Skip to content

Let a person take the wheel without being offered it - #166

Merged
davidmckayv merged 2 commits into
mainfrom
fix/a-person-can-always-take-the-wheel
Aug 22, 2026
Merged

Let a person take the wheel without being offered it#166
davidmckayv merged 2 commits into
mainfrom
fix/a-person-can-always-take-the-wheel

Conversation

@davidmckayv

Copy link
Copy Markdown
Contributor

Found while testing the four open PRs against latest main, and separate from all of them.

What was wrong

Take control only appeared once a Bot had called computer_request_help. So the button a person needs depended on the Bot getting one instruction right — and it does not always.

Asked, in a fresh channel, a completely neutral question:

Open drive.google.com and tell me the name of the first file listed in my Drive.

the built-in General Assistant answered:

"I cannot access your Google Drive directly right now because you need to be signed in. Please sign in to your Google account… Let me know if you would like me to prompt you to sign in."

Three times out of three. control.requested stayed false, so the computer panel offered nothing. The person is told to take control, and there is no control to take — a dead end reached by following the Bot's own instruction.

Why not fix the prompt

COMPUTER_GUIDANCE already forbids this in as many words, and names three example sentences it must never write:

Calling it IS how you ask. Words in your answer are not… So NEVER write "please sign in and let me know", "would you like to proceed", or "once you have signed in, tell me" instead of calling it. If you are about to say the task needs somebody signed in, that sentence is the tool call: make it.

The Bot wrote "would you like me to prompt you to sign in" anyway. More prose is not the answer to prose that is already being ignored.

The answer is that somebody who wants their own browser should not have to be offered it first. That makes the Bot's failure an annoyance rather than a dead end — which is what it should always have been.

What changes

The button is there whenever the Bot holds the wheel, in the inline card and in the full-size view.

The amber row stays the Bot asking — a different thing, and it keeps its colour and its reason. Without a request the row is quiet and simply says who is driving: "The assistant is driving. You can take over whenever you want."

Driven in Chrome

  1. Reproduced the prose answer on a fresh channel (three times, neutral prompt).
  2. Pressed the new Take control from that state. GET /control{"holder":"human","requested":false} — a person holding a wheel the Bot never offered.
  3. The full-size view showed the live Google sign-in page, typeable, with "You have control. Click and type on the page as you normally would."
  4. Hand back to the assistant returned holder: bot and the row went back to the quiet offer.
  5. The control survives a page reload.

What this does not do

It leaves the Bot's behaviour alone. It should still call computer_request_help, and a prompt that cannot reliably make it do so is worth its own change — this one just stops that failure from trapping anybody.

Take control only appeared once a Bot had called computer_request_help,
so the button a person needs depended on the Bot getting one instruction
right. It does not always.

Asked to open a page behind a sign-in, the built-in Bot answered "If
you'd like, I can prompt you to take control so you can complete the
sign-in ... would you like to proceed with signing in?" and called
nothing. Three times out of three, on a neutral question. The person was
told to take control and there was no control to take, and the computer
panel offered nothing: a dead end reached by following the Bot's own
instruction.

The prompt already forbids that sentence in as many words, and names two
more like it, so the answer is not more prose. It is that somebody who
wants their own browser should not have to be offered it first. The
button is now there whenever the Bot holds the wheel.

The amber row stays the Bot ASKING, which is a different thing and keeps
its colour and its reason. Without a request the row is quiet and says
who is driving.

Driven in Chrome. Reproduced the prose answer, then took the wheel from
that state: holder went to human with requested still false, the sign-in
page was live and typeable in the full-size view, and handing back
returned the Bot to driving.

Leaves the Bot's behaviour alone. It should still call the tool, and a
prompt that cannot make it is worth its own change.
@davidmckayv
davidmckayv merged commit ad2eacc into main Aug 22, 2026
8 checks passed
@davidmckayv
davidmckayv deleted the fix/a-person-can-always-take-the-wheel branch August 22, 2026 16:23
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant