This page is docs/ANNOUNCEMENTS.md from the repository. Wordings marked [A] are hypotheses that still need testing with people who use screen readers.

Announcements: how gantt-a11y should speak

Status: design proposal, 16 November 2026. Nothing in this document is implemented by being written here; src/strings.js is unchanged. "Today" and "the current text" refer to the working tree as it stood that evening, including src/core/focus-label.js, which appeared while this document was being written.

This document specifies what the extension says through its live region, in what order, how long it may take, and why. A companion page, listen.html, puts every text below into a live region so the variants can be compared by ear.

Read this first. Neither the author of the extension nor the author of this document uses a screen reader daily. The document therefore separates three kinds of statement, and every rule and every catalogue row carries one of these labels:

Where no reliable source was found, the text says "not found" instead of reconstructing behaviour from memory.

How durations were estimated. Each duration is computed, not measured: words at 170 words per minute, numbers and spelled codes expanded to the words a synthesizer speaks, plus 0.4 s per sentence end and 0.2 s per comma or colon, rounded to half a second. The 170 words per minute figure is an assumption for an unhurried default voice. Real figures depend on the synthesizer and the user's rate, and experienced users listen much faster: a large online study found that blind and low-vision listeners had higher listening rates than sighted ones [26]. Use the numbers to compare texts with each other, not as absolute times.


1. Prior art

What existing products announce for comparable structures, and in what order, according to the vendors' documentation and the other cited pages. Only what those pages state is reported; none of it was tested for this project.

1.1 Microsoft Project (desktop and web)

WhatReported behaviourRef
Moving between cells in the Gantt Chart view's tableThe screen reader speaks the cell content first and the column name after it.[16b]
Outline levelAfter indenting a task, JAWS says "Level 2"; after outdenting, the level one equivalent.[16b]
Expanded or collapsedNarrator says so only when focus is on a summary task.[16b]
After editing a task in Team PlannerThe resource name, the task name, then start and finish dates are announced.[16a]
Overlap in timeIn Team Planner, scheduling tasks so that they overlap produces the single word "Overallocated".[16a]
Known gaps stated by MicrosoftJAWS does not announce resource or task names properly in Team Planner; JAWS announces a new column's name only after it is added. Testing was done with Narrator and JAWS; other screen readers "might work".[16a][16b]
Dependencies, link lines, the bar chart paneNot found. The screen reader articles describe the table part of the view. They list no shortcut for moving between bars and do not say how a link or a predecessor is announced.[16b][16c]

Microsoft's general documentation explains each of the four link types with a full sentence and a picture, for example that with finish-to-start the dependent task cannot begin until the other is complete, and states that finish-to-start is the default [17]. That the vendor explains the terms sentence by sentence is weak evidence that the bare terms are not self-explanatory.

1.2 Excel and Google Sheets

WhatReported behaviourRef
Excel, moving between cellsThe screen reader announces the cell's column and row and its content. The page does not state the order.[18a]
Excel, a cell with a commentNarrator and JAWS add a short notice about the comment after the cell's content.[18b]
Google Sheets, moving between cellsContent first, then the address: a cell C4 containing a city name is read as the name followed by "C4".[19]
Google Sheets, extra stateAnnouncements when a cell holds a link, a note or invalid data, and when focus enters a filtered range.[19]
Google Sheets, on demandA "Verbalize to screen reader" menu speaks selected information when asked.[19]

Pattern: content before coordinates or headers, and extra state after the content.

1.3 Outlook and Google Calendar

WhatReported behaviourRef
Outlook calendar, an eventTitle first, then time, date, location, organizer.[21]
Outlook, conflicts or overlapping eventsNot found in the screen reader articles that were read.[21]
Google Calendar, an open eventNumber shortcuts move to one field at a time without moving focus: title, date and time, guests, rooms and location, description, attachments, notifications. Google introduced them so details can be heard on demand instead of by navigating through everything.[20a][20b]
Google Calendar, overlapping eventsNot found. The help page says some views do not show every event and that a "More" control reveals the rest.[20a]

Pattern: title first, and one field per request for the rest.

1.4 Trees and treegrids in the ARIA Authoring Practices

WhatGuidanceRef
Role structuretreegrid contains row elements that contain gridcell, columnheader or rowheader. Rows, cells, or both can be focusable.[1]
Expanded or collapsedaria-expanded on parent rows only, never on leaf rows.[1][2]
Level and positionThe example sets aria-level (counted from one), aria-posinset and aria-setsize on each row. The page warns that the last two were not allowed on rows in ARIA 1.1 and that support varies.[2]
When the author must set themIn a tree whose nodes are not all in the DOM (dynamic loading), all three must be set on every node.[3]
Virtualized rowsaria-rowindex with aria-rowcount when rows are missing from the DOM.[1]
What is readScreen readers usually switch to application mode in a grid and then speak only focusable elements and their labels.[1]
What a screen reader says for level, position and stateNot specified by the APG. It depends on the screen reader and on user settings; NVDA, for example, has a setting for reporting an object's position in its group, such as 1 of 4.[22]

Consequence for this extension: level, position and expanded state should reach the user through ARIA attributes, so that the user's own screen reader settings decide how and whether they are spoken. They should not be written into announcement text by default.

1.5 Other Gantt components and charts

1.6 Verbosity in screen readers themselves

Pattern: users already expect to control how much is said, and they do it in their screen reader. Anything the extension says through a live region bypasses those settings, which is the reason section 7 turns several decisions into settings.


2. Principles

2.1 Order and length

#RuleLabel
P1Put the distinguishing words first. In observation, screen reader users listened to the first few words of a line and moved on if it seemed irrelevant; the resulting guideline is keywords first and short sentences [25]. The APG gives the same rule for accessible names [4].[S]
P2Answer first, context second. For a reply to a command on the focused task, the first words are the answer ("Saved", "Conflict", "2 of 5", a count). The task name is context the user already has.[A]
P3Say the task name only when it is news: a different task than the focused one, a task that no longer exists (deleted), or a task just created. Do not repeat the focused task's name in replies about it.[A]
P4Never answer with silence. A result, an empty result and a refusal each get a short message. Messages about the result of an action are status messages and must reach assistive technology without a focus change [5].[S]
P5One announcement, one thought, about 8 seconds at most. Live-region speech cannot be replayed by the user, so it must be short and clear on first hearing [11]. Voice-interface guidance gives the same test: speakable in one or two breaths [29b]. Longer content is split into pages or moved into real, navigable content.[S] for brevity; the 8-second figure is [A]
P6Count before list, and cap the list. Start a list by saying how many items it has; read a few; offer the rest on request [29a]. The cap used here is 3 names. The source's figure, about five simple items, is for voice assistants; 3 is chosen because task names are multi-word.[S] for the pattern; the cap is [A]
P7Every value carries its meaning. A bare "25 August" is not understood (the author's VoiceOver finding). Either a label precedes the value ("Start 13 June") or the phrase labels itself: a range ("13 to 17 June"), a unit ("5 days"), a verb ("Ends 20 June").[A], from the author's testing
P8Names, not numbers. A dependency is spoken by task name, never by id or row number (the author's finding; the APG: convey purpose, not form [4]).[S]
P9Do not say what the screen reader says. No role words, no "level 3", no "collapsed" in text when ARIA can carry it [4][2].[S]
P10One word for one thing. "End" everywhere, or "finish" everywhere; "depends on" everywhere, or "predecessor" everywhere. See 2.6.[P]: plain-language practice; not tested here
P11Hints are separate from facts. "Press again for the next 3" is a hint. Hints are spoken after the fact, can be switched off, and are never needed to understand the fact.[A]

