Expand description
agent_step: the questions a browser agent asks for one step, built from the page’s element
table the way jev-ultrafast’s choose builds them (jev_ultrafast/model.py, MIT, Browser Use).
One operation head picks CLICK, TYPE_TEXT, SELECT, a page control such as SCROLL_DOWN, DONE or
BLOCKED, and each operation with candidates gets a <op>_target head over the element indexes,
so one request answers both what to do and where. AgentStep::decide checks the answers as
jev-ultrafast’s validate_choice does and maps them back to the action to run.
The snapshot is what jev-ultrafast’s snapshot.js returns: url, title, text and
actions, where an action has id, kind (click, fill, select, or anything else for a
page control), and for elements node, label, role and value. The body is the same one
jev-ultrafast sends, key for key and in the same order, so answers can be compared with Jev’s.
Structs§
- Agent
Step - The questions for one step and what is needed to read the answers.
- Decision
- The action an answer picked.
- Invalid
- An answer jev-ultrafast would refuse to act on.
Constants§
- HISTORY
- How many past actions go into the state.
- NEXT_
ACTION - jev-ultrafast’s rules for the operation head.
- TARGET
- jev-ultrafast’s rules for a target head, added to
NEXT_ACTION.
Functions§
- agent_
step - Builds the questions for one step from a snapshot, the goal and the actions taken so far.