Codex app in-app browser: provide agent-accessible DevTools console / page-world execution

Open 💬 0 comments Opened May 17, 2026 by Alien-AV

Summary

Please provide an agent-accessible DevTools console or page-world execution surface for the Codex app in-app browser, so agents can inspect and interact with exposed app runtime objects during local UI debugging.

Environment

  • Codex desktop app on Windows
  • In-app browser target: local dev server at http://localhost:8787/
  • Browser Use / in-app browser plugin was connected and could inspect DOM, click buttons, and read captured console logs.

What works today

  • DOM snapshots work.
  • Read-only page evaluation works for DOM reads.
  • Console log reading works via the browser dev log API.
  • Normal button clicks work.

Limitation

The current evaluation surface appears isolated/read-only and is not equivalent to DevTools console/page-world execution:

  • A local app exposed window.game, globalThis.game, window.Game, and globalThis.Game during page bootstrap.
  • The automation-visible DOM marker from the same bootstrap ran successfully.
  • However, agent-side read-only evaluation still saw window.game/globalThis.game as unavailable.
  • The evaluator also could not access some page APIs such as localStorage in the tested environment.
  • javascript: navigation was blocked by browser policy, so there was no supported way to execute in the real page context.

Expected

Codex should offer a documented, safe DevTools/page-world capability for local development targets, such as:

  • evaluate JavaScript in the actual page context with user-approved safety boundaries,
  • inspect page globals exposed by the app for debugging,
  • interact with local app runtime objects intentionally exposed for testing,
  • read/write local storage/session storage when appropriate for local debugging.

Why this matters

For local app development, agents often need to validate state transitions that are not fully visible in the DOM, inspect framework/runtime objects, seed state, or trigger app-level test helpers. Passive DOM reads and console-log capture help, but they are not enough to replace a developer console during debugging.

View original on GitHub ↗