qwen_agent/agent/deep_assistant.py

import json
import logging
import time
import copy
import os
import tempfile
import asyncio
from pathlib import Path
from typing import Any, Dict
from langchain.chat_models import init_chat_model
from deepagents import create_deep_agent
from deepagents.backends import CompositeBackend, LocalShellBackend
from deepagents.backends.filesystem import FilesystemBackend
from deepagents.backends.sandbox import SandboxBackendProtocol
from deepagents_cli.agent import create_cli_agent
from langchain.agents import create_agent
from langgraph.store.base import BaseStore
from langchain.agents.middleware import SummarizationMiddleware as LangchainSummarizationMiddleware
from .summarization_middleware import SummarizationMiddleware
from langchain_mcp_adapters.client import MultiServerMCPClient
from utils.fastapi_utils import detect_provider, sanitize_model_kwargs
from .guideline_middleware import GuidelineMiddleware
from .tool_output_length_middleware import ToolOutputLengthMiddleware
from .tool_use_cleanup_middleware import ToolUseCleanupMiddleware
from .tool_metrics_middleware import ToolMetricsMiddleware
from .filepath_fix_middleware import FilePathFixMiddleware
from .mcp_trace_meta import patch_mcp_client_session_trace_meta
from utils.settings import (
    SUMMARIZATION_MAX_TOKENS,
    SUMMARIZATION_TOKENS_TO_KEEP,
    TOOL_OUTPUT_MAX_LENGTH,
    MCP_HTTP_TIMEOUT,
    MCP_SSE_READ_TIMEOUT,
    DEFAULT_TRIM_TOKEN_LIMIT,
    DAYTONA_API_KEY,
    DAYTONA_SERVER_URL,
    DAYTONA_ENABLED,
)
from utils.token_counter import create_token_counter
from agent.agent_config import AgentConfig
from .mem0_manager import get_mem0_manager
from .mem0_middleware import create_mem0_middleware
from .mem0_config import Mem0Config
from agent.prompt_loader import load_system_prompt_async, load_mcp_settings_async
from agent.agent_memory_cache import get_memory_cache_manager
from .subagent_loader import load_subagents
from agent.plugin_hook_loader import collect_main_agent_hidden_tools
from .checkpoint_manager import get_checkpointer_manager
from langgraph.checkpoint.postgres.aio import AsyncPostgresSaver
from langgraph.checkpoint.memory import InMemorySaver
from langchain.tools import BaseTool
from langchain_core.language_models import BaseChatModel
from langgraph.pregel import Pregel
# New version imports: MemoryMiddleware and SkillsMiddleware were moved to deepagents.middleware
from deepagents.middleware import MemoryMiddleware, SkillsMiddleware
from langchain.agents.middleware import AgentMiddleware, ModelRequest, ModelResponse
from langchain.agents.middleware import HumanInTheLoopMiddleware, InterruptOnConfig, TodoListMiddleware
from langchain_core.messages import AIMessage, HumanMessage
from langgraph.types import Checkpointer
from deepagents_cli.config import settings, get_default_coding_instructions
from deepagents.middleware.filesystem import FilesystemMiddleware
from deepagents.middleware.patch_tool_calls import PatchToolCallsMiddleware
from langchain_anthropic.middleware import AnthropicPromptCachingMiddleware
from deepagents.graph import BASE_AGENT_PROMPT
# Added: LocalContextMiddleware
from deepagents_cli.local_context import LocalContextMiddleware
# Custom: FilesystemMiddleware with full SKILL.md reading support
from .custom_filesystem_middleware import CustomFilesystemMiddleware
# Sub-agent support
from deepagents.middleware.subagents import SubAgent, SubAgentMiddleware

# Global MemorySaver instance
# from langgraph.checkpoint.memory import MemorySaver
# _global_checkpointer = MemorySaver()

logger = logging.getLogger('app')


class EmptyResponseRetryMiddleware(AgentMiddleware):
    """Automatically retry when the model returns an empty response with no text and no tool calls."""

