AI Helpers

Use decide and infer LLM statements in DSL, and the per-agent AI assistant to build workflows.

reArray integrates LLM capabilities directly into task scripts and provides a conversational assistant for building workflows.

`decide` — branching decisions

Use decide inside an if statement to let the model choose a branch based on page context:

flow {
  in portal {
    if decide "Is the user logged in?" with ($page) {
      store greeting extract { css ".welcome" }
    } else {
      fill { css "#email" } $secrets.username
      fill { css "#password" } $secrets.password
      click { css "button[type='submit']" }
      wait visible { css ".welcome" } timeout 30000
    }
  }
}

with is required and must use parentheses. The model evaluates the question against the provided context and picks a branch.

`infer` — value inference

Use infer to derive a value from context:

flow {
  in portal {
    $tomorrow = infer "tomorrow in YYYY-MM-DD" with ($context.startTime)
    fill { css "#date-field" } $tomorrow
  }
}

Assign the result to a local with $name = infer ... or persist it with store key infer ....

Allowed `with` context

Both decide and infer accept these context values inside the with (...) clause:

Context	Description
`$page`	Current page snapshot — only when a fragment cannot answer the question
`fragment { selector }`	DOM fragment around a selector (prefer this over `$page`)
`$name`	A local (prefer `with ($block)` over dumping `$store`)
`$store.key`	A single stored value (prefer over whole `$store`)
`$params.key`	A single execution parameter (prefer over whole `$params`)
`$settings.key`	A single agent setting
`$context.key`	A single metadata field (e.g. `$context.startTime`)

Combine multiple values in one clause: with ($block, $params.mode), with (fragment { css ".panel" }, $store.invoiceStatus). Whole bags ($page, $store, $params) are valid but oversized when a fragment, local, or key suffices.

Never include $secrets in with clauses. Vault credentials must not be sent to the LLM.

Agent AI assistant

Each agent has a built-in conversational assistant at Agents → select agent → Chat.

Use the assistant to:

Describe a workflow in plain language and receive DSL suggestions
Debug failing scripts with awareness of your platform configuration
Iterate on selectors and flow logic without writing every line manually

The assistant runs in a dedicated assistant session mode with access to the agent's browser context. It is separate from task executions — chat does not queue or run tasks directly, but helps you author them.

Tips for the assistant

Reference platform handlers by name so the assistant uses correct in handler { } blocks
Describe the expected page state at each step
Ask for session-aware login patterns when automating authenticated flows

When to use AI helpers vs. explicit logic

Approach	Best for
Explicit `if` / `wait` / `extract` / `extract … format`	Stable UIs, structured text, repeatable flows
`decide`	Ambiguous page states, visual judgment, branching on layout
`infer`	Classification or language tasks that `extract … format` cannot express
AI assistant	Initial script authoring, debugging, exploration

For production-critical automations, prefer explicit conditions where possible and use decide/infer for cases that are hard to express with selectors.