Skip to content

Browser subtasks

Core can delegate a bounded browser/UI goal to a compatible native Codex installation while ChatGPT remains the parent controller. This is useful for inspecting an actual rendered page or a permitted web workflow that needs UI evidence. It is not general desktop control or coding-agent delegation.

Prerequisites

Use a configured local machine with command execution enabled, an authorized workspace and a compatible signed-in Codex installation with its enabled native node_repl browser bridge. The adapter discovers that configured bridge and its installed browser Skill; it does not install a browser, library, plugin or login.

The supported authentication setup must include a safe existing file-backed Codex login. A missing/unsafe login, including a keyring-only setup unsupported by this adapter, is a blocker, not permission to copy credentials or change stores. CLI presence and passive discovery metadata do not prove usable browser access.

No browser section is added to Core TOML. The actual tool supports browser mode only and browser selectors auto, iab, chrome or edge, subject to the installed native bridge's capabilities and permissions.

Give a precise goal

Use OpenMachine's browser adapter only to inspect https://example.com in a new
task-owned tab. Allow only the https://example.com origin. Report the visible
page title and heading with native page evidence. Do not sign in, submit forms,
download files, change settings or inspect unrelated tabs. Stop on a permission
prompt or navigation outside the allowed origin.

The assistant supplies start_codex_task with the exact start URL, allowed origins, self-contained goal, local workspace, original execution ID and a bounded timeout. A structured parent task admits the underlying execution. Allowed origins are instructions and observed checks, not a network firewall or isolation from another browser user.

Confirm the actual result

Read the same ID using read_codex_task. Review semantic status, the final native page evidence and any pending action. Generic read_execution exit zero means the wrapper finished, not that the intended browser action succeeded.

blocked and needs_confirmation require their stated resolution. A new authorized follow-up retains the parent task and exact tab context but uses a new operation ID. Do not replay an uncertain form submission or infer a purchase, message or configuration change from a successful controller exit.

Limits and recovery

Timeouts are bounded to 1-1,800 seconds. Cancel through the original execution, then inspect the receipt; cancellation is not rollback of browser effects. The runtime's browser resource lease does not coordinate manual browsing or unrelated Codex sessions. Remote SSH browsers and desktop computer mode are not supplied.

The browser operating guide owns precise bridge, login, tab-evidence and delegation boundaries.