    MAX_RETRIES = 5

    def _is_empty_response(self, result: ModelResponse) -> bool:
        """Determine whether the response is empty."""
        if not result.result:
            return True
        msg = result.result[0]
        if not isinstance(msg, AIMessage):
            return False
        content = msg.content or ""
        has_text = bool(content.strip()) if isinstance(content, str) else bool(content)
        return not has_text and len(msg.tool_calls) == 0

    def wrap_model_call(self, request, handler):
        result = handler(request)
        retries = 0
        while self._is_empty_response(result) and retries < self.MAX_RETRIES:
            retries += 1
            logger.warning(f"Empty response detected, retrying ({retries}/{self.MAX_RETRIES})")
            retry_messages = list(request.messages) + [
                HumanMessage(content="Please continue your response.")
            ]
            request = request.override(messages=retry_messages)
            result = handler(request)
        return result

    async def awrap_model_call(self, request, handler):
        result = await handler(request)
        retries = 0
        while self._is_empty_response(result) and retries < self.MAX_RETRIES:
            retries += 1
            logger.warning(f"Empty response detected, retrying ({retries}/{self.MAX_RETRIES})")
            retry_messages = list(request.messages) + [
                HumanMessage(content="Please continue your response.")
            ]
            request = request.override(messages=retry_messages)
            result = await handler(request)
        return result


# Utility functions
def read_system_prompt():
    """Read the shared stateless system prompt."""
    with open("./prompt/system_prompt_default.md", "r", encoding="utf-8") as f:
        return f.read().strip()


async def get_tools_from_mcp(mcp):
    """Extract tools from MCP configuration with caching."""
    patch_mcp_client_session_trace_meta()
    start_time = time.time()
    # Defensive handling: ensure mcp is a non-empty list containing mcpServers
    if not isinstance(mcp, list) or len(mcp) == 0 or "mcpServers" not in mcp[0]:
        logger.info(f"get_tools_from_mcp: invalid mcp config, elapsed: {time.time() - start_time:.3f}s")
        return []

    # Try to load tools from cache
    cache_manager = get_memory_cache_manager()
    cached_tools = cache_manager.get_mcp_tools(mcp)
    if cached_tools is not None:
        logger.info(f"get_tools_from_mcp: cached {len(cached_tools)} tools, elapsed: {time.time() - start_time:.3f}s")
        return cached_tools

    # Deep-copy the MCP config to avoid mutating the original configuration and affecting the cache key.
    mcp_copy = copy.deepcopy(mcp)

    # Update mcp_copy[0]["mcpServers"] by renaming the type field to transport.
    # If transport is missing, default to http when url exists, otherwise stdio.
    for cfg in mcp_copy[0]["mcpServers"].values():
        if "type" in cfg:
            cfg.pop("type")
        if "transport" not in cfg:
            cfg["transport"] = "http" if "url" in cfg else "stdio"

        # Add timeout settings for MCP servers using HTTP or SSE transport.
        # If timeout values are not configured, use the global defaults.
        if cfg.get("transport") in ("http", "sse"):
            if "timeout" not in cfg:
                cfg["timeout"] = MCP_HTTP_TIMEOUT
            if "sse_read_timeout" not in cfg:
                cfg["sse_read_timeout"] = MCP_SSE_READ_TIMEOUT

    # Ensure mcp_copy[0]["mcpServers"] is a dictionary.
    if not isinstance(mcp_copy[0]["mcpServers"], dict):
        logger.info(f"get_tools_from_mcp: mcpServers is not dict, elapsed: {time.time() - start_time:.3f}s")
        return []

    try:
        mcp_client = MultiServerMCPClient(mcp_copy[0]["mcpServers"])
        mcp_tools = await mcp_client.get_tools()

        # Cache the result
        cache_manager.set_mcp_tools(mcp, mcp_tools)

        logger.info(f"get_tools_from_mcp: loaded {len(mcp_tools)} tools, elapsed: {time.time() - start_time:.3f}s")
        return mcp_tools
    except Exception as e:
        # Return an empty list on exception to avoid propagating errors to callers.
        logger.error(f"get_tools_from_mcp: error {e}, elapsed: {time.time() - start_time:.3f}s")
        return []


from utils.daytona_sync import init_daytona_sandbox, sync_sandbox_to_local


async def init_agent(config: AgentConfig):
    """
    Initialize the agent with persistent memory and conversation summarization support.

