shotgun-sh 0.1.0.dev13__py3-none-any.whl → 0.1.0.dev15__py3-none-any.whl

shotgun/agents/history/history_processors.py

@@ -1,213 +1,410 @@
 """History processors for managing conversation history in Shotgun agents."""
 
-from pydantic_ai import RunContext
-from pydantic_ai.direct import model_request
+from typing import TYPE_CHECKING, Any, Protocol
+
 from pydantic_ai.messages import (
-    BuiltinToolCallPart,
-    BuiltinToolReturnPart,
     ModelMessage,
     ModelRequest,
     ModelResponse,
-    ModelResponsePart,
-    RetryPromptPart,
     SystemPromptPart,
     TextPart,
-    ThinkingPart,
-    ToolCallPart,
-    ToolReturnPart,
     UserPromptPart,
 )
 
+from shotgun.agents.config.models import shotgun_model_request
 from shotgun.agents.models import AgentDeps
 from shotgun.logging_config import get_logger
 from shotgun.prompts import PromptLoader
 
+from .constants import SUMMARY_MARKER, TOKEN_LIMIT_RATIO
+from .context_extraction import extract_context_from_messages
+from .history_building import ensure_ends_with_model_request
+from .message_utils import (
+    get_first_user_request,
+    get_system_prompt,
+)
+from .token_estimation import (
+    calculate_max_summarization_tokens as _calculate_max_summarization_tokens,
+)
+from .token_estimation import (
+    estimate_post_summary_tokens,
+    estimate_tokens_from_messages,
+)
+
+if TYPE_CHECKING:
+    pass
+
+
+class ContextProtocol(Protocol):
+    """Protocol defining the interface needed by token_limit_compactor."""
+
+    deps: AgentDeps
+    usage: Any  # Optional usage information
+
+
 logger = get_logger(__name__)
 
 # Global prompt loader instance
 prompt_loader = PromptLoader()
 
 
+def is_summary_part(part: Any) -> bool:
+    """Check if a message part is a compacted summary."""
+    return isinstance(part, TextPart) and part.content.startswith(SUMMARY_MARKER)
+
+
+def find_last_summary_index(messages: list[ModelMessage]) -> int | None:
+    """Find the index of the last summary in the message history.
+
+    Args:
+        messages: List of messages in the conversation history
+    Returns:
+        Index of the last summary message, or None if no summary exists.
+    """
+    for i in range(len(messages) - 1, -1, -1):
+        if isinstance(messages[i], ModelResponse):
+            for part in messages[i].parts:
+                if is_summary_part(part):
+                    return i
+    return None
+
+
+def extract_summary_content(summary_part: Any) -> str:
+    """Extract the summary content without the marker prefix."""
+    if isinstance(summary_part, TextPart):
+        return summary_part.content[len(SUMMARY_MARKER) :].strip()
+    return ""
+
+
+def create_marked_summary_part(summary_response: Any) -> TextPart:
+    """Create a TextPart with the summary marker prefix.
+
+    This consolidates the duplicate summary creation logic.
+    """
+    first_part = summary_response.parts[0]
+    if isinstance(first_part, TextPart):
+        summary_content = f"{SUMMARY_MARKER} {first_part.content}"
+        return TextPart(content=summary_content)
+    else:
+        # Fallback in case the response part is not TextPart
+        summary_content = f"{SUMMARY_MARKER} Summary content unavailable"
+        return TextPart(content=summary_content)
+
+
+def log_summarization_request(
+    model: Any, max_tokens: int, prompt: str, context: str, request_type: str
+) -> None:
+    """Log detailed summarization request information.
+
+    Consolidates duplicate logging patterns across the codebase.
+    """
+    logger.debug(f"{request_type} SUMMARIZATION REQUEST - Model: {model}")
+    logger.debug(f"{request_type} SUMMARIZATION REQUEST - Max tokens: {max_tokens}")
+    logger.debug(f"{request_type} SUMMARIZATION REQUEST - Instructions: {prompt}")
+    logger.debug(f"{request_type} SUMMARIZATION REQUEST - Context: {context}")
+
+
+def log_summarization_response(response: Any, request_type: str) -> None:
+    """Log detailed summarization response information.
+
+    Consolidates duplicate logging patterns across the codebase.
+    """
+    logger.debug(f"{request_type} SUMMARIZATION RESPONSE - Full response: {response}")
+    logger.debug(
+        f"{request_type} SUMMARIZATION RESPONSE - Content: "
+        f"{response.parts[0] if response.parts else 'No content'}"
+    )
+    logger.debug(f"{request_type} SUMMARIZATION RESPONSE - Usage: {response.usage}")
+
+
+# Use centralized calculate_max_summarization_tokens function
+calculate_max_summarization_tokens = _calculate_max_summarization_tokens
+
+
 async def token_limit_compactor(
-    ctx: RunContext[AgentDeps],
+    ctx: ContextProtocol,
     messages: list[ModelMessage],
 ) -> list[ModelMessage]:
-    """Compact message history based on token limits.
+    """Compact message history based on token limits with incremental processing.
 
-    This context-aware processor monitors token usage and removes older messages
-    when the conversation history becomes too large. It preserves system messages
-    and recent context while removing older user/assistant exchanges.
+    This incremental compactor prevents cascading summarization by:
+    1. Preserving existing summaries
+    2. Only processing NEW messages since the last summary
+    3. Combining summaries incrementally
+    4. Never re-processing already compacted content
 
     Args:
         ctx: Run context with usage information and dependencies
-        messages: List of messages in the conversation history
+        messages: Current conversation history
 
     Returns:
         Compacted list of messages within token limits
     """
-    # Get current token usage from context
-    current_tokens = ctx.usage.total_tokens if ctx.usage else 0
-
-    # Get token limit from model configuration or use fallback
-    model_max_tokens = ctx.deps.llm_model.max_input_tokens
-    max_tokens = int(
-        model_max_tokens * 0.8
-    )  # Use 80% of max to leave room for response
-    percentage_of_limit_used = (
-        (current_tokens / max_tokens) * 100 if max_tokens > 0 else 0
-    )
-    logger.debug(
-        "History compactor: current tokens=%d, limit=%d, percentage used=%.2f%%",
-        current_tokens,
-        max_tokens,
-        percentage_of_limit_used,
-    )
+    # Extract dependencies from context
+    deps = ctx.deps
+
+    # Get token limit from model configuration
+    model_max_tokens = deps.llm_model.max_input_tokens
+    max_tokens = int(model_max_tokens * TOKEN_LIMIT_RATIO)
+
+    # Find existing summaries to determine compaction strategy
+    last_summary_index = find_last_summary_index(messages)
+
+    if last_summary_index is not None:
+        # Check if post-summary conversation exceeds threshold for incremental compaction
+        post_summary_tokens = estimate_post_summary_tokens(
+            messages, last_summary_index, deps.llm_model
+        )
+        post_summary_percentage = (
+            (post_summary_tokens / max_tokens) * 100 if max_tokens > 0 else 0
+        )
+
+        logger.debug(
+            f"Found existing summary at index {last_summary_index}. "
+            f"Post-summary tokens: {post_summary_tokens}, threshold: {max_tokens}, "
+            f"percentage: {post_summary_percentage:.2f}%%"
+        )
+
+        # Only do incremental compaction if post-summary conversation exceeds threshold
+        if post_summary_tokens < max_tokens:
+            logger.debug(
+                f"Post-summary conversation under threshold ({post_summary_tokens} < {max_tokens}), "
+                f"keeping all {len(messages)} messages"
+            )
+            return messages
+
+        # INCREMENTAL COMPACTION: Process new messages since last summary
+        logger.debug(
+            "Post-summary conversation exceeds threshold, performing incremental compaction"
+        )
+
+        # Extract existing summary content
+        summary_message = messages[last_summary_index]
+        existing_summary_part = None
+        for part in summary_message.parts:
+            if is_summary_part(part):
+                existing_summary_part = part
+                break
+
+        if not existing_summary_part:
+            logger.warning(
+                "Found summary index but no summary part, falling back to full compaction"
+            )
+            return await _full_compaction(deps, messages)
+
+        existing_summary = extract_summary_content(existing_summary_part)
+
+        # Get messages AFTER the last summary for incremental processing
+        messages_to_process = messages[last_summary_index + 1 :]
+
+        if not messages_to_process:
+            logger.debug(
+                "No new messages since last summary, returning existing history"
+            )
+            return messages
+
+        # Extract context from new messages only
+        new_context = extract_context_from_messages(messages_to_process)
+
+        # Check if there's meaningful content (responses) to summarize
+        has_meaningful_content = any(
+            isinstance(msg, ModelResponse) for msg in messages_to_process
+        )
+
+        # If there are only user requests and no responses, no need to summarize
+        if not has_meaningful_content or not new_context.strip():
+            logger.debug(
+                "No meaningful new content to summarize, returning existing history"
+            )
+            return messages
+
+        # Use incremental summarization prompt with proper template variables
+        try:
+            incremental_prompt = prompt_loader.render(
+                "history/incremental_summarization.j2",
+                existing_summary=existing_summary,
+                new_messages=new_context,
+            )
+        except Exception:
+            # Fallback to regular summarization if incremental template doesn't exist yet
+            logger.warning(
+                "Incremental summarization template not found, using regular template"
+            )
+            incremental_prompt = prompt_loader.render("history/summarization.j2")
+            # Combine existing and new context for fallback
+            new_context = (
+                f"EXISTING SUMMARY:\n{existing_summary}\n\nNEW MESSAGES:\n{new_context}"
+            )
+
+        # Create incremental summary
+        request_messages: list[ModelMessage] = [
+            ModelRequest.user_text_prompt(new_context, instructions=incremental_prompt)
+        ]
+
+        # Calculate optimal max_tokens for summarization
+        max_tokens = calculate_max_summarization_tokens(
+            deps.llm_model, request_messages
+        )
+
+        # Debug logging using shared utilities
+        log_summarization_request(
+            deps.llm_model, max_tokens, incremental_prompt, new_context, "INCREMENTAL"
+        )
+
+        # Use shotgun wrapper to ensure full token utilization
+        summary_response = await shotgun_model_request(
+            model_config=deps.llm_model,
+            messages=request_messages,
+            max_tokens=max_tokens,  # Use calculated optimal tokens for summarization
+        )
+
+        log_summarization_response(summary_response, "INCREMENTAL")
+
+        # Calculate token reduction (from new messages only)
+        new_tokens = len(new_context.split())  # Rough estimate
+        summary_tokens = (
+            summary_response.usage.output_tokens if summary_response.usage else 0
+        )
+        logger.debug(
+            f"Incremental compaction: processed {len(messages_to_process)} new messages, "
+            f"reduced ~{new_tokens} tokens to {summary_tokens} tokens"
+        )
 
