diff --git a/architecture.md b/architecture.md
index a17f590a..a3811e52 100644
--- a/architecture.md
+++ b/architecture.md
@@ -293,7 +293,8 @@ MCP Client → Calls → Open in Report Viewer → ReportViewerGUI.show_report(M
   `event_dict` maps every keyword to a callable, on `execute_action()` reading an object-shaped script's
   actions from the key `webdriver_wrapper`, `auto_control` or `load_density`, and on their own keywords
   starting `WR_`, `AC_` or `LD_` (`framework_profiles.py`). `test_metadata_probe.py` asks the installed
-  packages and reads their executors' source for the key.
+  packages and exercises their public `execute_files()` APIs in isolated child processes to verify
+  the profile's document key, bare action lists, and rejection of unrelated keys.
 - **The automation packages' reports**: the report viewer reads what `je_api_testka`, `je_auto_control`,
   `je_web_runner` and `je_load_density` write (`execution_report/record_reports.py`) and relies on the pair of
   files `<name>_success.json` / `<name>_failure.json` (or `.xml`, under an `<xml_data>` root), on their
diff --git a/architecture_explore.md b/architecture_explore.md
index 9c1ffcb9..8f703bf2 100644
--- a/architecture_explore.md
+++ b/architecture_explore.md
@@ -416,6 +416,8 @@ first_summary → first_code_review → judge_single_review ┐（評分前一
 
 **動作腳本的語言服務** `language_service/`（`docs/adr/0010`）：WebRunner、AutoControl、LoadDensity 的腳本形狀相同（JSON 的動作清單，`["關鍵字"]` 或 `["關鍵字", 引數]`；整份是清單，或放在框架自己的 key 底下），所以一個 adapter 服務三個框架。
 
+`test_metadata_probe.py` 在獨立子程序驗證這三個框架的文件契約：把執行器的指令表換成只記錄引數的測試函式，透過公開的 `execute_files()` 讀取 JSON，檢查 profile 指定的 key、直接使用清單，以及拒絕錯誤 key。物件同時帶有無關 key，驗證執行器確實選取正確清單；不依賴執行器原始碼使用 `.get()` 或共用解析器，也不在測試主程序匯入框架。
+
 | 模組 | 內容 |
 |---|---|
 | `service_adapter.py`(271) | Phase 0 的契約（`docs/adr/0004`）：LSP 形狀的型別、`LanguageServiceAdapter`、只問 adapter 說自己會的事並把它的例外留在裡面的 `LanguageService`、`LanguageServiceRegistry` |
diff --git a/docs/updates/2026-10.md b/docs/updates/2026-10.md
index 7dfe22ba..28d4557c 100644
--- a/docs/updates/2026-10.md
+++ b/docs/updates/2026-10.md
@@ -327,3 +327,12 @@ Validation: 3514 passed, 17 skipped in test/test_utils/; Ruff reports no finding
 - **Completed**: the platform-smoke verification portion of `progress.md` #125. Both Linux and macOS jobs pass on PR #142 at `bb87f28`; Linux supplies Xvfb, xauth and libpulse0. Windows tests remain the full-suite gate.
 - **Files**: `progress.md` and the update-log index.
 - **Open items**: complete Linux/macOS unit suites and release gates remain in #125.
+
+## U-20261012-01 · 2026-10-12 · Verify framework document keys through their public executors · #incident #tests #compatibility
+
+- **Failure**: every Windows CI leg failed because the LoadDensity executor no longer contains the literal `.get("load_density"` after moving dispatch to ActionCore. Its document key remains `load_density`; source spelling was not the contract.
+- **Fix**: replace the source-text assertion with an isolated child-process contract test for each installed framework. Replace its command registry with a recording function and execute a temporary JSON file through public `execute_files()`. Check the profile's key with an unrelated key beside it, a bare action list, and rejection of a document with only the unrelated key. The parent process never imports an automation framework, and no real automation command runs.
+- **Checks**: the complete local unit suite passes (5036 passed, 4 skipped) with the previously installed framework versions, including all 28 metadata-probe tests. The four LoadDensity cases also pass against its current source and ActionCore's shared parser. All 28 tests also pass in a fresh environment with LoadDensity 0.0.79, AutoControl 1.0.4 and WebRunner 0.0.95; `pip check` finds no broken requirements. Pinned Ruff 0.15.22 and `git diff --check` pass.
+- **Files**: `test/test_utils/test_metadata_probe.py`, `architecture.md`, `architecture_explore.md`, and this update-log index.
+- **Analysis follow-up**: split the existing combined built-in assertion into separate assertions after SonarCloud reported it through its GitHub check annotations. The organization plan prevents its API from returning data for a PR targeting `dev`; the GitHub annotations provide the finding without exposing credentials.
+- **Open items**: none for this incident.
diff --git a/docs/updates/README.md b/docs/updates/README.md
index fe4dbdf8..7455ba8b 100644
--- a/docs/updates/README.md
+++ b/docs/updates/README.md
@@ -58,6 +58,7 @@ In the same commit: delete the item from `progress.md`, add a `#done` entry here
 