    Note: the agent itself is no longer cached; only mcp_tools are cached.
    Returns an (agent, checkpointer) tuple, and the caller must release the checkpointer after use.

    Args:
        config: AgentConfig object containing all initialization parameters

    Returns:
        (agent, checkpointer) tuple
    """

    create_start = time.time()

# Load configuration concurrently
    final_system_prompt, final_mcp_settings = await asyncio.gather(
        load_system_prompt_async(config),
        load_mcp_settings_async(config),
    )
    logger.info(f"init_agent config loaded, elapsed: {time.time() - create_start:.3f}s")

    mcp_settings = final_mcp_settings if final_mcp_settings else []
    system_prompt = final_system_prompt if final_system_prompt else read_system_prompt()

    config.system_prompt = system_prompt
    config.mcp_settings = mcp_settings

    workspace_root = str(Path.cwd() / "projects" /"robot"/ config.bot_id)
    local_workspace_root = workspace_root

    # Run high-cost I/O in parallel: MCP tools loading + Daytona sandbox initialization
    mcp_tools_task = asyncio.create_task(get_tools_from_mcp(mcp_settings))
    sandbox_task = asyncio.create_task(asyncio.to_thread(init_daytona_sandbox, config.bot_id, local_workspace_root))

    # Detect the provider or use the explicitly specified one
    model_provider, base_url = detect_provider(config.model_name, config.model_server)
    model_kwargs, dropped_params, default_temperature_applied = sanitize_model_kwargs(
        model_name=config.model_name,
        model_provider=model_provider,
        base_url=base_url,
        api_key=config.api_key,
        generate_cfg=config.generate_cfg,
        source="init_agent"
    )
    if dropped_params:
        logger.info(
            "init_agent dropped_params=%s model=%s provider=%s default_temperature_applied=%s",
            dropped_params,
            config.model_name,
            model_provider,
            default_temperature_applied
        )
    llm_instance = init_chat_model(**model_kwargs)

    # Create a new agent instance without caching
    logger.info(f"Creating new agent for session: {getattr(config, 'session_id', 'no-session')}")

    checkpointer = None

    # Get the checkpointer from the connection pool
    # prepare_checkpoint_message has already been called in from_v1/from_v2_request
    if config.session_id:
        try:
            manager = get_checkpointer_manager()
            checkpointer = manager.checkpointer
        except Exception as e:
            logger.warning(f"Failed to load checkpointer: {e}")

    # Build the middleware list
    middleware = []
    middleware.append(EmptyResponseRetryMiddleware())
    middleware.append(ToolMetricsMiddleware(config))
    middleware.append(ToolUseCleanupMiddleware())
    # tool_output_middleware = ToolOutputLengthMiddleware(
    #     max_length=(getattr(config.generate_cfg, 'tool_output_max_length', None) if config.generate_cfg else None) or TOOL_OUTPUT_MAX_LENGTH,
    #     truncation_strategy=getattr(config.generate_cfg, 'tool_output_truncation_strategy', 'smart') if config.generate_cfg else 'smart',
    #     tool_filters=getattr(config.generate_cfg, 'tool_output_filters', None) if config.generate_cfg else None,
    #     exclude_tools=getattr(config.generate_cfg, 'tool_output_exclude', []) if config.generate_cfg else [],
    #     preserve_code_blocks=getattr(config.generate_cfg, 'preserve_code_blocks', True) if config.generate_cfg else True,
    #     preserve_json=getattr(config.generate_cfg, 'preserve_json', True) if config.generate_cfg else True
    # )
    # middleware.append(tool_output_middleware)

    if config.enable_memori:
        try:
            if not config.user_identifier:
                logger.warning("Mem0 enabled but user_identifier is missing, skipping Mem0")
            else:
                mem0_manager = get_mem0_manager()
                mem0_middleware = create_mem0_middleware(
                    bot_id=config.bot_id,
                    user_identifier=config.user_identifier,
                    session_id=config.session_id or "default",
                    agent_config=config,
                    enabled=config.enable_memori,
                    semantic_search_top_k=config.memori_semantic_search_top_k,
                    mem0_manager=mem0_manager,
                    llm_instance=llm_instance,
                )