2.2 Punctuation, numbers, dates, durations

#RuleLabel
P12Periods between fields, commas inside a field. Screen readers turn commas and periods into pauses instead of speaking them [31a][24b]; the author found periods between fields necessary with VoiceOver. A sentence per field also gives braille readers visible boundaries.[S] for the mechanism; [A] that a period per field is the right amount
P13No brackets, dashes, slashes, arrows or other symbols to carry meaning. How symbols are read differs between screen readers and settings; a test of common symbols found none of the major screen readers reliable across the set, and found a dash read as "dash" where "minus" was meant [31a][31b]. Write "to", "and", "percent" as words. So "13 to 17 June", not "13–17 June"; "Documentation, start to start", not "Documentation (start to start)".[S]
P14Dates: day number and full month name. Never an ISO date, a numeric date or an abbreviated month in an announcement. How "2026-06-13" or "13 Jun" is pronounced depends on the speech engine and was not verified; listen.html has probes for exactly this. Generate with Intl.DateTimeFormat using day: "numeric", month: "long", so the order follows the language.[A]; prior art: a separate speech date format exists in Highcharts [34c]
P15Year only when it is news: when it differs from the current year or the project spans more than one year. Weekday in results of moving or resizing, where landing on a weekend matters; not in row labels.[A]
P16Speak the last day of work as the end date. The library stores end dates non-inclusively: a task whose last day is 17 June has an end date of midnight, 18 June, and the library's own documentation shows a template that subtracts to display the inclusive date [33b]. A spoken "ends 18 June" for a task that visibly ends on the 17th is an error a listener cannot catch.[S] for the storage model; the requirement is [A]
P17Durations in full words with correct plurals: "5 days", "1 day", never "5d". Unusual abbreviations get mispronounced [25], and WCAG asks for a way to learn an abbreviation's expansion [8].[S]
P18Ranges with "to"; cross-month ranges repeat the month: "13 to 17 June", "28 June to 3 July".[A]
P19Percent as a word with a noun: "40 percent done".[A]
P20No link-type abbreviations (FS, SS, FF, SF) in any announcement.[S] [8][25]

2.3 Polite or assertive, and collisions with focus speech

Rule P21: every live-region announcement in the catalogue is polite. [S] (Some results are delivered through the row's label instead; see P23.) Reasons:

  1. Assertive is for time-critical information that needs immediate attention and should be used sparingly [9][11]. Everything this extension announces is the result of a key the user just pressed or the answer to a question the user just asked. Nothing arrives uninvited.
  2. Interruption cannot be relied on anyway. In a 2026 comparison, JAWS with Chrome, Narrator with Edge, TalkBack and Orca treated assertive regions as polite; NVDA interrupted only for role="alert"; VoiceOver on macOS and iOS did interrupt [13]. A design that needs interruption to work is broken for most Windows users.
  3. An assertive update can discard queued polite ones [11]. With one polite region nothing the extension says can cancel something else it said.
  4. NVDA and JAWS prefix an alert with the word "Alert" [12]. Text that also starts with "Warning:" doubles the prefix. The catalogue uses the plain word "Conflict." and no alert role.

Rule P22: reserve an assertive region, empty, for future use. [S] One polite and one assertive region is the recommended maximum [11]. Candidates, none of which exist in the extension today: a save that fails on the server after the user has moved on; data replaced under the user by another person's edit. WCAG's AAA criterion on interruptions asks that the user be able to postpone or suppress them [6], so even these should have a setting.

Rule P23: when focus moves, the screen reader's reading of the new focus and a live-region message compete, and with a live region the order is not under the page's control. [S] Polite regions wait for a pause, and when that pause comes differs per screen reader [13]; a focus change can interrupt or confuse an announcement [11].

The extension has two ways to deliver a result when focus lands on a row:

The second route removes the race, and it is a good answer to P25 and P26 because the result stays on a braille display instead of flashing past. Its costs, all [A] and untested outside VoiceOver:

Design consequences for the texts:

ConsequenceLabel
The text must make sense before, after, or instead of the row label, because the route differs by mode and screen reader. That is why no result repeats the task name or a value that did not change.[A]
On the row-label route, the normal label follows the result whenever the result alone does not identify the task or its new state: a jump (N1 to N3), a created task (F6), a save that changed the name, a save with three or more changes (F2).[A]
The result may stand alone when it is complete in itself and the user has not moved: a move or resize (E1 to E3), a save of one or two fields (F1, F4).[A]
On the live-region route, send the announcement after focus has settled (the 500 ms used for deletions). The published recommendation is a gap of roughly 350 to 500 ms between updates [11].[S]
Compose one message and deliver it once. A move result and its conflict are one string, not two updates [11][10].[S]
Provide "repeat last announcement" for everything that goes through the live region; those messages cannot otherwise be replayed [11].[A]

Rule P24: coalesce bursts. [S] for the mechanism, [A] for the numbers. When updates arrive faster than they can be spoken, only the final state should be announced; with polite regions and rapid changes this is also what tends to happen in practice [9]. The extension already cancels a pending announcement when a new one arrives (live-region.js). D9 defines what the surviving announcement says.

Note on ariaNotify(). [S], status to be verified. A script API for announcements now exists; MDN lists it as newly available since September 2026, with a priority option that corresponds to polite and assertive, and advises combining related messages into one call [10]. Sources found during this research disagree about which browsers ship it, and a January 2026 test report planned to keep a live-region fallback [13]. Recommendation: feature-detect it later, keep the live region as the fallback, and do not change texts for it.

2.4 Braille

#RuleLabel
P25Assume the announcement is a short-lived flash message on a braille display, or absent. Screen readers show such messages for a few seconds, user-configurable [15][22]; NVDA showed no live regions in braille at all until 2023 [14]; in emulator tests VoiceOver's braille output lost content for polite regions [13].[S]
P26Nothing may exist only in an announcement. Every fact must also be reachable on demand in a form the user can read at their own pace: the row label, a command that can be repeated, or a dialog.[A]
P27The first 40 characters must carry the result. "Conflict. 2 days too early for Implementation." fits one 40-cell line; "Warning: Testing now starts before…" does not get to the point.[A]
P28Punctuation is text. Braille shows it. Do not add characters to force pauses in speech (doubled periods, stray commas).[A]
P29Test with a braille display, not only with the on-screen braille viewers, which do not reproduce a display user's experience [14][11].[S]

2.5 Translation

All [P]: these follow from how the Intl APIs and CLDR plural rules are built [35a][35b]; the Ukrainian examples are the author's own language.

2.6 Terminology

DecisionChoiceLabel
"Depends on / Required by" or "Predecessors / Successors""Depends on" and "Required by" by default; "Predecessor" and "Successor" as a wording setting. Microsoft's documentation for Project uses predecessor and successor throughout [17], so trained users will expect them. For everyone else the verb phrase needs no vocabulary. Today strings.js uses one pair in the dependency announcement and the other in navigation; whichever is chosen must be used in both (P10).[A]
Spoken confusability"Dependency" and "dependent" are close to indistinguishable in fast speech, so the plain wording never uses them as nouns: "Launch depends on this", "This depends on Implementation".[A]
Link types in plain languageNot named by default. Finish to start is not spoken, being the default [17]. The other three are spoken by their standard names in the dependency list, and as a rule in plain words wherever a conflict is explained: "must not start before Implementation starts". The plain-wording setting replaces the names everywhere with sentences such as "Documentation starts after this starts".[A]
Are "finish to start" and the others understood without explanation?Unknown. Not by people without project-management training, in this author's expectation. Section 8, task T2.[A]
LagNever the word "lag" by default: "at least 3 days after Implementation ends". Negative lag: "at most 2 days before Implementation ends".[A]
"End" or "finish""End" for dates and verbs, because the library's field and column are labelled "End date"; "finish" survives only inside the three link-type names. Whether "ends" is heard clearly at high speech rates (it is close to "and") is an open question.[A]
"Task" or the task's typeSay the type only when it is not an ordinary task: "Milestone", "Project". Use the words the application shows in its Type field, so that blind and sighted colleagues share vocabulary.[A]
"Conflict" or "Warning""Conflict": it names the thing; "warning" names only the tone and collides with the "Alert" prefix [12].[A]

3. Announcement catalogue

3.1 How to read the table

All rows are for the Standard verbosity level (section 5). Hints are not part of the texts; with hints on, a hint sentence follows.

