Merge remote-tracking branch 'upstream/main' into pr-tool-output-naming

Author: lionpeloux

pydantic_ai_slim/pydantic_ai/_agent_graph.py

@@ -588,7 +588,11 @@ async def _run_stream() -> AsyncIterator[_messages.HandleResponseEvent]:  # noqa
                     # as the empty response and request will not create any items in the API payload,
                     # in the hope the model will return a non-empty response this time.
                     ctx.state.increment_retries(ctx.deps.max_result_retries, model_settings=ctx.deps.model_settings)
-                    self._next_node = ModelRequestNode[DepsT, NodeRunEndT](_messages.ModelRequest(parts=[]))
+                    run_context = build_run_context(ctx)
+                    instructions = await ctx.deps.get_instructions(run_context)
+                    self._next_node = ModelRequestNode[DepsT, NodeRunEndT](
+                        _messages.ModelRequest(parts=[], instructions=instructions)
+                    )
                     return
 
                 text = ''
@@ -652,7 +656,11 @@ async def _run_stream() -> AsyncIterator[_messages.HandleResponseEvent]:  # noqa
                     ctx.state.increment_retries(
                         ctx.deps.max_result_retries, error=e, model_settings=ctx.deps.model_settings
                     )
-                    self._next_node = ModelRequestNode[DepsT, NodeRunEndT](_messages.ModelRequest(parts=[e.tool_retry]))
+                    run_context = build_run_context(ctx)
+                    instructions = await ctx.deps.get_instructions(run_context)
+                    self._next_node = ModelRequestNode[DepsT, NodeRunEndT](
+                        _messages.ModelRequest(parts=[e.tool_retry], instructions=instructions)
+                    )
 
             self._events_iterator = _run_stream()
 

tests/test_a2a.py

@@ -560,11 +560,11 @@ def track_messages(messages: list[ModelMessage], info: AgentInfo) -> ModelRespon
             task1_id = result1['id']
             context_id = result1['context_id']
 
-            # Wait for first task to complete
-            await anyio.sleep(0.1)
-            task1 = await a2a_client.get_task(task1_id)
-            assert 'result' in task1
-            assert task1['result']['status']['state'] == 'completed'
+            while task1 := await a2a_client.get_task(task1_id):  # pragma: no branch
+                if 'result' in task1 and task1['result']['status']['state'] == 'completed':
+                    result1 = task1['result']
+                    break
+                await anyio.sleep(0.1)
 
             # Verify the model received at least one message
             assert len(messages_received) == 1
@@ -668,11 +668,13 @@ def return_thinking_response(_: list[ModelMessage], info: AgentInfo) -> ModelRes
             task_id = result['id']
 
             # Wait for completion
-            await anyio.sleep(0.1)
-            task = await a2a_client.get_task(task_id)
+            while task := await a2a_client.get_task(task_id):  # pragma: no branch
+                if 'result' in task and task['result']['status']['state'] == 'completed':
+                    result = task['result']
+                    break
+                await anyio.sleep(0.1)
 
-            assert 'result' in task
-            assert task['result'] == snapshot(
+            assert result == snapshot(
                 {
                     'id': IsStr(),
                     'context_id': IsStr(),

tests/test_agent.py

@@ -1872,7 +1872,14 @@ def call_tool(messages: list[ModelMessage], info: AgentInfo) -> ModelResponse:
                         tool_call_id=IsStr(),
                         timestamp=IsDatetime(),
                     )
-                ]
+                ],
+                instructions="""\
+Always respond with a JSON object that's compatible with this schema:
+
+{"additionalProperties": false, "properties": {"city": {"type": "string"}}, "required": ["city"], "type": "object", "title": "get_weather"}
+
+Don't include any text or Markdown fencing before or after.\
+""",
             ),
             ModelResponse(
                 parts=[TextPart(content='{"city": "Mexico City"}')],
@@ -3888,6 +3895,32 @@ def empty_instructions() -> str:
     )
 
 
+def test_multi_agent_instructions_with_structured_output():
+    """Test that Agent2 uses its own instructions when called with Agent1's history.
+
+    Reproduces issue #3207: when running agents sequentially with no user_prompt
+    and structured output, Agent2's instructions were ignored.
+    """
+
+    class Output(BaseModel):
+        text: str
+
+    agent1 = Agent('test', instructions='Agent 1 instructions')
+    agent2 = Agent('test', instructions='Agent 2 instructions', output_type=Output)
+
+    result1 = agent1.run_sync('Hello')
+
+    # TestModel doesn't support structured output, so this will fail with retries
+    # But we can still verify that Agent2's instructions are used in retry requests
+    with capture_run_messages() as messages:
+        with pytest.raises(UnexpectedModelBehavior):
+            agent2.run_sync(message_history=result1.new_messages())
+
+    # Verify Agent2's retry requests used Agent2's instructions (not Agent1's)
+    requests = [m for m in messages if isinstance(m, ModelRequest)]
+    assert any(r.instructions == 'Agent 2 instructions' for r in requests)
+
+
 def test_empty_final_response():
     def llm(messages: list[ModelMessage], info: AgentInfo) -> ModelResponse:
         if len(messages) == 1: