One of the biggest reliability gaps in browser agents isn't navigation — it's keyboard behavior.
The computer-use model emits sequences like:
Down Down Down Tab*5 Enter Tab Tab Enter
The agent previously treated everything as combinations, which caused form-filling failures.
I extended the agent to support:
- space-separated sequential keypresses
- repetition syntax (Tab*5)
- automatic detection between sequences and combos
- clearer validation and error reporting
This closes a major ergonomics gap for complex form automation.