nanobot/nanobot/agent/loop.py

"""Agent loop: the core processing engine."""

from __future__ import annotations

import asyncio
import dataclasses
import os
import time
from contextlib import AsyncExitStack, nullcontext, suppress
from dataclasses import dataclass, field
from enum import Enum, auto
from pathlib import Path
from typing import TYPE_CHECKING, Any, Awaitable, Callable

from loguru import logger

from nanobot.agent import model_presets as preset_helpers
from nanobot.agent.autocompact import AutoCompact
from nanobot.agent.context import ContextBuilder
from nanobot.agent.hook import AgentHook, CompositeHook
from nanobot.agent.memory import Consolidator, Dream
from nanobot.agent.progress_hook import AgentProgressHook
from nanobot.agent.runner import _MAX_INJECTIONS_PER_TURN, AgentRunner, AgentRunSpec
from nanobot.agent.subagent import SubagentManager
from nanobot.agent.tools.file_state import FileStateStore, bind_file_states, reset_file_states
from nanobot.agent.tools.message import MessageTool
from nanobot.agent.tools.registry import ToolRegistry
from nanobot.agent.tools.self import MyTool
from nanobot.bus.events import InboundMessage, OutboundMessage
from nanobot.bus.queue import MessageBus
from nanobot.command import CommandContext, CommandRouter, register_builtin_commands
from nanobot.config.schema import AgentDefaults, ModelPresetConfig
from nanobot.providers.base import LLMProvider
from nanobot.providers.factory import ProviderSnapshot
from nanobot.session.manager import Session, SessionManager
from nanobot.utils.artifacts import generated_image_paths_from_messages
from nanobot.utils.document import extract_documents
from nanobot.utils.helpers import image_placeholder_text
from nanobot.utils.helpers import truncate_text as truncate_text_fn
from nanobot.utils.image_generation_intent import image_generation_prompt
from nanobot.utils.runtime import EMPTY_FINAL_RESPONSE_MESSAGE
from nanobot.utils.webui_titles import mark_webui_session, maybe_generate_webui_title_after_turn

if TYPE_CHECKING:
    from nanobot.config.schema import (
        ChannelsConfig,
        ProviderConfig,
        ToolsConfig,
    )
    from nanobot.cron.service import CronService


UNIFIED_SESSION_KEY = "unified:default"


class TurnState(Enum):
    RESTORE = auto()
    COMPACT = auto()
    COMMAND = auto()
    BUILD = auto()
    RUN = auto()
    SAVE = auto()
    RESPOND = auto()
    DONE = auto()


@dataclass
class StateTraceEntry:
    state: TurnState
    started_at: float
    duration_ms: float
    event: str
    error: str | None = None


@dataclass
class TurnContext:
    msg: InboundMessage
    session_key: str
    state: TurnState
    turn_id: str
    session: Session | None = None

    history: list[dict[str, Any]] = field(default_factory=list)
    initial_messages: list[dict[str, Any]] = field(default_factory=list)

    final_content: str | None = None
    tools_used: list[str] = field(default_factory=list)
    all_messages: list[dict[str, Any]] = field(default_factory=list)
    stop_reason: str = ""
    had_injections: bool = False

    user_persisted_early: bool = False
    save_skip: int = 0

    outbound: OutboundMessage | None = None
    generated_media: list[str] = field(default_factory=list)

    on_progress: Callable[..., Awaitable[None]] | None = None
    on_stream: Callable[[str], Awaitable[None]] | None = None
    on_stream_end: Callable[..., Awaitable[None]] | None = None
    on_retry_wait: Callable[[str], Awaitable[None]] | None = None

    pending_queue: asyncio.Queue | None = None
    pending_summary: str | None = None

    trace: list[StateTraceEntry] = field(default_factory=list)


class AgentLoop:
    """
    The agent loop is the core processing engine.

    It:
    1. Receives messages from the bus
    2. Builds context with history, memory, skills
    3. Calls the LLM
    4. Executes tool calls
    5. Sends responses back
    """

    @property
    def current_iteration(self) -> int:
        return self._current_iteration

    @property
    def tool_names(self) -> list[str]:
        return self.tools.tool_names

    _RUNTIME_CHECKPOINT_KEY = "runtime_checkpoint"
    _PENDING_USER_TURN_KEY = "pending_user_turn"

    # Event-driven state transition table.
    # Handlers return an event string; the driver looks up the next state here.
    _TRANSITIONS: dict[tuple[TurnState, str], TurnState] = {
        (TurnState.RESTORE, "ok"): TurnState.COMPACT,
        (TurnState.COMPACT, "ok"): TurnState.COMMAND,
        (TurnState.COMMAND, "dispatch"): TurnState.BUILD,
        (TurnState.COMMAND, "shortcut"): TurnState.DONE,
        (TurnState.BUILD, "ok"): TurnState.RUN,
        (TurnState.RUN, "ok"): TurnState.SAVE,
        (TurnState.SAVE, "ok"): TurnState.RESPOND,
        (TurnState.RESPOND, "ok"): TurnState.DONE,
    }

    def __init__(
        self,
        bus: MessageBus,
        provider: LLMProvider,
        workspace: Path,
        model: str | None = None,
        max_iterations: int | None = None,
        context_window_tokens: int | None = None,
        context_block_limit: int | None = None,
        max_tool_result_chars: int | None = None,
        provider_retry_mode: str = "standard",
        tool_hint_max_length: int | None = None,
        cron_service: CronService | None = None,
        restrict_to_workspace: bool = False,
        session_manager: SessionManager | None = None,
        mcp_servers: dict | None = None,
        channels_config: ChannelsConfig | None = None,
        timezone: str | None = None,
        session_ttl_minutes: int = 0,
        consolidation_ratio: float = 0.5,
        max_messages: int = 120,
        hooks: list[AgentHook] | None = None,
        unified_session: bool = False,
        disabled_skills: list[str] | None = None,
        tools_config: ToolsConfig | None = None,
        image_generation_provider_config: ProviderConfig | None = None,
        image_generation_provider_configs: dict[str, ProviderConfig] | None = None,
        provider_snapshot_loader: Callable[..., ProviderSnapshot] | None = None,
        provider_signature: tuple[object, ...] | None = None,
        model_presets: dict[str, ModelPresetConfig] | None = None,
        model_preset: str | None = None,
        preset_snapshot_loader: preset_helpers.PresetSnapshotLoader | None = None,
        runtime_model_publisher: Callable[[str, str | None], None] | None = None,
    ):
        from nanobot.config.schema import ToolsConfig

        _tc = tools_config or ToolsConfig()
        defaults = AgentDefaults()
        self.bus = bus
        self.channels_config = channels_config
        self.provider = provider
        self._provider_snapshot_loader = provider_snapshot_loader
        self._preset_snapshot_loader = preset_snapshot_loader
        self._runtime_model_publisher = runtime_model_publisher
        self._provider_signature = provider_signature
        self._default_selection_signature = preset_helpers.default_selection_signature(provider_signature)
        self.workspace = workspace
        self.model = model or provider.get_default_model()
        self.max_iterations = (
            max_iterations if max_iterations is not None else defaults.max_tool_iterations
        )
        self.context_window_tokens = (
            context_window_tokens
            if context_window_tokens is not None
            else defaults.context_window_tokens
        )
        self.context_block_limit = context_block_limit
        self.max_tool_result_chars = (
            max_tool_result_chars
            if max_tool_result_chars is not None
            else defaults.max_tool_result_chars
        )
        self.provider_retry_mode = provider_retry_mode
        self.tool_hint_max_length = (
            tool_hint_max_length if tool_hint_max_length is not None
            else defaults.tool_hint_max_length
        )
        self.tools_config = _tc
        self.web_config = _tc.web
        self.exec_config = _tc.exec
        self._image_generation_provider_configs = dict(image_generation_provider_configs or {})
        if (
            image_generation_provider_config is not None
            and "openrouter" not in self._image_generation_provider_configs
        ):
            self._image_generation_provider_configs["openrouter"] = image_generation_provider_config
        self.cron_service = cron_service
        self.restrict_to_workspace = restrict_to_workspace
        self._start_time = time.time()
        self._last_usage: dict[str, int] = {}
        self._extra_hooks: list[AgentHook] = hooks or []

