Merge 9221aa8fbb into 2ad56a4d71

src/utils/summarizer.py

@@ -98,12 +98,23 @@ class SummaryType(Enum):
     LONG = "honcho_chat_summary_long"
 
 
+def _custom_instructions_section(custom_instructions: str | None) -> str:
+    if not custom_instructions or not custom_instructions.strip():
+        return ""
+    return c(f"""
+CUSTOM INSTRUCTIONS:
+{custom_instructions.strip()}
+""")
+
+
 def short_summary_prompt(
     formatted_messages: str,
     output_words: int,
     previous_summary_text: str,
+    custom_instructions: str | None = None,
 ) -> str:
     """Generate the short summary prompt."""
+    custom_instructions_section = _custom_instructions_section(custom_instructions)
     return c(f"""
 You are a system that summarizes parts of a conversation to create a concise and accurate summary. Focus on capturing:
 
@@ -118,6 +129,8 @@ Provide a concise, factual summary that captures the essence of the conversation
 
 Return only the summary without any explanation or meta-commentary.
 
+{custom_instructions_section}
+
 <previous_summary>
 {previous_summary_text}
 </previous_summary>
@@ -134,8 +147,10 @@ def long_summary_prompt(
     formatted_messages: str,
     output_words: int,
     previous_summary_text: str,
+    custom_instructions: str | None = None,
 ) -> str:
     """Generate the long summary prompt."""
+    custom_instructions_section = _custom_instructions_section(custom_instructions)
     return c(f"""
 You are a system that creates thorough, comprehensive summaries of conversations. Focus on capturing:
 
@@ -152,6 +167,8 @@ Provide a thorough and detailed summary that captures the essence of the convers
 
 Return only the summary without any explanation or meta-commentary.
 
+{custom_instructions_section}
+
 <previous_summary>
 {previous_summary_text}
 </previous_summary>
@@ -202,6 +219,7 @@ async def create_short_summary(
     input_tokens: int,
     previous_summary: str | None = None,
     *,
+    custom_instructions: str | None = None,
     workspace_name: str | None = None,
 ) -> HonchoLLMCallResponse[str]:
     # input_tokens indicates how many tokens the message list + previous summary take up
@@ -217,7 +235,10 @@ async def create_short_summary(
         previous_summary_text = "There is no previous summary -- the messages are the beginning of the conversation."
 
     prompt = short_summary_prompt(
-        formatted_messages, output_words, previous_summary_text
+        formatted_messages,
+        output_words,
+        previous_summary_text,
+        custom_instructions,
     )
 
     # Mint a root span id.
@@ -243,6 +264,7 @@ async def create_long_summary(
     formatted_messages: str,
     previous_summary: str | None = None,
     *,
+    custom_instructions: str | None = None,
     workspace_name: str | None = None,
 ) -> HonchoLLMCallResponse[str]:
     # the word/token ratio is roughly 4:3 so we multiply by 0.75.
@@ -255,7 +277,10 @@ async def create_long_summary(
         previous_summary_text = "There is no previous summary -- the messages are the beginning of the conversation."
 
     prompt = long_summary_prompt(
-        formatted_messages, output_words, previous_summary_text
+        formatted_messages,
+        output_words,
+        previous_summary_text,
+        custom_instructions,
     )
 
     # Mint a root span id.
@@ -466,6 +491,7 @@ async def _create_and_save_summary(
         last_message_id=last_message_id,
         last_message_content_preview=last_message_content_preview,
         message_count=message_count,
+        custom_instructions=configuration.reasoning.custom_instructions,
         workspace_name=workspace_name,
     )
 