#SituationTemplateExampleLiveTimeLabelRationale
Reading on demand
R1Task row label (accessible name of the row, not a live region){name}. {start} to {end}, {duration}. {progress} percent done.“Testing. 13 to 17 June, 5 days. 40 percent done.”n/a (focus)5 s[A]Name first; self-labelling range and unit. See Alternatives.
R2Dependencies: noneNo dependencies.(same)polite1 s[S]A negative answer is still an answer; never silence [5].
R3Dependencies: one in a directionDepends on: {name}. Required by: {name}.“Depends on: Software design. Required by: Testing.”polite3.5 s[A]Colon keeps the name in its dictionary form for translation. Default link type not spoken [17].
R4Dependencies: 2 or 3 in a direction, one with a non-default typeDepends on: {name}. Required by {count}. {name}. {name}, {type}. {name}, {type}.“Depends on: Software design. Required by 3. Testing. Documentation, start to start. Code review, finish to finish.”polite8.5 s[A]Count before list [29a]. One sentence per item when any item carries a type.
R5Dependencies: more than 3 in a direction (first page)Depends on {count}. First 3: {name}, {name}, {name}.“Depends on 8. First 3: Implementation, Budget approval, Code review.”polite5 s[A]Cap at 3 names [29a]; order: conflicting first, then latest to end. With hints on, "Press again for the next 3." follows.
R6Dependencies: next page (same key again within 5 seconds){from} to {to} of {count}: {name}, {name}, {name}.“4 to 6 of 8: Documentation, Legal review, Translation.”polite4 s[A]Position first so the user knows how much is left.
R7Dependencies: last page{from} and {to} of {count}: {name}, {name}. End of list.“7 and 8 of 8: Hosting contract, Security audit. End of list.”polite5.5 s[A]Explicit end so that silence is not mistaken for a failure.
R8Tasks at the same time: noneNo tasks at the same time.(same)polite2.5 s[S]Same as R2.
R9Tasks at the same time: 1 to 3{count} tasks at the same time. {name}, {extent}. {name}, {extent}.“2 tasks at the same time. Marketing prep & PR, the whole time. Documentation, first 2 days.”polite7.5 s[A]Extent of overlap instead of raw dates; largest overlap first.
R10Tasks at the same time: many{count} tasks at the same time. {n} the whole time: {name}, {name}, {name}. {m} partly.“12 tasks at the same time. 3 the whole time: Marketing prep & PR, Hosting contract, Translation. 9 partly.”polite8.5 s[A]Summary by kind of overlap first; names only for the strongest group.
R11Tasks that end before this starts / start after this ends{count} tasks end before this starts. Nearest: {name}, ends {date}.“4 tasks end before this starts. Nearest: Implementation, ends 12 June.”polite5 s[A]The nearest one is the one that matters for scheduling; the rest on a second press.
R12Where am I (proposed command)In {parent}, in {grandparent}, in {root}. Level {level}.“In Backend, in API, in Release 2, in Website redesign. Level 5.”polite5.5 s[A]Nearest ancestor first [25]. On demand only; the level itself belongs in aria-level [2].
Navigation
N1Jumped to a predecessor, several exist{n} of {total}. {origin} depends on this.“2 of 5. Launch depends on this.”row label, else polite3.5 s[A]See Alternatives.
N2Jumped to a successor, several exist{n} of {total}. This depends on {origin}.“2 of 3. This depends on Implementation.”row label, else polite3.5 s[A]Mirror of N1 with the same verb.
N3Jumped, only one exists{origin} depends on this.“Testing depends on this.”row label, else polite2 s[A]"1 of 1" adds nothing.
N4Nothing to jump toDepends on nothing. / Nothing depends on this.“Depends on nothing.”polite1.5 s[A]States the fact instead of an error; focus stays [7].
N5Target exists but is filtered out{n} of {total}, {name}, is filtered out. Focus did not move.“2 of 5, Budget approval, is filtered out. Focus did not move.”polite5.5 s[A]Names the task (the fact is still useful) and says where focus is.
Editing
E1Moved by one key pressStarts {weekday} {date}, ends {weekday} {date}.“Starts Sunday 14 June, ends Thursday 18 June.”row label, else polite3.5 s[A]See Alternatives.
E2Moved by several key presses in a row (spoken once, after the last press){n} days later. Starts {weekday} {date}, ends {weekday} {date}.“10 days later. Starts Tuesday 23 June, ends Saturday 27 June.”row label, else polite5.5 s[A]One summary instead of ten messages [9][11]. See D9.
E3Duration changed{duration}. Ends {weekday} {date}.“6 days. Ends Thursday 18 June.”row label, else polite3 s[A]New value first; the end date is the consequence.
E4Duration already at the minimumAlready {duration}, the minimum.“Already 1 day, the minimum.”polite2.5 s[A]Says why nothing happened [7].
E5Task is read-onlyRead-only. Not changed.(same)polite2 s[A]Reason first, then the outcome. No name: the user is on the task.
E6Task is a project (summary) taskNot changed. A project takes its dates from its subtasks.(same)polite4.5 s[A]Outcome, then the rule, so the user knows what to edit instead.
E7Duration key on a milestoneA milestone has no duration.(same)polite2 s[A]Kept from the current text, minus the name.
Conflicts
C1One conflict on the task just changed (appended to E1 to E3 in the same message)Conflict. {amount} too early for {other}.“Conflict. 2 days too early for Implementation.”row label, else polite3.5 s[A]See Alternatives.
C2One conflict on another task, caused by this changeConflict. {other} is now {amount} too early.“Conflict. Testing is now 3 days too early.”row label, else polite3.5 s[A]The task that is now wrong is named first; the cause is the task just changed.
C3Conflict on demand or at the Detailed level: with the ruleConflict. {amount} too early: must not {start or end} before {other} {starts or ends}.“Conflict. 2 days too early: must not start before Implementation ends.”polite5 s[A]The rule in plain words replaces the link-type name [17].
C4The same for a link with lagConflict. {amount} too early: must {start or end} at least {lag} after {other} {starts or ends}.“Conflict. 2 days too early: must start at least 3 days after Implementation ends.”polite6 s[A]The lag is spoken as the rule it creates, not as the word "lag".
C5Several conflicts{count} conflicts. Largest: {other}, {amount} too early.“3 conflicts. Largest: Testing, 4 days too early.”polite4 s[A]Count and the worst one; the full list on demand (proposed command).
C6A conflict that existed is gone after this changeConflict resolved.(same)row label, else polite1 s[A]Confirms the fix; otherwise the user has to ask.
Task form
F1Saved, one or two fields changedSaved. {field} {new}, was {old}. Ends {date}.“Saved. Duration 8 days, was 5. Ends 20 June.”row label, else polite4.5 s[A]See Alternatives.
F2Saved, three or more fields changedSaved, {count} changes: {field}, {field}, {field}, {field}, {field}.“Saved, 5 changes: name, start, duration, progress, type.”row label, else polite4.5 s[A]Field names only; the row label must follow, because it carries the values.
F3Saved, nothing changedSaved, no changes.(same)row label, else polite1.5 s[S]Confirms that Save was accepted [5].
F4Saved, a changed field is not a column in the tableSaved. {field} {new}, was {old}.“Saved. Priority high, was normal.”row label, else polite3 s[A]Always spoken with its value: the row label cannot confirm it.
F5CancelledCancelled. Nothing changed.(same)row label, else polite2 s[A]Answers the worry behind Escape: did my edits go anywhere.
F6CreatedCreated {name}.“Created Security audit.”row label, else polite1.5 s[A]The name is new information here.
F7New task discardedNew task discarded.(same)polite1.5 s[A]Kept.
F8DeletedDeleted {name}.“Deleted Retrospective.”polite1 s[A]The name is needed: the row no longer exists to be read.
F9Deleted a task with subtasks or linksDeleted {name} and {count} subtasks. {count} links removed.“Deleted Backend and 6 subtasks. 4 links removed.”polite3.5 s[A]Side effects a sighted user would see disappear.
F10Value changed while the form is open (plus/minus buttons){field} {value}. Ends {date}.“Duration 6 days. Ends 18 June.”polite3 s[A]Debounced; label before value because focus is on a button, not on the value.
Structure
S1Chart is emptyNo tasks.(same)n/a (description of the grid)1 s[A]Belongs in the grid's accessible description, read on entry.
S2Summary of the chart (proposed command){count} tasks, {shown} shown. {start} to {end}. {conflicts} conflicts.“5,000 tasks, 212 shown. 1 June 2026 to 30 March 2027. 3 conflicts.”polite8.5 s[A]The overview a sighted user gets from the scroll bar and the time scale [28].
Help
H1Keyboard help, first of three groups (quick form){group}, {n} of {total}. {action}, {key}. {action}, {key}. {action}, {key}. {action}, {key}.“Reading, 1 of 3. Dependencies, Shift D. Same time, Shift O. Ends before, Shift B. Starts after, Shift A.”polite9.5 s[A]At most four keys per group; action before key. The full list is a dialog. See D13.
Meta
M1Repeat the last announcement (proposed command)(the last announcement again)“Saved. Duration 8 days, was 5. Ends 20 June.”polite4.5 s[A]Live-region speech cannot be replayed by the screen reader [11].
M2Nothing to repeatNothing to repeat.(same)polite1.5 s[A]No silence.

