name	browser
description	Use when controlling the Modus in-app browser for navigation, page inspection, forms, search, screenshots, or browser-state verification with raw Chrome DevTools Protocol.
allow-implicit-invocation	true
allowed-tools	["browser_tabs","browser_cdp","browser_events","browser_snapshot","browser_screenshot"]

Browser

Use the Modus browser as a real page surface. Keep the tool layer low-level: tabs, raw CDP commands, event drain, accessibility snapshots, and screenshots.

Workflow

Start with browser_tabs using action: "list". Create a tab with action: "new" when no suitable tab exists.
Use browser_cdp with official Chrome DevTools Protocol method and params; do not invent wrapper tools.
After navigation, read page facts with browser_cdp or browser_snapshot before screenshots: title, URL, readyState, visible text, accessibility state, or DOM.
After mutation, verify with a fresh browser_snapshot, browser_events, or page facts. Use browser_screenshot only when pixels or coordinate calibration matter.
If the page state is unclear, inspect with browser_cdp first. Do not guess.

CDP Patterns

Navigate: Page.navigate with { "url": "https://example.com" }.
Read page state: Runtime.evaluate with returnByValue: true and, when useful, awaitPromise: true.
Read non-visual structure: browser_snapshot for the current accessibility tree, roles, names, and actionable refs.
Wait for page state: poll document.readyState, URL, elements, or drained events. Do not use shell sleeps for browser waits.
Batch extraction/filtering in one Runtime.evaluate when the page has lists or search results.
Locate elements semantically: DOM.getDocument, then DOM.querySelector or DOM.querySelectorAll.
Click only after locating or visually confirming the target: Input.dispatchMouseEvent for press and release.
CDP mouse x/y are CSS pixels. If you measure from a screenshot image, use the screenshot-reported CSS size, image size, and devicePixelRatio; convert image pixels to CSS pixels before clicking.
Type text: focus the element, then use Input.insertText; submit with Input.dispatchKeyEvent when needed.

Rules

Do not look for Modus wrapper tools such as browser_click, browser_fill, browser_scroll, or site-specific helpers.
Prefer DOM or Runtime.evaluate for stable targeting. Use coordinates only after a screenshot or element box confirms them.
After coordinate clicks, verify the state changed with at least two fresh signals such as browser_snapshot, DOM/page facts, or browser_events. A successful CDP send ({}) only means the event was dispatched.
A single disappearing selector is not enough to declare success. Confirm the user-visible or task-relevant state changed.
Treat browser_snapshot as the primary text-model view of the page. Treat browser_screenshot as visual verification or coordinate calibration, not the primary way to understand text pages.
CDP events are read through browser_events; event names like Page.loadEventFired are not CDP commands.
When a command fails, drain events or capture a fresh snapshot, then re-locate the target from current state.
Keep a short note in your answer about what page state you verified.

Browser

Use the Modus browser as a real page surface. Keep the tool layer low-level: tabs, raw CDP commands, event drain, accessibility snapshots, and screenshots.

Workflow

Start with browser_tabs using action: "list". Create a tab with action: "new" when no suitable tab exists.

Use browser_cdp with official Chrome DevTools Protocol method and params; do not invent wrapper tools.

After navigation, read page facts with browser_cdp or browser_snapshot before screenshots: title, URL, readyState, visible text, accessibility state, or DOM.

After mutation, verify with a fresh browser_snapshot, browser_events, or page facts. Use browser_screenshot only when pixels or coordinate calibration matter.

If the page state is unclear, inspect with browser_cdp first. Do not guess.

CDP Patterns

Navigate: Page.navigate with { "url": "https://example.com" }.

Read page state: Runtime.evaluate with returnByValue: true and, when useful, awaitPromise: true.

Read non-visual structure: browser_snapshot for the current accessibility tree, roles, names, and actionable refs.

Wait for page state: poll document.readyState, URL, elements, or drained events. Do not use shell sleeps for browser waits.

Batch extraction/filtering in one Runtime.evaluate when the page has lists or search results.

Locate elements semantically: DOM.getDocument, then DOM.querySelector or DOM.querySelectorAll.

Click only after locating or visually confirming the target: Input.dispatchMouseEvent for press and release.

CDP mouse x/y are CSS pixels. If you measure from a screenshot image, use the screenshot-reported CSS size, image size, and devicePixelRatio; convert image pixels to CSS pixels before clicking.

Type text: focus the element, then use Input.insertText; submit with Input.dispatchKeyEvent when needed.

Rules

Do not look for Modus wrapper tools such as browser_click, browser_fill, browser_scroll, or site-specific helpers.

Prefer DOM or Runtime.evaluate for stable targeting. Use coordinates only after a screenshot or element box confirms them.

After coordinate clicks, verify the state changed with at least two fresh signals such as browser_snapshot, DOM/page facts, or browser_events. A successful CDP send ({}) only means the event was dispatched.

A single disappearing selector is not enough to declare success. Confirm the user-visible or task-relevant state changed.

Treat browser_snapshot as the primary text-model view of the page. Treat browser_screenshot as visual verification or coordinate calibration, not the primary way to understand text pages.

CDP events are read through browser_events; event names like Page.loadEventFired are not CDP commands.

When a command fails, drain events or capture a fresh snapshot, then re-locate the target from current state.

Keep a short note in your answer about what page state you verified.