-    # If we're under the limit, return all messages
-    if current_tokens < max_tokens:
-        logger.debug("Under token limit, keeping all %d messages", len(messages))
-        return messages
+        # Build the new compacted history with the updated summary
+        new_summary_part = create_marked_summary_part(summary_response)
 
-    # Get current token usage from context
-    current_tokens = ctx.usage.total_tokens if ctx.usage else 0
+        # Extract essential context from messages before the last summary (if any)
+        system_prompt = ""
+        first_user_prompt = ""
+        if last_summary_index > 0:
+            # Get system and first user from original conversation
+            system_prompt = get_system_prompt(messages[:last_summary_index]) or ""
+            first_user_prompt = (
+                get_first_user_request(messages[:last_summary_index]) or ""
+            )
 
-    context = ""
+        # Create the updated summary message
+        updated_summary_message = ModelResponse(parts=[new_summary_part])
 
-    # Separate system messages from conversation messages
-    for msg in messages:
-        if isinstance(msg, ModelResponse) or isinstance(msg, ModelRequest):
-            for part in msg.parts:
-                message_content = get_context_from_message(part)
-                if not message_content:
-                    continue
-                context += get_context_from_message(part) + "\n"
-        else:
-            # Handle whatever this is
-            pass
+        # Build final compacted history with CLEAN structure
+        compacted_messages: list[ModelMessage] = []
 
