posthog 6.7.2__py3-none-any.whl → 6.9.0__py3-none-any.whl

posthog/__init__.py

@@ -10,7 +10,11 @@ from posthog.contexts import (
     tag as inner_tag,
     set_context_session as inner_set_context_session,
     identify_context as inner_identify_context,
+    set_capture_exception_code_variables_context as inner_set_capture_exception_code_variables_context,
+    set_code_variables_mask_patterns_context as inner_set_code_variables_mask_patterns_context,
+    set_code_variables_ignore_patterns_context as inner_set_code_variables_ignore_patterns_context,
 )
+from posthog.feature_flags import InconclusiveMatchError, RequiresServerEvaluation
 from posthog.types import FeatureFlag, FlagsAndPayloads, FeatureFlagResult
 from posthog.version import VERSION
 
@@ -19,13 +23,14 @@ __version__ = VERSION
 """Context management."""
 
 
-def new_context(fresh=False, capture_exceptions=True):
+def new_context(fresh=False, capture_exceptions=True, client=None):
     """
     Create a new context scope that will be active for the duration of the with block.
 
     Args:
         fresh: Whether to start with a fresh context (default: False)
         capture_exceptions: Whether to capture exceptions raised within the context (default: True)
+        client: Optional Posthog client instance to use for this context (default: None)
 
     Examples:
         ```python
@@ -38,7 +43,9 @@ def new_context(fresh=False, capture_exceptions=True):
     Category:
         Contexts
     """
-    return inner_new_context(fresh=fresh, capture_exceptions=capture_exceptions)
+    return inner_new_context(
+        fresh=fresh, capture_exceptions=capture_exceptions, client=client
+    )
 
 
 def scoped(fresh=False, capture_exceptions=True):
@@ -102,6 +109,27 @@ def identify_context(distinct_id: str):
     return inner_identify_context(distinct_id)
 
 
+def set_capture_exception_code_variables_context(enabled: bool):
+    """
+    Set whether code variables are captured for the current context.
+    """
+    return inner_set_capture_exception_code_variables_context(enabled)
+
+
+def set_code_variables_mask_patterns_context(mask_patterns: list):
+    """
+    Variable names matching these patterns will be masked with *** when capturing code variables.
+    """
+    return inner_set_code_variables_mask_patterns_context(mask_patterns)
+
+
+def set_code_variables_ignore_patterns_context(ignore_patterns: list):
+    """
+    Variable names matching these patterns will be ignored completely when capturing code variables.
+    """
+    return inner_set_code_variables_ignore_patterns_context(ignore_patterns)
+
+
 def tag(name: str, value: Any):
     """
     Add a tag to the current context.

posthog/ai/anthropic/anthropic.py

@@ -10,7 +10,7 @@ import time
 import uuid
 from typing import Any, Dict, List, Optional
 
-from posthog.ai.types import StreamingContentBlock, ToolInProgress
+from posthog.ai.types import StreamingContentBlock, TokenUsage, ToolInProgress
 from posthog.ai.utils import (
     call_llm_and_track_usage,
     merge_usage_stats,
@@ -126,7 +126,7 @@ class WrappedMessages(Messages):
         **kwargs: Any,
     ):
         start_time = time.time()
-        usage_stats: Dict[str, int] = {"input_tokens": 0, "output_tokens": 0}
+        usage_stats: TokenUsage = TokenUsage(input_tokens=0, output_tokens=0)
         accumulated_content = ""
         content_blocks: List[StreamingContentBlock] = []
         tools_in_progress: Dict[str, ToolInProgress] = {}
@@ -210,14 +210,13 @@ class WrappedMessages(Messages):
         posthog_privacy_mode: bool,
         posthog_groups: Optional[Dict[str, Any]],
         kwargs: Dict[str, Any],
-        usage_stats: Dict[str, int],
+        usage_stats: TokenUsage,
         latency: float,
         content_blocks: List[StreamingContentBlock],
         accumulated_content: str,
     ):
         from posthog.ai.types import StreamingEventData
         from posthog.ai.anthropic.anthropic_converter import (
-            standardize_anthropic_usage,
             format_anthropic_streaming_input,
             format_anthropic_streaming_output_complete,
         )
@@ -236,7 +235,7 @@ class WrappedMessages(Messages):
             formatted_output=format_anthropic_streaming_output_complete(
                 content_blocks, accumulated_content
             ),
-            usage_stats=standardize_anthropic_usage(usage_stats),
+            usage_stats=usage_stats,
             latency=latency,
             distinct_id=posthog_distinct_id,
             trace_id=posthog_trace_id,

