codepp 0.0.437__py3-none-any.whl

code_puppy/agents/base_agent.py

@@ -0,0 +1,2156 @@
+"""Base agent configuration class for defining agent properties."""
+
+import asyncio
+import dataclasses
+import json
+import math
+import pathlib
+import signal
+import threading
+import time
+import traceback
+import uuid
+from abc import ABC, abstractmethod
+from typing import (
+    Any,
+    Callable,
+    Dict,
+    List,
+    Optional,
+    Sequence,
+    Set,
+    Tuple,
+    Type,
+    Union,
+)
+
+import mcp
+import pydantic
+import pydantic_ai.models
+from dbos import DBOS, SetWorkflowID
+from pydantic_ai import Agent as PydanticAgent
+from pydantic_ai import (
+    BinaryContent,
+    DocumentUrl,
+    ImageUrl,
+    RunContext,
+    UsageLimitExceeded,
+    UsageLimits,
+)
+from pydantic_ai.durable_exec.dbos import DBOSAgent
+from pydantic_ai.messages import (
+    ModelMessage,
+    ModelRequest,
+    ModelResponse,
+    TextPart,
+    ThinkingPart,
+    ToolCallPart,
+    ToolCallPartDelta,
+    ToolReturn,
+    ToolReturnPart,
+)
+from rich.text import Text
+
+from code_puppy.agents.event_stream_handler import event_stream_handler
+from code_puppy.callbacks import (
+    on_agent_run_end,
+    on_agent_run_start,
+    on_message_history_processor_end,
+    on_message_history_processor_start,
+)
+
+# Consolidated relative imports
+from code_puppy.config import (
+    get_agent_pinned_model,
+    get_compaction_strategy,
+    get_compaction_threshold,
+    get_global_model_name,
+    get_message_limit,
+    get_protected_token_count,
+    get_use_dbos,
+    get_value,
+)
+from code_puppy.error_logging import log_error
+from code_puppy.keymap import cancel_agent_uses_signal, get_cancel_agent_char_code
+from code_puppy.mcp_ import get_mcp_manager
+from code_puppy.messaging import (
+    emit_error,
+    emit_info,
+    emit_warning,
+)
+from code_puppy.messaging.spinner import (
+    SpinnerBase,
+    update_spinner_context,
+)
+from code_puppy.model_factory import ModelFactory, make_model_settings
+from code_puppy.summarization_agent import run_summarization_sync, SummarizationError
+from code_puppy.tools.agent_tools import _active_subagent_tasks
+from code_puppy.tools.command_runner import (
+    is_awaiting_user_input,
+)
+
+# Global flag to track delayed compaction requests
+_delayed_compaction_requested = False
+
+_reload_count = 0
+
+
+def _log_error_to_file(exc: Exception) -> Optional[str]:
+    """Log detailed error information to ~/.code_puppy/error_logs/log_{timestamp}.txt.
+
+    Args:
+        exc: The exception to log.
+
+    Returns:
+        The path to the log file if successful, None otherwise.
+    """
+    try:
+        from code_puppy.error_logging import get_logs_dir
+
+        error_logs_dir = pathlib.Path(get_logs_dir())
+        error_logs_dir.mkdir(parents=True, exist_ok=True)
+
+        timestamp = time.strftime("%Y%m%d_%H%M%S")
+        log_file = error_logs_dir / f"log_{timestamp}.txt"
+
+        with open(log_file, "w", encoding="utf-8") as f:
+            f.write(f"Timestamp: {time.strftime('%Y-%m-%d %H:%M:%S')}\n")
+            f.write(f"Exception Type: {type(exc).__name__}\n")
+            f.write(f"Exception Message: {str(exc)}\n")
+            f.write(f"Exception Args: {exc.args}\n")
+            f.write("\n--- Full Traceback ---\n")
+            f.write(traceback.format_exc())
+            f.write("\n--- Exception Chain ---\n")
+            # Walk the exception chain for chained exceptions
+            current = exc
+            chain_depth = 0
+            while current is not None and chain_depth < 10:
+                f.write(
+                    f"\n[Cause {chain_depth}] {type(current).__name__}: {current}\n"
+                )
+                f.write("".join(traceback.format_tb(current.__traceback__)))
+                current = (
+                    current.__cause__ if current.__cause__ else current.__context__
+                )
+                chain_depth += 1
+
+        return str(log_file)
+    except Exception:
+        # Don't let logging errors break the main flow
+        return None
+
+
+class BaseAgent(ABC):
+    """Base class for all agent configurations."""
+
+    def __init__(self):
+        self.id = str(uuid.uuid4())
+        self._message_history: List[Any] = []
+        self._compacted_message_hashes: Set[str] = set()
+        # Agent construction cache
+        self._code_generation_agent = None
+        self._last_model_name: Optional[str] = None
+        # Puppy rules loaded lazily
+        self._puppy_rules: Optional[str] = None
+        self.cur_model: pydantic_ai.models.Model
+        # Cache for MCP tool definitions (for token estimation)
+        # This is populated after the first successful run when MCP tools are retrieved
+        self._mcp_tool_definitions_cache: List[Dict[str, Any]] = []
+
+    def get_identity(self) -> str:
+        """Get a unique identity for this agent instance.
+
+        Returns:
+            A string like 'python-programmer-a3f2b1' combining name + short UUID.
+        """
+        return f"{self.name}-{self.id[:6]}"
+
+    def get_identity_prompt(self) -> str:
+        """Get the identity prompt suffix to embed in system prompts.
+
+        Returns:
+            A string instructing the agent about its identity for task ownership.
+        """
+        return (
+            f"\n\nYour ID is `{self.get_identity()}`. "
+            "Use this for any tasks which require identifying yourself "
+            "such as claiming task ownership or coordination with other agents."
+        )
+
+    def get_full_system_prompt(self) -> str:
+        """Get the complete system prompt with identity automatically appended.
+
+        This wraps get_system_prompt() and appends the agent's identity,
+        so subclasses don't need to worry about it.
+
+        Returns:
+            The full system prompt including identity information.
+        """
+        return self.get_system_prompt() + self.get_identity_prompt()
+
+    @property
+    @abstractmethod
+    def name(self) -> str:
+        """Unique identifier for the agent."""
+        pass
+
+    @property
+    @abstractmethod
+    def display_name(self) -> str:
+        """Human-readable name for the agent."""
+        pass
+
+    @property
+    @abstractmethod
+    def description(self) -> str:
+        """Brief description of what this agent does."""
+        pass
+
+    @abstractmethod
+    def get_system_prompt(self) -> str:
+        """Get the system prompt for this agent."""
+        pass
+
+    @abstractmethod
+    def get_available_tools(self) -> List[str]:
+        """Get list of tool names that this agent should have access to.
+
+        Returns:
+            List of tool names to register for this agent.
+        """
+        pass
+
+    def get_tools_config(self) -> Optional[Dict[str, Any]]:
+        """Get tool configuration for this agent.
+
+        Returns:
+            Dict with tool configuration, or None to use default tools.
+        """
+        return None
+
+    def get_user_prompt(self) -> Optional[str]:
+        """Get custom user prompt for this agent.
+
+        Returns:
+            Custom prompt string, or None to use default.
+        """
+        return None
+
+    # Message history management methods
+    def get_message_history(self) -> List[Any]:
+        """Get the message history for this agent.
+
+        Returns:
+            List of messages in this agent's conversation history.
+        """
+        return self._message_history
+
+    def set_message_history(self, history: List[Any]) -> None:
+        """Set the message history for this agent.
+
+        Args:
+            history: List of messages to set as the conversation history.
+        """
+        self._message_history = history
+
+    def clear_message_history(self) -> None:
+        """Clear the message history for this agent."""
+        self._message_history = []
+        self._compacted_message_hashes.clear()
+
+    def append_to_message_history(self, message: Any) -> None:
+        """Append a message to this agent's history.
+
+        Args:
+            message: Message to append to the conversation history.
+        """
+        self._message_history.append(message)
+
+    def extend_message_history(self, history: List[Any]) -> None:
+        """Extend this agent's message history with multiple messages.
+
+        Args:
+            history: List of messages to append to the conversation history.
+        """
+        self._message_history.extend(history)
+
+    def get_compacted_message_hashes(self) -> Set[str]:
+        """Get the set of compacted message hashes for this agent.
+
+        Returns:
+            Set of hashes for messages that have been compacted/summarized.
+        """
+        return self._compacted_message_hashes
+
+    def add_compacted_message_hash(self, message_hash: str) -> None:
+        """Add a message hash to the set of compacted message hashes.
+
+        Args:
+            message_hash: Hash of a message that has been compacted/summarized.
+        """
+        self._compacted_message_hashes.add(message_hash)
+
+    def get_model_name(self) -> Optional[str]:
+        """Get pinned model name for this agent, if specified.
+
+        Returns:
+            Model name to use for this agent, or global default if none pinned.
+        """
+        pinned = get_agent_pinned_model(self.name)
+        if pinned == "" or pinned is None:
+            return get_global_model_name()
+        return pinned
+
+    def _clean_binaries(self, messages: List[ModelMessage]) -> List[ModelMessage]:
+        """Remove BinaryContent items from message parts.
+
+        Note: This mutates the messages in-place by modifying part.content.
+        The return value is the same list for API consistency.
