Add Set-of-Marks overlay for VLM element grounding

README.md

@@ -13,6 +13,7 @@
 
 ## Table of Contents
 
+- [What's new (2026-06-19) — Set-of-Marks Overlay](#whats-new-2026-06-19--set-of-marks-overlay)
 - [What's new (2026-06-19) — Checkpoint & Resume](#whats-new-2026-06-19--checkpoint--resume)
 - [What's new (2026-06-19) — i18n / l10n Testing](#whats-new-2026-06-19--i18n--l10n-testing)
 - [What's new (2026-06-19) — Data Quality](#whats-new-2026-06-19--data-quality)
@@ -74,6 +75,13 @@
 
 ---
 
+## What's new (2026-06-19) — Set-of-Marks Overlay
+
+The standard VLM-grounding format, full stack. Full reference: [`docs/source/Eng/doc/new_features/v22_features_doc.rst`](docs/source/Eng/doc/new_features/v22_features_doc.rst).
+
+- **Number elements** — `mark_elements` / `render_marks` / `resolve_mark` (pure + Pillow): assign `1..N` to interactable elements (with centre/role/text), draw numbered red boxes on a screenshot, and map a chosen number back to its element — so a VLM picks a *number* instead of guessing pixels (directly strengthens the existing VLM locator).
+- **Mark-then-click loop** — `mark_screen(render_path=...)` / `mark_click(n)` (`AC_mark_screen` / `AC_mark_click`, `ac_*`): number the live a11y tree (+ optional overlay screenshot), feed marks+image to a model, then click mark `n`.
+
 ## What's new (2026-06-19) — Checkpoint & Resume
 
 Durable execution for long flows + a `py.typed` marker, full stack. Full reference: [`docs/source/Eng/doc/new_features/v21_features_doc.rst`](docs/source/Eng/doc/new_features/v21_features_doc.rst).

README/README_zh-CN.md

@@ -12,6 +12,7 @@
 
 ## 目录
 
+- [本次更新 (2026-06-19) — Set-of-Marks 叠图](#本次更新-2026-06-19--set-of-marks-叠图)
 - [本次更新 (2026-06-19) — 检查点与续跑](#本次更新-2026-06-19--检查点与续跑)
 - [本次更新 (2026-06-19) — i18n / l10n 测试](#本次更新-2026-06-19--i18n--l10n-测试)
 - [本次更新 (2026-06-19) — 数据质量](#本次更新-2026-06-19--数据质量)
@@ -73,6 +74,13 @@
 
 ---
 
+## 本次更新 (2026-06-19) — Set-of-Marks 叠图
+
+VLM 定位的标准格式,走完整五层。完整参考:[`docs/source/Zh/doc/new_features/v22_features_doc.rst`](../docs/source/Zh/doc/new_features/v22_features_doc.rst)。
+
+- **元素标号** — `mark_elements` / `render_marks` / `resolve_mark`(纯函数 + Pillow):为可交互元素指派 `1..N`(含中心/role/text),在截图上画编号红框,并把选到的编号对应回元素——让 VLM 挑*编号*而非猜像素(直接强化既有 VLM locator)。
+- **标号后点击循环** — `mark_screen(render_path=...)` / `mark_click(n)`(`AC_mark_screen` / `AC_mark_click`、`ac_*`):为实时 a11y 树标号(+可选叠图截图),把 marks+图像喂给模型,再点击第 `n` 号。
+
 ## 本次更新 (2026-06-19) — 检查点与续跑
 
 长流程的耐久执行 + `py.typed` 标记,走完整五层。完整参考:[`docs/source/Zh/doc/new_features/v21_features_doc.rst`](../docs/source/Zh/doc/new_features/v21_features_doc.rst)。

README/README_zh-TW.md

@@ -12,6 +12,7 @@
 
 ## 目錄
 
+- [本次更新 (2026-06-19) — Set-of-Marks 疊圖](#本次更新-2026-06-19--set-of-marks-疊圖)
 - [本次更新 (2026-06-19) — 檢查點與續跑](#本次更新-2026-06-19--檢查點與續跑)
 - [本次更新 (2026-06-19) — i18n / l10n 測試](#本次更新-2026-06-19--i18n--l10n-測試)
 - [本次更新 (2026-06-19) — 資料品質](#本次更新-2026-06-19--資料品質)
@@ -73,6 +74,13 @@
 
