decısıum
/

How a Step runs

Decisium reads the page, lists what is on it, and asks Jev questions whose answers are picks among that list. Jev never writes code or values; every answer comes back as a probability per option. Decisium's code then checks the confidence, performs the action with Playwright and decides pass or fail.

A question is asked in a request: one call to Jev carrying the Step's text, a text snapshot of the page (ui) and one or more questions. Questions in the same request read the same page, so their answers are consistent.

An action Step

An action Step runs in Turns. Each Turn is one request with three questions:

Question Options Answer used for
Given the actions taken for the Step so far, does ui show the Step is fully accomplished? yes/no, as a probability ending the Step
Which element in ui should be interacted with next? every actionable element, plus "the page itself" and "none" the target
How should that element be used? click, type, select, hover, press a key, scroll… the action

Then code decides, in this order:

  1. Done? done ≥ 0.8 ends the Step. After an action that changed the page, done ≥ 0.5 with target "none" is enough.
  2. Sure of the target? The top element needs ≥ 0.5 (confidence.target) and a lead of ≥ 0.1 over the next; otherwise AMBIGUOUS. Target "none" is BLOCKED.
  3. The right control? A click on a lookalike of a quoted control ("Publish again" when the Step quotes "Publish now") is AMBIGUOUS, and nothing is clicked.
  4. Act. Playwright performs it, waits for the requests it started, and the next Turn reads the new page.

A Step gets at most 8 Turns (maxTurns; then STEP_TIMEOUT). After each action Decisium waits up to 5 s (actionTimeout) for requests and DOM changes to settle. Some actions add one request:

When Extra question
The action types, selects or uploads Which value should be entered into the element? Options: the Step's quoted values, numbers and Doc string, plus "none".
A browser dialog is open How should it be answered? Options: OK / Cancel, or one of the quoted values for a prompt.
The action is a drag Onto which element should it be dropped?

I press Escape, I go back, I go to "/settings" ask nothing: code does them directly.

Example

Suppose you are testing a campaign-management app. You are viewing a campaign, and its delete command is inside a "More actions" menu. Your Step is When I delete the campaign.

Decisium can carry out the whole task without a separate Step for each click. Here is one recorded run; the probabilities illustrate what Jev answered, not fixed values every run will produce:

Turn Page done target Code
1 Campaign page 0.04 button "More actions" 0.99 click "More actions"
2 Menu open: "Delete campaign" 0.03 menuitem "Delete campaign" 1.00 click "Delete campaign"
3 Campaign list, toast "deleted" 0.98 none 0.86 DONE

This took 3 requests: open the menu, choose delete, then check whether the task is complete. Opening the menu alone did not finish the Step. You do not need to add I click "Delete campaign" afterward—the first Step already did that. To verify the result explicitly, follow it with a check that the campaign is no longer listed.

A check

A check is one request with four questions; Jev only says what to look at, code compares:

Question Options
What kind of check is it? value, presence, absence, state, fuzzy
Which element is the check about? every element with text, plus "none"
Which control state does it require? checked, disabled, expanded, selected…
Where does it say to look? the page's lists, tables, sections, dialogs; "anywhere" or "named area missing"

Code then reads the element and compares it with the quoted value. A check that does not pass yet is asked again each time the page changes, up to 5 s (checkTimeout): one more request per change. A check that quotes nothing first asks which value from an earlier Step it refers to (one more request). A judged check adds a fifth question: does ui satisfy the check?

Large pages

How many requests

Typical counts:

Step Requests
Action done in one click 2 (act, then confirm done)
Action through a menu or dialog 3 or more
Type into a field + 1 per value typed
Check that passes at once 1

Request counts depend on the path through your app, not just the number of Steps. Typing one value typically takes 3 requests: choose the field, choose the value, then confirm the task is complete. Menus, loading states and retries can add requests. Token usage also depends on how much page text each request includes.

Replays (--record none) send no request: the cassette holds each Step's answers. See Record and replay.

Where to see it