+        """
+        for message in messages:
+            for part in message.parts:
+                if hasattr(part, "content") and isinstance(part.content, list):
+                    part.content = [
+                        item
+                        for item in part.content
+                        if not isinstance(item, BinaryContent)
+                    ]
+        return messages
+
+    def ensure_history_ends_with_request(
+        self, messages: List[ModelMessage]
+    ) -> List[ModelMessage]:
+        """Ensure message history ends with a ModelRequest.
+
+        pydantic_ai requires that processed message history ends with a ModelRequest.
+        This can fail when swapping models mid-conversation if the history ends with
+        a ModelResponse from the previous model.
+
+        This method trims trailing ModelResponse messages to ensure compatibility.
+
+        Args:
+            messages: List of messages to validate/fix.
+
+        Returns:
+            List of messages guaranteed to end with ModelRequest, or empty list
+            if no ModelRequest is found.
+        """
+        messages = list(messages)  # defensive copy
+        if not messages:
+            return messages
+
+        # Trim trailing ModelResponse messages
+        while messages and isinstance(messages[-1], ModelResponse):
+            messages = messages[:-1]
+
+        return messages
+
+    # Message history processing methods (moved from state_management.py and message_history_processor.py)
+    def _stringify_part(self, part: Any) -> str:
+        """Create a stable string representation for a message part.
+
+        We deliberately ignore timestamps so identical content hashes the same even when
+        emitted at different times. This prevents status updates from blowing up the
+        history when they are repeated with new timestamps."""
+
+        attributes: List[str] = [part.__class__.__name__]
+
+        # Role/instructions help disambiguate parts that otherwise share content
+        if hasattr(part, "role") and part.role:
+            attributes.append(f"role={part.role}")
+        if hasattr(part, "instructions") and part.instructions:
+            attributes.append(f"instructions={part.instructions}")
+
+        if hasattr(part, "tool_call_id") and part.tool_call_id:
+            attributes.append(f"tool_call_id={part.tool_call_id}")
+
+        if hasattr(part, "tool_name") and part.tool_name:
+            attributes.append(f"tool_name={part.tool_name}")
+
+        content = getattr(part, "content", None)
+        if content is None:
+            attributes.append("content=None")
+        elif isinstance(content, str):
+            attributes.append(f"content={content}")
+        elif isinstance(content, pydantic.BaseModel):
+            attributes.append(
+                f"content={json.dumps(content.model_dump(), sort_keys=True)}"
+            )
+        elif isinstance(content, dict):
+            attributes.append(f"content={json.dumps(content, sort_keys=True)}")
+        elif isinstance(content, list):
+            for item in content:
+                if isinstance(item, str):
+                    attributes.append(f"content={item}")
+                if isinstance(item, BinaryContent):
+                    attributes.append(f"BinaryContent={hash(item.data)}")
+        else:
+            attributes.append(f"content={repr(content)}")
+        result = "|".join(attributes)
+        return result
+
+    def hash_message(self, message: Any) -> int:
+        """Create a stable hash for a model message that ignores timestamps."""
+        role = getattr(message, "role", None)
+        instructions = getattr(message, "instructions", None)
+        header_bits: List[str] = []
+        if role:
+            header_bits.append(f"role={role}")
+        if instructions:
+            header_bits.append(f"instructions={instructions}")
+
+        part_strings = [
+            self._stringify_part(part) for part in getattr(message, "parts", [])
+        ]
+        canonical = "||".join(header_bits + part_strings)
+        return hash(canonical)
+
+    def stringify_message_part(self, part) -> str:
+        """
+        Convert a message part to a string representation for token estimation or other uses.
+
+        Args:
+            part: A message part that may contain content or be a tool call
+
+        Returns:
+            String representation of the message part
+        """
+        result = ""
+        if hasattr(part, "part_kind"):
+            result += part.part_kind + ": "
+        else:
+            result += str(type(part)) + ": "
+
+        # Handle content
+        if hasattr(part, "content") and part.content:
+            # Handle different content types
+            if isinstance(part.content, str):
+                result = part.content
+            elif isinstance(part.content, pydantic.BaseModel):
+                result = json.dumps(part.content.model_dump())
+            elif isinstance(part.content, dict):
+                result = json.dumps(part.content)
+            elif isinstance(part.content, list):
+                result = ""
+                for item in part.content:
+                    if isinstance(item, str):
+                        result += item + "\n"
+                    if isinstance(item, BinaryContent):
+                        result += f"BinaryContent={hash(item.data)}\n"
+            else:
+                result = str(part.content)
+
+        # Handle tool calls which may have additional token costs
+        # If part also has content, we'll process tool calls separately
+        if hasattr(part, "tool_name") and part.tool_name:
+            # Estimate tokens for tool name and parameters
+            tool_text = part.tool_name
+            if hasattr(part, "args"):
+                tool_text += f" {str(part.args)}"
+            result += tool_text
+
+        return result
+
+    def estimate_token_count(self, text: str) -> int:
+        """
+        Simple token estimation using len(message) / 2.5.
+        This replaces tiktoken with a much simpler approach.
+        """
+        return max(1, math.floor((len(text) / 2.5)))
+
+    def estimate_tokens_for_message(self, message: ModelMessage) -> int:
+        """
+        Estimate the number of tokens in a message using len(message)
+        Simple and fast replacement for tiktoken.
+        """
+        total_tokens = 0
+
+        for part in message.parts:
+            part_str = self.stringify_message_part(part)
+            if part_str:
+                total_tokens += self.estimate_token_count(part_str)
+
+        return max(1, total_tokens)
+
+    def estimate_context_overhead_tokens(self) -> int:
+        """
+        Estimate the token overhead from system prompt and tool definitions.
+
+        This accounts for tokens that are always present in the context:
+        - System prompt (for non-Claude-Code models)
+        - Tool definitions (name, description, parameter schema)
+        - MCP tool definitions
+
+        Note: For Claude Code models, the system prompt is prepended to the first
+        user message, so it's already counted in the message history tokens.
+        We only count the short fixed instructions for Claude Code models.
+        """
+        total_tokens = 0
+
+        # 1. Estimate tokens for system prompt / instructions
+        # Use prepare_prompt_for_model() to get the correct instructions for token counting.
+        # For models that prepend system prompt to user message (claude-code, antigravity),
+        # this returns the short fixed instructions. For other models, returns full prompt.
+        try:
+            from code_puppy.model_utils import prepare_prompt_for_model
+
+            model_name = (
+                self.get_model_name() if hasattr(self, "get_model_name") else ""
+            )
+            system_prompt = self.get_full_system_prompt()
+
+            # Get the instructions that will be used (handles model-specific logic via hooks)
+            prepared = prepare_prompt_for_model(
+                model_name=model_name,
+                system_prompt=system_prompt,
+                user_prompt="",  # Empty - we just need the instructions
+                prepend_system_to_user=False,  # Don't modify prompt, just get instructions
+            )
+
+            if prepared.instructions:
+                total_tokens += self.estimate_token_count(prepared.instructions)
+        except Exception:
+            pass  # If we can't get system prompt, skip it
+
+        # 2. Estimate tokens for pydantic_agent tool definitions
+        pydantic_agent = getattr(self, "pydantic_agent", None)
+        if pydantic_agent:
+            tools = getattr(pydantic_agent, "_tools", None)
+            if tools and isinstance(tools, dict):
+                for tool_name, tool_func in tools.items():
+                    try:
+                        # Estimate tokens from tool name
+                        total_tokens += self.estimate_token_count(tool_name)
+
+                        # Estimate tokens from tool description
+                        description = getattr(tool_func, "__doc__", None) or ""
+                        if description:
+                            total_tokens += self.estimate_token_count(description)
+
+                        # Estimate tokens from parameter schema
+                        # Tools may have a schema attribute or we can try to get it from annotations
+                        schema = getattr(tool_func, "schema", None)
+                        if schema:
+                            schema_str = (
+                                json.dumps(schema)
+                                if isinstance(schema, dict)
+                                else str(schema)
+                            )
+                            total_tokens += self.estimate_token_count(schema_str)
+                        else:
+                            # Try to get schema from function annotations
+                            annotations = getattr(tool_func, "__annotations__", None)
+                            if annotations:
+                                total_tokens += self.estimate_token_count(
+                                    str(annotations)
+                                )
+                    except Exception:
+                        continue  # Skip tools we can't process
+
+        # 3. Estimate tokens for MCP tool definitions from cache
+        # MCP tools are fetched asynchronously, so we use a cache that's populated
+        # after the first successful run. See _update_mcp_tool_cache() method.
+        mcp_tool_cache = getattr(self, "_mcp_tool_definitions_cache", [])
+        if mcp_tool_cache:
+            for tool_def in mcp_tool_cache:
+                try:
+                    # Estimate tokens from tool name
+                    tool_name = tool_def.get("name", "")
+                    if tool_name:
+                        total_tokens += self.estimate_token_count(tool_name)
+
+                    # Estimate tokens from tool description
+                    description = tool_def.get("description", "")
+                    if description:
+                        total_tokens += self.estimate_token_count(description)
+
+                    # Estimate tokens from parameter schema (inputSchema)
+                    input_schema = tool_def.get("inputSchema")
+                    if input_schema:
+                        schema_str = (
+                            json.dumps(input_schema)
+                            if isinstance(input_schema, dict)
+                            else str(input_schema)
+                        )
+                        total_tokens += self.estimate_token_count(schema_str)
+                except Exception:
+                    continue  # Skip tools we can't process
+
+        return total_tokens
+
+    async def _update_mcp_tool_cache(self) -> None:
+        """
+        Update the MCP tool definitions cache by fetching tools from running MCP servers.
