diff --git a/CHANGELOG.md b/CHANGELOG.md
index cc5c9a0..91db238 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -22,11 +22,23 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/) and this p
   - fixed the stale link to the context object documentation
 - the plugin talks to the deployment through `cmem-client` only - the deprecated
   `cmem-cmempy` dependency and all `setup_cmempy_user_access` calls are gone
+  - **breaking:** `setup_cmempy_user_access` also set `OAUTH_GRANT_TYPE` and
+    `OAUTH_ACCESS_TOKEN` in the environment of the whole process, which task code
+    calling `cmem.cmempy.*` could rely on without setting up access itself. That side
+    effect is gone, so such code now has to authenticate on its own - the documented
+    way is `get_client(context)`
 
 ### Fixed
 
 - the **Install missing dependencies** action reports the installed version of an
   already installed package instead of repeating its name
+- the **Validate execution phase** action no longer fails with an `AttributeError`
+  when the execution code assigns `result = None`, which is the documented way of
+  handing nothing to the next task
+- the **Validate execution phase** action starts from the `data` its initialization
+  code builds, instead of carrying the mutations of earlier action runs
+- the documented behavior of a failed dependency installation: it stops the task
+  before the execution code runs, rather than letting the import fail later
 
 ## [1.2.1] 2025-09-18
 
diff --git a/README.md b/README.md
index 05ab9ee..899b382 100644
--- a/README.md
+++ b/README.md
@@ -6,7 +6,7 @@ Do not use in production!
 
 ```
 eval $(cmemc -c my-cmem config eval)
-task clean check plugin:install
+task clean check install
 ```
 
 😈
diff --git a/cmem_plugin_python/workflow_task.py b/cmem_plugin_python/workflow_task.py
index caa9e38..a9049ba 100644
--- a/cmem_plugin_python/workflow_task.py
+++ b/cmem_plugin_python/workflow_task.py
@@ -249,8 +249,11 @@
 Dependencies are matched by package name only.
 A package which is already installed is left at the version which is there, and a version
 specifier is not understood and leads to a fresh installation attempt on every run.
-A failed installation does not stop the task: execution continues and the code fails later, at
-the import.
+
+An installation which fails stops the task before the execution code runs at all, because the
+deployment rejects the request and the error is not caught.
+A misspelled dependency therefore fails the workflow run rather than the import, and the
+**Install missing dependencies** action reports the error instead of its per package output.
 """
 
 
@@ -376,9 +379,11 @@ def validate_execute_action(self) -> str:
         """Run the execute code and report results."""
         init_scope = self.do_init()
         test_inputs = init_scope.get("test_inputs", [])
-        scope = self.do_execute(test_inputs, None, self.data)
-        if "result" in scope:
-            result: Entities = scope["result"]
+        # the data of the fresh init scope, not self.data: the constructor's dict is mutated by
+        # every run of the execution code, so re-using it accumulates across action calls
+        scope = self.do_execute(test_inputs, None, init_scope.get("data", {}))
+        result: Entities | None = scope.get("result")
+        if result is not None:
             output = "# Schema\n"
             output += "``` python\n"
             output += f"{result.schema!r}\n"
diff --git a/tests/test_workflow_task.py b/tests/test_workflow_task.py
index bbea2e3..4b0b3e9 100644
--- a/tests/test_workflow_task.py
+++ b/tests/test_workflow_task.py
@@ -173,6 +173,36 @@ def test_validate_execute_action() -> None:
     )
 
 
+def test_validate_execute_action_with_none_result() -> None:
+    """Test Validate execute action with an explicit result of None"""
+    assert (
+        "No result provided"
+        in PythonCodeWorkflowPlugin(
+            init_code=PythonCode(""),
+            execute_code=PythonCode("result = None"),
+        ).validate_execute_action()
+    )
+
+
+def test_validate_execute_action_resets_data() -> None:
+    """Test that the action starts from the data of a fresh initialization run"""
+    counting_code = PythonCode(
+        """from cmem_plugin_base.dataintegration import entity
+data["count"] += 1
+result = entity.Entities(
+    entities=[entity.Entity(uri="urn:x-example:run", values=[[str(data["count"])]])],
+    schema=entity.EntitySchema(
+        type_uri="urn:x-example:count", paths=[entity.EntityPath("count")]
+    ),
+)"""
+    )
+    plugin = PythonCodeWorkflowPlugin(
+        init_code=PythonCode('data["count"] = 0'), execute_code=counting_code
+    )
+    for _ in range(3):
+        assert "['1']" in plugin.validate_execute_action()
+
+
 def test_validate_execute_action_fail() -> None:
     """Test Validate execute action fails"""
     with pytest.raises(SyntaxError, match=r"'\[' was never closed"):