Screen-reader findings

On 25 and 26 September 2026 every example in the workspace was run through the reader harness with Orca, on main at 261a218f, and what a screen-reader user got was compared with what they should get. This page says how that was done, what was fixed, what stays open and what was not measured. The pages under it hold every finding, one entry each.

Measured with Orca 46.1, KWin 6.6 (kwin_wayland --virtual), libatspi 2.52, and AccessKit 0.25 through accesskit_consumer 0.39, accesskit_atspi_common 0.20 and accesskit_unix 0.23.

How the sweep was done

The examples were split into 23 groups, one page each below. For each group one agent read the example, wrote scenarios stating act by act what a reader should get (tools/reader/scenarios/<group>.py), ran them with Orca, most of them two or three times, and recorded what the reader got. A second agent then took every finding, ran it again (tools/reader/scenarios/verify_<group>.py), read the Teksilo and adapter source it pointed to, and either confirmed it, corrected it (its severity, its layer or its cause), or added what the first had missed. No finding survived that did not reproduce.

Every entry says:

• Act: what was done, as keys pressed through the compositor, an AT-SPI action, or a focus request, which are three different things to a reader;
• The reader should get and the reader gets;
• Platform: where it was measured. Everything was measured on Linux. A claim about Windows or macOS comes from reading accesskit_windows 0.35 or accesskit_macos 0.27 and says so;
• Severity: critical blocks a task or loses data for a screen-reader user; high leaves them without what a sighted user gets on a common path; medium degrades it with a way round; low is polish;
• Layer: framework (Teksilo), example (the example's own code), upstream (AccessKit, libatspi or Orca), harness;
• Evidence: event lines from the AT-SPI listener, lines of Orca's debug log, and the source (file:line) that explains them. The run directories were not kept; the scenario named in the entry makes them again;
• Status: the commit that fixed it, or why it stays open.

K1 and K2 in some entries are the two fixes made before the sweep, which the sweep ran against: K1 names each window by its title (89a567cd), K2 is the announcer fix (7545b207).

The count

494 findings: 107 fixed, 13 partly fixed, 253 open in Teksilo, 63 open in the examples' own code, and 57 upstream. Every sweep finding was reproduced by the verifier; 85 of them were corrected on the way, and the verifiers added 80 the sweep had missed. The last page, the examples swept last, was swept on the build with the fixes and has no second agent's pass.

What was fixed

Each fix is one commit on main, with its own CHANGELOG entry. Each commit's message gives the test that was written first and seen fail on the old code, the ways the fix was broken on purpose to see it fail again, and the harness runs with Orca before and after. Each fix was then reviewed by an agent that had not written it, which reran the tests and the scenarios; where the review found a gap, its correction is part of the same commit.

What stays open

In Teksilo itself 265 findings stay open or partly fixed: 6 critical, 84 high, 102 medium, 73 low.

The critical and high ones in Teksilo itself, by what the reader lacks. Each id links to its entry.

• No route by keyboard or screen reader. Lightweight scene items cannot be focused, selected or moved (catalog-b-04, sceneetc-01), nor can the previewer's navigator rows be activated (sceneetc-02); a cross-view drag (misc-08), a tab moved to another group (tabs-09), a table's sorting, filters and column changes (tables-05), the tab strip's overflow list (tabs-07), the notification log's entries (toast-12, catalog-c-16) and the colour swatch grids (catalog-b-07, misc-12, misc-13) need a pointer; the magnet connect flow gives no feedback (sceneetc-07).
• A message that never reaches speech. A snackbar, banner or toast is read by its title and not its message (dialogs-05, catalog-a-13, toast-03); the ColorPicker's "Color changed" (misc-02, catalog-b-06), LogView::announce_appends (console-08) and the terminal's output (console-04, console-05, console-V1) are not announced; a wizard's step (catalog-a-M3, catalog-a-M4), the outcome of a shortcut rebind (chrome-14, chrome-15), the unread count (toast-10, catalog-c-15) and a date segment's step (datetime-05) are silent. Messages raised as focus arrives that are not the announcer's, a field's validation message or a message box's own live region, are still cut (password-03, password-05, toast-05).
• State a reader cannot see. A list's cursor without a selection (collections-01, catalog-c-26), a row's check box (collections-03), a table's name, cursor and selection (tables-01, tables-03), a chart's selected datum and its legend's actions (charts-01, charts-02), a spin box's unit (spinbox-06), Caps Lock where it was on at launch (password-04) and on macOS at all (password-v01), and an accordion's content, which sits inside its button where Orca's flat review does not look (dialogs-v-01, catalog-c-18).
• No name. The ColorPicker's channels (catalog-b-05, misc-03), ColorEdit (misc-16, catalog-b-M1), the SceneView (sceneetc-09, sceneetc-04), the dock resize handles (docking-05), ShortcutSettings' buttons (chrome-13), InputDialog's field (newkit-03), the previewer's knobs (sceneetc-14), a composite tooltip read as "Tooltip" (catalog-a-04); names left in English in a French interface (catalog-b-09).
• Focus lost or thrown after an action. Hiding or moving a dock (docking-02, docking-04, docking-06, docking-M2), pinning or removing a recent project (misc-10), a snackbar timing out (dialogs-06, dialogs-07), a banner dismissed (newkit-01), Enter on a corkboard card (sceneetc-06), Alt+letter on a collapsed menu bar (chrome-16).
• Text without its structure. Rich text runs its paragraphs together and drops lists, links and formatting (text-03, text-04, text-05, text-06, catalog-b-10); terminal rows are glued (console-03) and its way out is published only as a shortcut no adapter exports (console-06); a wrapped Arabic paragraph's text is garbled (winintl-02).
• Other. A tab move that did not happen is announced as done (tabs-02); a registry tooltip is not heard on the first visit (chrome-02); a live chart floods speech (charts-03); a close asked by the compositor shows no confirmation until later input (radioclose-01); SearchField's suggestions are silent (newkit-02).