+
+        This should be called after a successful run to populate the cache for
+        accurate token estimation in subsequent runs.
+        """
+        mcp_servers = getattr(self, "_mcp_servers", None)
+        if not mcp_servers:
+            return
+
+        tool_definitions = []
+        for mcp_server in mcp_servers:
+            try:
+                # Check if the server has list_tools method (pydantic-ai MCP servers)
+                if hasattr(mcp_server, "list_tools"):
+                    # list_tools() returns list[mcp_types.Tool]
+                    tools = await mcp_server.list_tools()
+                    for tool in tools:
+                        tool_def = {
+                            "name": getattr(tool, "name", ""),
+                            "description": getattr(tool, "description", ""),
+                            "inputSchema": getattr(tool, "inputSchema", {}),
+                        }
+                        tool_definitions.append(tool_def)
+            except Exception:
+                # Server might not be running or accessible, skip it
+                continue
+
+        self._mcp_tool_definitions_cache = tool_definitions
+
+    def update_mcp_tool_cache_sync(self) -> None:
+        """
+        Synchronously clear the MCP tool cache.
+
+        This clears the cache so that token counts will be recalculated on the next
+        agent run. Call this after starting/stopping MCP servers.
+
+        Note: We don't try to fetch tools synchronously because MCP servers require
+        async context management that doesn't work well from sync code. The cache
+        will be repopulated on the next successful agent run.
+        """
+        # Simply clear the cache - it will be repopulated on the next agent run
+        # This is safer than trying to call async methods from sync context
+        self._mcp_tool_definitions_cache = []
+
+    def _is_tool_call_part(self, part: Any) -> bool:
+        if isinstance(part, (ToolCallPart, ToolCallPartDelta)):
+            return True
+
+        part_kind = (getattr(part, "part_kind", "") or "").replace("_", "-")
+        if part_kind == "tool-call":
+            return True
+
+        has_tool_name = getattr(part, "tool_name", None) is not None
+        has_args = getattr(part, "args", None) is not None
+        has_args_delta = getattr(part, "args_delta", None) is not None
+
+        return bool(has_tool_name and (has_args or has_args_delta))
+
+    def _is_tool_return_part(self, part: Any) -> bool:
+        if isinstance(part, (ToolReturnPart, ToolReturn)):
+            return True
+
+        part_kind = (getattr(part, "part_kind", "") or "").replace("_", "-")
+        if part_kind in {"tool-return", "tool-result"}:
+            return True
+
+        if getattr(part, "tool_call_id", None) is None:
+            return False
+
+        has_content = getattr(part, "content", None) is not None
+        has_content_delta = getattr(part, "content_delta", None) is not None
+        return bool(has_content or has_content_delta)
+
+    def filter_huge_messages(self, messages: List[ModelMessage]) -> List[ModelMessage]:
+        filtered = [m for m in messages if self.estimate_tokens_for_message(m) < 50000]
+        pruned = self.prune_interrupted_tool_calls(filtered)
+        return pruned
+
+    def _find_safe_split_index(
+        self, messages: List[ModelMessage], initial_split_idx: int
+    ) -> int:
+        """
+        Adjust split index to avoid breaking tool_use/tool_result pairs.
+
+        Ensures that if a tool_result is in the protected zone, its corresponding
+        tool_use is also included. Otherwise the LLM will error with
+        'tool_use ids found without tool_result blocks'.
+
+        Args:
+            messages: Full message list
+            initial_split_idx: The initial split point (messages before this go to summarize)
+
+        Returns:
+            Adjusted split index that doesn't break tool pairs
+        """
+        if initial_split_idx <= 1:
+            return initial_split_idx
+
+        # Collect tool_call_ids from messages AFTER the split (protected zone)
+        protected_tool_return_ids: Set[str] = set()
+        for msg in messages[initial_split_idx:]:
+            for part in getattr(msg, "parts", []) or []:
+                if getattr(part, "part_kind", None) == "tool-return":
+                    tool_call_id = getattr(part, "tool_call_id", None)
+                    if tool_call_id:
+                        protected_tool_return_ids.add(tool_call_id)
+
+        if not protected_tool_return_ids:
+            return initial_split_idx
+
+        # Scan backwards from split point to find any tool_uses that match protected returns
+        adjusted_idx = initial_split_idx
+        for i in range(
+            initial_split_idx - 1, 0, -1
+        ):  # Don't include system message at 0
+            msg = messages[i]
+            has_matching_tool_use = False
+            for part in getattr(msg, "parts", []) or []:
+                if getattr(part, "part_kind", None) == "tool-call":
+                    tool_call_id = getattr(part, "tool_call_id", None)
+                    if tool_call_id and tool_call_id in protected_tool_return_ids:
+                        has_matching_tool_use = True
+                        break
+
+            if has_matching_tool_use:
+                # This message has a tool_use whose return is in protected zone
+                # Move the split point back to include this message in protected zone
+                adjusted_idx = i
+            else:
+                # Once we find a message without matching tool_use, we can stop
+                # (tool calls and returns should be adjacent)
+                break
+
+        return adjusted_idx
+
+    def split_messages_for_protected_summarization(
+        self,
+        messages: List[ModelMessage],
+    ) -> Tuple[List[ModelMessage], List[ModelMessage]]:
+        """
+        Split messages into two groups: messages to summarize and protected recent messages.
+
+        Returns:
+            Tuple of (messages_to_summarize, protected_messages)
+
+        The protected_messages are the most recent messages that total up to the configured protected token count.
+        The system message (first message) is always protected.
+        All other messages that don't fit in the protected zone will be summarized.
+        """
+        if len(messages) <= 1:  # Just system message or empty
+            return [], messages
+
+        # Always protect the system message (first message)
+        system_message = messages[0]
+        system_tokens = self.estimate_tokens_for_message(system_message)
+
+        if len(messages) == 1:
+            return [], messages
+
+        # Get the configured protected token count
+        protected_tokens_limit = get_protected_token_count()
+
+        # Calculate tokens for messages from most recent backwards (excluding system message)
+        protected_messages = []
+        protected_token_count = system_tokens  # Start with system message tokens
+
+        # Go backwards through non-system messages to find protected zone
+        for i in range(
+            len(messages) - 1, 0, -1
+        ):  # Stop at 1, not 0 (skip system message)
+            message = messages[i]
+            message_tokens = self.estimate_tokens_for_message(message)
+
+            # If adding this message would exceed protected tokens, stop here
+            if protected_token_count + message_tokens > protected_tokens_limit:
+                break
+
+            protected_messages.append(message)
+            protected_token_count += message_tokens
+
+        # Messages that were added while scanning backwards are currently in reverse order.
+        # Reverse them to restore chronological ordering, then prepend the system prompt.
+        protected_messages.reverse()
+        protected_messages.insert(0, system_message)
+
+        # Messages to summarize are everything between the system message and the
+        # protected tail zone we just constructed.
+        protected_start_idx = max(1, len(messages) - (len(protected_messages) - 1))
+
+        # IMPORTANT: Adjust split point to avoid breaking tool_use/tool_result pairs
+        # The LLM requires every tool_use to have its tool_result immediately after
+        protected_start_idx = self._find_safe_split_index(messages, protected_start_idx)
+
+        messages_to_summarize = messages[1:protected_start_idx]
+
+        # Emit info messages
+        emit_info(
+            f"🔒 Protecting {len(protected_messages)} recent messages ({protected_token_count} tokens, limit: {protected_tokens_limit})"
+        )
+        emit_info(f"📝 Summarizing {len(messages_to_summarize)} older messages")
+
+        return messages_to_summarize, protected_messages
+
+    def summarize_messages(
+        self, messages: List[ModelMessage], with_protection: bool = True
+    ) -> Tuple[List[ModelMessage], List[ModelMessage]]:
+        """
+        Summarize messages while protecting recent messages up to PROTECTED_TOKENS.
+
+        Returns:
+            Tuple of (compacted_messages, summarized_source_messages)
+            where compacted_messages always preserves the original system message
+            as the first entry.
+        """
+        messages_to_summarize: List[ModelMessage]
+        protected_messages: List[ModelMessage]
+
+        if with_protection:
+            messages_to_summarize, protected_messages = (
+                self.split_messages_for_protected_summarization(messages)
+            )
+        else:
+            messages_to_summarize = messages[1:] if messages else []
+            protected_messages = messages[:1]
+
+        if not messages:
+            return [], []
+
+        system_message = messages[0]
+
+        if not messages_to_summarize:
+            # Nothing to summarize, so just return the original sequence
+            return self.prune_interrupted_tool_calls(messages), []
+
+        instructions = (
+            "The input will be a log of Agentic AI steps that have been taken"
+            " as well as user queries, etc. Summarize the contents of these steps."
+            " The high level details should remain but the bulk of the content from tool-call"
+            " responses should be compacted and summarized. For example if you see a tool-call"
+            " reading a file, and the file contents are large, then in your summary you might just"
+            " write: * used read_file on space_invaders.cpp - contents removed."
+            "\n Make sure your result is a bulleted list of all steps and interactions."
+            "\n\nNOTE: This summary represents older conversation history. Recent messages are preserved separately."