 ---
 
+## 本次更新 (2026-06-19) — Set-of-Marks 疊圖
+
+VLM 定位的標準格式,走完整五層。完整參考:[`docs/source/Zh/doc/new_features/v22_features_doc.rst`](../docs/source/Zh/doc/new_features/v22_features_doc.rst)。
+
+- **元素標號** — `mark_elements` / `render_marks` / `resolve_mark`(純函式 + Pillow):為可互動元素指派 `1..N`(含中心/role/text),在截圖上畫編號紅框,並把選到的編號對應回元素——讓 VLM 挑*編號*而非猜像素(直接強化既有 VLM locator)。
+- **標號後點擊迴圈** — `mark_screen(render_path=...)` / `mark_click(n)`(`AC_mark_screen` / `AC_mark_click`、`ac_*`):為即時 a11y 樹標號(+可選疊圖截圖),把 marks+影像餵給模型,再點擊第 `n` 號。
+
 ## 本次更新 (2026-06-19) — 檢查點與續跑
 
 長流程的耐久執行 + `py.typed` 標記,走完整五層。完整參考:[`docs/source/Zh/doc/new_features/v21_features_doc.rst`](../docs/source/Zh/doc/new_features/v21_features_doc.rst)。

docs/source/Eng/doc/new_features/v22_features_doc.rst

@@ -0,0 +1,51 @@
+==================================================
+New Features (2026-06-19) — Set-of-Marks Overlay
+==================================================
+
+Modern GUI agents ground far more reliably when shown a screenshot with
+**numbered boxes** over the interactable elements plus an ``id -> bbox``
+legend ("Set-of-Marks" prompting): the model picks a *number* instead of
+guessing pixel coordinates. This turns AutoControl's existing element
+sources into that two-stage "mark then pick a number" loop and resolves the
+chosen number back to a click. Pure standard library + Pillow (already a
+dependency); wired through the full stack.
+
+.. contents::
+   :local:
+   :depth: 2
+
+
+Numbering and the legend
+=======================
+
+::
+
+    from je_auto_control import mark_elements, render_marks, resolve_mark
+
+    marks = mark_elements(elements)   # [{id, bbox, center, role, text}, ...]
+    legend = [(m["id"], m["text"]) for m in marks]
+    annotated_png = render_marks(screenshot_png_bytes, marks)
+    chosen = resolve_mark(marks, 3)   # the element the model picked
+
+``mark_elements`` assigns ``1..N`` to every element with a valid bounds and
+records its centre; ``render_marks`` draws numbered red boxes on a PNG;
+``resolve_mark`` maps a number back to its mark. These are pure and
+unit-testable with synthetic elements.
+
+
+Live "mark then click" loop
+==========================
+
+::
+
+    from je_auto_control import mark_screen, mark_click
+
+    result = mark_screen(render_path="marked.png")   # numbers the live a11y tree
+    # ... feed result["marks"] + marked.png to a VLM, get back a number ...
+    mark_click(3)                                     # click mark #3
+
+``mark_screen`` numbers the live accessibility elements (and optionally
+saves a numbered-box overlay screenshot), caching the marks; ``mark_click``
+resolves a number from that cache and clicks the element's centre. Exposed
+as ``AC_mark_screen`` / ``AC_mark_click`` (and ``ac_mark_screen`` /
+``ac_mark_click``).

docs/source/Eng/eng_index.rst

@@ -44,6 +44,7 @@ Comprehensive guides for all AutoControl features.
    doc/new_features/v19_features_doc
    doc/new_features/v20_features_doc
    doc/new_features/v21_features_doc
+   doc/new_features/v22_features_doc
    doc/ocr_backends/ocr_backends_doc
    doc/observability/observability_doc
    doc/operations_layer/operations_layer_doc

