browser-dvr-mcp GEMINI.md

Best Browser MCP Developer Guidelines & Agent Rules

To ensure high reliability of autonomous browser automation and avoid regressions, all developers and coding agents modifying the best-browser MCP server must adhere strictly to the following architectural design principles.

---

1. Unified Viewport Coordinate Space Principle

All tools that return visual coordinates, bounding boxes, or click targets must return them normalized to the main viewport pixel space.

• The Rule: No tool should ever return raw layout-relative or iframe-local coordinates. All calculations must resolve the target frame offsets using the helper getFrameOffset(frame) and add them to the coordinates before outputting values to the agent.
• Impact: This ensures that geometry surfaced by perception (e.g. get_semantic_surface) is immediately compatible with interaction tools (e.g. atomic_interact), preventing coordinate shift deadlocks.

---

2. Universal Subframe/Iframe Support

Enterprise web applications and design canvases utilize nested, scaled, same-origin, and cross-origin iframes. All perception and inspection tools must pierce these boundaries.

• The Rule: Any tool querying element properties, structures, styles, or listeners (e.g., evaluate_in_context, browser_get_listeners, get_element_tree) must support frame context switching.
• Protocol Targeting: Use findFrameForBackendNodeId to locate the target subframe CDP session, and pass the target frameId explicitly to CDP calls (such as Accessibility.getFullAXTree) to avoid falling back to the main document context.

---

3. Atomic "Locate-and-Act" Operations

To eliminate Virtual DOM element detachment and stale references, keep interaction operations atomic.

• The Rule: Never cache backend element handles across agent reasoning cycles. Prefer the atomic_interact tool, which locates the node (using backendNodeId) and dispatches native browser events in the exact same execution tick.