+        )
+
+        try:
+            # Prune any orphaned tool calls from messages before sending to LLM
+            # The LLM requires every tool_use to have a matching tool_result
+            pruned_messages_to_summarize = self.prune_interrupted_tool_calls(
+                messages_to_summarize
+            )
+
+            if not pruned_messages_to_summarize:
+                # After pruning, nothing left to summarize
+                return self.prune_interrupted_tool_calls(messages), []
+
+            new_messages = run_summarization_sync(
+                instructions, message_history=pruned_messages_to_summarize
+            )
+
+            if not isinstance(new_messages, list):
+                emit_warning(
+                    "Summarization agent returned non-list output; wrapping into message request"
+                )
+                new_messages = [ModelRequest([TextPart(str(new_messages))])]
+
+            compacted: List[ModelMessage] = [system_message] + list(new_messages)
+
+            # Drop the system message from protected_messages because we already included it
+            protected_tail = [
+                msg for msg in protected_messages if msg is not system_message
+            ]
+
+            compacted.extend(protected_tail)
+
+            return self.prune_interrupted_tool_calls(compacted), messages_to_summarize
+        except SummarizationError as e:
+            # SummarizationError has detailed error info
+            emit_error(f"Summarization failed: {e}")
+            if e.original_error:
+                emit_warning(
+                    f"💡 Tip: Underlying error was {type(e.original_error).__name__}. "
+                    "Consider using '/set compaction_strategy=truncation' as a fallback."
+                )
+            return messages, []  # Return original messages on failure
+        except Exception as e:
+            # Catch-all for unexpected errors
+            error_type = type(e).__name__
+            error_msg = str(e) if str(e) else "(no error details)"
+            emit_error(
+                f"Unexpected error during compaction: [{error_type}] {error_msg}"
+            )
+            return messages, []  # Return original messages on failure
+
+    def get_model_context_length(self) -> int:
+        """
+        Return the context length for this agent's effective model.
+
+        Honors per-agent pinned model via `self.get_model_name()`; falls back
+        to global model when no pin is set. Defaults conservatively on failure.
+        """
+        try:
+            model_configs = ModelFactory.load_config()
+            # Use the agent's effective model (respects /pin_model)
+            model_name = self.get_model_name()
+            model_config = model_configs.get(model_name, {})
+            context_length = model_config.get("context_length", 128000)
+            return int(context_length)
+        except Exception:
+            # Be safe; don't blow up status/compaction if model lookup fails
+            return 128000
+
+    def has_pending_tool_calls(self, messages: List[ModelMessage]) -> bool:
+        """
+        Check if there are any pending tool calls in the message history.
+
+        A pending tool call is one that has a ToolCallPart without a corresponding
+        ToolReturnPart. This indicates the model is still waiting for tool execution.
+
+        Returns:
+            True if there are pending tool calls, False otherwise
+        """
+        if not messages:
+            return False
+
+        tool_call_ids: Set[str] = set()
+        tool_return_ids: Set[str] = set()
+
+        # Collect all tool call and return IDs
+        for msg in messages:
+            for part in getattr(msg, "parts", []) or []:
+                tool_call_id = getattr(part, "tool_call_id", None)
+                if not tool_call_id:
+                    continue
+
+                if part.part_kind == "tool-call":
+                    tool_call_ids.add(tool_call_id)
+                elif part.part_kind == "tool-return":
+                    tool_return_ids.add(tool_call_id)
+
+        # Pending tool calls are those without corresponding returns
+        pending_calls = tool_call_ids - tool_return_ids
+        return len(pending_calls) > 0
+
+    def request_delayed_compaction(self) -> None:
+        """
+        Request that compaction be attempted after the current tool calls complete.
+
+        This sets a global flag that will be checked during the next message
+        processing cycle to trigger compaction when it's safe to do so.
+        """
+        global _delayed_compaction_requested
+        _delayed_compaction_requested = True
+        emit_info(
+            "🔄 Delayed compaction requested - will attempt after tool calls complete",
+            message_group="token_context_status",
+        )
+
+    def should_attempt_delayed_compaction(self) -> bool:
+        """
+        Check if delayed compaction was requested and it's now safe to proceed.
+
+        Returns:
+            True if delayed compaction was requested and no tool calls are pending
+        """
+        global _delayed_compaction_requested
+        if not _delayed_compaction_requested:
+            return False
+
+        # Check if it's now safe to compact
+        messages = self.get_message_history()
+        if not self.has_pending_tool_calls(messages):
+            _delayed_compaction_requested = False  # Reset the flag
+            return True
+
+        return False
+
+    def get_pending_tool_call_count(self, messages: List[ModelMessage]) -> int:
+        """
+        Get the count of pending tool calls for debugging purposes.
+
+        Returns:
+            Number of tool calls waiting for execution
+        """
+        if not messages:
+            return 0
+
+        tool_call_ids: Set[str] = set()
+        tool_return_ids: Set[str] = set()
+
+        for msg in messages:
+            for part in getattr(msg, "parts", []) or []:
+                tool_call_id = getattr(part, "tool_call_id", None)
+                if not tool_call_id:
+                    continue
+
+                if part.part_kind == "tool-call":
+                    tool_call_ids.add(tool_call_id)
+                elif part.part_kind == "tool-return":
+                    tool_return_ids.add(tool_call_id)
+
+        pending_calls = tool_call_ids - tool_return_ids
+        return len(pending_calls)
+
+    def prune_interrupted_tool_calls(
+        self, messages: List[ModelMessage]
+    ) -> List[ModelMessage]:
+        """
+        Remove any messages that participate in mismatched tool call sequences.
+
+        A mismatched tool call id is one that appears in a ToolCall (model/tool request)
+        without a corresponding tool return, or vice versa. We preserve original order
+        and only drop messages that contain parts referencing mismatched tool_call_ids.
+        """
+        if not messages:
+            return messages
+
+        tool_call_ids: Set[str] = set()
+        tool_return_ids: Set[str] = set()
+
+        # First pass: collect ids for calls vs returns
+        for msg in messages:
+            for part in getattr(msg, "parts", []) or []:
+                tool_call_id = getattr(part, "tool_call_id", None)
+                if not tool_call_id:
+                    continue
+                # Heuristic: if it's an explicit ToolCallPart or has a tool_name/args,
+                # consider it a call; otherwise it's a return/result.
+                if part.part_kind == "tool-call":
+                    tool_call_ids.add(tool_call_id)
+                else:
+                    tool_return_ids.add(tool_call_id)
+
+        mismatched: Set[str] = tool_call_ids.symmetric_difference(tool_return_ids)
+        if not mismatched:
+            return messages
+
+        pruned: List[ModelMessage] = []
+        dropped_count = 0
+        for msg in messages:
+            has_mismatched = False
+            for part in getattr(msg, "parts", []) or []:
+                tcid = getattr(part, "tool_call_id", None)
+                if tcid and tcid in mismatched:
+                    has_mismatched = True
+                    break
+            if has_mismatched:
+                dropped_count += 1
+                continue
+            pruned.append(msg)
+        return pruned
+
+    def message_history_processor(
+        self, ctx: RunContext, messages: List[ModelMessage]
+    ) -> List[ModelMessage]:
+        # First, prune any interrupted/mismatched tool-call conversations
+        model_max = self.get_model_context_length()
+
+        message_tokens = sum(self.estimate_tokens_for_message(msg) for msg in messages)
+        context_overhead = self.estimate_context_overhead_tokens()
+        total_current_tokens = message_tokens + context_overhead
+        proportion_used = total_current_tokens / model_max
+
+        context_summary = SpinnerBase.format_context_info(
+            total_current_tokens, model_max, proportion_used
+        )
+        update_spinner_context(context_summary)
+
+        # Get the configured compaction threshold
+        compaction_threshold = get_compaction_threshold()
+
+        # Get the configured compaction strategy
+        compaction_strategy = get_compaction_strategy()
+
+        if proportion_used > compaction_threshold:
+            # RACE CONDITION PROTECTION: Check for pending tool calls before summarization
+            if compaction_strategy == "summarization" and self.has_pending_tool_calls(
+                messages
+            ):
+                pending_count = self.get_pending_tool_call_count(messages)
+                emit_warning(
+                    f"⚠️  Summarization deferred: {pending_count} pending tool call(s) detected. "
+                    "Waiting for tool execution to complete before compaction.",
+                    message_group="token_context_status",
+                )
+                # Request delayed compaction for when tool calls complete
+                self.request_delayed_compaction()
+                # Return original messages without compaction
+                return messages, []
+
+            if compaction_strategy == "truncation":
+                # Use truncation instead of summarization
+                protected_tokens = get_protected_token_count()
+                filtered_messages = self.filter_huge_messages(messages)
+                result_messages = self.truncation(filtered_messages, protected_tokens)
+                # Track dropped messages by hash so message_history_accumulator
+                # won't re-inject them from pydantic-ai's full message list on
+                # subsequent calls within the same run (fixes ghost-task bug).