docs/source/Zh/doc/new_features/v22_features_doc.rst

@@ -0,0 +1,47 @@
+==========================================
+新功能 (2026-06-19) — Set-of-Marks 疊圖
+==========================================
+
+現代 GUI agent 在看到「畫上**編號方框**的截圖 + ``id -> bbox`` 圖例」時
+定位會可靠得多(Set-of-Marks prompting):模型挑一個*編號*,而不是猜
+像素座標。本功能把 AutoControl 既有的元素來源轉成這種「先標號、再挑號」
+的兩階段流程,並把選到的編號解析回一次點擊。純標準庫 + Pillow(已是相依);
+走完整五層。
+
+.. contents::
+   :local:
+   :depth: 2
+
+
+標號與圖例
+==========
+
+::
+
+    from je_auto_control import mark_elements, render_marks, resolve_mark
+
+    marks = mark_elements(elements)   # [{id, bbox, center, role, text}, ...]
+    legend = [(m["id"], m["text"]) for m in marks]
+    annotated_png = render_marks(screenshot_png_bytes, marks)
+    chosen = resolve_mark(marks, 3)   # 模型挑中的元素
+
+``mark_elements`` 會為每個有有效 bounds 的元素指派 ``1..N`` 並記錄中心點;
+``render_marks`` 在 PNG 上畫出編號紅框;``resolve_mark`` 把編號對應回該
+標記。這些都是純函式,可用合成元素做單元測試。
+
+
+即時「標號後點擊」迴圈
+======================
+
+::
+
+    from je_auto_control import mark_screen, mark_click
+
+    result = mark_screen(render_path="marked.png")   # 為即時 a11y 樹標號
+    # ... 把 result["marks"] + marked.png 餵給 VLM,取回一個編號 ...
+    mark_click(3)                                     # 點擊第 3 號標記
+
+``mark_screen`` 為即時 accessibility 元素標號(並可另存編號方框疊圖截圖),
+並快取這些標記;``mark_click`` 從快取解析編號並點擊該元素中心。對應
+``AC_mark_screen`` / ``AC_mark_click``(以及 ``ac_mark_screen`` /
+``ac_mark_click``)。

docs/source/Zh/zh_index.rst

@@ -44,6 +44,7 @@ AutoControl 所有功能的完整使用指南。
    doc/new_features/v19_features_doc
    doc/new_features/v20_features_doc
    doc/new_features/v21_features_doc
+   doc/new_features/v22_features_doc
    doc/ocr_backends/ocr_backends_doc
    doc/observability/observability_doc
    doc/operations_layer/operations_layer_doc

je_auto_control/__init__.py

@@ -165,6 +165,10 @@
 from je_auto_control.utils.checkpoint import (
     Checkpoint, CheckpointStore, run_resumable,
 )
+# Set-of-Marks overlay (number elements for VLM grounding)
+from je_auto_control.utils.set_of_marks import (
+    mark_click, mark_elements, mark_screen, render_marks, resolve_mark,
+)
 # Background popup/interrupt watchdog (unattended automation)
 from je_auto_control.utils.watchdog import (
     PopupWatchdog, WatchdogRule, default_popup_watchdog,
@@ -588,6 +592,8 @@ def start_autocontrol_gui(*args, **kwargs):
     "check_catalog", "check_overflow", "pseudo_localize",
     "pseudo_localize_catalog",
     "Checkpoint", "CheckpointStore", "run_resumable",
+    "mark_click", "mark_elements", "mark_screen", "render_marks",
+    "resolve_mark",
     # MCP server
     "AuditLogger", "HttpMCPServer", "MCPContent", "MCPPrompt",
     "MCPPromptArgument", "MCPResource", "MCPServer", "MCPTool",

je_auto_control/gui/script_builder/command_schema.py

@@ -577,7 +577,7 @@
         FieldSpec("automation_id", FieldType.STRING, optional=True),
     )
     specs.append(CommandSpec(
         "AC_control_get_value", "Native UI", "Get Control Value",
         fields=fields,
         description="Read a native control's value via the accessibility API.",
     ))
@@ -661,6 +661,24 @@
     _add_data_quality_specs(specs)
     _add_i18n_specs(specs)
     _add_checkpoint_specs(specs)
+    _add_set_of_marks_specs(specs)
+
+
+def _add_set_of_marks_specs(specs: List[CommandSpec]) -> None:
+    specs.append(CommandSpec(
+        "AC_mark_screen", "Native UI", "Set-of-Marks: Number Elements",
+        fields=(
+            FieldSpec("app_name", FieldType.STRING, optional=True),
+            FieldSpec("render_path", FieldType.FILE_PATH, optional=True),
+        ),
+        description="Number live UI elements (id->bbox legend) for VLM "
+                    "grounding; optional numbered-box overlay screenshot.",
+    ))
+    specs.append(CommandSpec(
+        "AC_mark_click", "Native UI", "Set-of-Marks: Click Number",
+        fields=(FieldSpec("mark_id", FieldType.INT),),
+        description="Click the element behind a numbered mark.",
+    ))
 
 
 def _add_checkpoint_specs(specs: List[CommandSpec]) -> None:

je_auto_control/utils/executor/action_executor.py

@@ -2743,6 +2743,19 @@ def _checkpoint_clear(run_id: str, db: str) -> Dict[str, Any]:
     return {"cleared": CheckpointStore(db).clear(run_id)}
 
