Upload README.md with huggingface_hub

README.md

@@ -59,11 +59,32 @@ Each interactive element is one line:
 
 Fields per element:
 - `[n]` — element index (used in action codes)
-- type: `btn`=button, `ed`=text-input, `img`=image, `chk`=checkbox, `swt`=switch, etc.
-- zone: approximate screen position (tl/tc/tr/ml/mc/mr/bl/bc/br)
-- `fc=focused` — this element has keyboard focus (K action types here)
-- `sr=scrollable` — this element is scrollable
-- label/content text follows
+- type: `bt`=Button, `ed`=EditText, `tx`=TextView, `im`=ImageView, `ck`=CheckBox, `sw`=Switch, `rd`=RadioButton, `sp`=Spinner, `sc`=ScrollView, `ls`=ListView/RecyclerView, `bar`=Toolbar, `tab`=TabLayout, `dw`=DrawerLayout, `vw`=other
+- zone: screen position of element center — row (`t`=top, `m`=mid, `b`=bottom) + col (`l`=left, `c`=center, `r`=right), e.g. `tl`=top-left, `mc`=mid-center
+- `fc` — element has keyboard focus (K action types here)
+- `ed` — element is editable (text input)
+- `sr` — element is scrollable
+- `hd` — element supports long-press
+- `ds` — element is disabled
+- `ck`/`uc` — checkbox checked/unchecked
+- `sl` — element is selected
+- `pw` — password field
+- `...above (N nodes, scroll up)` / `...below (N nodes, scroll down)` — off-screen content indicators
+
+### Tree Compression Rules
+
+The raw Android accessibility tree is compressed before being passed to the model:
+
+1. **Node filtering** — nodes without text, content description, resource ID, clickability, or scrollability are collapsed (their children are promoted up)
+2. **Off-screen filtering** — nodes fully outside the screen bounds (`x2<=0`, `y1>=screen_height`, etc.) are excluded; replaced with `...above (N)` / `...below (N)` scroll indicators
+3. **Sibling deduplication** — identical sibling subtrees are rendered only once (handles repeated list items)
+4. **Multi-window deduplication** — Android can return multiple root windows; duplicate root blocks are dropped. Roots whose tappable element IDs are fully covered by a larger root are also dropped (handles split-screen / overlay artifacts)
+5. **Largest-first ordering** — when multiple roots exist, the most complete window (most lines) is rendered first
+6. **Element indexing** — only `clickable=true AND enabled=true` nodes get a numeric index `[n]`. Non-clickable nodes are rendered without an index. Index order is top-to-bottom, left-to-right by element position
+7. **Type abbreviation** — class names are mapped to short tags (e.g. `android.widget.Button` → `bt`)
+8. **Zone encoding** — element center is bucketed into a 3×3 grid zone string (`tl/tc/tr/ml/mc/mr/bl/bc/br`)
+9. **Label selection** — `text` is preferred; falls back to `content_desc`; falls back to `resource_id` (last component after `/`)
+10. **Hard truncation** — tree is truncated at 4000 characters before tokenization to prevent OOM on dense screens
 
 ---