        self.context = ContextBuilder(workspace, timezone=timezone, disabled_skills=disabled_skills)
        self.sessions = session_manager or SessionManager(workspace)
        self.tools = ToolRegistry()
        # One file-read/write tracker per logical session. The tool registry is
        # shared by this loop, so tools resolve the active state via contextvars.
        self._file_state_store = FileStateStore()
        self.runner = AgentRunner(provider)
        self.subagents = SubagentManager(
            provider=provider,
            workspace=workspace,
            bus=bus,
            model=self.model,
            tools_config=_tc,
            max_tool_result_chars=self.max_tool_result_chars,
            restrict_to_workspace=restrict_to_workspace,
            disabled_skills=disabled_skills,
            max_iterations=self.max_iterations,
        )
        self._unified_session = unified_session
        self._max_messages = max_messages if max_messages > 0 else 120
        self._running = False
        self._mcp_servers = mcp_servers or {}
        self._mcp_stacks: dict[str, AsyncExitStack] = {}
        self._mcp_connected = False
        self._mcp_connecting = False
        self._active_tasks: dict[str, list[asyncio.Task]] = {}  # session_key -> tasks
        self._background_tasks: list[asyncio.Task] = []
        self._session_locks: dict[str, asyncio.Lock] = {}
        # Per-session pending queues for mid-turn message injection.
        # When a session has an active task, new messages for that session
        # are routed here instead of creating a new task.
        self._pending_queues: dict[str, asyncio.Queue] = {}
        # NANOBOT_MAX_CONCURRENT_REQUESTS: <=0 means unlimited; default 3.
        _max = int(os.environ.get("NANOBOT_MAX_CONCURRENT_REQUESTS", "3"))
        self._concurrency_gate: asyncio.Semaphore | None = (
            asyncio.Semaphore(_max) if _max > 0 else None
        )
        self.consolidator = Consolidator(
            store=self.context.memory,
            provider=provider,
            model=self.model,
            sessions=self.sessions,
            context_window_tokens=self.context_window_tokens,
            build_messages=self.context.build_messages,
            get_tool_definitions=self.tools.get_definitions,
            max_completion_tokens=provider.generation.max_tokens,
            consolidation_ratio=consolidation_ratio,
        )
        self.auto_compact = AutoCompact(
            sessions=self.sessions,
            consolidator=self.consolidator,
            session_ttl_minutes=session_ttl_minutes,
        )
        self.dream = Dream(
            store=self.context.memory,
            provider=provider,
            model=self.model,
        )
        self.model_presets: dict[str, ModelPresetConfig] = model_presets or {}
        self._active_preset: str | None = None
        if model_preset:
            self.set_model_preset(model_preset, publish_update=False)
        self._register_default_tools()
        self._runtime_vars: dict[str, Any] = {}
        self._current_iteration: int = 0
        self.commands = CommandRouter()
        register_builtin_commands(self.commands)

    @classmethod
    def from_config(
        cls,
        config: Any,
        bus: MessageBus | None = None,
        **extra: Any,
    ) -> AgentLoop:
        """Create an AgentLoop from config with the common parameter set.

        Extra keyword arguments are forwarded to ``AgentLoop.__init__``,
        allowing callers to override or extend the standard config-derived
        parameters (e.g. ``cron_service``, ``session_manager``).
        """
        from nanobot.providers.factory import make_provider

        if bus is None:
            bus = MessageBus()
        defaults = config.agents.defaults
        provider = extra.pop("provider", None) or make_provider(config)
        resolved = config.resolve_preset()
        model = extra.pop("model", None) or resolved.model
        context_window_tokens = extra.pop("context_window_tokens", None) or resolved.context_window_tokens
        provider_snapshot_loader = extra.pop("provider_snapshot_loader", None)
        preset_snapshot_loader = extra.pop("preset_snapshot_loader", None) or preset_helpers.make_preset_snapshot_loader(
            config,
            provider_snapshot_loader,
        )
        return cls(
            bus=bus,
            provider=provider,
            workspace=config.workspace_path,
            model=model,
            max_iterations=defaults.max_tool_iterations,
            context_window_tokens=context_window_tokens,
            context_block_limit=defaults.context_block_limit,
            max_tool_result_chars=defaults.max_tool_result_chars,
            provider_retry_mode=defaults.provider_retry_mode,
            tool_hint_max_length=defaults.tool_hint_max_length,
            restrict_to_workspace=config.tools.restrict_to_workspace,
            mcp_servers=config.tools.mcp_servers,
            channels_config=config.channels,
            timezone=defaults.timezone,
            unified_session=defaults.unified_session,
            disabled_skills=defaults.disabled_skills,
            session_ttl_minutes=defaults.session_ttl_minutes,
            consolidation_ratio=defaults.consolidation_ratio,
            max_messages=defaults.max_messages,
            tools_config=config.tools,
            model_presets=preset_helpers.configured_model_presets(config),
            model_preset=defaults.model_preset,
            provider_snapshot_loader=provider_snapshot_loader,
            preset_snapshot_loader=preset_snapshot_loader,
            **extra,
        )

    def _sync_subagent_runtime_limits(self) -> None:
        """Keep subagent runtime limits aligned with mutable loop settings."""
        self.subagents.max_iterations = self.max_iterations

