[chore] Fix detach failure of context (#112)

Author: Cirilla-zmh

instrumentation-loongsuite/loongsuite-instrumentation-dashscope/src/opentelemetry/instrumentation/dashscope/patch/image_synthesis.py

@@ -59,13 +59,14 @@ def wrap_image_synthesis_call(wrapped, instance, args, kwargs, handler=None):
         # Create invocation object
         invocation = _create_invocation_from_image_synthesis(kwargs, model)
 
-        # Start LLM invocation (creates span)
-        handler.start_llm(invocation)
-
         # In sync call scenario, set a flag in context to skip span creation in async_call and wait
+        # NOTE: This must be attached BEFORE start_llm to ensure correct LIFO detach order
         ctx = context.set_value(_SKIP_INSTRUMENTATION_KEY, True)
         token = context.attach(ctx)
 
+        # Start LLM invocation (creates span)
+        handler.start_llm(invocation)
+
         try:
             # Execute the wrapped call (internal will call async_call + wait)
             result = wrapped(*args, **kwargs)
@@ -83,7 +84,7 @@ def wrap_image_synthesis_call(wrapped, instance, args, kwargs, handler=None):
             handler.fail_llm(invocation, error)
             raise
         finally:
-            # Restore context
+            # Restore _SKIP_INSTRUMENTATION_KEY context
             if token is not None:
                 context.detach(token)
 

instrumentation-loongsuite/loongsuite-instrumentation-dashscope/src/opentelemetry/instrumentation/dashscope/patch/video_synthesis.py

@@ -59,13 +59,14 @@ def wrap_video_synthesis_call(wrapped, instance, args, kwargs, handler=None):
         # Create invocation object
         invocation = _create_invocation_from_video_synthesis(kwargs, model)
 
-        # Start LLM invocation (creates span)
-        handler.start_llm(invocation)
-
         # Set context flag to skip span creation in async_call and wait
+        # Note: This must be attached BEFORE start_llm to ensure correct LIFO detach order
         ctx = context.set_value(_SKIP_INSTRUMENTATION_KEY, True)
         token = context.attach(ctx)
 
+        # Start LLM invocation (creates span)
+        handler.start_llm(invocation)
+
         try:
             # Execute the wrapped call (internal will call async_call + wait)
             result = wrapped(*args, **kwargs)
@@ -83,7 +84,7 @@ def wrap_video_synthesis_call(wrapped, instance, args, kwargs, handler=None):
             handler.fail_llm(invocation, error)
             raise
         finally:
-            # Restore context
+            # Restore _SKIP_INSTRUMENTATION_KEY context
             if token is not None:
                 context.detach(token)