Open-source AI coworkers that each get a computer of their own: a browser, files and tools, with every action decided before it happens and recorded after. Bring any AG-UI agent.
Answer a conversation in which a skill was picked, on Anthropic, from the Python LangGraph Bot (#601)
kevin439ffa169717
diff --git a/CHANGELOG.md b/CHANGELOG.mdindex fc6dd01e..6fcac1fe 100644--- a/CHANGELOG.md+++ b/CHANGELOG.md@@ -8,6 +8,14 @@ Newest first. `Unreleased` is what is on `main` and not yet tagged.## Unreleased+### The Python LangGraph Bot on Anthropic answers after a skill was picked++A skill somebody picks reaches the Bot as a system message just ahead of their message, and it+stays in the conversation. Anthropic takes one system prompt, and its LangChain integration refuses+a system message that comes after a turn of the conversation, so with `BOT_PROVIDER=anthropic`+every run in that conversation failed from then on, before the model was asked. On Anthropic the+Python LangGraph Bot now puts every system message ahead of the conversation, in the order given,+where the integration joins them into its one prompt. Runs on OpenAI and Gemini are unchanged.### The LangGraph Bot answers on Anthropic and GeminiWith `BOT_PROVIDER=anthropic` or `BOT_PROVIDER=google`, the LangGraph Bot answered nothing: every rundiff --git a/agent-langgraph-agui/src/main.py b/agent-langgraph-agui/src/main.pyindex 14cd8bf7..925b8ba9 100644--- a/agent-langgraph-agui/src/main.py+++ b/agent-langgraph-agui/src/main.py@@ -157,9 +157,51 @@ def _model():)+def _system_turns_first(messages):+ """Every system message ahead of the conversation, in the order it was given.++ ANTHROPIC TAKES ONE SYSTEM PROMPT, AT THE TOP. `langchain-anthropic` joins system messages that+ follow one another into it, and refuses one that comes after a conversation turn: "Received+ multiple non-consecutive system messages." A skill somebody picks arrives as a system turn just+ ahead of their message and stays in the thread, so from then on every run in that conversation+ failed on Anthropic before the model was asked.++ Moved rather than merged, because the integration already joins the ones that are adjacent.+ Only for Anthropic: OpenAI takes a system turn anywhere, so a skill stays beside the message it+ was picked for, and `langchain-google-genai` already gathers every system message into Gemini's+ one system instruction by itself.+ """+ from langchain_core.messages import SystemMessage, convert_to_messages++ messages = convert_to_messages(messages)+ return [+ *(message for message in messages if isinstance(message, SystemMessage)),+ *(message for message in messages if not isinstance(message, SystemMessage)),+ ]+++def _wants_system_turns_first(model) -> bool:+ """+ Anthropic is the provider that refuses non-consecutive system turns, and it is asked for by+ class rather than by `BOT_PROVIDER` because `BOT_MODEL` can name a provider on its own.++ The import is here rather than at the top of the file because this module is imported to read+ a deployment's provider configuration in places that install no provider package at all, and a+ missing import there would answer that question with a crash.+ """+ try:+ from langchain_anthropic import ChatAnthropic+ except ImportError:+ return False+ return isinstance(model, ChatAnthropic)++async def answer(state: MessagesState):+ model = _model()messages = model_messages(state["messages"])- return {"messages": [await bind_tools(_model()).ainvoke(messages)]}+ if _wants_system_turns_first(model):+ messages = _system_turns_first(messages)+ return {"messages": [await bind_tools(model).ainvoke(messages)]}builder = StateGraph(MessagesState)diff --git a/agent-langgraph-agui/tests/test_provider_boundaries.py b/agent-langgraph-agui/tests/test_provider_boundaries.pyindex bd3c307a..5be6fcf8 100644--- a/agent-langgraph-agui/tests/test_provider_boundaries.py+++ b/agent-langgraph-agui/tests/test_provider_boundaries.py@@ -173,7 +173,7 @@ def provider_environment(monkeypatch):monkeypatch.delenv(name, raising=False)-async def _run_answer_with_httpx2_capture(monkeypatch, response_json):+async def _run_answer_with_httpx2_capture(monkeypatch, response_json, messages=None):captured = []async def send(self, request, **kwargs):@@ -188,7 +188,7 @@ async def _run_answer_with_httpx2_capture(monkeypatch, response_json):monkeypatch.setattr(httpx2.AsyncClient, "send", send)result = await main.answer(- {"messages": [{"role": "user", "content": "Say hello."}]}+ {"messages": messages or [{"role": "user", "content": "Say hello."}]})return result, captured@@ -344,6 +344,76 @@ async def test_anthropic_selection_reaches_anthropic_boundary_without_openai_key]+# A conversation as the server sends it after somebody picked a skill: the coworker's standing role+# at the head, and the skill's instruction as a system turn just ahead of the message it was picked+# for. The turn stays in the thread's history, so every later run carries it too.+_HISTORY_WITH_A_SKILL = [+ {"role": "system", "content": "You are Ada, Analyst."},+ {"role": "user", "content": "Summarise the Q3 filing."},+ {"role": "assistant", "content": "Revenue rose 4%."},+ {"role": "system", "content": "Answer in bullet points."},+ {"role": "user", "content": "Again, shorter."},+]+++async def test_anthropic_answers_a_conversation_in_which_a_skill_was_picked(+ monkeypatch,+):+ # Anthropic takes one system prompt, and langchain-anthropic refuses a system message that+ # does not follow the others: "Received multiple non-consecutive system messages."+ monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-ant-openbot-ci")+ monkeypatch.setenv("ANTHROPIC_BASE_URL", "http://127.0.0.1:4311")+ monkeypatch.setenv("BOT_PROVIDER", "anthropic")++ result, captured = await _run_answer_with_httpx2_capture(+ monkeypatch,+ {+ "id": "msg-openbot-ci",+ "type": "message",+ "role": "assistant",+ "model": "claude-sonnet-4-5",+ "content": [{"type": "text", "text": "anthropic proof"}],+ "stop_reason": "end_turn",+ "stop_sequence": None,+ "usage": {"input_tokens": 1, "output_tokens": 1},+ },+ _HISTORY_WITH_A_SKILL,+ )++ assert result["messages"][0].content == "anthropic proof"+ assert [block["text"] for block in captured[0]["body"]["system"]] == [+ "You are Ada, Analyst.",+ "Answer in bullet points.",+ ]+ assert captured[0]["body"]["messages"] == [+ {"role": "user", "content": "Summarise the Q3 filing."},+ {"role": "assistant", "content": "Revenue rose 4%."},+ {"role": "user", "content": "Again, shorter."},+ ]+++async def test_openai_keeps_a_skill_turn_beside_the_message_it_was_picked_for(+ monkeypatch,+):+ monkeypatch.setenv("OPENAI_API_KEY", "sk-openbot-ci")++ result, captured = await _run_answer_with_httpx2_capture(+ monkeypatch,+ _openai_response("gpt-4o-mini", "openai proof"),+ _HISTORY_WITH_A_SKILL,+ )++ assert result["messages"][0].content == "openai proof"+ assert [+ (message["role"], message["content"])+ for message in captured[0]["body"]["messages"]+ ] == [+ (message["role"], message["content"]) for message in _HISTORY_WITH_A_SKILL+ ]++@pytest.mark.asyncio@pytest.mark.parametrize(("provider", "model"),