[Feature] Give agents control over JS dialogs (confirm/prompt/beforeunload): dismiss/text/status instead of forced auto-accept

[Wraient]

What I found

bsk 0.3.1 (daemon protocol 1.3, extension 0.3.1, Chrome 153, Linux) silently
auto-accepts every JavaScript dialog and gives the agent no way to do
anything else. There is no dialog command, no accept/dismiss/status
subcommand, and the words dialog/alert/beforeunload appear zero times in
SKILL.md, so agents can't even discover the behavior.

Repro (all on https://example.com, fresh Agent Window session):

bsk evaluate --session <id> "alert('x')"
# null
# dialog: type=alert handled=accepted message=x        <- auto-OK, reported. Fine.

bsk evaluate --session <id> "confirm('proceed with delete?')"
# true
# dialog: type=confirm handled=accepted ...            <- agent CANNOT cancel.

bsk evaluate --session <id> "prompt('your name?','anonymous')"
# anonymous
# dialog: type=prompt handled=accepted ...             <- agent CANNOT type or cancel.

bsk evaluate --session <id> "window.onbeforeunload = () => true; 1"
bsk navigate --session <id> "https://www.iana.org/"    # needs prior user gesture or Chrome suppresses it
# dialog: type=beforeunload handled=accepted ...       <- "Leave" forced, "Stay" impossible.

Credit where due: the dialog: type=… handled=… message=… report lines are
good — the agent at least learns what happened after the fact.

Why this matters

One confirm("Delete all 200 rows?") on a hostile or sloppy page auto-returns
true and the agent marches on believing the user said yes. Any flow that
needs typed prompt() input or a beforeunload "Stay" is simply unusable —
there is no control surface at all, only the post-hoc report. Forced accept is
safer than a hang, but for confirm/prompt it trades a stall for a wrong
answer, which is worse for destructive actions.

What good looks like elsewhere

• vercel-labs/agent-browser: dialog accept/dismiss [text],
  dialog status, and auto-dismiss limited to alert+beforeunload
  (the two types with no real decision); confirm/prompt stay pending for
  an explicit agent decision, every response carries the pending dialog as a
  warning, --no-auto-dialog opts out (PR #1075 there).
• browser-use harness: exposes CDP Page.handleJavaScriptDialog
  (accept + promptText) and documents reactive (CDP) vs proactive (JS
  stub) handling, including the beforeunload navigate-then-dismiss recipe.

Proposal

1. bsk dialog accept [text] | bsk dialog dismiss [--session …] and
   bsk dialog status (pending type/message/url), wired to the same
   interception that already produces the dialog: lines.
2. Limit auto-accept to alert (+beforeunload with an opt-out flag);
   leave confirm/prompt pending for an explicit decision instead of
   answering for the agent.
3. Document the dialog: report lines and the new commands in SKILL.md
   (currently zero mentions of dialog/alert/beforeunload).

Happy to re-test any build that adds this.