Merge branch 'master' into webb/openai-separate-input-handling

Author: alexander-alderman-webb

.github/workflows/changelog-preview.yml

@@ -11,6 +11,7 @@ on:
 permissions:
   contents: write
   pull-requests: write
+  statuses: write
 
 jobs:
   changelog-preview:

sentry_sdk/_types.py

@@ -359,3 +359,7 @@ class SDKInfo(TypedDict):
     )
 
     HttpStatusCodeRange = Union[int, Container[int]]
+
+    class TextPart(TypedDict):
+        type: Literal["text"]
+        content: str

sentry_sdk/consts.py

@@ -542,6 +542,12 @@ class SPANDATA:
     Example: 2048
     """
 
+    GEN_AI_SYSTEM_INSTRUCTIONS = "gen_ai.system_instructions"
+    """
+    The system instructions passed to the model.
+    Example: [{"type": "text", "text": "You are a helpful assistant."},{"type": "text", "text": "Be concise and clear."}]
+    """
+
     GEN_AI_REQUEST_MESSAGES = "gen_ai.request.messages"
     """
     The messages passed to the model. The "content" can be a string or an array of objects.

sentry_sdk/integrations/anthropic.py

@@ -39,13 +39,14 @@
     from anthropic.resources import AsyncMessages, Messages
 
     if TYPE_CHECKING:
-        from anthropic.types import MessageStreamEvent
+        from anthropic.types import MessageStreamEvent, TextBlockParam
 except ImportError:
     raise DidNotEnable("Anthropic not installed")
 
 if TYPE_CHECKING:
     from typing import Any, AsyncIterator, Iterator, List, Optional, Union
     from sentry_sdk.tracing import Span
+    from sentry_sdk._types import TextPart
 
 
 class AnthropicIntegration(Integration):
@@ -177,44 +178,53 @@ def _transform_anthropic_content_block(
     return result if result is not None else content_block
 
 
+def _transform_system_instructions(
+    system_instructions: "Union[str, Iterable[TextBlockParam]]",
+) -> "list[TextPart]":
+    if isinstance(system_instructions, str):
+        return [
+            {
+                "type": "text",
+                "content": system_instructions,
+            }
+        ]
+
+    return [
+        {
+            "type": "text",
+            "content": instruction["text"],
+        }
+        for instruction in system_instructions
+        if isinstance(instruction, dict) and "text" in instruction
+    ]
+
+
 def _set_input_data(
     span: "Span", kwargs: "dict[str, Any]", integration: "AnthropicIntegration"
 ) -> None:
     """
     Set input data for the span based on the provided keyword arguments for the anthropic message creation.
     """
     set_data_normalized(span, SPANDATA.GEN_AI_OPERATION_NAME, "chat")
-    system_prompt = kwargs.get("system")
+    system_instructions: "Union[str, Iterable[TextBlockParam]]" = kwargs.get("system")  # type: ignore
     messages = kwargs.get("messages")
     if (
         messages is not None
         and len(messages) > 0
         and should_send_default_pii()
         and integration.include_prompts
     ):
-        normalized_messages = []
-        if system_prompt:
-            system_prompt_content: "Optional[Union[str, List[dict[str, Any]]]]" = None
-            if isinstance(system_prompt, str):
-                system_prompt_content = system_prompt
-            elif isinstance(system_prompt, Iterable):
-                system_prompt_content = []
-                for item in system_prompt:
-                    if (
-                        isinstance(item, dict)
-                        and item.get("type") == "text"
-                        and item.get("text")
-                    ):
-                        system_prompt_content.append(item.copy())
-
-            if system_prompt_content:
-                normalized_messages.append(
-                    {
-                        "role": GEN_AI_ALLOWED_MESSAGE_ROLES.SYSTEM,
-                        "content": system_prompt_content,
-                    }
-                )
+        if isinstance(system_instructions, str) or isinstance(
+            system_instructions, Iterable
+        ):
+            set_data_normalized(
+                span,
+                SPANDATA.GEN_AI_SYSTEM_INSTRUCTIONS,
+                _transform_system_instructions(system_instructions),
+                unpack=False,
+            )
 