Partly fixed, with the part left open in each entry: dialogs-04 (the popover's own focus), catalog-b-03 and chrome-17 (one Tab stop per menu bar item), collections-m3 (context menus stay unnamed), text-v03, datetime-06, datetime-07, toast-02, toast-07, sceneetc-05, and radioclose-M1: a modal that opens as a native window, the default on Windows and macOS, still lets a screen reader act on the window behind it.

In the examples' own code, the most common open finding is a control the example leaves unnamed; two are critical or near it: the menus example's context menus cannot be reached by keyboard (menus-11), and the native-menu example's in-window commands do nothing on Linux (chrome-18).

Upstream

The sweep traced 66 findings to code outside Teksilo. None has been reported upstream yet: publishing an issue is a decision for the project, so this is the list to report from. Where Teksilo can work round one, the entry says so; the node-id fix above is such a work-round.

What was not measured

• Windows and macOS. Nothing was run there. What NVDA, JAWS, Narrator or VoiceOver get is read from the adapters' source only, and where the adapters differ from AT-SPI the entries say so. Two differences that bear on the fixes: accesskit_windows raises an announcement as LiveRegionChanged with no text of its own (adapter.rs:256-263), so a screen reader reads the node's name; accesskit_macos carries the announcement's text (event.rs:65-99). The key report (teksilo_platform::key_report) does nothing on either: NVDA and VoiceOver read the keyboard themselves.
• Orca reading the keyboard itself. On an X11 session, or with a libatspi that uses KWin's org.freedesktop.a11y.KeyboardMonitor, Orca hears keys without the application; neither was run. On this Wayland session Orca hears exactly the keys the application reports.
• Other versions: Orca before or after 46.1, another libatspi, GNOME's compositor. Braille: off. Orca's own language: where a distribution installs Orca's translations under /usr/share/locale-langpack, Orca 46.1 stays English whatever LANG says, so every Orca line quoted here is English.
• Pointer, touch and pen. Acts were keys, AT-SPI actions and focus requests. A pointer was moved only where a hover mattered.
• Timing on another machine. A speech cut is estimated from Orca's log (see the guide); a race lost here may be won on a faster machine.
• Three examples, whole or in part. telemetry-codegen opens no window. telemetry-teksilo does not build: the workspace excludes it while its manifest inherits its version and dependencies from the workspace, its adapter crate pins teksilo-core and teksilo-telemetry at 0.6.2 against the tree's 0.13.1, and it expects teksilo-collector-proto outside the repository. web-view-demo was swept around its web view only: the private session has no DISPLAY, and the engine cannot embed under Wayland, so the page's own tree, and going into and out of it, were not observed. automation_bridge_smoke was swept with --serve; without it, it exits after its own smoke test.

What the sweep found in the harness

The sweep's agents recorded 222 notes on the harness itself. They came down to a handful of defects, all fixed in the harness before it landed: a key client per key that made the seat flap until winit panicked (now one client for a run); Orca's speech credited by the clock when its loop stalled for seconds (now credited to the event that caused it, with a bounded wait for an application that never goes quiet); speech from the settle before an act credited to the act; a stopped run reported as a pass; a tree read from libatspi's stale cache; an observation raised on every overlay close for a focus loss Orca drops harmlessly; and a module that did not import taking the whole scenario registry down with it.

The findings, by example

• Dialogs and popovers: 15 findings, 1 critical, 10 high, 2 medium, 2 low
• Date and time pickers: 20 findings, 3 critical, 5 high, 7 medium, 5 low
• Menus and drop-downs: 22 findings, 3 critical, 8 high, 7 medium, 4 low
• Toasts and the notification log: 19 findings, 1 critical, 10 high, 5 medium, 3 low
• Lists and trees: 18 findings, 9 high, 6 medium, 3 low
• Widget catalog, first eight tabs: 28 findings, 3 critical, 12 high, 8 medium, 5 low
• Widget catalog, middle eight tabs: 24 findings, 4 critical, 11 high, 8 medium, 1 low
• Widget catalog, last six tabs: 37 findings, 1 critical, 14 high, 17 medium, 5 low
• New widgets kit: 14 findings, 5 high, 4 medium, 5 low
• Spin boxes: 13 findings, 2 critical, 3 high, 3 medium, 5 low
• Password field: 19 findings, 8 high, 5 medium, 6 low
• Tabs: 18 findings, 1 critical, 5 high, 7 medium, 5 low
• Tables: 19 findings, 5 high, 9 medium, 5 low
• Grid view: 17 findings, 1 critical, 5 high, 5 medium, 6 low
• Radio tiles and close confirmation: 7 findings, 3 high, 1 medium, 3 low
• Windows and languages: 22 findings, 2 critical, 4 high, 6 medium, 10 low
• Charts: 16 findings, 6 high, 5 medium, 5 low
• Rich text and input methods: 22 findings, 2 critical, 12 high, 4 medium, 4 low
• Docking: 20 findings, 3 critical, 7 high, 5 medium, 5 low
• Terminal, log view and code editor: 21 findings, 2 critical, 8 high, 5 medium, 6 low
• Tooltips, menu bars and window chrome: 34 findings, 1 critical, 11 high, 13 medium, 9 low
• Pickers, files and drag and drop: 27 findings, 2 critical, 13 high, 8 medium, 4 low
• Scenes, animation and the rest: 32 findings, 2 critical, 13 high, 9 medium, 8 low
• The examples swept last: 10 findings, 5 medium, 5 low