posthog/ai/anthropic/anthropic_async.py

@@ -11,17 +11,12 @@ import uuid
 from typing import Any, Dict, List, Optional
 
 from posthog import setup
-from posthog.ai.types import StreamingContentBlock, ToolInProgress
+from posthog.ai.types import StreamingContentBlock, TokenUsage, ToolInProgress
 from posthog.ai.utils import (
     call_llm_and_track_usage_async,
-    extract_available_tool_calls,
-    get_model_params,
-    merge_system_prompt,
     merge_usage_stats,
-    with_privacy_mode,
 )
 from posthog.ai.anthropic.anthropic_converter import (
-    format_anthropic_streaming_content,
     extract_anthropic_usage_from_event,
     handle_anthropic_content_block_start,
     handle_anthropic_text_delta,
@@ -131,7 +126,7 @@ class AsyncWrappedMessages(AsyncMessages):
         **kwargs: Any,
     ):
         start_time = time.time()
-        usage_stats: Dict[str, int] = {"input_tokens": 0, "output_tokens": 0}
+        usage_stats: TokenUsage = TokenUsage(input_tokens=0, output_tokens=0)
         accumulated_content = ""
         content_blocks: List[StreamingContentBlock] = []
         tools_in_progress: Dict[str, ToolInProgress] = {}
@@ -215,71 +210,39 @@ class AsyncWrappedMessages(AsyncMessages):
         posthog_privacy_mode: bool,
         posthog_groups: Optional[Dict[str, Any]],
         kwargs: Dict[str, Any],
-        usage_stats: Dict[str, int],
+        usage_stats: TokenUsage,
         latency: float,
         content_blocks: List[StreamingContentBlock],
         accumulated_content: str,
     ):
-        if posthog_trace_id is None:
-            posthog_trace_id = str(uuid.uuid4())
-
-        # Format output using converter
-        formatted_content = format_anthropic_streaming_content(content_blocks)
-        formatted_output = []
-
-        if formatted_content:
-            formatted_output = [{"role": "assistant", "content": formatted_content}]
-        else:
-            # Fallback to accumulated content if no blocks
-            formatted_output = [
-                {
-                    "role": "assistant",
-                    "content": [{"type": "text", "text": accumulated_content}],
-                }
-            ]
-
-        event_properties = {
-            "$ai_provider": "anthropic",
-            "$ai_model": kwargs.get("model"),
-            "$ai_model_parameters": get_model_params(kwargs),
-            "$ai_input": with_privacy_mode(
-                self._client._ph_client,
-                posthog_privacy_mode,
-                sanitize_anthropic(merge_system_prompt(kwargs, "anthropic")),
-            ),
-            "$ai_output_choices": with_privacy_mode(
-                self._client._ph_client,
-                posthog_privacy_mode,
-                formatted_output,
-            ),
-            "$ai_http_status": 200,
-            "$ai_input_tokens": usage_stats.get("input_tokens", 0),
-            "$ai_output_tokens": usage_stats.get("output_tokens", 0),
-            "$ai_cache_read_input_tokens": usage_stats.get(
-                "cache_read_input_tokens", 0
-            ),
-            "$ai_cache_creation_input_tokens": usage_stats.get(
-                "cache_creation_input_tokens", 0
+        from posthog.ai.types import StreamingEventData
+        from posthog.ai.anthropic.anthropic_converter import (
+            format_anthropic_streaming_input,
+            format_anthropic_streaming_output_complete,
+        )
+        from posthog.ai.utils import capture_streaming_event
+
+        # Prepare standardized event data
+        formatted_input = format_anthropic_streaming_input(kwargs)
+        sanitized_input = sanitize_anthropic(formatted_input)
+
+        event_data = StreamingEventData(
+            provider="anthropic",
+            model=kwargs.get("model", "unknown"),
+            base_url=str(self._client.base_url),
+            kwargs=kwargs,
+            formatted_input=sanitized_input,
+            formatted_output=format_anthropic_streaming_output_complete(
+                content_blocks, accumulated_content
             ),
-            "$ai_latency": latency,
-            "$ai_trace_id": posthog_trace_id,
-            "$ai_base_url": str(self._client.base_url),
-            **(posthog_properties or {}),
-        }
-
-        # Add tools if available
-        available_tools = extract_available_tool_calls("anthropic", kwargs)
-
-        if available_tools:
-            event_properties["$ai_tools"] = available_tools
-
-        if posthog_distinct_id is None:
-            event_properties["$process_person_profile"] = False
-
-        if hasattr(self._client._ph_client, "capture"):
-            self._client._ph_client.capture(
-                distinct_id=posthog_distinct_id or posthog_trace_id,
-                event="$ai_generation",
-                properties=event_properties,
-                groups=posthog_groups,
-            )
+            usage_stats=usage_stats,
+            latency=latency,
+            distinct_id=posthog_distinct_id,
+            trace_id=posthog_trace_id,
+            properties=posthog_properties,
+            privacy_mode=posthog_privacy_mode,
+            groups=posthog_groups,
+        )
+
+        # Use the common capture function
+        capture_streaming_event(self._client._ph_client, event_data)