+                result_hashes = {self.hash_message(m) for m in result_messages}
+                summarized_messages = [
+                    m
+                    for m in filtered_messages
+                    if self.hash_message(m) not in result_hashes
+                ]
+            else:
+                # Default to summarization (safe to proceed - no pending tool calls)
+                result_messages, summarized_messages = self.summarize_messages(
+                    self.filter_huge_messages(messages)
+                )
+
+            final_token_count = sum(
+                self.estimate_tokens_for_message(msg) for msg in result_messages
+            )
+            # Update spinner with final token count
+            final_summary = SpinnerBase.format_context_info(
+                final_token_count, model_max, final_token_count / model_max
+            )
+            update_spinner_context(final_summary)
+
+            self.set_message_history(result_messages)
+            for m in summarized_messages:
+                self.add_compacted_message_hash(self.hash_message(m))
+            return result_messages
+        return messages
+
+    def truncation(
+        self, messages: List[ModelMessage], protected_tokens: int
+    ) -> List[ModelMessage]:
+        """
+        Truncate message history to manage token usage.
+
+        Protects:
+        - The first message (system prompt) - always kept
+        - The second message if it contains a ThinkingPart (extended thinking context)
+        - The most recent messages up to protected_tokens
+
+        Args:
+            messages: List of messages to truncate
+            protected_tokens: Number of tokens to protect
+
+        Returns:
+            Truncated list of messages
+        """
+        import queue
+
+        emit_info("Truncating message history to manage token usage")
+        result = [messages[0]]  # Always keep the first message (system prompt)
+
+        # Check if second message exists and contains a ThinkingPart
+        # If so, protect it (extended thinking context shouldn't be lost)
+        skip_second = False
+        if len(messages) > 1:
+            second_msg = messages[1]
+            has_thinking = any(
+                isinstance(part, ThinkingPart) for part in second_msg.parts
+            )
+            if has_thinking:
+                result.append(second_msg)
+                skip_second = True
+
+        num_tokens = 0
+        stack = queue.LifoQueue()
+
+        # Determine which messages to consider for the recent-tokens window
+        # Skip first message (already added), and skip second if it has thinking
+        start_idx = 2 if skip_second else 1
+        messages_to_scan = messages[start_idx:]
+
+        # Put messages in reverse order (most recent first) into the stack
+        # but break when we exceed protected_tokens
+        for msg in reversed(messages_to_scan):
+            num_tokens += self.estimate_tokens_for_message(msg)
+            if num_tokens > protected_tokens:
+                break
+            stack.put(msg)
+
+        # Pop messages from stack to get them in chronological order
+        while not stack.empty():
+            result.append(stack.get())
+
+        result = self.prune_interrupted_tool_calls(result)
+        return result
+
+    def run_summarization_sync(
+        self,
+        instructions: str,
+        message_history: List[ModelMessage],
+    ) -> Union[List[ModelMessage], str]:
+        """
+        Run summarization synchronously using the configured summarization agent.
+        This is exposed as a method so it can be overridden by subclasses if needed.
+
+        Args:
+            instructions: Instructions for the summarization agent
+            message_history: List of messages to summarize
+
+        Returns:
+            Summarized messages or text
+        """
+        return run_summarization_sync(instructions, message_history)
+
+    # ===== Agent wiring formerly in code_puppy/agent.py =====
+    def load_puppy_rules(self) -> Optional[str]:
+        """Load AGENT(S).md from both global config and project directory.
+
+        Checks for AGENTS.md/AGENT.md/agents.md/agent.md in this order:
+        1. Global config directory (~/.code_puppy/ or XDG config)
+        2. Current working directory (project-specific)
+
+        If both exist, they are combined with global rules first, then project rules.
+        This allows project-specific rules to override or extend global rules.
+        """
+        if self._puppy_rules is not None:
+            return self._puppy_rules
+        from pathlib import Path
+
+        possible_paths = ["AGENTS.md", "AGENT.md", "agents.md", "agent.md"]
+
+        # Load global rules from CONFIG_DIR
+        global_rules = None
+        from code_puppy.config import CONFIG_DIR
+
+        for path_str in possible_paths:
+            global_path = Path(CONFIG_DIR) / path_str
+            if global_path.exists():
+                global_rules = global_path.read_text(encoding="utf-8-sig")
+                break
+
+        # Load project-local rules from current working directory
+        project_rules = None
+        for path_str in possible_paths:
+            project_path = Path(path_str)
+            if project_path.exists():
+                project_rules = project_path.read_text(encoding="utf-8-sig")
+                break
+
+        # Combine global and project rules
+        # Global rules come first, project rules second (allowing project to override)
+        rules = [r for r in [global_rules, project_rules] if r]
+        self._puppy_rules = "\n\n".join(rules) if rules else None
+        return self._puppy_rules
+
+    def load_mcp_servers(self, extra_headers: Optional[Dict[str, str]] = None):
+        """Load MCP servers through the manager and return pydantic-ai compatible servers.
+
+        Note: The manager automatically syncs from mcp_servers.json during initialization,
+        so we don't need to sync here. Use reload_mcp_servers() to force a re-sync.
+        """
+
+        mcp_disabled = get_value("disable_mcp_servers")
+        if mcp_disabled and str(mcp_disabled).lower() in ("1", "true", "yes", "on"):
+            return []
+
+        manager = get_mcp_manager()
+        return manager.get_servers_for_agent()
+
+    def reload_mcp_servers(self):
+        """Reload MCP servers and return updated servers.
+
+        Forces a re-sync from mcp_servers.json to pick up any configuration changes.
+        """
+        # Clear the MCP tool cache when servers are reloaded
+        self._mcp_tool_definitions_cache = []
+
+        # Force re-sync from mcp_servers.json
+        manager = get_mcp_manager()
+        manager.sync_from_config()
+
+        return manager.get_servers_for_agent()
+
+    def _load_model_with_fallback(
+        self,
+        requested_model_name: str,
+        models_config: Dict[str, Any],
+        message_group: str,
+    ) -> Tuple[Any, str]:
+        """Load the requested model, applying a friendly fallback when unavailable."""
+        try:
+            model = ModelFactory.get_model(requested_model_name, models_config)
+            return model, requested_model_name
+        except ValueError as exc:
+            available_models = list(models_config.keys())
+            available_str = (
+                ", ".join(sorted(available_models))
+                if available_models
+                else "no configured models"
+            )
+            emit_warning(
+                (
+                    f"Model '{requested_model_name}' not found. "
+                    f"Available models: {available_str}"
+                ),
+                message_group=message_group,
+            )
+
+            fallback_candidates: List[str] = []
+            global_candidate = get_global_model_name()
+            if global_candidate:
+                fallback_candidates.append(global_candidate)
+
+            for candidate in available_models:
+                if candidate not in fallback_candidates:
+                    fallback_candidates.append(candidate)
+
+            for candidate in fallback_candidates:
+                if not candidate or candidate == requested_model_name:
+                    continue
+                try:
+                    model = ModelFactory.get_model(candidate, models_config)
+                    emit_info(
+                        f"Using fallback model: {candidate}",
+                        message_group=message_group,
+                    )
+                    return model, candidate
+                except ValueError:
+                    continue
+
+            friendly_message = (
+                "No valid model could be loaded. Update the model configuration or set "
+                "a valid model with `config set`."
+            )
+            emit_error(
+                friendly_message,
+                message_group=message_group,
+            )
+            raise ValueError(friendly_message) from exc
+
+    def reload_code_generation_agent(self, message_group: Optional[str] = None):
+        """Force-reload the pydantic-ai Agent based on current config and model."""
+        from code_puppy.tools import (
+            EXTENDED_THINKING_PROMPT_NOTE,
+            has_extended_thinking_active,
+            register_tools_for_agent,
+        )
+
+        # Invalidate the project-local rules cache so a fresh read from the
+        # current working directory is performed on the next load_puppy_rules()
+        # call.  This is critical for /cd: the user may have switched to a
+        # different project that has its own AGENT.md (or none at all).
+        self._puppy_rules = None
+
+        if message_group is None:
+            message_group = str(uuid.uuid4())
+
+        model_name = self.get_model_name()
+
+        models_config = ModelFactory.load_config()
+        model, resolved_model_name = self._load_model_with_fallback(
+            model_name,
+            models_config,
+            message_group,
+        )
+
+        instructions = self.get_full_system_prompt()
+        puppy_rules = self.load_puppy_rules()
+        if puppy_rules:
+            instructions += f"\n{puppy_rules}"
+
+        mcp_servers = self.load_mcp_servers()
+
+        model_settings = make_model_settings(resolved_model_name)
+
+        # Handle claude-code models: swap instructions (prompt prepending happens in run_with_mcp)
+        from code_puppy.model_utils import prepare_prompt_for_model
+
+        # When extended thinking is active, nudge the model to think between
+        # tool calls (the share_your_reasoning tool is stripped in this case).
