Vom Algorithms

VOM Algorithms

Core Concepts

The Visual Object Model treats a terminal as a 2D grid of styled cells and identifies UI elements through a two-stage pipeline:

ScreenBuffer → Segmentation → Clusters → Classification → Components
   (cells)        (RLE)      (regions)   (heuristics)    (UI elements)

Key data structures (see cli/src/vom/mod.rs):

• ScreenBuffer: 2D grid of Cell (char + style)
• Cluster: Style-homogeneous text region with bounds
• Component: Classified UI element with role and hash

Algorithm Selection Guide

Quick Implementation Patterns

Adding a New Role

1. Add variant to Role enum in cli/src/vom/mod.rs
2. Add detection function in cli/src/vom/classifier.rs
3. Insert in priority order within infer_role()
4. Add tests

// classifier.rs
fn is_progress_bar(text: &str) -> bool {
    let bar_chars = ['█', '▓', '▒', '░', '─', '━'];
    let count = text.chars().filter(|c| bar_chars.contains(c)).count();
    count > text.len() / 2
}

fn infer_role(cluster: &Cluster, cursor_row: u16, cursor_col: u16) -> Role {
    // ... existing checks ...
    if is_progress_bar(&cluster.text) {
        return Role::ProgressBar;
    }
    // ... rest of cascade ...
}

Modifying Segmentation

Read 01-run-length-encoding.md first. Key file: cli/src/vom/segmentation.rs

Current predicate: style equality. To change grouping logic:

fn should_merge(current: &Cluster, cell: &Cell) -> bool {
    current.style == cell.style
    // Add additional conditions here
}

Implementing Element Tracking

Read 07-content-hashing.md and 12-incremental-updates.md.

// Track elements across frames
let prev_hash = component.visual_hash;
// After re-segmentation, find by hash:
let same_element = new_components.iter().find(|c| c.visual_hash == prev_hash);

Code Locations

Complexity Targets

• Segmentation: O(W×H) single pass
• Classification: O(clusters) with O(text_len) per cluster
• Full snapshot: < 5ms for 80×24 terminal

Testing Patterns

#[test]
fn test_new_element_detection() {
    let cluster = make_cluster("█████░░░░░", CellStyle::default(), 0, 0);
    let role = infer_role(&cluster, 99, 99);
    assert_eq!(role, Role::ProgressBar);
}

Always test:

1. Positive detection (element recognized)
2. Negative cases (similar but different elements)
3. Boundary conditions (edge of screen, empty text)
4. Style variations (bold, inverse, colored)