+        normalized_messages = []
         for message in messages:
             if (
                 message.get("role") == GEN_AI_ALLOWED_MESSAGE_ROLES.USER

sentry_sdk/integrations/google_genai/utils.py

@@ -32,16 +32,20 @@
     event_from_exception,
     safe_serialize,
 )
-from google.genai.types import GenerateContentConfig
+from google.genai.types import GenerateContentConfig, Part, Content
+from itertools import chain
 
 if TYPE_CHECKING:
     from sentry_sdk.tracing import Span
+    from sentry_sdk._types import TextPart
     from google.genai.types import (
         GenerateContentResponse,
         ContentListUnion,
+        ContentUnionDict,
         Tool,
         Model,
         EmbedContentResponse,
+        ContentUnion,
     )
 
 
@@ -720,6 +724,62 @@ def extract_finish_reasons(
     return finish_reasons if finish_reasons else None
 
 
+def _transform_system_instruction_one_level(
+    system_instructions: "Union[ContentUnionDict, ContentUnion]",
+    can_be_content: bool,
+) -> "list[TextPart]":
+    text_parts: "list[TextPart]" = []
+
+    if isinstance(system_instructions, str):
+        return [{"type": "text", "content": system_instructions}]
+
+    if isinstance(system_instructions, Part) and system_instructions.text:
+        return [{"type": "text", "content": system_instructions.text}]
+
+    if can_be_content and isinstance(system_instructions, Content):
+        if isinstance(system_instructions.parts, list):
+            for part in system_instructions.parts:
+                if isinstance(part.text, str):
+                    text_parts.append({"type": "text", "content": part.text})
+        return text_parts
+
+    if isinstance(system_instructions, dict) and system_instructions.get("text"):
+        return [{"type": "text", "content": system_instructions["text"]}]
+
+    elif can_be_content and isinstance(system_instructions, dict):
+        parts = system_instructions.get("parts", [])
+        for part in parts:
+            if isinstance(part, Part) and isinstance(part.text, str):
+                text_parts.append({"type": "text", "content": part.text})
+            elif isinstance(part, dict) and isinstance(part.get("text"), str):
+                text_parts.append({"type": "text", "content": part["text"]})
+        return text_parts
+
+    return text_parts
+
+
+def _transform_system_instructions(
+    system_instructions: "Union[ContentUnionDict, ContentUnion]",
+) -> "list[TextPart]":
+    text_parts: "list[TextPart]" = []
+
+    if isinstance(system_instructions, list):
+        text_parts = list(
+            chain.from_iterable(
+                _transform_system_instruction_one_level(
+                    instructions, can_be_content=False
+                )
+                for instructions in system_instructions
+            )
+        )
+
+        return text_parts
+
+    return _transform_system_instruction_one_level(
+        system_instructions, can_be_content=True
+    )
+
+
 def set_span_data_for_request(
     span: "Span",
     integration: "Any",
@@ -741,27 +801,19 @@ def set_span_data_for_request(
         messages = []
 
         # Add system instruction if present
+        system_instructions = None
         if config and hasattr(config, "system_instruction"):
-            system_instruction = config.system_instruction
-            if system_instruction:
-                system_messages = extract_contents_messages(system_instruction)
-                # System instruction should be a single system message
-                # Extract text from all messages and combine into one system message
-                system_texts = []
-                for msg in system_messages:
-                    content = msg.get("content")
-                    if isinstance(content, list):
-                        # Extract text from content parts
-                        for part in content:
-                            if isinstance(part, dict) and part.get("type") == "text":
-                                system_texts.append(part.get("text", ""))
-                    elif isinstance(content, str):
-                        system_texts.append(content)
-
-                if system_texts:
-                    messages.append(
-                        {"role": "system", "content": " ".join(system_texts)}
-                    )
+            system_instructions = config.system_instruction
+        elif isinstance(config, dict) and "system_instruction" in config:
+            system_instructions = config.get("system_instruction")
+
+        if system_instructions is not None:
+            set_data_normalized(
+                span,
+                SPANDATA.GEN_AI_SYSTEM_INSTRUCTIONS,
+                _transform_system_instructions(system_instructions),
+                unpack=False,
+            )
 