 
+def _mark_screen(app_name: Optional[str] = None,
+                 render_path: Optional[str] = None) -> Dict[str, Any]:
+    """Adapter: number live UI elements (Set-of-Marks) for VLM grounding."""
+    from je_auto_control.utils.set_of_marks import mark_screen
+    return mark_screen(app_name=app_name, render_path=render_path)
+
+
+def _mark_click(mark_id: int) -> Dict[str, Any]:
+    """Adapter: click the element behind a numbered mark."""
+    from je_auto_control.utils.set_of_marks import mark_click
+    return {"clicked": mark_click(int(mark_id))}
+
+
 class Executor:
     """
     Executor
@@ -2953,6 +2966,8 @@ def __init__(self):
             "AC_run_resumable": _run_resumable,
             "AC_checkpoint_status": _checkpoint_status,
             "AC_checkpoint_clear": _checkpoint_clear,
+            "AC_mark_screen": _mark_screen,
+            "AC_mark_click": _mark_click,
             "AC_a11y_record_start": _a11y_record_start,
             "AC_a11y_record_stop": _a11y_record_stop,
             "AC_a11y_record_events": _a11y_record_events,

je_auto_control/utils/mcp_server/tools/_factories.py

@@ -2303,6 +2303,32 @@ def checkpoint_tools() -> List[MCPTool]:
     ]
 
 
+def set_of_marks_tools() -> List[MCPTool]:
+    return [
+        MCPTool(
+            name="ac_mark_screen",
+            description=("Set-of-Marks: number the live UI elements (a11y "
+                         "tree) and return an id->bbox/center/role/text "
+                         "legend for VLM grounding — the model picks a number "
+                         "instead of pixels. Optionally render a numbered-box "
+                         "overlay screenshot to 'render_path'."),
+            input_schema=schema({"app_name": {"type": "string"},
+                                 "render_path": {"type": "string"}}),
+            handler=h.mark_screen,
+            annotations=SIDE_EFFECT_ONLY,
+        ),
+        MCPTool(
+            name="ac_mark_click",
+            description=("Click the element behind a numbered mark from the "
+                         "last ac_mark_screen. Returns {clicked}."),
+            input_schema=schema({"mark_id": {"type": "integer"}},
+                                required=["mark_id"]),
+            handler=h.mark_click,
+            annotations=SIDE_EFFECT_ONLY,
+        ),
+    ]
+
+
 def unattended_tools() -> List[MCPTool]:
     return [
         MCPTool(
@@ -3357,7 +3383,7 @@ def media_assert_tools() -> List[MCPTool]:
     skill_library_tools, guardrail_tools, a2a_tools, office_tools,
     agent_memory_tools, determinism_tools, observer_tools,
     sbom_tools, sharding_tools, data_quality_tools, i18n_tools,
-    checkpoint_tools,
+    checkpoint_tools, set_of_marks_tools,
     screen_record_tools,
     process_and_shell_tools, remote_desktop_tools, gamepad_tools,
     usb_passthrough_tools, assertion_tools, data_source_tools,

je_auto_control/utils/mcp_server/tools/_handlers.py

@@ -1133,6 +1133,16 @@ def checkpoint_clear(run_id, db):
     return {"cleared": CheckpointStore(db).clear(run_id)}
 
 
+def mark_screen(app_name=None, render_path=None):
+    from je_auto_control.utils.set_of_marks import mark_screen as _ms
+    return _ms(app_name=app_name, render_path=render_path)
+
+
+def mark_click(mark_id):
+    from je_auto_control.utils.set_of_marks import mark_click as _mc
+    return {"clicked": _mc(int(mark_id))}
+
+
 def vlm_locate(description: str,
                screen_region: Optional[List[int]] = None,
                model: Optional[str] = None) -> Optional[List[int]]:

je_auto_control/utils/set_of_marks/__init__.py

@@ -0,0 +1,10 @@
+"""Set-of-Marks overlay — number on-screen elements for VLM grounding."""
+from je_auto_control.utils.set_of_marks.set_of_marks import (
+    last_marks, mark_click, mark_elements, mark_screen, render_marks,
+    resolve_mark,
+)
+
+__all__ = [
+    "last_marks", "mark_click", "mark_elements", "mark_screen",
+    "render_marks", "resolve_mark",
+]