 | ID | Date | Title | Tags | Batch |
 |---|---|---|---|---|
+| U-20261012-01 | 2026-10-12 | Verify framework document keys through their public executors | #incident #tests #compatibility | [2026-10](2026-10.md) |
 | U-20261008-20 | 2026-10-08 | Linux and macOS platform smoke checks pass | #done #ci #platforms | [2026-10](2026-10.md) |
 | U-20261008-19 | 2026-10-08 | Keep FileAutomation backends installed after its extras split | #done #dependencies | [2026-10](2026-10.md) |
 | U-20261008-18 | 2026-10-08 | Name independent deterministic results and normalize hosts with HTTPS | #done #tests #security | [2026-10](2026-10.md) |
@@ -722,4 +723,4 @@ In the same commit: delete the item from `progress.md`, add a `#done` entry here
 | [2026-09-c.md](2026-09-c.md) | 2026-09 | 115 |
 | [2026-09-d.md](2026-09-d.md) | 2026-09 | 91 |
 | [2026-09-e.md](2026-09-e.md) | 2026-09 | 83 |
-| [2026-10.md](2026-10.md) | 2026-10 | 22 |
+| [2026-10.md](2026-10.md) | 2026-10 | 23 |
diff --git a/pyproject.toml b/pyproject.toml
index d3a118be..75a0cdbe 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -6,7 +6,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "pybreeze"
-version = "1.0.35"
+version = "1.0.36"
 authors = [
     { name = "JE-Chen", email = "jechenmailman@gmail.com" },
 ]
diff --git a/test/test_utils/test_metadata_probe.py b/test/test_utils/test_metadata_probe.py
index 1c99bb1d..9fcbe7af 100644
--- a/test/test_utils/test_metadata_probe.py
+++ b/test/test_utils/test_metadata_probe.py
@@ -9,6 +9,8 @@
 
 import ast
 import importlib.util
+import json
+import subprocess
 import sys
 from pathlib import Path
 
@@ -18,6 +20,35 @@
 from pybreeze.utils.language_service.framework_profiles import PROFILES, FrameworkProfile
 from pybreeze.utils.language_service.keyword_metadata import ParameterKind
 from pybreeze.utils.language_service.metadata_probe import PROBE_SCRIPT, probe_command, read_metadata
+from pybreeze.utils.subprocess_util import utf8_subprocess_env
+
+
+_EXECUTOR_MARKER = "pybreeze-executor-contract:"
+_EXECUTOR_SCRIPT = r'''
+import contextlib
+import importlib
+import json
+import sys
+
+calls = []
+
+
+def record(value):
+    calls.append(value)
+    return value
+
+
+error = None
+with contextlib.redirect_stdout(sys.stderr):
+    try:
+        module = importlib.import_module(sys.argv[1])
+        module.executor.event_dict = {"pybreeze_contract_probe": record}
+        module.execute_files([sys.argv[2]])
+    except Exception as exc:
+        error = type(exc).__name__
+        print(repr(exc), file=sys.stderr)
+print(sys.argv[3] + json.dumps({"calls": calls, "error": error}))
+'''
 
 FAKE = FrameworkProfile(
     framework="fake_runner", label="FakeRunner", document_key="fake_runner", executor_module="fake_runner.executor",
@@ -122,7 +153,8 @@ def test_a_method_is_described_without_its_self(installed):
 def test_pythons_built_ins_are_marked_and_what_is_not_callable_is_left_out(installed):
     metadata = read_metadata(FAKE)
 
-    assert metadata.keyword("print").builtin and metadata.keyword("len").builtin
+    assert metadata.keyword("print").builtin
+    assert metadata.keyword("len").builtin
     assert not metadata.keyword("FR_open").builtin
     assert metadata.keyword("FR_table") is None
 
@@ -240,15 +272,31 @@ def test_each_framework_lists_its_keywords_under_its_prefix(profile):
 
 
 @pytest.mark.parametrize("profile", PROFILES, ids=lambda profile: profile.framework)
-def test_each_frameworks_executor_reads_its_actions_from_the_key_the_profile_names(profile):
-    # Found without importing the package: je_auto_control changes the process as it is imported
-    package = importlib.util.find_spec(profile.framework)
-    if package is None:
+@pytest.mark.parametrize("document_form", ["wrapped", "list", "wrong-key"])
+def test_each_frameworks_executor_reads_its_actions_from_the_key_the_profile_names(
+        profile, document_form, tmp_path):
+    # Import and execute in a child: AutoControl changes the process on import.
+    if importlib.util.find_spec(profile.framework) is None:
         pytest.skip(f"{profile.framework} is not installed here")
-    module_path = Path(package.submodule_search_locations[0]).joinpath(
-        *profile.executor_module.split(".")[1:]).with_suffix(".py")
-
-    source = module_path.read_text(encoding="utf-8")
-
-    assert f'.get("{profile.document_key}"' in source
-    assert "event_dict" in source and "executor = Executor()" in source
+    actions = [["pybreeze_contract_probe", ["expected"]]]
+    document = {"pybreeze_unrelated_actions": [["pybreeze_contract_probe", ["wrong"]]]}
+    if document_form == "wrapped":
+        document[profile.document_key] = actions
+    elif document_form == "list":
+        document = actions
+    action_file = tmp_path / "actions.json"
+    action_file.write_text(json.dumps(document), encoding="utf-8")
+    result = subprocess.run(
+        [sys.executable, "-c", _EXECUTOR_SCRIPT, profile.executor_module,
+         str(action_file), _EXECUTOR_MARKER],
+        cwd=Path(sys.executable).parent, env=utf8_subprocess_env(),
+        capture_output=True, text=True, encoding="utf-8", timeout=60, shell=False,
+        check=False)
+    assert result.returncode == 0, result.stderr
+    answer = json.loads(result.stdout.removeprefix(_EXECUTOR_MARKER))
+    if document_form == "wrong-key":
+        assert answer["error"] is not None, result.stderr
+        assert answer["calls"] == [], result.stderr
+    else:
+        assert answer["error"] is None, result.stderr
+        assert answer["calls"] == ["expected"], result.stderr