All agents now track which variables were explicitly present in their prompt
versus fetched from environment variables or git remote. When constructing
subagent prompts, only explicitly-present variables are included. Fetched
variables are omitted, allowing each subagent to fetch them independently.

This prevents credentials and other fetched values from being garbled as they
propagate through multiple LLM prompt layers.

Affected agents:
- auto-agents (primary orchestrator)
- implementation-supervisor, pr-merge-supervisor, pr-review-supervisor
- supervisor (generic)
- implementation-worker, pr-merge-worker, pr-review-worker
- task-implementor, tier-dispatcher
- work-group-util, git-clone-util, git-push-util, git-checkout-util

@@ -0,0 +21,4 @@
+from cleveragents.core.exceptions import ValidationError
+
+
+def _make_actor_yaml(name: str, provider: str = "openai", model: str = "gpt-4") -> str:

@@ -0,0 +36,4 @@
+    tmp = tempfile.mkdtemp()
+    root = Path(tmp) / "actors"
+    root.mkdir(parents=True)
+

@@ -0,0 +65,4 @@
+
+
+@given("the loader discovers all actors")
+@when("the loader discovers all actors")

@@ -0,0 +96,4 @@
+    assert len(result) == 2
+    names = {a.name for a in result}
+    assert names == {"local/alpha", "local/beta"}
+

@@ -0,0 +27,4 @@
+    When I create an ActorLoader with initial actors from multiple namespaces
+    And another thread concurrently adds and removes YAML files (triggering ``discover()``)
+    And a worker repeatedly calls ``list_actors(namespace=...)`` in parallel
+    Then no exceptions should be raised

@@ -0,0 +1,171 @@
+"""Unit tests for ActorLoader.thread-safety during namespace-filtered list_actors.