ta-studio-mcp 1.0.1 → 1.0.2

package/README.md

@@ -90,28 +90,32 @@ Ask your AI agent:
 
 ### Coordinate Scaling (Critical Fix)
 
-Mobile MCP screenshots are JPEG-compressed at ~45% of native resolution (486×1080 vs 1080×2400), but element coordinates from `list_elements_on_screen` are in native space. The fix:
+Mobile MCP screenshots are JPEG-compressed at ~45% of native resolution (486×1080 vs 1080×2400), but element coordinates from `list_elements_on_screen` are in native space. The fix involves parsing the native resolution and applying scale factors:
 
-```
-scale_x = img.width / screen_width   → 486/1080 = 0.45
-scale_y = img.height / screen_height  → 1080/2400 = 0.45
-```
+- **Regex**: `re.search(r'(\d+)\s*x\s*(\d+)', get_screen_size())`
+- **Scaling**: `target_x = raw_x * (img_width / screen_width)`
 
-### OAVR Navigation Pattern
+### Set-of-Mark (SoM) Annotation
 
-```
-Observe  → Screen Classifier analyzes current screen
-Act      → Execute action (click, swipe, type) via Mobile MCP
-Verify   → Action Verifier confirms success
-Reason   → Failure Diagnosis suggests recovery if failed
-```
+The server provides methodology for high-performance screenshot tagging:
+- **PIL Threading**: Uses `asyncio.to_thread` for CPU-intensive drawing to keep the event loop responsive.
+- **TOON Format**: Token Optimized Object Notation strips 40% of redundant metadata from screen hierarchies before LLM processing.
+- **Priority Logic**: Class matching for `radiobutton` is prioritized over `button` to prevent classification collisions.
+
+### Flicker Detection (4-Layer)
+
+1. **Trigger**: `adb shell screenrecord`
+2. **Extraction**: `ffmpeg -vf "select='gt(scene,0.003)'"`
+3. **Analysis**: SSIM (Structural Similarity Index) pairs
+4. **LLM**: GPT-5.2 Vision verification
 
 ### Agent Configuration
 
-| Agent | Model | parallel_tool_calls | Reasoning |
-|-------|-------|-------------------|-----------|
-| Coordinator | gpt-5.2 | `true` | high |
-| Device Testing | gpt-5-mini | `false` | medium |
+| Agent | Model | parallel_tool_calls | Reasoning | Why? |
+|-------|-------|-------------------|-----------|------|
+| Coordinator | gpt-5.2 | `true` | high | Orchestration tasks can run in parallel. |
+| Device Testing | gpt-5-mini | `false` | medium | Navigation is sequential; parallel calls cause session race conditions. |
+
 
 ## Tech Stack
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ta-studio-mcp",
-  "version": "1.0.1",
+  "version": "1.0.2",
   "description": "TA Studio MCP — Domain knowledge, patterns, bug fixes, and workflows for AI agents working on the TA Studio mobile test automation platform.",
   "type": "module",
   "bin": {