Rows over 8 seconds, and why they stay:

3.2 Where the texts were deliberately shortened

CutSaved in the exampleRisk accepted
Task name removed from replies about the focused task (save, move, duration, errors)1 to 3 words per messageIf the row label is not read, the task is unidentified.
Year removed unless it differs3 spoken words per dateAmbiguity in multi-year projects; rule P15 restores it there.
"Start date:" / "End date:" replaced by a range in the row labelabout 2 s per rowUnlabelled range.
"finish to start" not spoken3 words per linkSilence must be learned as "default type".
The rule removed from the conflict that follows a changeabout 1.5 sThe user must ask to learn which rule is broken.
Old-and-new values replaced by field names when 3 or more fields changeabout 10 s for five fieldsValues come only from the row label.
Names after the third replaced by a countunboundedThe name the user wants may be on a later page.
"Warning:" and "now" removed from conflicts2 wordsNone identified.

3.3 What the catalogue needs that the extension does not do today

These are behaviour changes, listed so the texts are not read as drop-in replacements for strings.js:

  1. Paging: the same reading command pressed again within about 5 seconds continues the list (R6, R7, R10).
  2. A list dialog for long results (D2, D13).
  3. Commands that do not exist: repeat last announcement (M1), where am I (R12), conflicts of this task (C3 to C5), chart summary (S2), back to the task a jump started from (D10).
  4. Ordering of dependencies by conflict and by end date instead of link order (R5).
  5. Extent of overlap for concurrent tasks (R9).
  6. Size of a conflict in duration units (C1), and detection that a conflict disappeared (C6).
  7. A burst counter for moves (E2), and the choice per result between "result alone" and "result, then row label" (2.3).
  8. Plural categories, a speech date formatter, inclusive end dates (P14 to P17, 2.5).
  9. Disambiguation of duplicate task names (D4).
  10. Verbosity, hints and wording settings (sections 5 and 7).

4. Alternatives for the most important announcements

For each: the scenario, two to four variants, what each puts first, whom it suits, what it risks, and the choice. Every variant is on listen.html under the same letter.

4.1 Task row label

Scenario. Focus lands on the task Testing: an ordinary task, 13 to 17 June 2026, 5 days, 40% complete. This text is the row's accessible name (the library builds it from gantt.templates.tooltip_text), not a live-region message.

TextTimePuts firstSuitsRisk
A“Testing. Start date: 13 June 2026. End date: 17 June 2026.”7 sName, then every value behind its full field label (today's demo template).Newcomers: each value is labelled with the same words as the form.Longest. The year and the words "date" repeat on every row; no duration or progress.
B (chosen)“Testing. 13 to 17 June, 5 days. 40 percent done.”5 sName, then the time span as one phrase, then duration, then progress.Regular users who move through many rows."13 to 17 June" has no field label; relies on the range reading as start and end.
C“Testing. Start 13 June. End 17 June. 5 days. 40 percent done.”6 sName, then short labelled fields, one sentence per field.Newcomers and braille readers: one field per sentence, each with a label.Five sentence pauses per row make fast row-by-row scanning slower.

Choice: B. The name is the only part that differs on every row, so it goes first [25][4]. A range ("13 to 17 June") and a unit ("5 days") label themselves, which keeps the author's rule that no bare value is spoken, at about 70% of the length of A. C is the Detailed-level form. Assumption: that a spoken range is understood as start and end without labels.

4.2 Dependencies on demand (Shift+D)

Scenario. Focus is on Implementation. It depends on Software design (finish to start). Three tasks depend on it: Testing (finish to start), Documentation (start to start), Code review (finish to finish).

TextTimePuts firstSuitsRisk
A“Depends on: Software design (finish to start). Required by: Testing (finish to start), Documentation (start to start), Code review (finish to finish).”9.5 sDirection, then every name with its link type in brackets (today's text).Users trained in project management who want every link type every time.Longest. Brackets are not spoken reliably, so a type can attach to the wrong name; the default type is repeated.
B (chosen)“Depends on: Software design. Required by 3. Testing. Documentation, start to start. Code review, finish to finish.”8.5 sDirection and count, then names; a link type only when it is not the default finish to start.Regular users; anyone listening for one specific name.Silence about the type has to be learned as meaning finish to start. "Start to start" is still jargon.
C“Starts after Software design ends. 3 tasks wait for this. Testing starts after this ends. Documentation starts after this starts. Code review ends after this ends.”11 sThe scheduling rule itself in plain words, one sentence per link.Newcomers and people without project-management training.Longest to hear in full. "Starts after it starts" is exact but easy to mishear at speed.

Choice: B. Count first lets the listener decide whether to keep listening [29a]; names carry the meaning (the author's finding); finish to start is the default link type [17], so saying it each time adds words and no information. One sentence per item when any item has a qualifier keeps the type attached to its own name. C is offered as the plain-language wording setting.

4.3 Result of saving the task form

Scenario. The user changed the duration of Testing from 5 to 8 days and pressed Save. The form closes and focus returns to the row. Today the result is spoken in place of the row label; on the fallback route it follows the row label.

TextTimePuts firstSuitsRisk
A“Saved Testing. Duration: 5 days to 8 days. Ends 20 June 2026.”6.5 sConfirmation with the task name, then old value to new value (today's text).Newcomers who want the full sentence."5 days to 8 days" sounds like a range. The name is spoken twice whenever the row label is also read.
B (chosen)“Saved. Duration 8 days, was 5. Ends 20 June.”4.5 sConfirmation word, then the new value, then the old one.Everyone; the first word already answers "did it work".No task name: when the result replaces the row label, nothing identifies the task.
C“Saved, 1 change: duration.”2 sConfirmation and which field changed; values are left to the row label, which must follow.Experienced users; braille readers (fits a 40-cell line).The user must read the row to get the new value; a wrong value typed by mistake is not echoed.

Choice: B. Silence after an action is the failure the author observed, and a save result is a status message in the WCAG sense [5]. "Saved" comes first so that even an interrupted announcement confirms the action. New value before old, because the new one is the current truth; "was" removes the range ambiguity of "5 days to 8 days". The name is dropped because the user has not left the row; when the name itself changed, the row label follows the result.

Scenario. Testing depends on Implementation (finish to start). Implementation ends 12 June. The user moved Testing two days earlier, to 11 June.

TextTimePuts firstSuitsRisk
A“Warning: Testing now starts before Implementation ends.”3 sA warning word, then both task names and the broken rule (today's text).Newcomers: a full sentence with both names.Does not say by how much. "Warning" doubles with the "Alert" prefix some screen readers add to alerts [12].
B (chosen)“Conflict. 2 days too early for Implementation.”3.5 sCategory word, then the size of the problem, then the other task.Everyone; tells how many key presses fix it, and is short enough to follow a move result.Does not say which rule is broken (start or end); "too early" leaves the subject implicit.
C“Conflict. 2 days too early: must not start before Implementation ends.”5 sCategory word, size, then the rule in plain words with the other task.Newcomers; anyone who asks for the conflict on demand.Too long to append to a move result without passing 8 seconds.
D“Conflict with Implementation. Earliest start 13 June.”3.5 sCategory word, the other task, then the date that fixes it.Users who think in dates and go straight to the fix.Hides the rule and the size; "earliest start" becomes "earliest end" for two of the four link types.

