Codex Computer Use

Codex Computer Use

Use Codex as a separate local verification agent when the task needs real UI interaction, screenshots, simulator/browser/device state, or an independent runtime check outside the current context. For code reading, typechecking, linting, or tests, run them directly instead — no computer use needed.

Launching apps, simulators, or browsers to verify the requested work is fine without asking; ask first only if the run could disrupt the user's environment beyond that (closing their apps, changing system settings, acting on real accounts or data).

Environment

Capability depends on where this runs — check before dispatching:

• Browser automation (the common case): the codex CLI drives Chrome via CDP (browser_use feature, stable and enabled). Needs Chrome and a display; works on the laptop, not on a headless devbox.
• Full OS control (native apps, simulators, cross-app flows): requires the Codex desktop app on macOS or Windows — macOS needs Screen Recording + Accessibility permissions. Linux is unsupported.
• Headless environment (no display/Chrome): report that computer use is unavailable and fall back to non-GUI verification — curl the endpoints, run the test suite, read logs.

Workflow

1. Confirm the target is reachable: app running, URL or app name known, test credentials if the flow needs them.
2. Create a temporary artifact directory; write the verification prompt to $ARTIFACT_DIR/prompt.md.
3. Run codex exec:

ARTIFACT_DIR="$(mktemp -d "${TMPDIR:-/tmp}/codex-verify.XXXXXX")"

codex exec -m gpt-5.6-sol -c model_reasoning_effort=medium -s workspace-write \
  -C "$ARTIFACT_DIR" --skip-git-repo-check \
  - < "$ARTIFACT_DIR/prompt.md" > "$ARTIFACT_DIR/log.md"

Set the Codex working root to the temporary artifact directory. This keeps workspace-write available for screenshots and the report without making the source repository writable.

Computer-use runs are long: pass an explicit Bash timeout, or run in the background and poll for $ARTIFACT_DIR/report.md.

Verification Prompt

Keep it simple and self-contained:

Verify this flow in <app / URL>:

1. <step>
2. <step>
3. <expected outcome>

Use the browser to walk the flow. Save a screenshot per step to <artifact dir> and write <artifact dir>/report.md with pass/fail per step, each referencing its screenshot.

Report only — do not edit any files outside <artifact dir>. If a step cannot be completed, capture the failing state and say exactly where it stopped. If everything passes, say so explicitly.

Keep "do not edit" — the 5.6 models (sol especially) are eager to start fixing what they find broken.

Reporting Back

Read the report and look at the screenshots yourself before relaying anything; separate confirmed behavior from Codex claims you did not verify, and relay results with their evidence (which steps, which screenshots).