Agents can click.
Can they prove what changed?

Use the browser you already have open—then return evidence an agent can reason about.

chrome-cdp-ex · Codex route

00:07—00:15

One CLI skill

Give Codex eyes and hands.

The same zero-dependency runtime works from a normal Codex skill directory.

$ mkdir -p ~/.codex/skills
$ cp -R skills/chrome-cdp-ex ~/.codex/skills/
$ node scripts/setup.mjs --for codex

✓ CLI route ready
MCP remains optional

Step 1 · Perceive

00:15—00:24

Structured perception

See controls, layout, and health.

doctor

→

perceive

→

act

→

report

$ ./bin/chrome-cdp perceive 35D7E4B7 -C -d 8 --format json

schema: chrome-cdp-ex.perceive.v1
page: chrome-cdp-ex long-session smoke
console.errors: 0
refs: @1 textbox · @2 button · …

Step 2 · Act

00:24—00:33

Mutation with automatic evidence

Act once. Get the receipt automatically.

$ ./bin/chrome-cdp fill 35D7E4B7 '#cmd' 'look trainer' --format json --compact

dispatch.ok: true settlement.state: settled
outcome.status: changed receipt.sequence: 1

Step 3 · Action Receipt

00:33—00:42

Causal evidence

Dispatch is not the same as success.

A receipt separates what was sent, how the page settled, what changed, and what to do next.

schema: chrome-cdp-ex.action-receipt.v1
dispatch: ok · fill
settlement: settled
outcome: changed
observed delta: textbox value updated
next step: report / record-actions

Step 4 · Since action

00:42—00:51

Ask a causal question

What changed since that action?

$ ./bin/chrome-cdp perceive 35D7E4B7 --since-action --format json

schema: chrome-cdp-ex.perceive-diff.v1
mode: since-action
recommendation.commands: report · record-actions

Step 5 · Report and handoff

00:51—01:00

v2.15.0 live-validated with Codex

A complete trail, ready for the next agent.

10/10 mixed runtime campaign rounds passed. 34/34 checks passed in each of five safe real-app fixture rounds.

Claim boundary

Live validation covers the Codex CLI-skill route on one disposable local fixture. Other hosts remain documented—not claimed as live-tested. Benchmark proof is not user-adoption proof.

2actions in report

5/5fixture profiles

0personal tabs used