Choice: B. One short category word first works like an earcon and fits the start of a braille line. "N days too early" is true for all four link types and for links with lag, so one pattern covers every case and the user never has to decode a link-type name. B is what follows a change; C is what the conflicts command and the Detailed level speak.

4.5 Tasks at the same time (Shift+O)

Scenario. Focus is on Testing, 13 to 17 June. Marketing prep & PR runs 10 to 18 June; Documentation runs 5 to 14 June.

TextTimePuts firstSuitsRisk
A“2 tasks at the same time: Documentation, Marketing prep & PR.”4.5 sCount, then names in start order (today's text).Quick check of who else is active.Says nothing about how much they overlap; a one-day and a full overlap sound the same.
B (chosen)“2 tasks at the same time. Marketing prep & PR, the whole time. Documentation, first 2 days.”7.5 sCount, then names with the extent of overlap, largest overlap first.Everyone; answers "how much" without date arithmetic."First 2 days" is relative to the focused task, which has to be understood once.
C“2 tasks at the same time. Documentation, 5 to 14 June. Marketing prep & PR, 10 to 18 June.”8.5 sCount, then names with their own dates.Users who want raw dates and will compare them themselves.The listener has to compute the overlap in their head while the next item is already being spoken.

Choice: B. A sighted user sees the extent of overlap at a glance, so the extent is the information to speak, not the raw dates. Largest overlap first puts the most relevant task where an impatient listener will still hear it [25].

4.6 Moving a task with the keyboard (Alt+Shift+Right)

Scenario. The user moves Testing one day later: from 13 to 17 June to 14 to 18 June.

TextTimePuts firstSuitsRisk
A“Testing moved to 14 June 2026, ends 18 June 2026.”5.5 sTask name, the verb, then start and end with the year (today's text).Newcomers; a complete sentence.Name and year repeat on every key press; four words pass before the first new fact.
B (chosen)“Starts Sunday 14 June, ends Thursday 18 June.”3.5 sThe new start with its weekday, then the new end.Everyone; the weekday exposes a move onto a weekend.Does not say how far the task moved.
C“1 day later. 14 to 18 June.”3.5 sThe size and direction of the move, then the new span.Experienced users; the shortest form.No weekday; "14 to 18 June" unlabelled.

Choice: B. The user pressed the key and is focused on the task, so name and verb are known; the new start is the news. For a burst of presses the chosen form gains C's first sentence as a summary (see D9).

4.7 Go to predecessor (Shift+P) when there are several

Scenario. Launch depends on five tasks. The user presses Shift+P a second time; focus moves to the second of them and the screen reader reads that row.

TextTimePuts firstSuitsRisk
A“Predecessor 2 of 5, finish to start.”3 sThe relation as a noun, position, then link type (today's text).Users who know the term predecessor.Does not say whose predecessor; after a few jumps the starting task is forgotten.
B (chosen)“2 of 5. Launch depends on this.”3.5 sPosition, then the relation as a sentence that names the task the user started from.Everyone; no jargon, and the origin is restated.The link type is left out unless it is unusual.
C“Predecessor 2 of 5 of Launch.”2.5 sThe relation as a noun, position, then the origin.Users who chose project-management wording."Of 5 of Launch" is awkward to hear.

Choice: B. The new row's own label is read after this text, so the text only adds what the row cannot say: position in the set and what the set is. Naming the origin matters because the next press continues around the origin's list.


5. Verbosity levels

Are three levels needed? The pattern is familiar to users: JAWS and VoiceOver both have three [23][24c]. The argument against is cost: every level multiplies the texts to translate and test. The proposal is three levels defined by rules applied to the same templates, not three sets of strings, plus a separate switch for hints. Whether users want this inside one component, in addition to their screen reader's own verbosity, is open (section 8, T10). All of this section is [A].

BriefStandard (default)Detailed
ForDaily users who know the chartEveryone until they chooseNew users; careful review
Row label“Testing. 13 June, 5 days.” (3 s)“Testing. 13 to 17 June, 5 days. 40 percent done.” (5 s)“Testing. Start 13 June. End 17 June. 5 days. 40 percent done.” (6 s)
Names in a list before "and more"335
Link typesOnly non-defaultOnly non-defaultAlways, as plain sentences
Save“Saved, 1 change: duration.” (2 s)“Saved. Duration 8 days, was 5. Ends 20 June.” (4.5 s)“Saved. Duration 8 days, was 5 days. End 20 June, was 17 June.” (6 s) Every change with old and new value, at most two changes per message
Move“14 to 18 June.” (2 s)“Starts Sunday 14 June, ends Thursday 18 June.” (3.5 s)“1 day later. Starts Sunday 14 June, ends Thursday 18 June.” (5 s)
Conflict after a change“Conflict, Implementation.” (1.5 s)“Conflict. 2 days too early for Implementation.” (3.5 s)“Conflict. 2 days too early: must not start before Implementation ends. Earliest start 13 June.” (6.5 s)
Weekday in edit resultsNoYesYes
YearOnly when differentOnly when differentAlways
Parent name in row labelNoNoYes
Errors“Read-only.” (1 s)“Read-only. Not changed.” (2 s)“Read-only. Not changed. Ask the project owner to unlock it.” (5 s) (application-supplied advice)

Not affected by level: counts, the first word of results ("Saved", "Conflict", "Created", "Deleted"), and refusals. A Brief user must still learn that something failed.

Hints (how to get more, which key to press) are a separate on/off switch, on by default, and each distinct hint is spoken at most three times per session and then drops out by itself. The decay is an assumption modelled on the idea that hints are for learning.


6. Edge cases

Each case gives the text, its estimated time, and the reasoning. The five cases the brief marks as most important (D2, D6, D7, D8, D9) are worked through in most detail. Everything here is [A] unless a reference is given.

D1. Deep nesting

A task on level 5 with four ancestors.

D2. Many dependencies

The limit. Up to 3 names in one direction are spoken in full. From 4 upward the announcement gives the count and the first 3. Reasoning: a voice-design source recommends opening a list with its count, keeping simple one-word items to about five, reading long items one at a time, and paging the rest with "more" [29a]. Task names are usually two to five words, so three names already take 4 to 6 seconds. Screen reader users are faster and more practised listeners than voice-assistant customers [26], so the limit may be too low; it is a setting (section 7).

CountTextTime
0“No dependencies.”1 s
1“Depends on: Software design.”2 s
3“Depends on 3: Software design, Budget approval, Code review.”4 s
8, first press“Depends on 8. First 3: Implementation, Budget approval, Code review.”5 s
8, second press“4 to 6 of 8: Documentation, Legal review, Translation.”4 s
8, third press“7 and 8 of 8: Hosting contract, Security audit. End of list.”5.5 s
25“Depends on 25. Last to end: Security audit, 12 June. 1 in conflict: Legal review.”7.5 s

Which three first. Not link order (creation order means nothing to the listener). First any dependency in conflict, then the ones that end last, because the predecessor that ends last is the one that decides when this task can start.

When both directions are long, the first press covers "Depends on" only and ends with the count of the other direction: "Required by 6." The second direction has its own pages.

Access to the rest, in order of cost:

  1. Press the same key again within about 5 seconds for the next three. "Press twice" is an established convention in screen readers; NVDA uses a second press of the same command to spell what it just read [36]. After the last page the next press starts over.
  2. Walk them one at a time with go-to-predecessor (Shift+P), which already cycles around the list and says "2 of 8".
  3. A list dialog from 10 upward: a real dialog with a heading and a list, one task per item ("Legal review. Ends 10 June. In conflict."), which the user reads at their own speed with ordinary screen reader commands and on a braille display, and from which Enter jumps to the task. Long or structured content does not belong in a live region: it is flattened to plain text and cannot be re-read [11]. This is the only form that works for 25.