+        if has_extended_thinking_active(resolved_model_name):
+            instructions += EXTENDED_THINKING_PROMPT_NOTE
+
+        prepared = prepare_prompt_for_model(
+            model_name, instructions, "", prepend_system_to_user=False
+        )
+        instructions = prepared.instructions
+
+        self.cur_model = model
+        p_agent = PydanticAgent(
+            model=model,
+            instructions=instructions,
+            output_type=str,
+            retries=3,
+            toolsets=mcp_servers,
+            history_processors=[self.message_history_accumulator],
+            model_settings=model_settings,
+        )
+
+        agent_tools = self.get_available_tools()
+        register_tools_for_agent(p_agent, agent_tools, model_name=resolved_model_name)
+
+        # Get existing tool names to filter out conflicts with MCP tools
+        existing_tool_names = set()
+        try:
+            # Get tools from the agent to find existing tool names
+            tools = getattr(p_agent, "_tools", None)
+            if tools:
+                existing_tool_names = set(tools.keys())
+        except Exception:
+            # If we can't get tool names, proceed without filtering
+            pass
+
+        # Filter MCP server toolsets to remove conflicting tools
+        filtered_mcp_servers = []
+        if mcp_servers and existing_tool_names:
+            for mcp_server in mcp_servers:
+                try:
+                    # Get tools from this MCP server
+                    server_tools = getattr(mcp_server, "tools", None)
+                    if server_tools:
+                        # Filter out conflicting tools
+                        filtered_tools = {}
+                        for tool_name, tool_func in server_tools.items():
+                            if tool_name not in existing_tool_names:
+                                filtered_tools[tool_name] = tool_func
+
+                        # Create a filtered version of the MCP server if we have tools
+                        if filtered_tools:
+                            # Create a new toolset with filtered tools
+                            from pydantic_ai.tools import ToolSet
+
+                            filtered_toolset = ToolSet()
+                            for tool_name, tool_func in filtered_tools.items():
+                                filtered_toolset._tools[tool_name] = tool_func
+                            filtered_mcp_servers.append(filtered_toolset)
+                        else:
+                            # No tools left after filtering, skip this server
+                            pass
+                    else:
+                        # Can't get tools from this server, include as-is
+                        filtered_mcp_servers.append(mcp_server)
+                except Exception:
+                    # Error processing this server, include as-is to be safe
+                    filtered_mcp_servers.append(mcp_server)
+        else:
+            # No filtering needed or possible
+            filtered_mcp_servers = mcp_servers if mcp_servers else []
+
+        if len(filtered_mcp_servers) != len(mcp_servers):
+            emit_info(
+                Text.from_markup(
+                    f"[dim]Filtered {len(mcp_servers) - len(filtered_mcp_servers)} conflicting MCP tools[/dim]"
+                )
+            )
+
+        self._last_model_name = resolved_model_name
+        # expose for run_with_mcp
+        # Wrap it with DBOS, but handle MCP servers separately to avoid serialization issues
+        global _reload_count
+        _reload_count += 1
+        if get_use_dbos():
+            # Don't pass MCP servers to the agent constructor when using DBOS
+            # This prevents the "cannot pickle async_generator object" error
+            # MCP servers will be handled separately in run_with_mcp
+            agent_without_mcp = PydanticAgent(
+                model=model,
+                instructions=instructions,
+                output_type=str,
+                retries=3,
+                toolsets=[],  # Don't include MCP servers here
+                history_processors=[self.message_history_accumulator],
+                model_settings=model_settings,
+            )
+
+            # Register regular tools (non-MCP) on the new agent
+            agent_tools = self.get_available_tools()
+            register_tools_for_agent(
+                agent_without_mcp, agent_tools, model_name=resolved_model_name
+            )
+
+            # Wrap with DBOS - pass event_stream_handler at construction time
+            # so DBOSModel gets the handler for streaming output
+            dbos_agent = DBOSAgent(
+                agent_without_mcp,
+                name=f"{self.name}-{_reload_count}",
+                event_stream_handler=event_stream_handler,
+            )
+            self.pydantic_agent = dbos_agent
+            self._code_generation_agent = dbos_agent
+
+            # Store filtered MCP servers separately for runtime use
+            self._mcp_servers = filtered_mcp_servers
+        else:
+            # Normal path without DBOS - include filtered MCP servers in the agent
+            # Re-create agent with filtered MCP servers
+            p_agent = PydanticAgent(
+                model=model,
+                instructions=instructions,
+                output_type=str,
+                retries=3,
+                toolsets=filtered_mcp_servers,
+                history_processors=[self.message_history_accumulator],
+                model_settings=model_settings,
+            )
+            # Register regular tools on the agent
+            agent_tools = self.get_available_tools()
+            register_tools_for_agent(
+                p_agent, agent_tools, model_name=resolved_model_name
+            )
+
+            self.pydantic_agent = p_agent
+            self._code_generation_agent = p_agent
+            self._mcp_servers = filtered_mcp_servers
+        return self._code_generation_agent
+
+    def _create_agent_with_output_type(self, output_type: Type[Any]) -> PydanticAgent:
+        """Create a temporary agent configured with a custom output_type.
+
+        This is used when structured output is requested via run_with_mcp.
+        The agent is created fresh with the same configuration as the main agent
+        but with the specified output_type instead of str.
+
+        Args:
+            output_type: The Pydantic model or type for structured output.
+
+        Returns:
+            A configured PydanticAgent (or DBOSAgent wrapper) with the custom output_type.
+        """
+        from code_puppy.model_utils import prepare_prompt_for_model
+        from code_puppy.tools import (
+            EXTENDED_THINKING_PROMPT_NOTE,
+            has_extended_thinking_active,
+            register_tools_for_agent,
+        )
+
+        model_name = self.get_model_name()
+        models_config = ModelFactory.load_config()
+        model, resolved_model_name = self._load_model_with_fallback(
+            model_name, models_config, str(uuid.uuid4())
+        )
+
+        instructions = self.get_full_system_prompt()
+        puppy_rules = self.load_puppy_rules()
+        if puppy_rules:
+            instructions += f"\n{puppy_rules}"
+
+        mcp_servers = getattr(self, "_mcp_servers", []) or []
+        model_settings = make_model_settings(resolved_model_name)
+
+        prepared = prepare_prompt_for_model(
+            model_name, instructions, "", prepend_system_to_user=False
+        )
+        instructions = prepared.instructions
+
+        # When extended thinking is active, nudge the model to think between
+        # tool calls (the share_your_reasoning tool is stripped in this case).
+        if has_extended_thinking_active(resolved_model_name):
+            instructions += EXTENDED_THINKING_PROMPT_NOTE
+
+        global _reload_count
+        _reload_count += 1
+
+        if get_use_dbos():
+            temp_agent = PydanticAgent(
+                model=model,
+                instructions=instructions,
+                output_type=output_type,
+                retries=3,
+                toolsets=[],
+                history_processors=[self.message_history_accumulator],
+                model_settings=model_settings,
+            )
+            agent_tools = self.get_available_tools()
+            register_tools_for_agent(
+                temp_agent, agent_tools, model_name=resolved_model_name
+            )
+            # Pass event_stream_handler at construction time for streaming output
+            dbos_agent = DBOSAgent(
+                temp_agent,
+                name=f"{self.name}-structured-{_reload_count}",
+                event_stream_handler=event_stream_handler,
+            )
+            return dbos_agent
+        else:
+            temp_agent = PydanticAgent(
+                model=model,
+                instructions=instructions,
+                output_type=output_type,
+                retries=3,
+                toolsets=mcp_servers,
+                history_processors=[self.message_history_accumulator],
+                model_settings=model_settings,
+            )
+            agent_tools = self.get_available_tools()
+            register_tools_for_agent(
+                temp_agent, agent_tools, model_name=resolved_model_name
+            )
+            return temp_agent
+
+    # It's okay to decorate it with DBOS.step even if not using DBOS; the decorator is a no-op in that case.
+    @DBOS.step()
+    def message_history_accumulator(self, ctx: RunContext, messages: List[Any]):
+        _message_history = self.get_message_history()
+
+        # Hook: on_message_history_processor_start - dump the message history before processing
+        on_message_history_processor_start(
+            agent_name=self.name,
+            session_id=getattr(self, "session_id", None),
+            message_history=list(_message_history),  # Copy to avoid mutation issues
+            incoming_messages=list(messages),
+        )
+        message_history_hashes = set([self.hash_message(m) for m in _message_history])
+        messages_added = 0
+        last_msg_index = len(messages) - 1
+        for i, msg in enumerate(messages):
+            msg_hash = self.hash_message(msg)
+            if msg_hash not in message_history_hashes:
+                # Always preserve the last message (the user's new prompt) even
+                # if its hash matches a previously compacted/summarized message.
+                # Short or repeated prompts (e.g. "yes", "1") can collide with
+                # compacted hashes, which would silently drop the user's input
+                # and leave the history ending with a ModelResponse.  That
+                # triggers an Anthropic API error: "This model does not support
+                # assistant message prefill."
+                if (
+                    i == last_msg_index
+                    or msg_hash not in self.get_compacted_message_hashes()
+                ):
+                    _message_history.append(msg)
+                    messages_added += 1
+
+        # Apply message history trimming using the main processor
+        # This ensures we maintain global state while still managing context limits
+        self.message_history_processor(ctx, _message_history)
+        result_messages_filtered_empty_thinking = []
+        filtered_count = 0
+        for msg in self.get_message_history():
+            # Filter out single-part messages that are empty ThinkingParts
+            if len(msg.parts) == 1 and isinstance(msg.parts[0], ThinkingPart):
+                if not msg.parts[0].content:
+                    filtered_count += 1
+                    continue
+            # For multi-part messages, strip empty ThinkingParts but keep the message
+            elif any(isinstance(p, ThinkingPart) and not p.content for p in msg.parts):
+                msg = dataclasses.replace(
+                    msg,
+                    parts=[
+                        p
+                        for p in msg.parts
+                        if not (isinstance(p, ThinkingPart) and not p.content)
+                    ],
+                )
+                if not msg.parts:
+                    filtered_count += 1
+                    continue
+            result_messages_filtered_empty_thinking.append(msg)
+        self.set_message_history(result_messages_filtered_empty_thinking)
+
+        # Safety net: ensure history always ends with a ModelRequest.
