[Feat] Recover from stalled browser and desktop tasks with bounded replanning

[Yunaik]

Why

Browser and desktop tasks can stall after repeated ineffective actions. The existing rethink mechanism provides a starting point for detecting stalled execution and requesting a revised plan.

Scope

Detect repeated ineffective actions, refresh the observation, and ask the existing chat model for a revised plan. Limit recovery attempts and total recovery time. Stop or ask the user when the budget is exhausted.

Reuse the existing rethink mechanism where applicable. Preserve permission checks during retries. Calibrate any confidence thresholds separately for each model using labelled examples.

Acceptance criteria

1. Behavioural tests cover repeated ineffective actions, delayed progress, observation refresh, successful recovery, and exhausted recovery budgets.
2. Deliberately blocked browser and desktop tasks either recover with verified completion or end with a clear explanation and an actionable escalation.
3. Recovery records include the triggering actions, refreshed observation, revised plan, attempts, and termination reason.
4. The shared task suite compares recovery enabled and disabled, reporting completion rate, extra model calls, added time, and wasted actions.
5. Retries stop within the configured attempt and time limits. Tests verify that permission checks remain active.

Dependencies and starting points

Use the repeatable browser and desktop task suite linked below. Begin with existing operations, then include the desktop extensions as they become available.

s1a/tool/rethink.py, evals/README.md:33, and docs/browser-front.md.

Evaluation dependency: #13.

Related desktop extension: #15.