posthog/ai/anthropic/anthropic_converter.py

@@ -14,7 +14,6 @@ from posthog.ai.types import (
     FormattedMessage,
     FormattedTextContent,
     StreamingContentBlock,
-    StreamingUsageStats,
     TokenUsage,
     ToolInProgress,
 )
@@ -164,7 +163,68 @@ def format_anthropic_streaming_content(
     return formatted
 
 
-def extract_anthropic_usage_from_event(event: Any) -> StreamingUsageStats:
+def extract_anthropic_web_search_count(response: Any) -> int:
+    """
+    Extract web search count from Anthropic response.
+
+    Anthropic provides exact web search counts via usage.server_tool_use.web_search_requests.
+
+    Args:
+        response: The response from Anthropic API
+
+    Returns:
+        Number of web search requests (0 if none)
+    """
+    if not hasattr(response, "usage"):
+        return 0
+
+    if not hasattr(response.usage, "server_tool_use"):
+        return 0
+
+    server_tool_use = response.usage.server_tool_use
+
+    if hasattr(server_tool_use, "web_search_requests"):
+        return max(0, int(getattr(server_tool_use, "web_search_requests", 0)))
+
+    return 0
+
+
+def extract_anthropic_usage_from_response(response: Any) -> TokenUsage:
+    """
+    Extract usage from a full Anthropic response (non-streaming).
+
+    Args:
+        response: The complete response from Anthropic API
+
+    Returns:
+        TokenUsage with standardized usage
+    """
+    if not hasattr(response, "usage"):
+        return TokenUsage(input_tokens=0, output_tokens=0)
+
+    result = TokenUsage(
+        input_tokens=getattr(response.usage, "input_tokens", 0),
+        output_tokens=getattr(response.usage, "output_tokens", 0),
+    )
+
+    if hasattr(response.usage, "cache_read_input_tokens"):
+        cache_read = response.usage.cache_read_input_tokens
+        if cache_read and cache_read > 0:
+            result["cache_read_input_tokens"] = cache_read
+
+    if hasattr(response.usage, "cache_creation_input_tokens"):
+        cache_creation = response.usage.cache_creation_input_tokens
+        if cache_creation and cache_creation > 0:
+            result["cache_creation_input_tokens"] = cache_creation
+
+    web_search_count = extract_anthropic_web_search_count(response)
+    if web_search_count > 0:
+        result["web_search_count"] = web_search_count
+
+    return result
+
+
+def extract_anthropic_usage_from_event(event: Any) -> TokenUsage:
     """
     Extract usage statistics from an Anthropic streaming event.
 
@@ -175,7 +235,7 @@ def extract_anthropic_usage_from_event(event: Any) -> StreamingUsageStats:
         Dictionary of usage statistics
     """
 
-    usage: StreamingUsageStats = {}
+    usage: TokenUsage = TokenUsage()
 
     # Handle usage stats from message_start event
     if hasattr(event, "type") and event.type == "message_start":
@@ -192,6 +252,16 @@ def extract_anthropic_usage_from_event(event: Any) -> StreamingUsageStats:
     if hasattr(event, "usage") and event.usage:
         usage["output_tokens"] = getattr(event.usage, "output_tokens", 0)
 
+        # Extract web search count from usage
+        if hasattr(event.usage, "server_tool_use"):
+            server_tool_use = event.usage.server_tool_use
+            if hasattr(server_tool_use, "web_search_requests"):
+                web_search_count = int(
+                    getattr(server_tool_use, "web_search_requests", 0)
+                )
+                if web_search_count > 0:
+                    usage["web_search_count"] = web_search_count
+
     return usage
 
 
@@ -329,26 +399,6 @@ def finalize_anthropic_tool_input(
             del tools_in_progress[block["id"]]
 
