Accessibility

Stella's interactive surface is the Command Deck, a tabbed terminal UI. By default it draws the way full-window terminal apps do: on the alternate screen, repainting a whole grid many times a second. That is a good surface for an eye and a poor one for a screen reader, which is handed cell churn with no notion of a new line appeared — and which loses the entire conversation the moment you quit, because the alternate screen is torn down with the program.

--accessible changes how the deck talks to your terminal so that a reader can follow it.

Turning it on

# One session
stella --accessible

# The same thing, spelled out
stella chat --accessible

# Resuming a session works the same way
stella resume --accessible

--accessible is a global flag, and stella chat and stella resume are the two commands that open the deck.

To make it your default, set the environment variable in your shell profile rather than remembering a flag on every invocation:

export STELLA_ACCESSIBLE=1

The variable is on when it is present and set to anything other than 0 — so STELLA_ACCESSIBLE=1 and STELLA_ACCESSIBLE=yes both enable it, while STELLA_ACCESSIBLE=0 and an empty value leave it off.

--simple is accepted as an alias for --accessible. It is the spelling an earlier, short-lived separate surface used; it now selects this mode.

What changes

The deck never takes over the screen

The session draws inline, underneath your prompt, on your own screen. The alternate screen is never entered and never left. Nothing above the deck is hidden while it runs, and nothing is destroyed when it exits.

Finished messages move into scrollback

Each completed transcript entry is written into your terminal's ordinary scrollback exactly once — so your reader announces it as it arrives, your review cursor can go back through it, and it is still there after you quit. The live pane skips whatever has already been written, so nothing is read out twice.

Entries move per agent lane, and the newest entry of each lane is held in the pane until the session ends, because streaming deltas keep coalescing into it until it settles.

Motion is frozen

Accessible mode implies --no-anim: the progress bar's shimmer and pulse and the caret blink all stop. No region of the screen repaints on a clock, so there is nothing for a reader to pick up as churn.

Panels stack into one column

The three places the deck drew two panes side by side — the GRAPH tab, the SKILLS tab, and the session's work rail — stack vertically instead. No rendered row splices content from two logical panes, so reading order is top to bottom. For the work rail, the deck takes the path a narrow terminal already takes: a one-row state strip, with ⌃S opening the whole thing.

Grid views read as labelled records

A table is a compression scheme that only works on an eye. $0.05 sitting under a Cost header two rows up is a labelled value if you can see the column; read aloud it is the number five with nothing attached, because the alignment carrying the meaning is whitespace — exactly what a reader collapses.

So TRACES, ISSUES, TOOLS, AGENTS and INSTALLED render one record per row, with every value carrying its own label:

status running · cost $0.05 · cpu 3%

Nothing is dropped — the same fields, in the same order, saying what they are. A row too long for the terminal is clipped with an ellipsis rather than wrapped, because the deck's scrolling is line-exact and a record that wrapped to two rows would break ↑/↓.

Moving somewhere is announced

On the alternate screen, changing tabs is self-evident: the whole grid changes. Drawn inline and read aloud, the same transition is silent — the pane repaints in place and nothing says you are somewhere else.

So each move emits a line into scrollback:

The overlays that announce themselves are help, state, queue editor, sessions, context, inbox, inspect, transcript search, and the graph file picker. Chrome that opens and closes on nearly every keystroke — the slash menu, the composer's own popups — deliberately stays quiet, because announcing it would bury the moves that matter.

The program's own voice is marked

Messages Stella itself speaks — "conversation cleared", driver and chrome notices — open with ▸. In the default deck those are distinguished by which rail they render on, which is a visual distinction, and therefore exactly the one that does not survive being read aloud. Without the marker the transcript asserts that the model said "conversation cleared".

The cursor and the mouse

The hardware cursor sits on the real insertion point in the composer, which is what a reader follows and what a CJK/Japanese/Korean IME anchors its candidate window to.

Mouse capture is forced off, even if something asked for it: a captured mouse takes your terminal's own selection away, and selection is how several assistive technologies read a terminal.

What does not change

Everything else. All nine tabs, the prompt queue, sub-agents, steering a running turn, permission gates, slash commands, resume — the same program, the same keys, the same behavior. If a feature works in the deck, it works here.

When your terminal will not cooperate

Drawing inline means anchoring to the cursor's current position, which Stella has to ask for: it writes a cursor-position request and waits for the terminal to answer. Every real emulator answers. Some minimal terminals and most test harnesses do not, and the request times out.

When that happens the deck still starts. It draws on your own screen — never the alternate one — with the scrollback flush disabled, and it says so:

accessible mode: this terminal did not report its cursor position, so messages stay in
the deck pane instead of moving into scrollback

The degrade is announced rather than silent on purpose. The promise of the mode is that finished messages become durable output, and quietly not delivering it is worse than not offering it. If you see that notice, everything else about the mode still applies — stacked panels, labelled records, frozen motion, marked notices — but the transcript stays in the pane, and announcements have nowhere to go.

--accessible or --plain?

They answer different questions, and they compose with different things.

Use --accessible when you are at a terminal with a screen reader. Use --plain for pipes, CI logs, TERM=dumb, and anywhere there is no interactive terminal at all — the deck steps aside for the REPL automatically in those cases anyway.

Accessible mode takes no part in that deck-or-REPL decision. If stdin or stdout is not a terminal you get the REPL, whether or not you asked for accessible mode, because an inline viewport needs a real terminal just as much as the alternate screen does.

Related terminal controls

These are independent of --accessible and work on their own:

Checking that it is working

Start a session with stella --accessible and look for these, in order:

1. Your prompt is still visible above the deck. If the screen cleared and your scrollback disappeared, you are on the alternate screen and the mode is not on — check for a STELLA_ACCESSIBLE=0 in your environment overriding the profile you expected.
2. No degrade notice. If you see the cursor-position message quoted above, your terminal did not answer, and finished messages will stay in the pane.
3. Ask something, and let the turn finish. The completed message should be in ordinary scrollback — scroll your terminal up, or use your reader's review cursor, and it should be reachable as normal text.
4. Press ⇥. You should hear ▸ AGENTS tab or similar.
5. Quit. The conversation should still be on screen.

Reporting a gap

Accessibility here means ordinary sequential terminal output, which every screen reader already handles — there is no terminal accessibility API to target, so there is nothing reader-specific to configure. If some part of the deck still does not read correctly, that is a bug worth filing on the issue tracker, with your terminal emulator and reader named.

See also

• stella chat — the command reference, including every flag
• Agent Modes — the other ways to drive Stella