fix/system prompt sometimes missing (#326)

* fix: always capture system prompt

* chore: bump version

* fix: gemini system prompt capture

* chore: imports at top

* fix: test

The mock we were passing from this test
reporetd that it had a `system instruction` field,
breaking assumptions

* chore: lint

* fix: better code organization

* chore: lint

Author: carlos-marchal-ph

CHANGELOG.md

@@ -1,3 +1,7 @@
+# 6.7.4 - 2025-09-05
+
+- fix: Missing system prompts for some providers
+
 # 6.7.3 - 2025-09-04
 
 - fix: missing usage tokens in Gemini

posthog/ai/gemini/gemini.py

@@ -3,7 +3,8 @@
 import uuid
 from typing import Any, Dict, Optional
 
-from posthog.ai.types import TokenUsage
+from posthog.ai.types import TokenUsage, StreamingEventData
+from posthog.ai.utils import merge_system_prompt
 
 try:
     from google import genai
@@ -19,7 +20,6 @@
     merge_usage_stats,
 )
 from posthog.ai.gemini.gemini_converter import (
-    format_gemini_input,
     extract_gemini_usage_from_chunk,
     extract_gemini_content_from_chunk,
     format_gemini_streaming_output,
@@ -356,10 +356,8 @@ def _capture_streaming_event(
         latency: float,
         output: Any,
     ):
-        from posthog.ai.types import StreamingEventData
-
         # Prepare standardized event data
-        formatted_input = self._format_input(contents)
+        formatted_input = self._format_input(contents, **kwargs)
         sanitized_input = sanitize_gemini(formatted_input)
 
         event_data = StreamingEventData(
@@ -381,10 +379,12 @@ def _capture_streaming_event(
         # Use the common capture function
         capture_streaming_event(self._ph_client, event_data)
 
-    def _format_input(self, contents):
+    def _format_input(self, contents, **kwargs):
         """Format input contents for PostHog tracking"""
 
-        return format_gemini_input(contents)
+        # Create kwargs dict with contents for merge_system_prompt
+        input_kwargs = {"contents": contents, **kwargs}
+        return merge_system_prompt(input_kwargs, "gemini")
 
     def generate_content_stream(
         self,

posthog/ai/gemini/gemini_converter.py

@@ -220,6 +220,30 @@ def format_gemini_response(response: Any) -> List[FormattedMessage]:
     return output
 
 
+def extract_gemini_system_instruction(config: Any) -> Optional[str]:
+    """
+    Extract system instruction from Gemini config parameter.
+
+    Args:
+        config: Config object or dict that may contain system instruction
+
+    Returns:
+        System instruction string if present, None otherwise
+    """
+    if config is None:
+        return None
+
+    # Handle different config formats
+    if hasattr(config, "system_instruction"):
+        return config.system_instruction
+    elif isinstance(config, dict) and "system_instruction" in config:
+        return config["system_instruction"]
+    elif isinstance(config, dict) and "systemInstruction" in config:
+        return config["systemInstruction"]
+
+    return None
+
+
 def extract_gemini_tools(kwargs: Dict[str, Any]) -> Optional[Any]:
     """
     Extract tool definitions from Gemini API kwargs.
@@ -237,6 +261,38 @@ def extract_gemini_tools(kwargs: Dict[str, Any]) -> Optional[Any]:
     return None
 
 
+def format_gemini_input_with_system(
+    contents: Any, config: Any = None
+) -> List[FormattedMessage]:
+    """
+    Format Gemini input contents into standardized message format, including system instruction handling.
+
+    Args:
+        contents: Input contents in various possible formats
+        config: Config object or dict that may contain system instruction
+
+    Returns:
+        List of formatted messages with role and content fields, with system message prepended if needed
+    """
+    formatted_messages = format_gemini_input(contents)
+
+    # Check if system instruction is provided in config parameter
+    system_instruction = extract_gemini_system_instruction(config)
+
+    if system_instruction is not None:
+        has_system = any(msg.get("role") == "system" for msg in formatted_messages)
+        if not has_system:
+            from posthog.ai.types import FormattedMessage
+
+            system_message: FormattedMessage = {
+                "role": "system",
+                "content": system_instruction,
+            }
+            formatted_messages = [system_message] + list(formatted_messages)
+
+    return formatted_messages
+
+
 def format_gemini_input(contents: Any) -> List[FormattedMessage]:
     """
     Format Gemini input contents into standardized message format for PostHog tracking.