For 25 the first press does not list names at all. It gives the count and the two facts a sighted user would look for in a fan of 25 link lines: which one drives the start, and whether any is broken. For comparison, listen.html includes all 25 names in one message (23.5 s).

D3. Many columns

D4. Long names, HTML, emoji, codes

D5. Special kinds of task

CaseTextTimeNote
No dates“Translation. Not scheduled.”2 sTime queries skip it and say so: "1 task has no dates."
Zero duration, ordinary task“Sign-off. 20 June, no duration.”3 sOne date, not a range from a day to the same day.
Milestone“Go-live. Milestone, 20 June.”3 sName first, type second; the type explains the missing duration.
Project (summary)“Website redesign. Project, 1 to 22 June, 10 tasks. 35 percent done.”6.5 sThe count of contained tasks replaces the duration.
Collapsed parentThe same text; "collapsed" comes from aria-expanded [1].The count tells what is hidden.
Collapsed parent hiding a conflict“Website redesign. Project, 1 to 22 June, 10 tasks. 35 percent done. 2 conflicts inside.”8 sOtherwise a conflict is unreachable without opening every branch.

In lists a milestone is introduced by its type: "milestone Go-live, on 15 June".

D6. Tasks at the same time

What to exclude. The task itself, its ancestors and its descendants (the extension does this). Proposal: also project (summary) tasks from other branches, because a summary bar overlaps everything under it and adds a name without adding a fact.

How to say how much they overlap. In words relative to the focused task, in the chart's duration unit:

OverlapPhrase
The other task covers all of this one"the whole time"
It covers the beginning of this one"first 2 days"
It covers the end of this one"last 3 days", "last day"
It lies inside this one"2 days in the middle"
A milestone inside"milestone Go-live, on 15 June"

Days are counted the way durations are counted in the chart (working days if the chart uses a work calendar). No percentages: "40 percent of the time" requires arithmetic to become days.

CountTextTime
0“No tasks at the same time.”2.5 s
1“1 task at the same time: Documentation, first 2 days.”4.5 s
2“2 tasks at the same time. Marketing prep & PR, the whole time. Documentation, first 2 days.”7.5 s
12, first press“12 tasks at the same time. 3 the whole time: Marketing prep & PR, Hosting contract, Translation. 9 partly.”8.5 s
12, second press“Partly, 1 to 3 of 9: Documentation, first 2 days. Code review, 2 days in the middle. Launch, last day.”9.5 s

Order: whole-time overlaps first, then by length of overlap, longest first. Today's order is by start date.

For 12, the first message groups by kind of overlap. A sighted user looking at twelve parallel bars sees at once which run alongside for the full length and which only touch; that grouping is the summary. From about 10 upward the list dialog from D2 is the better tool. The 12-task second press is at 9.5 s; it stays because each of its three sentences is complete on its own.

Prior art for the one-word version: according to Microsoft's support article, Project says only "Overallocated" when tasks overlap in Team Planner [16a]. That is a different question (one person, too much work), but it shows a shipped product answering overlap with a state word and leaving the detail to navigation.

One pattern for all four link types. In every type the conflict is the same fact: a date of the dependent task comes before the date the link allows. So the task is "too early", by an amount:

Link typeShort form after a changeFull form on demand
Finish to start“Conflict. 2 days too early for Implementation.”“Conflict. 2 days too early: must not start before Implementation ends.”
Start to startsame pattern“Conflict. 1 day too early: must not start before Implementation starts.”
Finish to finishsame pattern“Conflict. 3 days too early: must not end before Implementation ends.”
Start to finishsame pattern“Conflict. 1 day too early: must not end before Launch starts.”
With lagsame pattern“Conflict. 2 days too early: must start at least 3 days after Implementation ends.”
With leadsame pattern“Conflict. 1 day too early: may start at most 2 days before Implementation ends.”

The short form takes 3.5 s; the full forms 5 s to 6 s.

Why the short form after a change. The conflict is appended to the result of the change in the same message, because two updates in quick succession can collide or be dropped [11]: “Starts Thursday 11 June, ends Monday 15 June. Conflict. 2 days too early for Implementation.” (6.5 s). With the full form that message passes 8 seconds, and a burst summary with a conflict comes close to 10. The short form still gives what the user needs to act: something is wrong, with which task, and by how many days, which for a move is also the number of key presses that fixes it.

Which task is the subject. The task the user just changed is implicit. When the change breaks a link on another task, that task is named first: “Conflict. Testing is now 3 days too early.” (3.5 s).

Several conflicts: “3 conflicts. Largest: Testing, 4 days too early.” (4 s). Count first, then the largest. The list is on demand through a conflicts command, one sentence per conflict in the full form, three per page.

When a conflict goes away: “Starts Saturday 13 June, ends Wednesday 17 June. Conflict resolved.” (4.5 s). Without it, the user who moved a task to fix a conflict has to ask whether it worked.

Are the link-type names understood? The design avoids needing an answer: no conflict text contains a link-type name. For comparison, listen.html includes a jargon version (“Conflict: finish-to-start link from Implementation violated by 2 days.”). Microsoft's own help explains each type with a sentence [17]. Whether "start to start" in the dependency list is understood is task T2.

Politeness: polite (2.3). The conflict is the consequence of the key the user just pressed and travels with that key's result.

Not a prompt. The announcement never asks a question ("Move back?") because a live region cannot take an answer, and voice guidance warns that a listener has only seconds to respond to spoken options [29a].

D8. Saving the form

After Save the form closes and focus returns to the row. Depending on the delivery route (2.3) the result is heard instead of the row label, before it, or after it. The texts are written to work in all three positions (rule P23).

CaseTextTimeReasoning
One field“Saved. Duration 8 days, was 5. Ends 20 June.”4.5 sNew value, then old. The end date is spoken because it changed as a consequence.
Two fields“Saved. Start 15 June, was 13 June. Progress 60 percent, was 40.”6 sStill one message under 8 s.
Five fields“Saved, 5 changes: name, start, duration, progress, type.”4.5 sFive old-and-new pairs would take about 15 s. The row label must follow this message, because it is what carries the new values.
Five fields, Detailed“Saved, 5 changes. Name Acceptance testing, was Testing. Start 15 June, was 13 June.” then “Duration 8 days, was 5. Progress 60 percent, was 40. Type project, was task.”6.5 s + 6.5 sTwo messages; the second is sent only after a pause, and is what the repeat command pages through.
Nothing changed“Saved, no changes.”1.5 sThe user pressed Save and must hear that it was accepted [5].
Field that is not a column“Saved. Priority high, was normal.”3 sAlways with its value, even among five changes: the row label cannot confirm it.
Save caused a conflict“Saved. Start 11 June, was 13 June. Conflict. 2 days too early for Implementation.”6.5 sOne message (P23).
Save refused by validation“Not saved. Name is empty.”2.5 sThe item in error is identified and described in text [7]. Focus should go to the field.

"Was", not "to". Today's "Duration: 5 days to 8 days" is heard as a range. "8 days, was 5" cannot be.

The threshold of three. One or two changes are spoken with values, three or more as field names followed by the row label. Changes to fields outside the table are always spoken with values and do not count toward the three.

A changed name is always followed by the row label, so the new name is heard as the row's name and not only as a value.

Consequential changes. If the user changed the duration and the end date moved, the end date is not a second "change"; it is spoken once as "Ends 20 June".

Cancel: “Cancelled. Nothing changed.” Delete: “Deleted Retrospective.” with the name, because the row is gone and the screen reader will read a different row next.

D9. Many key presses in a row

The user holds or repeats Alt+Shift+Right ten times.

What must not happen: ten announcements queued behind each other, each naming a date that is already wrong.

The answer depends on the delivery route (2.3).

