Abstract:

FastRender is an experimental, high-performance browser engine written in Rust, designed to explore parallel AI coding agent architectures. The project implements a multiprocess, sandboxed renderer with goals for spec-compliant HTML/CSS rendering and JavaScript execution via a custom runtime. The repository establishes a rigorous operational framework for AI agents, emphasizing "correct pixels" as the primary product and enforcing strict non-negotiables: no page-specific hacks, mandatory spec-compliance, and no production panics.

Technically, the system is feature-rich, supporting AVIF decoding, hardware-accelerated UI via WGPU/egui, and a comprehensive media stack. Development is partitioned into specialized workstreams (Rendering, Browser App, and JavaScript) to facilitate parallel contributions. A critical component of the project is its resource safety protocol, which mandates hard external limits (Time, Memory, and Scope) for all execution paths to prevent pathological behavior in the engine or build system. The project maintains an extensive benchmarking suite to monitor regressions in layout, paint, and cascading logic.

System Architecture and Operational Protocol Summary

• [README.md] Core Engine Goals: FastRender is a multiprocess browser engine developed in Rust. It prioritizes spec-compliant rendering, sandboxed processes, and cross-platform support (Linux/macOS).
• [AGENTS.md] Agent Workstream Partitioning: Development is divided into parallel workstreams including capability_buildout (CSS/layout), browser_chrome (UI), and js_engine (execution/GC). Each stream has strict ownership boundaries to prevent code-drift.
• [AGENTS.md] Technical Non-Negotiables: The architecture forbids post-layout "pixel nudging" and page-specific hacks. Implementation must follow a staged pipeline: Parse → Style → Box Tree → Layout → Paint.
• [AGENTS.md] Resource Safety & Execution Limits: The "Cardinal Rule" assumes all inputs are hostile. All commands must run under hard OS-level caps using timeout -k (SIGKILL fallback) and memory limits via run_limited.sh. Unscoped cargo commands are strictly forbidden to prevent host resource exhaustion.
• [Cargo.toml] Feature-Gated Capabilities: The engine uses an extensive feature system. vmjs enables the JavaScript/WebIDL runtime, while browser_ui pulls in the WGPU/egui desktop stack. Multimedia support includes H.264 (OpenH264), VP9 (libvpx), and AAC/Opus audio.
• [Cargo.toml] Optimization Profiles: The project defines two release tiers: release (opt-level 1) for fast agent iteration and release-max (LTO + single codegen unit) for final distribution.
• [Assets/Chrome] UI Prototyping: Internal UI (chrome) is built using standard web tech. The chrome:// scheme is served via trusted fetchers, providing a baseline for tab strips, omniboxes, and modal dialogs.
• [Benches/Cascade] Performance Stress Testing: The cascading engine is benchmarked against "Layer-Heavy" and "Deep DOM" scenarios to measure allocator pressure and rule-candidate explosion.
• [Benches/Layout] Parallel Layout Engine: Testing highlights performance gains from LayoutParallelism, comparing serial vs. parallel execution for complex Grid and Flexbox structures.
• [Benches/Paint] Display List Optimization: The paint subsystem includes a DisplayListOptimizer that performs culling, transparent item removal, and fill merging to ensure a "60fps feel" for the browser chrome.
• [Benches/History] Optimized History Searching: The GlobalHistoryStore utilizes optimized ASCII-only case folding and memchr-accelerated scans to maintain responsiveness with 10,000+ entries.
• [AGENTS.md] Regression Philosophy: Unit tests in src/ are required for internal algorithms, while integration tests in tests/ are reserved for public API and data-driven runners (WPT). One-off renders must be added before implementing fixes to prevent regressions.