Read a half-ticked box as not ticked, not as ticked

[kevin9327]

Ask a Bot to tick every row on a page whose "select all" box is half-ticked — some rows already
selected, some not — and it clicks nothing and tells the person the rows are selected. They are not.
The same happens on any tri-state checkbox: a consent tree, a permissions matrix, a partly expanded
filter list.

Why

The snapshot a Bot reads before it acts carries checked per control. Here is what
ariaSnapshot({ mode: "ai" }) actually produces, captured against Chromium 151 (playwright-core
1.62.1's own build) for a fieldset of three boxes with the first one's indeterminate set:

- group "Toppings" [ref=e2]:
  - generic [ref=e4]:
    - checkbox "Select all" [checked=mixed] [ref=e5]
    - text: Select all
  - generic [ref=e6]:
    - checkbox "Bacon" [checked] [ref=e7]
    - text: Bacon
  - generic [ref=e8]:
    - checkbox "Extra Cheese" [ref=e9]
    - text: Extra Cheese

An ordinary tick is the bare [checked]; the third state is [checked=mixed] — the only flag in
that output carrying a value. Playwright's renderer says the same thing in two lines:

if (ariaNode.checked === "mixed")
  key += ` [checked=mixed]`;
if (ariaNode.checked === true)
  key += ` [checked]`;

toElement read the flag like this:

element.checked = descriptor.flags.get("checked") !== "false";

"false" is a spelling Playwright does not emit. "mixed", the one it does, is not "false", so a
half-ticked box arrived as checked: true and the Bot went on from a snapshot that said the job was
done.

The fix

const state = descriptor.flags.get("checked");
element.checked = state !== "mixed" && state !== "false";

The defensive "false" branch stays, since removing it would be a second change in one line.

Why false and not "leave it off": the published contract is checked?: boolean
(server/src/computer/schema.ts), and for a checkable role this file always fills it in on purpose —
its own comment says absence would otherwise be ambiguous between an unticked box and a control that
does not tick. false is the true half of a yes-or-no answer about a box that is not ticked, and it
is also the answer that produces the right action, because clicking a half-ticked box ticks it.

Measured

bun test agent-computer/tests/aria-snapshot.test.ts

• Against origin/main (1c7bd92) with only the three new tests applied: 26 pass, 1 fail.

expect(byName.get("Select all")?.checked).toBe(false);
error: expect(received).toBe(expected)
Expected: false
Received: true
(fail) parseAriaSnapshot, against captured output > a half-ticked box is not reported as ticked

• With the change: 27 pass, 0 fail.

The fixture is the captured output above rather than a hand-written entry, for the reason the note at
the top of that file gives: a guess loses the generic wrapper each box sits inside and the text
node beside it, and would pass while the shipped parser walked something else.

The other two new tests are the guard against over-correcting and pass before and after: the fixture
is valid YAML, and a half-ticked box is still a control with a usable ref — its flag carries a value
and ref comes after it, which is the parse that would break if this were done with a pattern. The
existing case, an ordinary [checked] and an ordinary empty box, is asserted in the same test and is
unchanged.

bun test agent-computer/tests — 234 pass, 15 skip, 20 fail. The same 20 fail on origin/main
(231 pass there, so the three new ones are the whole difference): 11 in shell.test.ts, which spawns
/bin/bash, and 9 in workspace.test.ts, which creates symlinks. Neither is available on the Windows
machine these were run on and neither is touched here.

Also run, clean: bunx biome check on both changed files, and bun run typecheck in
agent-computer.

Note on the changelog

A Bot behaves differently afterwards, so there is an entry. It goes at the top of ## Unreleased,
which is the line every open PR touching the changelog inserts at, so it will conflict with any other
that lands first. Happy to rebase.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.