On the row-label route (the code today). Every key press re-renders the row and focuses it again, so every press is spoken as focus speech. No waiting is needed and a single press is confirmed at once. The proposal is that the text for each press is the running total:

  1. First press: “Starts Sunday 14 June, ends Thursday 18 June.” (3.5 s).
  2. Second press within about a second: “2 days later. Starts Monday 15 June, ends Friday 19 June.”
  3. Tenth press: “10 days later. Starts Tuesday 23 June, ends Saturday 27 June.” (5.5 s).
  4. A conflict that exists at that moment is part of the same text: “4 days earlier. Starts Tuesday 9 June, ends Saturday 13 June. Conflict. 4 days too early for Implementation.” (8 s). It is at the 8-second limit and stays one message because the alternative is two competing updates.
  5. Presses in opposite directions net out: 6 later and 2 earlier is "4 days later". A net of zero is "Back where it started. 13 to 17 June."
  6. The counter resets after about a second without a press.

Whatever the user hears last is complete and correct, however many presses came before it. This works well only if each new focus announcement cuts off the one before it; if a screen reader queues them instead, the user hears ten sentences. That behaviour was not verified for any screen reader and is the first thing to check on listen.html, which simulates this route on a focusable row.

On the live-region route (cell navigation, fallback): trailing announcement with a summary.

  1. Each key press cancels the pending announcement and restarts a quiet-time timer of about 350 ms. The published range for spacing updates is 350 to 500 ms [11].
  2. When the timer fires, the text is the same running total as above.
  3. Conflicts that appeared and disappeared during the burst are not mentioned.

The cost of the live-region route: a single press is confirmed about a third of a second later than it could be. That delay may feel like a silent failure, which is the complaint this project started with; task T4 measures it.

Alternatives, all on listen.html as a simulation:

The same rule for duration keys and for the plus and minus buttons in the form (already debounced at 350 ms in form.js).

Reaching a limit during a burst stops the count and says so: "3 days shorter. 1 day, the minimum. Ends Saturday 13 June."

D10. Going to a predecessor when there are several

D11. Errors and refusals

Pattern: what did not happen, then why, then (if there is one) what to do instead. No task name, no "Error:" prefix, no apology. WCAG asks that the item in error be identified and the error described in text [7].

CaseTextTime
Read-only“Read-only. Not changed.”2 s
Project task“Not changed. A project takes its dates from its subtasks.”4.5 s
Milestone, duration key“A milestone has no duration.”2 s
Minimum duration“Already 1 day, the minimum.”2.5 s
No predecessor / successor“Depends on nothing.” / “Nothing depends on this.”1.5 s
Target filtered out“2 of 5, Budget approval, is filtered out. Focus did not move.”5.5 s
Link refused: loop“Link not created. It would make a loop: Implementation already depends on Testing.”5.5 s
Link refused: rule“Link not created. A project cannot be linked to its own subtask.”5 s

D12. Empty chart, one task, 5000 tasks

D13. Keyboard help

Ten key bindings in one message take about half a minute (31.5 s), cannot be replayed, and vanish from a braille display. Two forms are proposed.

1. A help dialog (primary). A real dialog: a heading, three short headed groups, each a list of "action, key". The user reads it with their own commands, at their own pace, in speech or braille, and can search it. This follows directly from the limits of live regions [11] and from P26.

2. Spoken help in groups (quick form), on the help key without opening anything; the same key again continues:

GroupTextTime
1“Reading, 1 of 3. Dependencies, Shift D. Same time, Shift O. Ends before, Shift B. Starts after, Shift A.”9.5 s
2“Moving focus, 2 of 3. To what this depends on, Shift P. To what depends on this, Shift S.”8.5 s
3“Changing, 3 of 3. Earlier, Alt Shift Left. Later, Alt Shift Right. Longer, Alt Shift Up. Shorter, Alt Shift Down.”10 s

7. Settings, not rules

These decisions depend most on personal habit, experience and equipment. They should be options with the catalogue's choice as the default. Order: most likely to matter first. All [A]; the argument for having settings at all is that announcements bypass the verbosity controls users already have in their screen readers (1.6), and that WCAG's AAA criterion wants interruptions to be suppressible [6].

SettingValuesDefaultWhy it is personal
VerbosityBrief, Standard, DetailedStandardExperience with the chart; speech rate.
HintsOn, offOn, each hint at most three timesNeeded once, noise afterwards.
Wording for linksPlain ("depends on"), project management ("predecessor")PlainTraining.
Link typesOnly unusual, always, never; as names or as sentencesOnly unusual, namesWhether the user schedules with all four types.
Names before "and more"1 to 103Speech rate and memory; braille users may want more.
Feedback while repeating a keySummary after a pause, a tick per press plus summary, every press in fullSummaryScreen reader behaviour differs [13]; some users want rhythm.
Weekday in datesNever, edit results, alwaysEdit resultsCalendar-driven or not.
Year in datesWhen different, alwaysWhen differentMulti-year projects.
Task codes in namesAs written, at the end, omitted in listsAs writtenWhether the team uses codes as names.
Row label fields and orderA list of fieldsName, span, duration, progressThe question the user asks most differs per role.
Order of ancestorsNearest first, root firstNearest firstHabit from file paths.
Automatic conflict announcementsWith the change, only on demandWith the changeSome users will want silence while rearranging many tasks.
Delay before announcing after focus moves0 to 1000 ms500 msDepends on screen reader and speech rate.
Announcements at allOn, off per featureOnAlready possible through options.features.

Not settings: the first word of a result, the politeness level, and the rule that refusals are always spoken.

Settings belong to the person, not the page: they should be stored per user and be changeable from the keyboard without leaving the chart.


8. Open questions for user testing

Everything labelled [A] is a hypothesis. This section turns the important ones into tasks.

Who. Five to eight people who use a screen reader daily; a mix of NVDA, JAWS and VoiceOver; at least two who use a braille display; at least two with project-planning experience and two without. Participants use their own machine, speech rate and settings. Recruiting through organisations of blind people and paying participants is the norm.

Material. The demo chart for real tasks; listen.html for comparing wordings side by side. Rotate the order of variants between participants.

Method note. Ask for the answer to a factual question after each announcement ("when does it end?"), not only for a preference. Preference and comprehension often differ.

#QuestionTask for the participantWhat to record
T1Row label: range or labelled fields? (P7, row A/B/C)Arrow through 12 rows in each variant. "Which task ends last? Which is longest?"Correct answers, time, where speech was cut off, preference.
T2Are link-type names understood? (2.6)Hear "Documentation, start to start". "Explain in your own words what that means for Documentation." Then the plain sentence.Correct explanations per variant, by planning experience.
T3Is "finish to start" as silence learnable?After ten minutes of use: "What type of link is there between Testing and Implementation?"Whether the unspoken default is inferred.
T4Burst feedback and its delay (D9)"Move Testing ten days later. What day does it start now?" in each of the three strategies. Then one single press: "Did anything happen?"Correct date; whether the single-press delay is noticed; preference.
T5Save result and the row label (P23, D8)Change the duration, save. "What changed, and when does the task end now? Which task are you on?" with the result instead of the label, before it, and after it through the live region; with and without the task name.What was actually heard, in what order, per screen reader; missed announcements; whether the participant can still name the task.
T6Conflict wording (D7)Move a task into a conflict. "What is wrong, and what would you do?" for variants A to D.Whether the participant names the other task and the amount; number of key presses to fix.
T7"Too early": is the implicit subject understood?As T6, with a conflict on another task. "Which task is in the wrong place?"Correct task named.
T8Overlap wording (D6)"Which tasks run alongside Testing for its whole length? Which only partly?" variants A to C.Correct answers; whether "first 2 days" is understood as relative to the focused task.
T9List limit and paging (D2)On a task with 8 predecessors: "Does this task depend on Legal review?" (the fifth).Whether paging is discovered; whether the dialog or one-by-one navigation is preferred; preferred limit.
T10Is a verbosity setting inside the component wanted? (5)Use Brief and Detailed for five minutes each.Which level each participant keeps; whether they expected their screen reader's setting to govern this.
T11Where am I: order of ancestors (D1)On a level-5 task: "What is this task part of?" in both orders.Speed and preference.
T12Pronunciation in the participant's own setup (P13, P14, P17)Play the probes on listen.html (ISO date, "13 Jun", "5d", "PRJ-1423", "FS", en dash, emoji, ampersand).What was actually spoken, per screen reader and voice. This replaces assumptions with a table.
T13Help: dialog or spoken groups? (D13)"Find out how to hear which tasks run at the same time" with each form.Time to the key; whether the spoken groups are remembered.
T14Braille (2.4)T5, T6 and T9 with speech off and a braille display.What appeared on the display, for how long, what was readable.
T15"Ends" versus "finishes" at the participant's speech rateHear "Starts Sunday 14 June, ends Thursday 18 June" at their own rate. "What is the second date?"Mishearings.
T16Depends on / required by versus predecessor / successorFree choice after trying both; ask what they would call the relation unprompted before hearing either.The words people use themselves.
T17Weekday in move resultsMove a task onto a weekend without being told.Whether the weekday is noticed and wanted.
T19Is a row without its name disorienting? (2.3)After a save, ask the participant to read the current row again with their screen reader's own command.Whether hearing the result in place of the name confuses, and whether they expected the name.
T18Does anything need to interrupt? (P21)At the end: "Was there a moment when you learned something too late?"Candidates for assertive, or confirmation that there are none.