                if mem0_middleware:
                    middleware.append(mem0_middleware)
                    logger.info("Mem0 middleware added to agent")

        except Exception as e:
            logger.error(f"Failed to create Mem0 middleware: {e}, continuing without Mem0")

    if config.enable_thinking:
        middleware.append(GuidelineMiddleware(llm_instance, config, system_prompt))

    summarization_middleware = SummarizationMiddleware(
        model=llm_instance,
        trigger=('tokens', SUMMARIZATION_MAX_TOKENS),
        trim_tokens_to_summarize=DEFAULT_TRIM_TOKEN_LIMIT,
        keep=('tokens', SUMMARIZATION_TOKENS_TO_KEEP),
        token_counter=create_token_counter(config.model_name)
    )
    middleware.append(summarization_middleware)
    logger.info(f"init_agent middleware ready, elapsed: {time.time() - create_start:.3f}s")

    mcp_tools = await mcp_tools_task
    logger.info(f"Loaded {len(mcp_tools)} MCP tools")
    logger.info(f"init_agent mcp tools ready, elapsed: {time.time() - create_start:.3f}s")

    # Build the main agent's tool list by hiding tools blacklisted in plugin.json.
    # Sub-agents still receive the full mcp_tools set, so hidden tools remain usable by them.
    hidden_tools = collect_main_agent_hidden_tools(config.bot_id)
    main_tools = [t for t in mcp_tools if t.name not in hidden_tools] if hidden_tools else mcp_tools
    if hidden_tools:
        logger.info(
            f"Main agent hides {len(mcp_tools) - len(main_tools)} tools: {sorted(hidden_tools)}"
        )

    sandbox, sandbox_type, workspace_root = await sandbox_task
    logger.info(f"init_agent sandbox ready, elapsed: {time.time() - create_start:.3f}s")

    # Inject shell_env into Daytona sandbox via BASH_ENV file
    if sandbox is not None and sandbox_type == "daytona":
        _shell_env = {
            "ASSISTANT_ID": config.bot_id,
            "USER_IDENTIFIER": config.user_identifier,
            "TRACE_ID": config.trace_id,
            "ENABLE_SELF_KNOWLEDGE": str(config.enable_self_knowledge).lower(),
            **(config.shell_env or {}),
        }
        env_lines = "\n".join(f'export {k}="{v}"' for k, v in _shell_env.items() if v is not None)
        if env_lines:
            from utils.daytona_sync import REMOTE_BASH_ENV_PATH, REMOTE_WORKSPACE_ROOT
            bash_env_content = f"cd {REMOTE_WORKSPACE_ROOT}\n{env_lines}"
            sandbox.execute(f"cat > {REMOTE_BASH_ENV_PATH} << 'ENVEOF'\n{bash_env_content}\nENVEOF")
            logger.info(f"Injected {len(_shell_env)} env vars into Daytona BASH_ENV")

    # Load sub-agents from skill directories
    subagents = await load_subagents(
        bot_id=config.bot_id,
        tools=mcp_tools,
        model=llm_instance,
    )
    if subagents:
        logger.info(f"Loaded {len(subagents)} sub-agents: {[s['name'] for s in subagents]}")

    agent, composite_backend = create_custom_cli_agent(
        model=llm_instance,
        assistant_id=config.bot_id,
        system_prompt=system_prompt,
        tools=main_tools,
        auto_approve=True,
        workspace_root=workspace_root,
        middleware=middleware,
        checkpointer=checkpointer,
        sandbox=sandbox,
        sandbox_type=sandbox_type,
        subagents=subagents if subagents else None,
        shell_env={
            k: v for k, v in {
                "ASSISTANT_ID": str(config.bot_id),
                "USER_IDENTIFIER": str(config.user_identifier) if config.user_identifier else None,
                "TRACE_ID": str(config.trace_id) if config.trace_id else None,
                "ENABLE_SELF_KNOWLEDGE": str(config.enable_self_knowledge).lower(),
                **(config.shell_env or {}),
            }.items() if v is not None
        }
    )

    logger.info(f"create agent elapsed: {time.time() - create_start:.3f}s")
    return agent, checkpointer, sandbox


class CustomSkillsMiddleware(SkillsMiddleware):
    """Custom SkillsMiddleware using the new signature format."""

    def before_agent(self, state, runtime, config):
        """Load skills metadata before agent execution.

