Merge leading system messages for OpenAI-compatible endpoints

haiku_rag_slim/haiku/rag/utils.py

@@ -7,6 +7,7 @@ from typing import TYPE_CHECKING, Any, cast
 from packaging.version import Version, parse
 
 if TYPE_CHECKING:
+    from pydantic_ai.profiles.openai import OpenAIModelProfile
     from rich.console import RenderableType
 
     from haiku.rag.client import HaikuRAG
@@ -75,6 +76,16 @@ def apply_common_settings(
     return settings_dict
 
 
+# Strict OpenAI-compatible backends (some vLLM chat templates, e.g. Qwen's)
+# reject more than one leading system message. Instructions from multiple
+# sources (agent preamble, capability instructions, dynamic notices) map to
+# one system message each, so have pydantic-ai merge them for any endpoint
+# that is not api.openai.com. Harmless on backends that allow multiples.
+_OPENAI_COMPAT_PROFILE: "OpenAIModelProfile" = {
+    "openai_chat_supports_multiple_system_messages": False
+}
+
+
 def get_model(
     model_config: "ModelConfig",
     app_config: "AppConfig | None" = None,
@@ -125,6 +136,7 @@ def get_model(
             model_name=model,
             provider=OllamaProvider(base_url=base_url),
             settings=model_settings,
+            profile=_OPENAI_COMPAT_PROFILE,
         )
 
     elif provider == "openai":
@@ -154,6 +166,7 @@ def get_model(
                 model_name=model,
                 provider=OpenAIProvider(base_url=model_config.base_url),
                 settings=openai_settings,
+                profile=_OPENAI_COMPAT_PROFILE,
             )
 
         return OpenAIChatModel(model_name=model, settings=openai_settings)

tests/test_utils.py

@@ -229,6 +229,45 @@ def test_get_model_openai_extra_body_forwarded():
     assert result._settings.get("extra_body") == extra
 
 
+@pytest.mark.asyncio
+async def test_get_model_merges_system_messages_for_openai_compatible():
+    """Instruction parts from multiple sources (agent preamble, capability
+    instructions, dynamic notices) each map to their own system message.
+    Strict OpenAI-compatible templates (e.g. Qwen on vLLM) reject more than
+    one leading system message, so OpenAI-compatible endpoints merge them;
+    plain OpenAI keeps them separate."""
+    from pydantic_ai.messages import InstructionPart, ModelRequest, UserPromptPart
+    from pydantic_ai.models import ModelRequestParameters
+
+    parts = [
+        InstructionPart(content="Base instructions."),
+        InstructionPart(content="Limit notice.", dynamic=True),
+    ]
+    messages = [ModelRequest(parts=[UserPromptPart("hi")])]
+
+    async def mapped_for(model):
+        return await model._map_messages(
+            messages, ModelRequestParameters(instruction_parts=parts)
+        )
+
+    vllm = get_model(
+        ModelConfig(provider="openai", name="qwen3.6", base_url="http://vllm:1/v1")
+    )
+    mapped = await mapped_for(vllm)
+    assert [m["role"] for m in mapped] == ["system", "user"]
+    assert mapped[0]["content"] == "Base instructions.\n\nLimit notice."
+
+    ollama = get_model(ModelConfig(provider="ollama", name="llama3"))
+    assert [m["role"] for m in await mapped_for(ollama)] == ["system", "user"]
+
+    openai_native = get_model(ModelConfig(provider="openai", name="gpt-4o"))
+    assert [m["role"] for m in await mapped_for(openai_native)] == [
+        "system",
+        "system",
+        "user",
+    ]
+
+
 def test_get_model_ollama_extra_body_forwarded():
     """`extra_body` is forwarded through the Ollama (openai-compatible) branch too."""
     extra = {"chat_template_kwargs": {"enable_thinking": False}}