What can be checked without participants (by the author, with VoiceOver, NVDA and JAWS): whether a re-rendered, re-focused row is read again, and whether a new focus announcement cuts off the previous one, in each screen reader (the row-label route of 2.3 and D9 depends on both); what a braille display shows while a result is the row's label; the order of row label and announcement on the live-region route (T5's mechanics); what listen.html's probes produce in each screen reader (T12's raw data); what ARIA attributes the library puts on rows (1.5, D1, D12); whether queued polite announcements pile up during a burst (D9).

What stays unknown after this research: how Microsoft Project presents dependencies and the bar chart to screen readers; how Outlook and Google Calendar convey overlapping events; default speech rates of the screen readers; how each synthesizer reads ISO dates and codes.


9. References

Only sources that were opened and read for this document in November 2026. Where a source was seen only through a summary, that is stated.

W3C

Live regions

Products

Screen readers

Research

Speech and voice design

Internationalisation

Numbers [27] and [30] are intentionally unused: two sources considered during research were dropped because they were seen only through third-party summaries.


Implementation notes

Added 16 November 2026, after this document was turned into code. Everything above is unchanged; this section records where the implementation decides something the document leaves open, or deliberately departs from it. The decisions were made by the project owner.

File paths in the sections above refer to the code as it stood on 16 November 2026. The code now has an engine in src/core/ and an adapter in src/adapters/dhtmlx/ (see ADAPTERS.md): src/strings.js is src/core/strings.js, src/features/form.js became part of src/core/engine.js and src/adapters/dhtmlx/adapter.js, and src/core/focus-label.js is src/adapters/dhtmlx/focus-label.js.

Deliberate deviation

Decisions where the document is open or inconsistent

#WhereDecision
aD8, two-field example without "Ends""Ends {date}." is spoken after one or two changed table fields whenever the end date changed.
bR4 and D2, list of 2 or 3 dependenciesWithout unusual link types: "Depends on 3: A, B, C." With any unusual type: "Required by 3." followed by one sentence per task.
cD2, threshold for a summary without namesFrom 10 dependencies in one direction, the first press gives the count, the deciding task and the conflicts; names follow on further presses. 10 is the threshold the document gives for the list dialog.
dN3 with an unusual link type"This depends on Build, finish to finish." / "Testing depends on this, start to start."
eMoving a milestone (not in the catalogue)"On Sunday 21 June."
fC5 when the changed task itself is the one too early"2 conflicts. Largest: 2 days too early for Implementation."
gOrder of "Required by" (only "Depends on" is specified)Tasks in conflict first, then the one that starts first.
hD6, projects from other branchesNot counted as "at the same time". (The option features.time.includeProjects was removed on 2026-11-18, with the other fine-tuning options, before the first release.)
iSection 5, Detailed level, link typesSpoken by name every time; the plain-sentence wording (4.2 C) is not implemented.
kS2, R12, C5, D10: keysShift+V overview, Shift+W where am I, Shift+C next conflict, Shift+Z back. Help gains a group "Orientation" (Shift+W, Shift+V); Shift+C and Shift+Z join "Moving focus".
lS2, chart summary"{count} tasks, {shown} shown." only when some tasks have no row; otherwise "{count} tasks." The period, then "No conflicts." or "{count} conflicts.", then "{count} tasks without dates." when there are any. Brief: tasks and period, conflicts only when there are some. Detailed: also the number of projects and milestones, and the largest conflict.
mR12, where am IThe task's name first, then ancestors nearest first, then "Level {level}, {n} of {total}." with the position among the siblings that have a row; a top-level task: "Top level, {n} of {total}." Brief: name and nearest parent. Detailed: also "{count} subtasks inside."
nC5, conflicts on demandNot a spoken list: each press moves focus to the next task that is too early (largest conflict first) and speaks "{n} of {total}." with the full form (C3) before the row label, like the jumps. With one conflict the position is left out.
oD10, back to originThe origin is the task where a series of Shift+P, Shift+S or Shift+C presses started; arrowing to another task and jumping again starts a new series. Without one: "Nothing to go back to."
pKeys for moving a task and changing its duration (4.6, D9, D13 examples use Alt+Shift+arrows)Since 2026-11-17: Shift+J earlier, Shift+K later, Shift+I longer, Shift+M shorter. According to the vendors' documentation (sources in docs/KEYS.md), Alt/Option+Shift+arrows are indent and outdent in Microsoft Project and Bryntum, and group and ungroup in Google Sheets; that Left Alt+Shift can switch the input language on Windows is a common Windows setting reported by the project owner; Microsoft's pages read for docs/KEYS.md do not document it. The texts are unchanged; the examples in this document keep the old keys. Sources and reasons: docs/KEYS.md.
q2.3 and D9: the row-label route for keyboard editsSince 2026-11-18 keyboard edits (and Shift+C when the conflict is on the focused task) use the live region, after the row is re-rendered; the row label is left as the task's own label. Focus stays on the same row, and VoiceOver in Safari did not announce the re-focus of that row, so the result in the label was not heard (found by the project owner with VoiceOver). The result is now heard after the row instead of before it, 500 ms after the re-render (fixed; the option features.edit.delay was removed on 2026-11-18): an immediate live region update was not heard either, most likely because the screen reader was still handling the replaced row. With the 500 ms wait the result is heard (confirmed by the project owner with VoiceOver in Safari, macOS 26.2, 2026-11-18). The form results (saved, created, cancelled) keep the row-label route, but the label is set when Save or Cancel is pressed, before the library re-renders the row and focuses it; the extension no longer re-renders the row itself after the form closes, which had moved focus too early and replaced the focused row (VoiceOver silent). Tried in between on 2026-11-18: the live region, which the project owner heard after the row, the wrong order. With the label set at Save, "Saved, no changes." is heard first (confirmed by the project owner with VoiceOver in Safari, macOS 26.2, saving with Enter and with Space). Two earlier checks had run older JavaScript modules kept by Safari's cache; the demo server now forbids caching and the demo's event log (?debug=1) says which version runs. Jumps, back and the next conflict on another task keep the row-label route too, because focus moves to another row.
rCell navigation mode (2.3)From 2026-11-18 (morning) the row-label route also worked on cells, by wrapping the focused column's template; removed the same day after the pre-release review: it was a second replacement of a shared template on top of the first. In cell mode every result, jumps and the task form included, now goes through the live region and is heard after the cell.
s2.3, the result in the row's labelChanged on 2026-11-18 after the pre-release review. The result is always followed by the row's own label, so the row never loses the task's name ("Cancelled. Nothing changed." alone did). It is removed when focus leaves the row (to another element, to nothing, or to another window) and at the latest 5 seconds after focus landed; it is never spoken twice (the 600 ms fallback could repeat a result that had been heard). With the tooltip plugin, which shows the same template to sighted users, results go to the live region instead. After a keyboard edit or a link change the data is read again and what was stored is spoken; a change refused by the application gets "Not changed. The change was not accepted." (new text).
jD13, spoken help itemsEach item ends with the word "key": "Dependencies, Shift D key." Found by the project owner with VoiceOver: after a single capital letter and a period ("Shift D. Same time") VoiceOver heard an initial and made no pause between items.

Not implemented