n" 60x24, with 15 targets under 44px on the board. The drawn sizes are unchanged. Each control now carries a transparent ::after that reaches 44x44 (the arrows, the key button, the panel and message close buttons), and the two arrows on a card sit 1rem apart so their hit areas meet without overlapping. The top bar links, the theme toggle and the notice buttons rise to 44px tall on a coarse pointer. Measured at 390 and 360: every card control reports a 44px hit area and elementFromPoint at 21px from its centre returns the control.

A moved card landed out of sight (severity 2). At 1440x900 the arrow put FAULT-121 at y 902 to 1076 with focus on its off-screen arrow, and a drag did the same. applyMove now scrolls the card into view (scrollIntoView, nearest, no smooth scrolling so reduced motion is honoured) before it places focus; cards carry a scroll-margin that clears the sticky top bar and the message at the foot (11rem on a phone, where the message is taller). After the move the card sits at y 606 to 780 and focus is on its visible arrow, by arrow and by drag, at 1440 and 1920. On a phone the card can leave the stage on screen; that case still goes through "Show <stage>" from round ten.

"Keep my view" left a stale board (severity 2). Dismissing removed the notice and the mark, so Investigating read 4/4 while the saved state was 3, and the way back was Ctrl+Z. Keep now leaves the notice up in a kept state: "Priya Raman saved FAULT-118 in Mitigating. Your board shows it in Investigating.", then "You kept your view, so this board differs from the saved one. Saved: Investigating 3 of 4, Mitigating 4 of 3. Mitigating is over its limit there." The counts are worked out from the cards, not written in. "Keep my view" goes (it is done); "Show their version" stays as the visible way back. The card mark reads "Priya Raman saved this in Mitigating" and the panel line says "You kept it in Investigating." The toast and Ctrl+Z still undo the keep, and undoing a move made over a kept notice restores it kept (prior.kept). Closing the toast no longer removes anything.

The toast covered a modal sheet (severity 2). The message (z-index 30) sat over the phone sheet (z-index 20) and hid its Activity comment. While a modal sheet is open syncMode sets html.sheet-open and the message is not drawn; it comes back when the sheet closes, with its Undo. Verified at 390 and 360: hidden with the sheet open, shown again after Close.

Inventions (no product facts): the wording of the kept notice, mark and panel line above; that the notice persists until the conflict is resolved; that the message waits behind a sheet rather than moving into it. Left: a move made inside the sheet shows its message only after the sheet closes (the live region still announces it); the summary save and cancel buttons in the panel keep their drawn hit area; the phone stage switch still looks like plain text apart from the selected tab (severity 1).

Round thirteen (2026-11-10)

An independent reviewer shipped the page and left one nit: on a phone the stage switch reads as plain text. Reproduced at 390 dark and 360 light before any edit: only the selected tile had a ground, the other three were borderless text on the page, and each carried a bare "4/4" with no limit marker.

The fix. The switch is now an ADS tab row drawn as bordered controls (--ds-background-input ground, --ds-border-input edge, hovered and pressed input tokens), so all four read as buttons before one is chosen. The selected tab takes --ds-background-selected, --ds-border-selected and a two-pixel selected rule along its foot, with --ds-text-selected. Each tab carries what the laptop column head carries: the stage name, the count, and under them the same limit marker (the slot bars and "Limit 3", "At limit of 3" or the "Over by 1" lozenge, from the same paintWip). An over-limit stage's name and count go to --ds-text-danger. Tiles are 58px tall (44px minimum) and the grid stays two by two.

Semantics. The group is a tablist; each stage is a role="tab" with aria-selected and a roving tabindex (only the chosen one is in the Tab order). Arrow keys, Home and End move to a stage and show it (automatic activation), and focus follows. The accessible name is built from the live counts, for example "Mitigating, 4 cards, over its limit of 3", so the limit state is announced with the selection. Laptop widths are unchanged; the whole switch stays hidden above 48rem.

Invention (no product facts): only the tab labels' wording "N cards" and "limit N", which restate the existing counts and limits; Triage has no limit and so shows no marker.

Round twenty-eight (2026-11-10)

An independent reviewer scored 36/40 and shipped with a note. Reproduced at 390 dark and 360 light before any edit.

Touch targets on the card arrows (note). The arrows draw 28x28 and the round-eleven ::after reached 8px beyond the drawn box on each side. Its inset is measured from the padding edge, inside the 1px border, so 9px of inset was only 8px beyond the box: exactly 44px, and a touch on the last pixel (elementFromPoint at 22px from the centre) fell through. Now the inset is 10px from the padding edge (9px beyond the box, 46px), and the two arrows on a card sit 18px apart so their hit areas touch and never overlap. The drawn size is unchanged. The sheet and message close buttons share the rule. Measured at 390 and 360 on all four stages: every point of each arrow's 44x44 box (1px steps, both edges included) returns that arrow, 0 failures in 12150 to 16200 samples per stage, and no neighbouring target takes any of them.

Focus off screen after a phone move (severity 1). Tapping an arrow in Investigating moved the card to a stage that is not drawn, and focus fell to the stage tab, 1066px above the viewport (scrollY 2043), because the search for the next control looked at the first arrow in the whole board, which sits in a hidden stage. Now applyMove looks for the next card's arrow in the shown stage (same lane and same direction first, then the same direction anywhere, then any arrow), focuses it and scrolls its card into view with the existing scroll-margin; when the stage is empty it focuses the stage tab, which now carries a scroll-margin under the top bar, so the tab and its "Nothing under investigation" line are in view. After a move the focused arrow sits at y 222 to 250 of 844 at 390 and 360. Laptop widths are unchanged (the moved card is still on screen).

Blank empty lane cell (severity 1). An empty cell is now a quiet line, "Nothing here" (.cellnote, body-small, --ds-text-subtle), shown while the cell is empty and hidden during a drag, where the stage name and dashed outline already label it. It sits inside the cell's labelled group, so a screen reader reads it with "Investigating in platform".

Inventions (no product facts): the wording "Nothing here"; that the next focus lands on the next card's arrow in the same direction. Left: on the phone an empty stage shows both the stage line ("Nothing under investigation") and a "Nothing here" in each lane. The sheet's own arrows are unchanged.