        Adjust path display to use relative paths.

        Args:
            state: Current agent state.
            runtime: Runtime context.

        Returns:
            Updated state with skills_metadata populated.
        """
        if "skills_metadata" in state and len(state["skills_metadata"]) ==0 :
            del state["skills_metadata"]
        state = super().before_agent(state, runtime, config)
        return state

    async def abefore_agent(self, state, runtime, config):
        """Load skills metadata before agent execution.

        Adjust path display to use relative paths.

        Args:
            state: Current agent state.
            runtime: Runtime context.

        Returns:
            Updated state with skills_metadata populated.
        """
        if "skills_metadata" in state and len(state["skills_metadata"]) ==0 :
            del state["skills_metadata"]
        state = await super().abefore_agent(state, runtime, config)
        return state

def create_custom_cli_agent(
    model: str | BaseChatModel,
    assistant_id: str,
    *,
    tools: list[BaseTool] | None = None,
    sandbox: SandboxBackendProtocol | None = None,
    sandbox_type: str | None = None,
    system_prompt: str | None = None,
    auto_approve: bool = False,
    enable_skills: bool = True,
    enable_shell: bool = True,
    middleware: list[AgentMiddleware] = [],
    workspace_root: str | None = None,
    checkpointer: Checkpointer | None = None,
    store: BaseStore | None = None,
    shell_env: dict[str, str] | None = None,
    subagents: list[SubAgent] | None = None,
) -> tuple[Pregel, CompositeBackend]:
    """Create a CLI-configured agent with custom workspace_root for shell commands.

    This is a custom version of create_cli_agent that allows specifying a custom
    workspace_root for shell commands instead of using Path.cwd().

    Args:
        model: LLM model to use (e.g., "anthropic:claude-sonnet-4-5-20250929")
        assistant_id: Agent identifier for memory/state storage
        tools: Additional tools to provide to agent (default: empty list)
        sandbox: Optional sandbox backend for remote execution (e.g., ModalBackend).
                 If None, uses local filesystem + shell.
        sandbox_type: Type of sandbox provider ("modal", "runloop", "daytona").
                     Used for system prompt generation.
        system_prompt: Override the default system prompt. If None, generates one
                      based on sandbox_type and assistant_id.
        auto_approve: If True, automatically approves all tool calls without human
                     confirmation. Useful for automated workflows.
        enable_memory: Enable AgentMemoryMiddleware for persistent memory
        enable_skills: Enable SkillsMiddleware for custom agent skills
        enable_shell: Enable ShellMiddleware for local shell execution (only in local mode)
        workspace_root: Working directory for shell commands. If None, uses Path.cwd().
        checkpointer: Optional checkpointer for persisting conversation state
        store: Optional BaseStore for persisting user preferences and agent memory
        shell_env: Optional custom environment variables to pass to ShellMiddleware.
                   These will be merged with os.environ. Custom vars take precedence.

    Returns:
        2-tuple of (agent_graph, composite_backend)
        - agent_graph: Configured LangGraph Pregel instance ready for execution
        - composite_backend: CompositeBackend for file operations
    """
    if tools is None:
        tools = []

    # Build middleware stack based on enabled features
    agent_middleware = middleware

    # Prepare workspace root
    if workspace_root is None:
        workspace_root = str(Path.cwd())

    # CONDITIONAL SETUP: Local vs Remote Sandbox
    if sandbox is None:
        # ========== LOCAL MODE ==========
        if enable_shell:
            # Create environment for shell commands
            final_shell_env = os.environ.copy()
            if shell_env:
                final_shell_env.update(shell_env)

            # Use LocalShellBackend for filesystem + shell execution
            backend = LocalShellBackend(
                root_dir=workspace_root,
                virtual_mode=False,
                inherit_env=True,
                env=final_shell_env,
            )
        else:
            # No shell access - use plain FilesystemBackend
            backend = FilesystemBackend(root_dir=workspace_root, virtual_mode=False)