+        # If compaction or filtering somehow leaves a trailing ModelResponse,
+        # the Anthropic API will reject it with a prefill error.
+        final_history = self.ensure_history_ends_with_request(
+            self.get_message_history()
+        )
+        if final_history != self.get_message_history():
+            self.set_message_history(final_history)
+
+        # Hook: on_message_history_processor_end - dump the message history after processing
+        messages_filtered = len(messages) - messages_added + filtered_count
+        on_message_history_processor_end(
+            agent_name=self.name,
+            session_id=getattr(self, "session_id", None),
+            message_history=list(final_history),  # Copy to avoid mutation issues
+            messages_added=messages_added,
+            messages_filtered=messages_filtered,
+        )
+
+        return final_history
+
+    def _spawn_ctrl_x_key_listener(
+        self,
+        stop_event: threading.Event,
+        on_escape: Callable[[], None],
+        on_cancel_agent: Optional[Callable[[], None]] = None,
+    ) -> Optional[threading.Thread]:
+        """Start a keyboard listener thread for CLI sessions.
+
+        Listens for Ctrl+X (shell command cancel) and optionally the configured
+        cancel_agent_key (when not using SIGINT/Ctrl+C).
+
+        Args:
+            stop_event: Event to signal the listener to stop.
+            on_escape: Callback for Ctrl+X (shell command cancel).
+            on_cancel_agent: Optional callback for cancel_agent_key (only used
+                when cancel_agent_uses_signal() returns False).
+        """
+        try:
+            import sys
+        except ImportError:
+            return None
+
+        stdin = getattr(sys, "stdin", None)
+        if stdin is None or not hasattr(stdin, "isatty"):
+            return None
+        try:
+            if not stdin.isatty():
+                return None
+        except Exception:
+            return None
+
+        def listener() -> None:
+            try:
+                if sys.platform.startswith("win"):
+                    self._listen_for_ctrl_x_windows(
+                        stop_event, on_escape, on_cancel_agent
+                    )
+                else:
+                    self._listen_for_ctrl_x_posix(
+                        stop_event, on_escape, on_cancel_agent
+                    )
+            except Exception:
+                emit_warning(
+                    "Key listener stopped unexpectedly; press Ctrl+C to cancel."
+                )
+
+        thread = threading.Thread(
+            target=listener, name="code-puppy-key-listener", daemon=True
+        )
+        thread.start()
+        return thread
+
+    def _listen_for_ctrl_x_windows(
+        self,
+        stop_event: threading.Event,
+        on_escape: Callable[[], None],
+        on_cancel_agent: Optional[Callable[[], None]] = None,
+    ) -> None:
+        import msvcrt
+        import time
+
+        # Get the cancel agent char code if we're using keyboard-based cancel
+        cancel_agent_char: Optional[str] = None
+        if on_cancel_agent is not None and not cancel_agent_uses_signal():
+            cancel_agent_char = get_cancel_agent_char_code()
+
+        while not stop_event.is_set():
+            try:
+                if msvcrt.kbhit():
+                    key = msvcrt.getwch()
+                    if key == "\x18":  # Ctrl+X
+                        try:
+                            on_escape()
+                        except Exception:
+                            emit_warning(
+                                "Ctrl+X handler raised unexpectedly; Ctrl+C still works."
+                            )
+                    elif (
+                        cancel_agent_char
+                        and on_cancel_agent
+                        and key == cancel_agent_char
+                    ):
+                        try:
+                            on_cancel_agent()
+                        except Exception:
+                            emit_warning("Cancel agent handler raised unexpectedly.")
+            except Exception:
+                emit_warning(
+                    "Windows key listener error; Ctrl+C is still available for cancel."
+                )
+                return
+            time.sleep(0.05)
+
+    def _listen_for_ctrl_x_posix(
+        self,
+        stop_event: threading.Event,
+        on_escape: Callable[[], None],
+        on_cancel_agent: Optional[Callable[[], None]] = None,
+    ) -> None:
+        import select
+        import sys
+        import termios
+        import tty
+
+        # Get the cancel agent char code if we're using keyboard-based cancel
+        cancel_agent_char: Optional[str] = None
+        if on_cancel_agent is not None and not cancel_agent_uses_signal():
+            cancel_agent_char = get_cancel_agent_char_code()
+
+        stdin = sys.stdin
+        try:
+            fd = stdin.fileno()
+        except (AttributeError, ValueError, OSError):
+            return
+        try:
+            original_attrs = termios.tcgetattr(fd)
+        except Exception:
+            return
+
+        try:
+            tty.setcbreak(fd)
+            while not stop_event.is_set():
+                try:
+                    read_ready, _, _ = select.select([stdin], [], [], 0.05)
+                except Exception:
+                    break
+                if not read_ready:
+                    continue
+                data = stdin.read(1)
+                if not data:
+                    break
+                if data == "\x18":  # Ctrl+X
+                    try:
+                        on_escape()
+                    except Exception:
+                        emit_warning(
+                            "Ctrl+X handler raised unexpectedly; Ctrl+C still works."
+                        )
+                elif (
+                    cancel_agent_char and on_cancel_agent and data == cancel_agent_char
+                ):
+                    try:
+                        on_cancel_agent()
+                    except Exception:
+                        emit_warning("Cancel agent handler raised unexpectedly.")
+        finally:
+            termios.tcsetattr(fd, termios.TCSADRAIN, original_attrs)
+
+    async def run_with_mcp(
+        self,
+        prompt: str,
+        *,
+        attachments: Optional[Sequence[BinaryContent]] = None,
+        link_attachments: Optional[Sequence[Union[ImageUrl, DocumentUrl]]] = None,
+        output_type: Optional[Type[Any]] = None,
+        **kwargs,
+    ) -> Any:
+        """Run the agent with MCP servers, attachments, and full cancellation support.
+
+        Args:
+            prompt: Primary user prompt text (may be empty when attachments present).
+            attachments: Local binary payloads (e.g., dragged images) to include.
+            link_attachments: Remote assets (image/document URLs) to include.
+            output_type: Optional Pydantic model or type for structured output.
+                When provided, creates a temporary agent configured to return
+                this type instead of the default string output.
+            **kwargs: Additional arguments forwarded to `pydantic_ai.Agent.run`.
+
+        Returns:
+            The agent's response (typed according to output_type if specified).
+
+        Raises:
+            asyncio.CancelledError: When execution is cancelled by user.
+        """
+        # Sanitize prompt to remove invalid Unicode surrogates that can cause
+        # encoding errors (especially common on Windows with copy-paste)
+        if prompt:
+            try:
+                prompt = prompt.encode("utf-8", errors="surrogatepass").decode(
+                    "utf-8", errors="replace"
+                )
+            except (UnicodeEncodeError, UnicodeDecodeError):
+                # Fallback: filter out surrogate characters directly
+                prompt = "".join(
+                    char if ord(char) < 0xD800 or ord(char) > 0xDFFF else "\ufffd"
+                    for char in prompt
+                )
+
+        group_id = str(uuid.uuid4())
+        # Avoid double-loading: reuse existing agent if already built
+        pydantic_agent = (
+            self._code_generation_agent or self.reload_code_generation_agent()
+        )
+
+        # If a custom output_type is specified, create a temporary agent with that type
+        if output_type is not None:
+            pydantic_agent = self._create_agent_with_output_type(output_type)
+
+        # Handle model-specific prompt transformations via prepare_prompt_for_model()
+        # This uses the get_model_system_prompt hook, so plugins can register their own handlers
+        from code_puppy.model_utils import prepare_prompt_for_model
+
+        # Only prepend system prompt on first message (empty history)
+        should_prepend = len(self.get_message_history()) == 0
+        if should_prepend:
+            system_prompt = self.get_full_system_prompt()
+            puppy_rules = self.load_puppy_rules()
+            if puppy_rules:
+                system_prompt += f"\n{puppy_rules}"
+
+            prepared = prepare_prompt_for_model(
+                model_name=self.get_model_name(),
+                system_prompt=system_prompt,
+                user_prompt=prompt,
+                prepend_system_to_user=True,
+            )
+            prompt = prepared.user_prompt
+
+        # Build combined prompt payload when attachments are provided.