    def _apply_provider_snapshot(
        self,
        snapshot: ProviderSnapshot,
        *,
        publish_update: bool = True,
        model_preset: str | None = None,
    ) -> None:
        """Swap model/provider for future turns without disturbing an active one."""
        provider = snapshot.provider
        model = snapshot.model
        context_window_tokens = snapshot.context_window_tokens
        old_model = self.model
        self.provider = provider
        self.model = model
        self.context_window_tokens = context_window_tokens
        self.runner.provider = provider
        self.subagents.set_provider(provider, model)
        self.consolidator.set_provider(provider, model, context_window_tokens)
        self.dream.set_provider(provider, model)
        self._provider_signature = snapshot.signature
        if publish_update and self._runtime_model_publisher is not None:
            self._runtime_model_publisher(
                self.model,
                model_preset if model_preset is not None else self.model_preset,
            )
        logger.info("Runtime model switched for next turn: {} -> {}", old_model, model)

    def _refresh_provider_snapshot(self) -> None:
        if self._provider_snapshot_loader is None:
            return
        try:
            snapshot = self._provider_snapshot_loader()
        except Exception:
            logger.exception("Failed to refresh provider config")
            return
        default_selection = preset_helpers.default_selection_signature(snapshot.signature)
        if self._active_preset and self._default_selection_signature in (None, default_selection):
            self._default_selection_signature = default_selection
            try:
                snapshot = self._build_model_preset_snapshot(self._active_preset)
            except Exception:
                logger.exception("Failed to refresh active model preset")
                return
        else:
            self._active_preset = None
            self._default_selection_signature = default_selection
        if snapshot.signature == self._provider_signature:
            return
        self._default_selection_signature = preset_helpers.default_selection_signature(snapshot.signature)
        self._apply_provider_snapshot(snapshot)

    @property
    def model_preset(self) -> str | None:
        return self._active_preset

    @model_preset.setter
    def model_preset(self, name: str | None) -> None:
        self.set_model_preset(name)

    def _build_model_preset_snapshot(self, name: str) -> ProviderSnapshot:
        return preset_helpers.build_runtime_preset_snapshot(
            name=name,
            presets=self.model_presets,
            provider=self.provider,
            loader=self._preset_snapshot_loader,
        )

    def set_model_preset(self, name: str | None, *, publish_update: bool = True) -> None:
        """Resolve a preset by name and apply all runtime model dependents."""
        name = preset_helpers.normalize_preset_name(name, self.model_presets)
        snapshot = self._build_model_preset_snapshot(name)
        self._apply_provider_snapshot(snapshot, publish_update=publish_update, model_preset=name)
        self._active_preset = name

    def _register_default_tools(self) -> None:
        """Register the default set of tools via plugin loader."""
        from nanobot.agent.tools.context import ToolContext
        from nanobot.agent.tools.loader import ToolLoader

        ctx = ToolContext(
            config=self.tools_config,
            workspace=str(self.workspace),
            bus=self.bus,
            subagent_manager=self.subagents,
            cron_service=self.cron_service,
            provider_snapshot_loader=self._provider_snapshot_loader,
            image_generation_provider_configs=self._image_generation_provider_configs,
            timezone=self.context.timezone or "UTC",
        )
        loader = ToolLoader()
        registered = loader.load(ctx, self.tools)

        # MyTool needs runtime state reference — manual registration
        if self.tools_config.my.enable:
            self.tools.register(
                MyTool(runtime_state=self, modify_allowed=self.tools_config.my.allow_set)
            )
            registered.append("my")

        logger.info("Registered {} tools: {}", len(registered), registered)

    async def _connect_mcp(self) -> None:
        """Connect to configured MCP servers (one-time, lazy)."""
        if self._mcp_connected or self._mcp_connecting or not self._mcp_servers:
            return
        self._mcp_connecting = True
        from nanobot.agent.tools.mcp import connect_mcp_servers

        try:
            self._mcp_stacks = await connect_mcp_servers(self._mcp_servers, self.tools)
            if self._mcp_stacks:
                self._mcp_connected = True
            else:
                logger.warning("No MCP servers connected successfully (will retry next message)")
        except asyncio.CancelledError:
            logger.warning("MCP connection cancelled (will retry next message)")
            self._mcp_stacks.clear()
        except BaseException as e:
            logger.warning("Failed to connect MCP servers (will retry next message): {}", e)
            self._mcp_stacks.clear()
        finally:
            self._mcp_connecting = False

    def _set_tool_context(
        self, channel: str, chat_id: str,
        message_id: str | None = None, metadata: dict | None = None,
        session_key: str | None = None,
    ) -> None:
        """Update context for all tools that need routing info."""
        from nanobot.agent.tools.context import ContextAware, RequestContext

        if session_key is not None:
            effective_key = session_key
        elif self._unified_session:
            effective_key = UNIFIED_SESSION_KEY
        else:
            effective_key = f"{channel}:{chat_id}"

        request_ctx = RequestContext(
            channel=channel,
            chat_id=chat_id,
            message_id=message_id,
            session_key=effective_key,
            metadata=dict(metadata or {}),
        )

        for name in self.tools.tool_names:
            tool = self.tools.get(name)
            if tool and isinstance(tool, ContextAware):
                tool.set_context(request_ctx)

