Ulka
A browser-agent extension that combines language-model planning with Jev-selected browser actions and an evidence-checked execution runtime.
Added to Jevfast
How it uses Jev
FX plans browser subgoals and handles writing and verification. For each action decision, the runtime builds a Choice over currently available operations and, for operations with multiple candidates, a second Choice over compatible observed targets. IDs and labels are generated from the fresh accessibility snapshot and target filters. Ulka validates target freshness, applies approval rules, executes, and verifies.
What Jev decides
Representative action-space fixture decisionExample answers · not a recorded Jev response · Source ↗
Question 1 · operation
YOUR APP
INSTRUCTION
Choose a safe next action from observed state; DONE requires visible goal completion and BLOCKED means no supported action can progress.
STATE
Representative source-derived page with one clickable button and one editable field, no scrollable content or observed tabs. The action-space builder offers the shown operations; target choices are omitted because each operation has one target. Other pages produce different options. Jev also receives the goal, page meaning, time context, and recent actions.
JEV · CHOICE
- CLICK
- TYPE_TEXT
- WAIT
- GO_BACK
- GO_FORWARD
- RELOAD
- OPEN_TAB
- DONE
- BLOCKED
App workflow
- Plan task
FX converts a user request into browser subgoals when the task needs multiple milestones.
- Observe and decide
Ulka reads the current accessibility state and action catalog; Jev selects a compatible operation and target.
- Guard and execute
The runtime checks target freshness and obtains required approval before executing the browser action.
- Verify outcome
The configured text model checks completion against the original user request using observed evidence.
Why it is interesting
The separation between semantic action choice and host-enforced freshness, approval, execution, and verification makes the model decision inspectable and bounded.