 
-def standardize_anthropic_usage(usage: Dict[str, Any]) -> TokenUsage:
-    """
-    Standardize Anthropic usage statistics to common TokenUsage format.
-
-    Anthropic already uses standard field names, so this mainly structures the data.
-
-    Args:
-        usage: Raw usage statistics from Anthropic
-
-    Returns:
-        Standardized TokenUsage dict
-    """
-    return TokenUsage(
-        input_tokens=usage.get("input_tokens", 0),
-        output_tokens=usage.get("output_tokens", 0),
-        cache_read_input_tokens=usage.get("cache_read_input_tokens"),
-        cache_creation_input_tokens=usage.get("cache_creation_input_tokens"),
-    )
-
-
 def format_anthropic_streaming_input(kwargs: Dict[str, Any]) -> Any:
     """
     Format Anthropic streaming input using system prompt merging.

posthog/ai/gemini/gemini.py

@@ -3,6 +3,9 @@ import time
 import uuid
 from typing import Any, Dict, Optional
 
+from posthog.ai.types import TokenUsage, StreamingEventData
+from posthog.ai.utils import merge_system_prompt
+
 try:
     from google import genai
 except ImportError:
@@ -17,7 +20,6 @@ from posthog.ai.utils import (
     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,
@@ -294,7 +296,7 @@ class Models:
         **kwargs: Any,
     ):
         start_time = time.time()
-        usage_stats: Dict[str, int] = {"input_tokens": 0, "output_tokens": 0}
+        usage_stats: TokenUsage = TokenUsage(input_tokens=0, output_tokens=0)
         accumulated_content = []
 
         kwargs_without_stream = {"model": model, "contents": contents, **kwargs}
@@ -350,15 +352,12 @@ class Models:
         privacy_mode: bool,
         groups: Optional[Dict[str, Any]],
         kwargs: Dict[str, Any],
-        usage_stats: Dict[str, int],
+        usage_stats: TokenUsage,
         latency: float,
         output: Any,
     ):
-        from posthog.ai.types import StreamingEventData
-        from posthog.ai.gemini.gemini_converter import standardize_gemini_usage
-
         # 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(
@@ -368,7 +367,7 @@ class Models:
             kwargs=kwargs,
             formatted_input=sanitized_input,
             formatted_output=format_gemini_streaming_output(output),
-            usage_stats=standardize_gemini_usage(usage_stats),
+            usage_stats=usage_stats,
             latency=latency,
             distinct_id=distinct_id,
             trace_id=trace_id,
@@ -380,10 +379,12 @@ class Models:
         # 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

@@ -10,7 +10,6 @@ from typing import Any, Dict, List, Optional, TypedDict, Union
 from posthog.ai.types import (
     FormattedContentItem,
     FormattedMessage,
-    StreamingUsageStats,
     TokenUsage,
 )
 
@@ -221,6 +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.
@@ -238,6 +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.
@@ -283,7 +338,116 @@ def format_gemini_input(contents: Any) -> List[FormattedMessage]:
     return [_format_object_message(contents)]
 
