workspace/tests/test_delegation_sync_via_polling.py

@@ -30,6 +30,10 @@ import pytest
 os.environ.setdefault("WORKSPACE_ID", "00000000-0000-0000-0000-000000000001")
 os.environ.setdefault("PLATFORM_URL", "http://localhost:8080")
 
+# OFFSEC-003: tool_delegate_task wraps non-error results in boundary markers
+# so the agent can distinguish trusted own output from untrusted peer content.
+from a2a_tools_delegation import _A2A_BOUNDARY_END, _A2A_BOUNDARY_START
+
 
 def _resp(status_code, payload, text=None):
     r = MagicMock()
@@ -88,7 +92,10 @@ class TestFlagOffLegacyPath:
                 "ws-target", "task body", source_workspace_id="ws-self"
             )
 
-        assert result == "legacy ok", f"expected legacy passthrough, got {result!r}"
+        # OFFSEC-003: boundary wrapping is applied by tool_delegate_task even on
+        # the legacy send_a2a_message path (sanitize then wrap at line 333-334).
+        assert _A2A_BOUNDARY_START in result and "legacy ok" in result, \
+            f"expected boundary-wrapped legacy ok, got {result!r}"
         assert send_calls == [("ws-target", "task body", "ws-self")]
         poll_mock.assert_not_called()
 
@@ -153,7 +160,10 @@ class TestPollModeAutoFallback:
         assert poll_calls[0] == ("ws-target", "task body", "ws-self")
         # Caller sees the real reply, NOT the queued sentinel and NOT
         # a DELEGATION FAILED string.
-        assert result == "real response from poll-mode peer"
+        # OFFSEC-003: _delegate_sync_via_polling returns sanitized plain text;
+        # tool_delegate_task wraps it in boundary markers before returning.
+        assert _A2A_BOUNDARY_START in result and "real response from poll-mode peer" in result, \
+            f"expected boundary-wrapped response, got {result!r}"
 
     async def test_non_queued_send_result_does_not_trigger_fallback(self, monkeypatch):
         # Push-mode peer returns a normal text reply — fallback path
@@ -179,7 +189,9 @@ class TestPollModeAutoFallback:
                 "ws-target", "task", source_workspace_id="ws-self"
             )
 
-        assert result == "normal reply"
+        # OFFSEC-003: boundary wrapping applied by tool_delegate_task before return.
+        assert _A2A_BOUNDARY_START in result and "normal reply" in result, \
+            f"expected boundary-wrapped normal reply, got {result!r}"
         poll_mock.assert_not_called()
 
     async def test_error_send_result_does_not_trigger_fallback(self, monkeypatch):