        # Set up composite backend with routing based on the new implementation
        # NOTE: virtual_mode=True anchors all paths to root_dir. This is required for
        # these offload-only backends: CompositeBackend strips the route prefix and
        # forwards "/" to grep, so virtual_mode=False would resolve "/" to the real
        # filesystem root and scan the whole disk (hitting /usr, /var, other sessions'
        # temp dirs), causing 45-152s grep calls. virtual_mode=True confines grep to
        # the temp dir and filters out-of-root results.
        large_results_backend = FilesystemBackend(
            root_dir=tempfile.mkdtemp(prefix="deepagents_large_results_"),
            virtual_mode=True,
        )
        conversation_history_backend = FilesystemBackend(
            root_dir=tempfile.mkdtemp(prefix="deepagents_conversation_history_"),
            virtual_mode=True,
        )
        composite_backend = CompositeBackend(
            default=backend,
            routes={
                "/large_tool_results/": large_results_backend,
                "/conversation_history/": conversation_history_backend,
            },
        )

        # Add skills middleware (using new signature)
        if enable_skills:
            skills_sources = ["./skills"]

            agent_middleware.append(
                CustomSkillsMiddleware(
                    backend=FilesystemBackend(root_dir=workspace_root, virtual_mode=False),
                    sources=skills_sources,
                )
            )

        # Add LocalContextMiddleware (new in latest version)
        # Check if backend supports execute (is _ExecutableBackend)
        if enable_shell:
            from deepagents_cli.local_context import LocalContextMiddleware, _ExecutableBackend
            if isinstance(backend, _ExecutableBackend):
                agent_middleware.append(LocalContextMiddleware(backend=backend))

    else:
        # ========== REMOTE SANDBOX MODE ==========
        composite_backend = CompositeBackend(
            default=sandbox,  # Remote sandbox (ModalBackend, etc.)
            routes={},  # No virtualization
        )

        # Add skills middleware
        if enable_skills:
            skills_sources = ["/workspace/skills"]

            agent_middleware.append(
                CustomSkillsMiddleware(
                    backend=sandbox,
                    sources=skills_sources,
                )
            )

        # Note: Shell middleware not used in sandbox mode
        # File operations and execute tool are provided by the sandbox backend

    # Get or use custom system prompt
    if system_prompt is None:
        from deepagents_cli.agent import get_system_prompt as _get_system_prompt
        system_prompt = _get_system_prompt(assistant_id=assistant_id, sandbox_type=sandbox_type)

    # Configure interrupt_on based on auto_approve setting
    if auto_approve:
        # No interrupts - all tools run automatically
        interrupt_on = {}
    else:
        # Full HITL for destructive operations
        from deepagents_cli.agent import _add_interrupt_on
        interrupt_on = _add_interrupt_on()

    deepagent_middleware = [
        TodoListMiddleware(),
        FilePathFixMiddleware(),  # Fix extra spaces in CJK file names within tool call arguments
        CustomFilesystemMiddleware(backend=composite_backend),  # Use the custom FilesystemMiddleware with full SKILL.md reading support
    ]
    # Insert SubAgentMiddleware after FilesystemMiddleware (matches create_deep_agent ordering)
    if subagents:
        subagent_middleware = SubAgentMiddleware(
            backend=composite_backend,
            subagents=subagents,
        )
        deepagent_middleware.append(subagent_middleware)
        logger.info(f"SubAgentMiddleware added with {len(subagents)} sub-agents: {[s['name'] for s in subagents]}")
    deepagent_middleware.extend([
        AnthropicPromptCachingMiddleware(unsupported_model_behavior="ignore"),
        PatchToolCallsMiddleware(),
    ])
    if agent_middleware:
        deepagent_middleware.extend(agent_middleware)
    if interrupt_on is not None:
        deepagent_middleware.append(HumanInTheLoopMiddleware(interrupt_on=interrupt_on))

    agent = create_agent(
        model,
        system_prompt=system_prompt + "\n\n" + BASE_AGENT_PROMPT if system_prompt else BASE_AGENT_PROMPT,
        tools=tools,
        middleware=deepagent_middleware,
        checkpointer=checkpointer,
        store=store,
    ).with_config({"recursion_limit": 1000})
    return agent, composite_backend