gateway-browser

name	gateway-browser
description	Use when tasks require operating web pages. Run the bundled browser script in this skill so browser actions stay observable and reversible.

Gateway Browser Skill

When the user asks for browser operations, run the bundled browser script in this skill.

Workflow:

1. Read current page state with node ./.codex/skills/gateway-browser/scripts/gateway-browser.mjs snapshot before deciding the next action.
2. Execute one minimal action at a time (click/type/select-option/navigate/wait-for).
3. Re-run snapshot after each key action; refs are not stable across navigations.
4. If needed, use screenshot, tabs, evaluate, or recording commands to collect evidence.
5. Report action, evidence, result, and next step.

Environment:

• GATEWAY_BROWSER_API_BASE
• GATEWAY_INTERNAL_API_TOKEN

Examples:

• node ./.codex/skills/gateway-browser/scripts/gateway-browser.mjs snapshot
• node ./.codex/skills/gateway-browser/scripts/gateway-browser.mjs navigate --url "https://example.com"
• node ./.codex/skills/gateway-browser/scripts/gateway-browser.mjs click --ref e1
• node ./.codex/skills/gateway-browser/scripts/gateway-browser.mjs type --ref e2 --text "hello" --submit true
• node ./.codex/skills/gateway-browser/scripts/gateway-browser.mjs tabs --action list

Report format:

• Action: what changed on the page in this step.
• Evidence: snapshot/screenshot/console/network findings that support the conclusion.
• Result: success, blocked state, or failure reason.
• Next step: the next minimal action or the exact user takeover request.

Status templates:

• In progress: report the latest action, current evidence, and immediate next step.
• Blocked: report the blocker, risk, and the exact decision needed from the user.
• Handoff: report why user takeover is required, what state is preserved, and how to resume.
• Done: report what was completed, the final result/artifact, and any follow-up suggestion.

Stop conditions:

• Stop when the page intent is ambiguous, multiple similar targets exist, or the expected page state did not appear.
• Stop when a modal, redirect, or permission prompt changes the task scope unexpectedly.
• Stop when the action would send external data, upload files, or submit content the user did not explicitly approve.
• When stopping, report the last confirmed page state and the exact decision needed from the user.

Rules:

• Reuse existing tabs whenever possible.
• Prefer visible, reversible actions over hidden shortcuts.
• Do not run Playwright directly or invent another browser wrapper; use only the script bundled in this skill.
• On login/OTP/captcha/payment confirmation, request user takeover.
• Before submit/delete/publish/payment or other irreversible actions, capture evidence and confirm intent if the user did not state it explicitly.
• If the page is visually unclear or the user asks what is on screen, capture a screenshot instead of guessing from stale refs.
• If an action fails twice, stop and ask for user decision.
• If a click/type fails, refresh the snapshot first, then inspect console/network before retrying.
• During user takeover, keep the current tab/state intact and report the exact resume point.
• Treat logged-in browser profiles, cookies, and storage as sensitive data.
• Avoid arbitrary page-context JS unless it is necessary to inspect or unblock the task.