+        attachment_parts: List[Any] = []
+        if attachments:
+            attachment_parts.extend(list(attachments))
+        if link_attachments:
+            attachment_parts.extend(list(link_attachments))
+
+        if attachment_parts:
+            prompt_payload: Union[str, List[Any]] = []
+            if prompt:
+                prompt_payload.append(prompt)
+            prompt_payload.extend(attachment_parts)
+        else:
+            prompt_payload = prompt
+
+        async def run_agent_task():
+            try:
+                self.set_message_history(
+                    self.prune_interrupted_tool_calls(self.get_message_history())
+                )
+
+                # DELAYED COMPACTION: Check if we should attempt delayed compaction
+                if self.should_attempt_delayed_compaction():
+                    emit_info(
+                        "🔄 Attempting delayed compaction (tool calls completed)",
+                        message_group="token_context_status",
+                    )
+                    current_messages = self.get_message_history()
+                    compacted_messages, _ = self.compact_messages(current_messages)
+                    if compacted_messages != current_messages:
+                        self.set_message_history(compacted_messages)
+                        emit_info(
+                            "✅ Delayed compaction completed successfully",
+                            message_group="token_context_status",
+                        )
+
+                usage_limits = UsageLimits(request_limit=get_message_limit())
+
+                # Handle MCP servers - add them temporarily when using DBOS
+                if (
+                    get_use_dbos()
+                    and hasattr(self, "_mcp_servers")
+                    and self._mcp_servers
+                ):
+                    # Temporarily add MCP servers to the DBOS agent using internal _toolsets
+                    original_toolsets = pydantic_agent._toolsets
+                    pydantic_agent._toolsets = original_toolsets + self._mcp_servers
+
+                    try:
+                        # Set the workflow ID for DBOS context so DBOS and Code Puppy ID match
+                        with SetWorkflowID(group_id):
+                            result_ = await pydantic_agent.run(
+                                prompt_payload,
+                                message_history=self.get_message_history(),
+                                usage_limits=usage_limits,
+                                event_stream_handler=event_stream_handler,
+                                **kwargs,
+                            )
+                            return result_
+                    finally:
+                        # Always restore original toolsets
+                        pydantic_agent._toolsets = original_toolsets
+                elif get_use_dbos():
+                    with SetWorkflowID(group_id):
+                        result_ = await pydantic_agent.run(
+                            prompt_payload,
+                            message_history=self.get_message_history(),
+                            usage_limits=usage_limits,
+                            event_stream_handler=event_stream_handler,
+                            **kwargs,
+                        )
+                        return result_
+                else:
+                    # Non-DBOS path (MCP servers are already included)
+                    result_ = await pydantic_agent.run(
+                        prompt_payload,
+                        message_history=self.get_message_history(),
+                        usage_limits=usage_limits,
+                        event_stream_handler=event_stream_handler,
+                        **kwargs,
+                    )
+                    return result_
+            except* UsageLimitExceeded as ule:
+                emit_info(f"Usage limit exceeded: {str(ule)}", group_id=group_id)
+                emit_info(
+                    "The agent has reached its usage limit. You can ask it to continue by saying 'please continue' or similar.",
+                    group_id=group_id,
+                )
+            except* mcp.shared.exceptions.McpError as mcp_error:
+                emit_info(f"MCP server error: {str(mcp_error)}", group_id=group_id)
+                emit_info(f"{str(mcp_error)}", group_id=group_id)
+                emit_info(
+                    "Try disabling any malfunctioning MCP servers", group_id=group_id
+                )
+            except* asyncio.exceptions.CancelledError:
+                emit_info("Cancelled")
+                if get_use_dbos():
+                    await DBOS.cancel_workflow_async(group_id)
+            except* InterruptedError as ie:
+                emit_info(f"Interrupted: {str(ie)}")
+                if get_use_dbos():
+                    await DBOS.cancel_workflow_async(group_id)
+            except* Exception as other_error:
+                # Filter out CancelledError and UsageLimitExceeded from the exception group - let it propagate
+                remaining_exceptions = []
+
+                def collect_non_cancelled_exceptions(exc):
+                    if isinstance(exc, ExceptionGroup):
+                        for sub_exc in exc.exceptions:
+                            collect_non_cancelled_exceptions(sub_exc)
+                    elif not isinstance(
+                        exc, (asyncio.CancelledError, UsageLimitExceeded)
+                    ):
+                        remaining_exceptions.append(exc)
+                        emit_info(f"Unexpected error: {str(exc)}", group_id=group_id)
+                        emit_info(f"{str(exc.args)}", group_id=group_id)
+                        # Log to file for debugging
+                        log_error(
+                            exc,
+                            context=f"Agent run (group_id={group_id})",
+                            include_traceback=True,
+                        )
+
+                collect_non_cancelled_exceptions(other_error)
+
+                # If there are CancelledError exceptions in the group, re-raise them
+                cancelled_exceptions = []
+
+                def collect_cancelled_exceptions(exc):
+                    if isinstance(exc, ExceptionGroup):
+                        for sub_exc in exc.exceptions:
+                            collect_cancelled_exceptions(sub_exc)
+                    elif isinstance(exc, asyncio.CancelledError):
+                        cancelled_exceptions.append(exc)
+
+                collect_cancelled_exceptions(other_error)
+            finally:
+                self.set_message_history(
+                    self.prune_interrupted_tool_calls(self.get_message_history())
+                )
+
+        # Create the task FIRST
+        agent_task = asyncio.create_task(run_agent_task())
+
+        # Fire agent_run_start hook - plugins can use this to start background tasks
+        # (e.g., token refresh heartbeats for OAuth models)
+        try:
+            await on_agent_run_start(
+                agent_name=self.name,
+                model_name=self.get_model_name(),
+                session_id=group_id,
+            )
+        except Exception:
+            pass  # Don't fail agent run if hook fails
+
+        loop = asyncio.get_running_loop()
+
+        def schedule_agent_cancel() -> None:
+            from code_puppy.tools.command_runner import _RUNNING_PROCESSES
+
+            if len(_RUNNING_PROCESSES):
+                emit_warning(
+                    "Refusing to cancel Agent while a shell command is currently running - press Ctrl+X to cancel the shell command."
+                )
+                return
+            if agent_task.done():
+                return
+
+            # Cancel all active subagent tasks
+            if _active_subagent_tasks:
+                emit_warning(
+                    f"Cancelling {len(_active_subagent_tasks)} active subagent task(s)..."
+                )
+                for task in list(
+                    _active_subagent_tasks
+                ):  # Create a copy since we'll be modifying the set
+                    if not task.done():
+                        loop.call_soon_threadsafe(task.cancel)
+            loop.call_soon_threadsafe(agent_task.cancel)
+
+        def keyboard_interrupt_handler(_sig, _frame):
+            # If we're awaiting user input (e.g., file permission prompt),
+            # don't cancel the agent - let the input() call handle the interrupt naturally
+            if is_awaiting_user_input():
+                # Don't do anything here - let the input() call raise KeyboardInterrupt naturally
+                return
+
+            schedule_agent_cancel()
+
+        def graceful_sigint_handler(_sig, _frame):
+            # When using keyboard-based cancel, SIGINT should be a no-op
+            # (just show a hint to user about the configured cancel key)
+            # Also reset terminal to prevent bricking on Windows+uvx
+            from code_puppy.keymap import get_cancel_agent_display_name
+            from code_puppy.terminal_utils import reset_windows_terminal_full
+
+            # Reset terminal state first to prevent bricking
+            reset_windows_terminal_full()
+
+            cancel_key = get_cancel_agent_display_name()
+            emit_info(f"Use {cancel_key} to cancel the agent task.")
+
+        original_handler = None
+        key_listener_stop_event = None
+        _key_listener_thread = None
+
+        try:
+            if cancel_agent_uses_signal():
+                # Use SIGINT-based cancellation (default Ctrl+C behavior)
+                original_handler = signal.signal(
+                    signal.SIGINT, keyboard_interrupt_handler
+                )
+            else:
+                # Use keyboard listener for agent cancellation
+                # Set a graceful SIGINT handler that shows a hint
+                original_handler = signal.signal(signal.SIGINT, graceful_sigint_handler)
+                # Spawn keyboard listener with the cancel agent callback
+                key_listener_stop_event = threading.Event()
+                _key_listener_thread = self._spawn_ctrl_x_key_listener(
+                    key_listener_stop_event,
+                    on_escape=lambda: None,  # Ctrl+X handled by command_runner
+                    on_cancel_agent=schedule_agent_cancel,
+                )
+
+            # Wait for the task to complete or be cancelled
+            result = await agent_task
+
+            # Update MCP tool cache after successful run for accurate token estimation
+            if hasattr(self, "_mcp_servers") and self._mcp_servers:
+                try:
+                    await self._update_mcp_tool_cache()
+                except Exception:
+                    pass  # Don't fail the run if cache update fails
+
+            # Extract response text for the callback
+            _run_response_text = ""
+            if result is not None:
+                if hasattr(result, "data"):
+                    _run_response_text = str(result.data) if result.data else ""
+                elif hasattr(result, "output"):
+                    _run_response_text = str(result.output) if result.output else ""
+                else:
+                    _run_response_text = str(result)
+
+            _run_success = True
+            _run_error = None
+            return result
+        except asyncio.CancelledError:
+            _run_success = False
+            _run_error = None  # Cancellation is not an error
+            _run_response_text = ""
+            agent_task.cancel()
+        except KeyboardInterrupt:
+            _run_success = False
+            _run_error = None  # User interrupt is not an error
+            _run_response_text = ""
+            if not agent_task.done():
+                agent_task.cancel()
+        except Exception as e:
+            _run_success = False
+            _run_error = e
+            _run_response_text = ""
+            raise
+        finally:
+            # Fire agent_run_end hook - plugins can use this for:
+            # - Stopping background tasks (token refresh heartbeats)
+            # - Workflow orchestration (Ralph's autonomous loop)
+            # - Logging/analytics
+            try:
+                await on_agent_run_end(
+                    agent_name=self.name,
+                    model_name=self.get_model_name(),
+                    session_id=group_id,
+                    success=_run_success,
+                    error=_run_error,
+                    response_text=_run_response_text,
+                    metadata={"model": self.get_model_name()},
+                )
+            except Exception:
+                pass  # Don't fail cleanup if hook fails
+
+            # Stop keyboard listener if it was started
+            if key_listener_stop_event is not None:
+                key_listener_stop_event.set()
+            # Restore original signal handler
+            if (
+                original_handler is not None
+            ):  # Explicit None check - SIG_DFL can be 0/falsy!
+                signal.signal(signal.SIGINT, original_handler)