axtree

Documentation

Index

• func BuildIDMap(nodes []*Node) map[int]int64
• func Count(nodes []*Node) int
• func DumpRaw(w io.Writer, nodes []*accessibility.Node)
• func EnrichWithOCR(ctx context.Context, nodes []*Node, engine *ocr.Engine)
• func Extract(ctx context.Context) ([]*accessibility.Node, error)
• func Format(nodes []*Node) string
• func FormatWithOptions(nodes []*Node, opts FormatOptions) string
• type FormatOptions
• type Node
  • func BuildAndFilter(rawNodes []*accessibility.Node) []*Node

Functions

func BuildIDMap

func BuildIDMap(nodes []*Node) map[int]int64

BuildIDMap traverses the tree using the same counter logic as Format(), returning a map from display ID (1, 2, ...) to BackendDOMNodeID. This is essential for translating user-facing element numbers into CDP-level identifiers for actions like click/type.

func Count

func Count(nodes []*Node) int

Count returns the total number of displayable nodes in the tree. Root document nodes are excluded from the count (they are flattened).

func DumpRaw

func DumpRaw(w io.Writer, nodes []*accessibility.Node)

DumpRaw prints the raw AX node list for debugging purposes.

func EnrichWithOCR

func EnrichWithOCR(ctx context.Context, nodes []*Node, engine *ocr.Engine)

EnrichWithOCR traverses the filtered tree and applies OCR to image/interactive elements that have no accessible name. This fills in the gap where AX Tree provides no text info for image buttons/links.

Patterns that trigger OCR (to minimize overhead):

• Any image/img node with no name (images are visual; missing alt = OCR)
• link/button nodes with no name and no text children (likely icon-only)

The OCR engine should be pre-initialized and reused across calls.

func Extract

func Extract(ctx context.Context) ([]*accessibility.Node, error)

Extract fetches the full accessibility tree from the current page via CDP. This is a single CDP call (Accessibility.getFullAXTree) that typically completes in < 50ms, returning a flat list of AX nodes.

func Format

func Format(nodes []*Node) string

Format renders the filtered tree as structured text suitable for LLM consumption. Output format:

1: heading "Page Title"
2: button "Submit" disabled
3: navigation "Main Nav"
  4: link "Home"
  5: link "About"
6: textbox "Search" focused

func FormatWithOptions

func FormatWithOptions(nodes []*Node, opts FormatOptions) string

FormatWithOptions renders the filtered tree with additional formatting controls.

Types

type FormatOptions

type FormatOptions struct {
	InteractiveOnly bool // only show interactive elements
	Compact         bool // compact mode: omit structural wrappers without names
	MaxDepth        int  // 0 = unlimited
	Cursor          bool // annotate the focused element with [cursor]
}

FormatOptions controls the formatting behavior.

type Node

type Node struct {
	Role      string   // semantic role: button, link, heading, textbox, etc.
	Name      string   // accessible name (label)
	Value     string   // current value (e.g. text in an input field)
	States    []string // active states: focused, checked, unchecked, disabled, etc.
	Level     int      // heading level (1-6), 0 if not applicable
	BackendID int64    // backendDOMNodeId — for future CDP click/type operations
	Children  []*Node
}

Node is a simplified, filtered accessibility tree node. It carries only the information needed for LLM consumption.

func BuildAndFilter

func BuildAndFilter(rawNodes []*accessibility.Node) []*Node

BuildAndFilter converts a flat CDP AX node list into a filtered, hierarchical tree. This is the core of Layer 2 — deterministic local processing, zero LLM tokens.