Bellow is the code I generated for a test database. You're right, that it has to be more robust. E.g. if triggerlevel <>1 I need to check, if it is the same parent initiated it.
But in this code I do not understand, why it is telling about Insert trigger violated while it is an Update trigger?

********************************************************************************
procedure __RI_UPDATE_calls
** "Referential integrity update trigger for" calls
LOCAL llRetVal
llRetVal = .t.
PRIVATE pcParentDBF,pnParentRec,pcChildDBF,pnChildRec,pcParentID,pcChildID
PRIVATE pcParentExpr,pcChildExpr
STORE "" TO pcParentDBF,pcChildDBF,pcParentID,pcChildID,pcParentExpr,pcChildExpr
STORE 0 TO pnParentRec,pnChildRec
IF _triggerlevel=1
  BEGIN TRANSACTION
  PRIVATE pcRIcursors,pcRIwkareas,pcRIolderror,pnerror,;
  pcOldDele,pcOldExact,pcOldTalk,pcOldCompat,PcOldDBC
  pcOldTalk=SET("TALK")
  SET TALK OFF
  pcOldDele=SET("DELETED")
  pcOldExact=SET("EXACT")
  pcOldCompat=SET("COMPATIBLE")
  SET COMPATIBLE OFF
  SET DELETED ON
  SET EXACT OFF
  pcRIcursors=""
  pcRIwkareas=""
  pcRIolderror=ON("error")
  pnerror=0
  ON ERROR pnerror=rierror(ERROR(),message(),message(1),program())
  IF TYPE('gaErrors(1)')<>"U"
    release gaErrors
  ENDIF
  PUBLIC gaErrors(1,12)
  pcOldDBC=DBC()
  SET DATA TO ("CONTACTS")
ENDIF first trigger
LOCAL lcParentID && parent's value to be sought in child
LOCAL lcOldParentID && previous parent id value
LOCAL lcChildWkArea && child work area handle returned by riopen
LOCAL lcChildID && child's value to be sought in parent
LOCAL lcOldChildID && old child id value
LOCAL lcParentWkArea && parentwork area handle returned by riopen
LOCAL lcStartArea
lcStartArea=select()
llRetVal=.t.
lcChildWkArea=select()
IF _triggerlevel=1 or type("pccascadeparent")#"C" or (NOT pccascadeparent=="CONTACTS")
  SELECT (lcChildWkArea)
  lcChildID=CONTACT_ID
  lcOldChildID=oldval("CONTACT_ID")
  pcChildDBF=dbf(lcChildWkArea)
  pnChildRec=recno(lcChildWkArea)
  pcChildID=lcOldChildID
  pcChildExpr="CONTACT_ID"
  IF lcChildID<>lcOldChildID
    lcParentWkArea=riopen("contacts","contact_id")
    IF lcParentWkArea<=0
      IF _triggerlevel=1
        DO riend WITH .F.
      ENDIF at the end of the highest trigger level
      SELECT (lcStartArea)
      RETURN .F.
    ENDIF not able to open the child work area
    pcParentDBF=dbf(lcParentWkArea)
    llRetVal=SEEK(lcChildID,lcParentWkArea)
    pnParentRec=recno(lcParentWkArea)
    =rireuse("contacts",lcParentWkArea)
    IF NOT llRetVal
      pnError = rierror(-1,"Insert restrict rule violated.","","")
      IF _triggerlevel=1
        DO riend WITH llRetVal
      ENDIF at the end of the highest trigger level
      SELECT (lcStartArea)
      RETURN llRetVal
    ENDIF no parent
  ENDIF this value was changed
ENDIF not part of a cascade from "contacts"
lcParentWkArea=lcChildWkArea
IF _triggerlevel=1
  do riend with llRetVal
ENDIF at the end of the highest trigger level
SELECT (lcStartArea)
RETURN llRetVal
** "End of Referential integrity Update trigger for" calls

>>>>It is actually becoming more and more interesting. I made bunch of other changes to fix bugs in cascading deletes/updates. Now I hope it works correctly...
>>>
>>>Hope for you too ;-)
>>>
>>>You'll have to test all scenario's though to be sure.
>>
>>I tested a couple, but I'm now trying to understand, how it should work.
>>
>>If we have a cascaded update, what should be our logic for our table as a child? So far I put a condition, that we check for our table as a child only in it is a first level of trigger execution, e.g. if our trigger was called as a result of cascading update on a parent, I do not run the second SCAN. However, I'm not 100% certain I'm correct.
>>
>>Can somebody with the good understanding of how triggers should behave explain to me all possible scenarios?
>>
>>Thanks in advance.
>
>The trigger level is imo irrelevant here
>
>if there is a cascaded update, the child update fires
>
>The child would normally check if there's a parent with that key.
>
>Doing this will give an illegal recursion. It's also not necessary to test this, if we trust the code of the parent
>
>There's a simple test: do not check any relations with the table one level above, in case of a cascade
>
>Suppose you keep an array, say CascadeInitiator[]
>Each time a trigger fires
>

>dime __t__CascadeInitiator[_triggerlevel]
>__t__CascadeInitiator[_triggerlevel] = ''
>

>
>if you are doing a cascaded update
>

>__t__CascadeInitiator[_TriggerLevel] = m.thistable
>replace .....
>

>
>All relations with __t__CascadeInitiator[_TriggerLevel - 1] should not be tested (error 1887)
>
>think that's about it
>
>To be sure. Use standard RI, set up a cascade delete and a cascade update. Generate the code and look at it