         # Extract messages from contents
         contents_messages = extract_contents_messages(contents)

sentry_sdk/integrations/langchain.py

@@ -36,6 +36,7 @@
     from uuid import UUID
 
     from sentry_sdk.tracing import Span
+    from sentry_sdk._types import TextPart
 
 
 try:
@@ -189,6 +190,40 @@ def _get_current_agent() -> "Optional[str]":
     return None
 
 
+def _get_system_instructions(messages: "List[List[BaseMessage]]") -> "List[str]":
+    system_instructions = []
+
+    for list_ in messages:
+        for message in list_:
+            # type of content: str | list[str | dict] | None
+            if message.type == "system" and isinstance(message.content, str):
+                system_instructions.append(message.content)
+
+            elif message.type == "system" and isinstance(message.content, list):
+                for item in message.content:
+                    if isinstance(item, str):
+                        system_instructions.append(item)
+
+                    elif isinstance(item, dict) and item.get("type") == "text":
+                        instruction = item.get("text")
+                        if isinstance(instruction, str):
+                            system_instructions.append(instruction)
+
+    return system_instructions
+
+
+def _transform_system_instructions(
+    system_instructions: "List[str]",
+) -> "List[TextPart]":
+    return [
+        {
+            "type": "text",
+            "content": instruction,
+        }
+        for instruction in system_instructions
+    ]
+
+
 class LangchainIntegration(Integration):
     identifier = "langchain"
     origin = f"auto.ai.{identifier}"
@@ -430,9 +465,21 @@ def on_chat_model_start(
             _set_tools_on_span(span, all_params.get("tools"))
 
             if should_send_default_pii() and self.include_prompts:
+                system_instructions = _get_system_instructions(messages)
+                if len(system_instructions) > 0:
+                    set_data_normalized(
+                        span,
+                        SPANDATA.GEN_AI_SYSTEM_INSTRUCTIONS,
+                        _transform_system_instructions(system_instructions),
+                        unpack=False,
+                    )
+
                 normalized_messages = []
                 for list_ in messages:
                     for message in list_:
+                        if message.type == "system":
+                            continue
+
                         normalized_messages.append(
                             self._normalize_langchain_message(message)
                         )

sentry_sdk/integrations/openai_agents/spans/invoke_agent.py

@@ -18,6 +18,17 @@
     import agents
     from typing import Any, Optional
 
+    from sentry_sdk._types import TextPart
+
+
+def _transform_system_instruction(system_instructions: "str") -> "list[TextPart]":
+    return [
+        {
+            "type": "text",
+            "content": system_instructions,
+        }
+    ]
+
 
 def invoke_agent_span(
     context: "agents.RunContextWrapper", agent: "agents.Agent", kwargs: "dict[str, Any]"
@@ -35,16 +46,16 @@ def invoke_agent_span(
     if should_send_default_pii():
         messages = []
         if agent.instructions:
-            message = (
+            system_instruction = (
                 agent.instructions
                 if isinstance(agent.instructions, str)
                 else safe_serialize(agent.instructions)
             )
-            messages.append(
-                {
-                    "content": [{"text": message, "type": "text"}],
-                    "role": "system",
-                }
+            set_data_normalized(
+                span,
+                SPANDATA.GEN_AI_SYSTEM_INSTRUCTIONS,
+                _transform_system_instruction(system_instruction),
+                unpack=False,
             )
 
         original_input = kwargs.get("original_input")