[ahamsel]

Why

Real runs showed four problems:

• Late navigations. Sites often navigate a moment after an action (Kijiji submits its search once suggestions load), so Jev's next click landed on the page being left.
• Planner answers arrived all at once at the end. Gemini delivers tool arguments, including finish({result}), in one piece.
• Facebook Marketplace, "find me rent downtown under $1000":
  • Jev tried to click "Send message to " three times.
  • Each decline was honored at the click, so nothing was sent. But fast mode's step loop swallowed the decline as an ordinary failed step, so it moved on and asked again, and the flight log showed plain "click" steps.
  • It then ran out of fast-mode steps and handed off to the planner: 67s, 2 listings.
• Walmart: picked walmart.com for a user in Canada.

What changed

Late navigations

• Both drivers notice when the page starts changing after the last snapshot.
• Instead of acting on the old page, they wait for the new one and ask for a fresh snapshot.
• Same-document navigations and subframe loads don't hold clicks up.

Streaming

• The planner replies with its answer as plain text, which streams.
• Tool-chat turns are streamed and rebuilt exactly (content, tool calls, reasoning_details) for the next turn.
• Text written before a tool call is withdrawn.

Declines and irreversible actions

• Skip replaces Cancel. The action isn't done and isn't asked about again during the task; the task carries on, and the answer says what was skipped. The stop button ends the task.
• Find-only requests never send, buy or delete. Those controls aren't offered to Jev, and the planner refuses them without asking.
• A Stop during a confirmation ends the task.
• The flight log shows "declined:", "skipped:" and "failed:" steps.

Fast mode

• A field the task gives nothing to type into is skipped instead of handing off.
• When it runs out of steps with pages read, it answers from them (checked by Jev) before calling the planner.
• Option-finding answers list the best several matches.
• Pages keep 4k characters of text for the answer; Jev still sees 1.5k.

Resilience

• A model turn that fails mid-stream (an upstream abort hit a real run) is withdrawn and retried once.
• The same goes for errors inside 200 responses.

Locale

• The site picker, router and planner know the user's browser language and time zone (walmart.ca in Canada).

Traces

• Each Jev decision records its page (title, URL, start of text) and its top four options with probabilities.

Results

• Same Marketplace request (Edge): 16.6s, no prompts. A detailed best downtown match ($625, available October, link) plus 15 more listings under $1000. The answer starts showing at 12.2s. Before: 67s, 3 prompts, 2 listings.
• User's MUN request (Edge, from the MUN account): 15.9s, correct. The answer streams about 1.4s before the end.
• Kijiji (Chrome): 10.8s, no stale clicks.
• npm test: 71/71. New tests cover:
  • late-navigation driver behavior;
  • streamed tool turns (rebuilt calls and reasoning; withdrawn text);
  • retry after a mid-stream failure;
  • declined clicks skipped and not re-offered;
  • find-only requests never offering or clicking buy/send;
  • fields with nothing to type skipped.
• npm run extension:smoke: 11/11. Read-only cases now also fail on any confirmation prompt.
• Realistic suite in the user's browsers: Chrome 11/11, median 9.4s; Edge (see comment).
• npm run build passes.

🤖 Generated with Claude Code

[ahamsel]

Edge realistic suite: 11/11 passed, median 10.5s (mail_two_parts 14.6s, books_two_parts 8.1s, shop_buy 19.1s with 1 confirmation).