 
-def extract_gemini_usage_from_chunk(chunk: Any) -> StreamingUsageStats:
+def extract_gemini_web_search_count(response: Any) -> int:
+    """
+    Extract web search count from Gemini response.
+
+    Gemini bills per request that uses grounding, not per query.
+    Returns 1 if grounding_metadata is present with actual search data, 0 otherwise.
+
+    Args:
+        response: The response from Gemini API
+
+    Returns:
+        1 if web search/grounding was used, 0 otherwise
+    """
+
+    # Check for grounding_metadata in candidates
+    if hasattr(response, "candidates"):
+        for candidate in response.candidates:
+            if (
+                hasattr(candidate, "grounding_metadata")
+                and candidate.grounding_metadata
+            ):
+                grounding_metadata = candidate.grounding_metadata
+
+                # Check if web_search_queries exists and is non-empty
+                if hasattr(grounding_metadata, "web_search_queries"):
+                    queries = grounding_metadata.web_search_queries
+
+                    if queries is not None and len(queries) > 0:
+                        return 1
+
+                # Check if grounding_chunks exists and is non-empty
+                if hasattr(grounding_metadata, "grounding_chunks"):
+                    chunks = grounding_metadata.grounding_chunks
+
+                    if chunks is not None and len(chunks) > 0:
+                        return 1
+
+            # Also check for google_search or grounding in function call names
+            if hasattr(candidate, "content") and candidate.content:
+                if hasattr(candidate.content, "parts") and candidate.content.parts:
+                    for part in candidate.content.parts:
+                        if hasattr(part, "function_call") and part.function_call:
+                            function_name = getattr(
+                                part.function_call, "name", ""
+                            ).lower()
+
+                            if (
+                                "google_search" in function_name
+                                or "grounding" in function_name
+                            ):
+                                return 1
+
+    return 0
+
+
+def _extract_usage_from_metadata(metadata: Any) -> TokenUsage:
+    """
+    Common logic to extract usage from Gemini metadata.
+    Used by both streaming and non-streaming paths.
+
+    Args:
+        metadata: usage_metadata from Gemini response or chunk
+
+    Returns:
+        TokenUsage with standardized usage
+    """
+    usage = TokenUsage(
+        input_tokens=getattr(metadata, "prompt_token_count", 0),
+        output_tokens=getattr(metadata, "candidates_token_count", 0),
+    )
+
+    # Add cache tokens if present (don't add if 0)
+    if hasattr(metadata, "cached_content_token_count"):
+        cache_tokens = metadata.cached_content_token_count
+        if cache_tokens and cache_tokens > 0:
+            usage["cache_read_input_tokens"] = cache_tokens
+
+    # Add reasoning tokens if present (don't add if 0)
+    if hasattr(metadata, "thoughts_token_count"):
+        reasoning_tokens = metadata.thoughts_token_count
+        if reasoning_tokens and reasoning_tokens > 0:
+            usage["reasoning_tokens"] = reasoning_tokens
+
+    return usage
+
+
+def extract_gemini_usage_from_response(response: Any) -> TokenUsage:
+    """
+    Extract usage statistics from a full Gemini response (non-streaming).
+
+    Args:
+        response: The complete response from Gemini API
+
+    Returns:
+        TokenUsage with standardized usage statistics
+    """
+    if not hasattr(response, "usage_metadata") or not response.usage_metadata:
+        return TokenUsage(input_tokens=0, output_tokens=0)
+
+    usage = _extract_usage_from_metadata(response.usage_metadata)
+
+    # Add web search count if present
+    web_search_count = extract_gemini_web_search_count(response)
+    if web_search_count > 0:
+        usage["web_search_count"] = web_search_count
+
+    return usage
+
+
+def extract_gemini_usage_from_chunk(chunk: Any) -> TokenUsage:
     """
     Extract usage statistics from a Gemini streaming chunk.
 
@@ -291,21 +455,24 @@ def extract_gemini_usage_from_chunk(chunk: Any) -> StreamingUsageStats:
         chunk: Streaming chunk from Gemini API
 
     Returns:
-        Dictionary of usage statistics
+        TokenUsage with standardized usage statistics
     """
 
-    usage: StreamingUsageStats = {}
+    usage: TokenUsage = TokenUsage()
+
+    # Extract web search count from the chunk before checking for usage_metadata
+    # Web search indicators can appear on any chunk, not just those with usage data
+    web_search_count = extract_gemini_web_search_count(chunk)
+    if web_search_count > 0:
+        usage["web_search_count"] = web_search_count
 
     if not hasattr(chunk, "usage_metadata") or not chunk.usage_metadata:
         return usage
 
-    # Gemini uses prompt_token_count and candidates_token_count
-    usage["input_tokens"] = getattr(chunk.usage_metadata, "prompt_token_count", 0)
-    usage["output_tokens"] = getattr(chunk.usage_metadata, "candidates_token_count", 0)
+    usage_from_metadata = _extract_usage_from_metadata(chunk.usage_metadata)
 
-    # Calculate total if both values are defined (including 0)
-    if "input_tokens" in usage and "output_tokens" in usage:
-        usage["total_tokens"] = usage["input_tokens"] + usage["output_tokens"]
+    # Merge the usage from metadata with any web search count we found
+    usage.update(usage_from_metadata)
 
     return usage
 
@@ -417,22 +584,3 @@ def format_gemini_streaming_output(
 
     # Fallback for empty or unexpected input
     return [{"role": "assistant", "content": [{"type": "text", "text": ""}]}]
-
-
-def standardize_gemini_usage(usage: Dict[str, Any]) -> TokenUsage:
-    """
-    Standardize Gemini usage statistics to common TokenUsage format.
-
-    Gemini already uses standard field names (input_tokens/output_tokens).
-
-    Args:
-        usage: Raw usage statistics from Gemini
-
-    Returns:
-        Standardized TokenUsage dict
-    """
-    return TokenUsage(
-        input_tokens=usage.get("input_tokens", 0),
-        output_tokens=usage.get("output_tokens", 0),
-        # Gemini doesn't currently support cache or reasoning tokens
-    )