| name | codex-computer-use |
| description | Codex (GPT-5.x) computer-use verification — drives the browser or desktop, launches apps, captures screenshots. Use when the user asks to test a flow, verify UI behavior, inspect a running app, capture screenshots, or confirm implemented behavior end-to-end. |
Runtime verification via Codex computer use
Codex's computer use is the strongest available and runs on near-free usage — default to it for runtime verification instead of driving the browser yourself. The output you want is evidence: screenshots on disk plus an explicit pass/fail per expectation, not a vibe.
Command mechanics (flags, sandbox, resume) live in the shared wrapper's docs: C:\Users\shawn\.claude\skills\codex\SKILL.md.
Steps
-
State the expectations. List the concrete behaviors to verify (e.g. "clicking Save persists the row after reload") and how to reach the app — URL or launch command, plus any test credentials. Done when each expectation is independently checkable.
-
Run the verification with write access (screenshots need disk) and the maximum Bash timeout (600000ms) — computer-use runs are long:
node C:\Users\shawn\.claude\skills\codex\scripts\codex-run.mjs --write --cd <dir> \
"Launch/open <app>. Verify: <numbered expectations>. Take a screenshot at each step, save them under <dir>/verify/, and report pass/fail per expectation with the screenshot filename as evidence."
Prompt simply; Codex does not need Claude-style scaffolding. Done when the run exits with a per-expectation report.
-
Inspect the evidence. View the screenshots yourself and check they show what the report claims. Done when every expectation has a verdict backed by a screenshot you looked at — anything unverifiable gets reported as unverified, not passed.
-
Relay the pass/fail table and screenshot paths to the user. If a run times out or Codex can't reach the app, report that state and the last evidence captured — don't silently rerun.