@@ -562,6 +588,7 @@ async def _create_summary(
     last_message_content_preview: str,
     message_count: int,
     *,
+    custom_instructions: str | None = None,
     workspace_name: str | None = None,
 ) -> tuple[Summary, bool, int, int]:
     """
@@ -594,12 +621,14 @@ async def _create_summary(
                 formatted_messages,
                 input_tokens,
                 previous_summary_text,
+                custom_instructions=custom_instructions,
                 workspace_name=workspace_name,
             )
         else:
             response = await create_long_summary(
                 formatted_messages,
                 previous_summary_text,
+                custom_instructions=custom_instructions,
                 workspace_name=workspace_name,
             )
 

tests/conftest.py

@@ -85,6 +85,8 @@ _RUNTIME_MOCK_TEST_BLOCKLIST_PREFIXES = (
     "tests/llm/",
     # LLM transport tests mock providers directly and don't need database/runtime setup.
     "tests/utils/test_length_finish_reason.py",
+    # Summary unit tests patch LLM calls directly and should not require database setup.
+    "tests/utils/test_summarizer.py",
     "tests/utils/test_clients.py",
     # Session-scope SQL shape — asserts on compiled statements, never executes one.
     "tests/crud/test_session_scope_clauses.py",

tests/utils/test_summarizer.py

@@ -18,6 +18,8 @@ from src.utils.summarizer import (
     _create_summary,  # pyright: ignore[reportPrivateUsage]
     create_long_summary,
     create_short_summary,
+    long_summary_prompt,
+    short_summary_prompt,
 )
 
 # Common test arguments for _create_summary
@@ -222,6 +224,68 @@ class TestCreateSummary:
         assert summary["token_count"] == 0
 
 
+class TestSummaryCustomInstructions:
+    def test_short_summary_prompt_includes_custom_instructions(self):
+        prompt = short_summary_prompt(
+            formatted_messages=_FORMATTED_MESSAGES,
+            output_words=50,
+            previous_summary_text="previous context",
+            custom_instructions="Always write the summary in German.",
+        )
+
+        assert "CUSTOM INSTRUCTIONS:" in prompt
+        assert "Always write the summary in German." in prompt
+        assert "previous context" in prompt
+        assert _FORMATTED_MESSAGES in prompt
+
+    def test_long_summary_prompt_omits_blank_custom_instructions(self):
+        prompt = long_summary_prompt(
+            formatted_messages=_FORMATTED_MESSAGES,
+            output_words=250,
+            previous_summary_text="previous context",
+            custom_instructions="   \n\t   ",
+        )
+
+        assert "CUSTOM INSTRUCTIONS:" not in prompt
+        assert "previous context" in prompt
+        assert _FORMATTED_MESSAGES in prompt
+
+    @pytest.mark.asyncio
+    async def test_create_summary_passes_custom_instructions_to_short_summary(self):
+        mock_response = HonchoLLMCallResponse(
+            content="Benutzer bevorzugt kurze Antworten.",
+            input_tokens=100,
+            output_tokens=8,
+            finish_reasons=["STOP"],
+        )
+
+        with patch(
+            "src.utils.summarizer.create_short_summary",
+            new_callable=AsyncMock,
+            return_value=mock_response,
+        ) as mock_short:
+            summary, is_fallback, _, _ = await _create_summary(
+                formatted_messages=_FORMATTED_MESSAGES,
+                previous_summary_text=None,
+                summary_type=SummaryType.SHORT,
+                input_tokens=_INPUT_TOKENS,
+                message_public_id=_MESSAGE_PUBLIC_ID,
+                last_message_id=_LAST_MESSAGE_ID,
+                last_message_content_preview=_LAST_MESSAGE_CONTENT_PREVIEW,
+                message_count=_MESSAGE_COUNT,
+                custom_instructions="Always write summaries in German.",
+            )
+
+        assert is_fallback is False
+        assert "Benutzer" in summary["content"]
+        await_args = mock_short.await_args
+        if await_args is None:
+            raise AssertionError("Expected short summary call")
+        assert await_args.kwargs["custom_instructions"] == (
+            "Always write summaries in German."
+        )
+
+
 @pytest.mark.asyncio
 class TestSummaryCallerMigration:
     async def test_create_short_summary_uses_model_config(self):