Browser reference: tools, policy and limits

DataMind Oda drives the Browser tab of a chat's side panel with ten browser tools. This page is the reference for them: what each call does and takes, the references it uses for elements and tabs, the limits it works under, and the policy and site grants behind every call. For the browser itself, see The browser in Oda.

The browser tools

• Every call takes a tab handle. Leave it out and the call uses the tab used most recently in the chat, such as tab-2.
• A browser_tabs call that opens a tab without a url reads Pass url for action "open".
• A browser_navigate call with no address and no action reads Pass url, or action back, forward, or reload.
• A browser_click call with no reference and no coordinates, or with a reference and coordinates together, reads Pass either ref, or both x and y.
• A browser_scroll amount outside its range reads amount must be an integer from 1 to 20.
• The tools run in the DataMind Oda desktop app and the VS Code extension. Elsewhere a call reports The Oda browser is only available in the DataMind Oda desktop app and VS Code extension.

Element and tab references

• A snapshot lists the page's controls with references such as e1 and e2; browser_click and browser_type name one of them.
• References expire when the page navigates. Take a new snapshot when the page may have changed; a reference that no longer fits fails, and Oda is told to take a new snapshot and use its references.
• Elements you pick yourself keep the same shape, p1 and p2, and stay valid until the page navigates. See Picking page elements.
• Tab handles look like tab-1 and stay with their tab for its lifetime.

Typing and sensitive fields

• browser_type types into a field named by reference. It refuses password, payment, bank-account and one-time-code fields in every mode: Oda is told never to type these and to ask you to fill the field in yourself.
• Before typing, the call focuses the field and checks it again. A field that lost focus adds the detail focus moved to another field; an element that takes no text adds the detail that element is not a text field; click it, or find the input inside it with browser_snapshot.
• What Oda reads on a page stays page content: names, text and attributes come from the page, and Oda does not follow instructions that appear inside a page. When a step needs a password, a payment detail or a CAPTCHA, you complete it in the tab, and Oda continues after you confirm.

Addresses

• Only HTTP and HTTPS addresses open. An address without a scheme is treated as https://, and localhost:5173 is read as a host and a port, not as a scheme.
• An address with a username or a password, and anything that is not HTTP or HTTPS, is refused: Pass a complete http:// or https:// address without a username or password.
• A call that reaches an address that cannot be opened is answered with That address cannot be opened in the Oda browser.

Limits

What you see in the chat

• Each call appears as one step card, named after the action: List browser tabs, Open <url> or Close a browser tab; Go to <url> (or Browser back, Browser forward, Browser reload); Read the page; Read the page text; Click <ref> or Click at <x>,<y>; Type into <ref>; Press <key>; Scroll <direction>; Capture the page; Wait for the page or Wait for "<text>".
• Tabs are listed one per line, as tab-2: followed by the page's title, its address in angle brackets, and flags in brackets when they apply, such as [loading] or [loading, user has control]. A tab list opens with Browser tabs in this chat, most recent first:, or reads No browser tabs are open in this chat. when there are none.
• A wait ends with The awaited text is on the page. or The awaited text did not appear before the timeout.
• Text that runs past the size limit is cut, and the result says The page content was cut at the size limit; scroll or narrow the task to read the rest.
• A screenshot is saved through the app's image storage as one JPEG named browser-screenshot.jpg. Where that storage is not available, the call fails with Screenshots need image storage, which this app does not provide. Use browser_snapshot instead.

Policy and site grants

• The company's browser rules are checked before every call to a site and win over any answer you give; a refused site is not retried. The approval flow itself, including which addresses skip the question, is in The browser in Oda.
• Answers are remembered: Allow in this chat lasts until the chat ends or Oda restarts, and Always allow <site> is kept for the workspace folder, so every chat in that workspace reuses it. A chat without a workspace folder keeps an "always" answer only while Oda is open. If the stored answers cannot be read, they grant nothing, and the next answer you give rewrites them.
• The tools come with your account's browser permission: they appear in a chat while the account holds it and are removed when it is withdrawn.

When a call fails

When a browser call fails, Oda receives a short reason and guidance to act on. When the browser has more detail, the detail follows the guidance in parentheses.

Questions, answered

What do the element references in a snapshot mean?

A snapshot lists the page's controls, each with a reference such as e1. browser_click and browser_type name one of these references to act on that element. References expire when the page navigates, so take a new snapshot after the page changes; the elements you pick yourself use the same shape, such as p1, and are explained in Picking page elements.

Why did a click fail with "Pass either ref, or both x and y."?

The click did not carry a usable target. A click takes either an element reference, or both coordinates measured on the latest screenshot. A call with neither, or with a reference and coordinates together, is refused; use one form and run the call again.

Why does a reference fail as "not in the latest snapshot"?

The page changed since the snapshot the reference came from, most often because it navigated. Take a new snapshot and use its references; Oda is told That element reference is not in the latest snapshot. Take a new browser_snapshot and use its references.

What does "focus moved to another field" mean?

While typing, the call focuses the field and checks it again. When focus lands elsewhere, or on a field that takes a password, a payment detail or a one-time code, typing stops and the sensitive-field guidance follows: Oda never types these and asks you to fill the field in yourself. If the reference points at something that takes no text, the detail reads that element is not a text field; click it, or find the input inside it with browser_snapshot. Take a new snapshot, click the field yourself, then retry.

Which keys can Oda press?

browser_press_key presses one key from a fixed set of fourteen: Enter, Tab, Escape, Backspace, Delete, Space, ArrowUp, ArrowDown, ArrowLeft, ArrowRight, Home, End, PageUp and PageDown. A key outside the set, such as F12, is refused.

What are the limits for typing, scrolling and waiting?

• Typed text: up to 10000 characters per call.
• Scroll: 1 to 20 screens per call.
• Wait: 10 seconds by default and up to 30 seconds, and the text it waits for is up to 500 characters.
• A click's coordinates run from 0 to 100000, and a screenshot's width from 320 to 4096 pixels.

Why can't Oda take a screenshot here?

A screenshot is saved through the app's image storage, and where that storage is not available the call fails with Screenshots need image storage, which this app does not provide. Use browser_snapshot instead. Use a page snapshot or the page text instead, and keep screenshots for charts, images or layouts that an outline cannot describe.

Where is an "Always allow" answer remembered?

For the workspace folder: Always allow <site> is kept for that workspace, so every chat that works in the same folder reuses it. Allow in this chat lasts until the chat ends or Oda restarts, and a chat without a workspace folder keeps its always answer only while Oda is open. If the stored answers cannot be read, they grant nothing, and the next answer rewrites them.

Why did the browser tools disappear from a chat?

The tools come with your account's browser permission. While the account holds it, the tools are present in chats; when it is withdrawn, they are removed and browser calls stop. If the tools never appear, ask your administrator to give the account the browser permission (see Manage DataMind Oda for your company).

What does "The page content was cut at the size limit" mean?

A read of the page, a snapshot or the page text, was longer than the size limit, so the result was cut and the rest was not sent. Scroll the page or narrow the task, then read again; the cut affects that result, not the page.