+        # Only add system/user context if it exists and is meaningful
+        if system_prompt or first_user_prompt:
+            compacted_messages.append(
+                ModelRequest(
+                    parts=[
+                        SystemPromptPart(content=system_prompt),
+                        UserPromptPart(content=first_user_prompt),
+                    ]
+                )
+            )
+
+        # Add the summary
+        compacted_messages.append(updated_summary_message)
+
+        # Ensure history ends with ModelRequest for PydanticAI compatibility
+        compacted_messages = ensure_ends_with_model_request(
+            compacted_messages, messages
+        )
+
+        logger.debug(
+            f"Incremental compaction complete: {len(messages)} -> {len(compacted_messages)} messages"
+        )
+        return compacted_messages
+
+    else:
+        # Check if total conversation exceeds threshold for full compaction
+        total_tokens = estimate_tokens_from_messages(messages, deps.llm_model)
+        total_percentage = (total_tokens / max_tokens) * 100 if max_tokens > 0 else 0
+
+        logger.debug(
+            f"No existing summary found. Total tokens: {total_tokens}, threshold: {max_tokens}, "
+            f"percentage: {total_percentage:.2f}%%"
+        )
+
+        # Only do full compaction if total conversation exceeds threshold
+        if total_tokens < max_tokens:
+            logger.debug(
+                f"Total conversation under threshold ({total_tokens} < {max_tokens}), "
+                f"keeping all {len(messages)} messages"
+            )
+            return messages
+
+        # FIRST-TIME COMPACTION: Process all messages
+        logger.debug(
+            "Total conversation exceeds threshold, performing initial full compaction"
+        )
+        return await _full_compaction(deps, messages)
+
+
+async def _full_compaction(
+    deps: AgentDeps,
+    messages: list[ModelMessage],
+) -> list[ModelMessage]:
+    """Perform full compaction for first-time summarization."""
+    # Extract context from all messages
+    context = extract_context_from_messages(messages)
+
+    # Use regular summarization prompt
     summarization_prompt = prompt_loader.render("history/summarization.j2")
-    summary_response = await model_request(
-        model=ctx.model,
-        messages=[
-            ModelRequest.user_text_prompt(context, instructions=summarization_prompt)
-        ],
+    request_messages: list[ModelMessage] = [
+        ModelRequest.user_text_prompt(context, instructions=summarization_prompt)
+    ]
+
+    # Calculate optimal max_tokens for summarization
+    max_tokens = calculate_max_summarization_tokens(deps.llm_model, request_messages)
+
+    # Debug logging using shared utilities
+    log_summarization_request(
+        deps.llm_model, max_tokens, summarization_prompt, context, "FULL"
+    )
+
+    # Use shotgun wrapper to ensure full token utilization
+    summary_response = await shotgun_model_request(
+        model_config=deps.llm_model,
+        messages=request_messages,
+        max_tokens=max_tokens,  # Use calculated optimal tokens for summarization
     )
-    # Usage before and after
+
+    # Calculate token reduction
+    current_tokens = estimate_tokens_from_messages(messages, deps.llm_model)
     summary_usage = summary_response.usage
     reduction_percentage = (
-        (current_tokens - summary_usage.output_tokens) / current_tokens
-    ) * 100
+        ((current_tokens - summary_usage.output_tokens) / current_tokens) * 100
+        if current_tokens > 0 and summary_usage
+        else 0
+    )
+
+    log_summarization_response(summary_response, "FULL")
+
+    # Log token reduction (already calculated above)
     logger.debug(
-        "Compacted %s tokens into %s tokens for a %.2f percent reduction",
+        "Full compaction: %s tokens -> %s tokens (%.2f%% reduction)",
         current_tokens,
-        summary_usage.output_tokens,
+        summary_usage.output_tokens if summary_usage else 0,
         reduction_percentage,
     )
 
-    system_prompt = get_system_promt(messages) or ""
+    # Mark summary with special prefix
+    marked_summary_part = create_marked_summary_part(summary_response)
+
+    # Build compacted history structure
+    system_prompt = get_system_prompt(messages) or ""
     user_prompt = get_first_user_request(messages) or ""
-    # Extract content from the first response part safely
-    summarization_part = summary_response.parts[0]
-    return [
+
+    # Create base structure
+    compacted_messages: list[ModelMessage] = [
         ModelRequest(
             parts=[
                 SystemPromptPart(content=system_prompt),
                 UserPromptPart(content=user_prompt),
             ]
         ),
-        ModelResponse(
-            parts=[
-                summarization_part,
-            ]
-        ),
+        ModelResponse(parts=[marked_summary_part]),
     ]
 
+    # Ensure history ends with ModelRequest for PydanticAI compatibility
+    compacted_messages = ensure_ends_with_model_request(compacted_messages, messages)
 