posthog/ai/openai/openai_converter.py

@@ -606,7 +606,6 @@ def format_openai_streaming_input(
     Returns:
         Formatted input ready for PostHog tracking
     """
-    if api_type == "chat":
-        return kwargs.get("messages")
-    else:  # responses API
-        return kwargs.get("input")
+    from posthog.ai.utils import merge_system_prompt
+
+    return merge_system_prompt(kwargs, "openai")

posthog/ai/utils.py

@@ -1,9 +1,9 @@
 import time
 import uuid
-from typing import Any, Callable, Dict, Optional
+from typing import Any, Callable, Dict, List, Optional, cast
 
 from posthog.client import Client as PostHogClient
-from posthog.ai.types import StreamingEventData, TokenUsage
+from posthog.ai.types import FormattedMessage, StreamingEventData, TokenUsage
 from posthog.ai.sanitization import (
     sanitize_openai,
     sanitize_anthropic,
@@ -158,7 +158,9 @@ def extract_available_tool_calls(provider: str, kwargs: Dict[str, Any]):
     return None
 
 
-def merge_system_prompt(kwargs: Dict[str, Any], provider: str):
+def merge_system_prompt(
+    kwargs: Dict[str, Any], provider: str
+) -> List[FormattedMessage]:
     """
     Merge system prompts and format messages for the given provider.
     """
@@ -169,10 +171,11 @@ def merge_system_prompt(kwargs: Dict[str, Any], provider: str):
         system = kwargs.get("system")
         return format_anthropic_input(messages, system)
     elif provider == "gemini":
-        from posthog.ai.gemini.gemini_converter import format_gemini_input
+        from posthog.ai.gemini.gemini_converter import format_gemini_input_with_system
 
         contents = kwargs.get("contents", [])
-        return format_gemini_input(contents)
+        config = kwargs.get("config")
+        return format_gemini_input_with_system(contents, config)
     elif provider == "openai":
         from posthog.ai.openai.openai_converter import format_openai_input
 
@@ -187,9 +190,11 @@ def merge_system_prompt(kwargs: Dict[str, Any], provider: str):
         if kwargs.get("system") is not None:
             has_system = any(msg.get("role") == "system" for msg in messages)
             if not has_system:
-                messages = [
-                    {"role": "system", "content": kwargs.get("system")}
-                ] + messages
+                system_msg = cast(
+                    FormattedMessage,
+                    {"role": "system", "content": kwargs.get("system")},
+                )
+                messages = [system_msg] + messages
 
         # For Responses API, add instructions to the system prompt if provided
         if kwargs.get("instructions") is not None:
@@ -207,9 +212,11 @@ def merge_system_prompt(kwargs: Dict[str, Any], provider: str):
                 )
             else:
                 # Create a new system message with instructions
-                messages = [
-                    {"role": "system", "content": kwargs.get("instructions")}
-                ] + messages
+                instruction_msg = cast(
+                    FormattedMessage,
+                    {"role": "system", "content": kwargs.get("instructions")},
+                )
+                messages = [instruction_msg] + messages
 
         return messages
 

posthog/test/ai/gemini/test_gemini.py

@@ -616,6 +616,8 @@ def test_tool_use_response(mock_client, mock_google_genai_client, mock_gemini_re
 
     mock_config = MagicMock()
     mock_config.tools = [mock_tool]
+    # Explicitly specify this config doesn't have system_instruction
+    del mock_config.system_instruction
 
     response = client.models.generate_content(
         model="gemini-2.5-flash",