-def get_first_user_request(messages: list[ModelMessage]) -> str | None:
-    """Extract first user request from messages.
-
-    Args:
-        messages: List of messages in the conversation history
-    Returns:
-        The first user request as a string.
-    """
-    for msg in messages:
-        if isinstance(msg, ModelRequest):
-            for part in msg.parts:
-                if isinstance(part, UserPromptPart):
-                    if isinstance(part.content, str):
-                        return part.content
-    return None
-
-
-def get_system_promt(messages: list[ModelMessage]) -> str | None:
-    """Extract system prompt from messages.
-
-    Args:
-        messages: List of messages in the conversation history
-
-    Returns:
-        The system prompt as a string.
-    """
-    for msg in messages:
-        if isinstance(msg, ModelRequest):
-            for part in msg.parts:
-                if isinstance(part, SystemPromptPart):
-                    return part.content
-    return None
-
-
-def get_context_from_message(
-    message_part: SystemPromptPart
-    | UserPromptPart
-    | ToolReturnPart
-    | RetryPromptPart
-    | ModelResponsePart,
-) -> str:
-    """Extract context from a message part.
-
-    Args:
-        message: The message part to extract context from.
-
-    Returns:
-        The extracted context as a string.
-    """
-
-    if isinstance(message_part, SystemPromptPart):
-        return ""  # We do not include system prompts in the summary
-    elif isinstance(message_part, UserPromptPart):
-        if isinstance(message_part.content, str):
-            return "<USER_PROMPT>\n" + message_part.content + "\n</USER_PROMPT>"
-        else:
-            return ""
-    elif isinstance(message_part, ToolReturnPart):
-        return "<TOOL_RETURN>\n" + str(message_part.content) + "\n</TOOL_RETURN>"
-    elif isinstance(message_part, RetryPromptPart):
-        if isinstance(message_part.content, str):
-            return "<RETRY_PROMPT>\n" + message_part.content + "\n</RETRY_PROMPT>"
-        return ""
-
-    # TextPart | ToolCallPart | BuiltinToolCallPart | BuiltinToolReturnPart | ThinkingPart
-    if isinstance(message_part, TextPart):
-        return "<ASSISTANT_TEXT>\n" + message_part.content + "\n</ASSISTANT_TEXT>"
-    elif isinstance(message_part, ToolCallPart):
-        if isinstance(message_part.args, dict):
-            args_str = ", ".join(f"{k}={repr(v)}" for k, v in message_part.args.items())
-            tool_call_str = f"{message_part.tool_name}({args_str})"
-        else:
-            tool_call_str = f"{message_part.tool_name}({message_part.args})"
-        return "<TOOL_CALL>\n" + tool_call_str + "\n</TOOL_CALL>"
-    elif isinstance(message_part, BuiltinToolCallPart):
-        return (
-            "<BUILTIN_TOOL_CALL>\n" + message_part.tool_name + "\n</BUILTIN_TOOL_CALL>"
-        )
-    elif isinstance(message_part, BuiltinToolReturnPart):
-        return (
-            "<BUILTIN_TOOL_RETURN>\n"
-            + message_part.tool_name
-            + "\n</BUILTIN_TOOL_RETURN>"
-        )
-    elif isinstance(message_part, ThinkingPart):
-        return "<THINKING>\n" + message_part.content + "\n</THINKING>"
-
-    return ""
+    return compacted_messages

shotgun/agents/history/message_utils.py

@@ -0,0 +1,46 @@
+"""Utility functions for working with PydanticAI messages."""
+
+from pydantic_ai.messages import (
+    ModelMessage,
+    ModelRequest,
+    SystemPromptPart,
+    UserPromptPart,
+)
+
+
+def get_first_user_request(messages: list[ModelMessage]) -> str | None:
+    """Extract first user request content from messages."""
+    for msg in messages:
+        if isinstance(msg, ModelRequest):
+            for part in msg.parts:
+                if isinstance(part, UserPromptPart) and isinstance(part.content, str):
+                    return part.content
+    return None
+
+
+def get_last_user_request(messages: list[ModelMessage]) -> ModelRequest | None:
+    """Extract the last user request from messages."""
+    for msg in reversed(messages):
+        if isinstance(msg, ModelRequest):
+            for part in msg.parts:
+                if isinstance(part, UserPromptPart):
+                    return msg
+    return None
+
+
+def get_user_content_from_request(request: ModelRequest) -> str | None:
+    """Extract user prompt content from a ModelRequest."""
+    for part in request.parts:
+        if isinstance(part, UserPromptPart) and isinstance(part.content, str):
+            return part.content
+    return None
+
+
+def get_system_prompt(messages: list[ModelMessage]) -> str | None:
+    """Extract system prompt from messages."""
+    for msg in messages:
+        if isinstance(msg, ModelRequest):
+            for part in msg.parts:
+                if isinstance(part, SystemPromptPart):
+                    return part.content
+    return None