    @staticmethod
    def _runtime_chat_id(msg: InboundMessage) -> str:
        """Return the chat id shown in runtime metadata for the model."""
        return str(msg.metadata.get("context_chat_id") or msg.chat_id)

    async def _build_bus_progress_callback(
        self, msg: InboundMessage
    ) -> Callable[..., Awaitable[None]]:
        """Build a progress callback that publishes to the message bus."""

        async def _bus_progress(
            content: str,
            *,
            tool_hint: bool = False,
            tool_events: list[dict[str, Any]] | None = None,
            reasoning: bool = False,
            reasoning_end: bool = False,
        ) -> None:
            meta = dict(msg.metadata or {})
            meta["_progress"] = True
            meta["_tool_hint"] = tool_hint
            if reasoning:
                meta["_reasoning_delta"] = True
            if reasoning_end:
                meta["_reasoning_end"] = True
            if tool_events:
                meta["_tool_events"] = tool_events
            await self.bus.publish_outbound(
                OutboundMessage(
                    channel=msg.channel,
                    chat_id=msg.chat_id,
                    content=content,
                    metadata=meta,
                )
            )

        return _bus_progress

    async def _build_retry_wait_callback(
        self, msg: InboundMessage
    ) -> Callable[[str], Awaitable[None]]:
        """Build a retry-wait callback that publishes to the message bus."""

        async def _on_retry_wait(content: str) -> None:
            meta = dict(msg.metadata or {})
            meta["_retry_wait"] = True
            await self.bus.publish_outbound(
                OutboundMessage(
                    channel=msg.channel,
                    chat_id=msg.chat_id,
                    content=content,
                    metadata=meta,
                )
            )

        return _on_retry_wait

    def _persist_user_message_early(
        self,
        msg: InboundMessage,
        session: Session,
        **kwargs: Any,
    ) -> bool:
        """Persist the triggering user message before the turn starts.

        Returns True if the message was persisted.
        """
        media_paths = [p for p in (msg.media or []) if isinstance(p, str) and p]
        has_text = isinstance(msg.content, str) and msg.content.strip()
        if has_text or media_paths:
            extra: dict[str, Any] = {"media": list(media_paths)} if media_paths else {}
            extra.update(kwargs)
            text = msg.content if isinstance(msg.content, str) else ""
            session.add_message("user", text, **extra)
            self._mark_pending_user_turn(session)
            self.sessions.save(session)
            return True
        return False

    def _build_initial_messages(
        self,
        msg: InboundMessage,
        session: Session,
        history: list[dict[str, Any]],
        pending_summary: str | None,
    ) -> list[dict[str, Any]]:
        """Build the initial message list for the LLM turn."""
        return self.context.build_messages(
            history=history,
            current_message=image_generation_prompt(msg.content, msg.metadata),
            media=msg.media if msg.media else None,
            channel=msg.channel,
            chat_id=self._runtime_chat_id(msg),
            sender_id=msg.sender_id,
            session_summary=pending_summary,
        )

    async def _dispatch_command_inline(
        self,
        msg: InboundMessage,
        key: str,
        raw: str,
        dispatch_fn: Callable[[CommandContext], Awaitable[OutboundMessage | None]],
    ) -> None:
        """Dispatch a command directly from the run() loop and publish the result."""
        ctx = CommandContext(msg=msg, session=None, key=key, raw=raw, loop=self)
        result = await dispatch_fn(ctx)
        if result:
            await self.bus.publish_outbound(result)
        else:
            logger.warning("Command '{}' matched but dispatch returned None", raw)

    async def _cancel_active_tasks(self, key: str) -> int:
        """Cancel and await all active tasks and subagents for *key*.

        Returns the total number of cancelled tasks + subagents.
        """
        tasks = self._active_tasks.pop(key, [])
        cancelled = sum(1 for t in tasks if not t.done() and t.cancel())
        for t in tasks:
            with suppress(asyncio.CancelledError, Exception):
                await t
        sub_cancelled = await self.subagents.cancel_by_session(key)
        return cancelled + sub_cancelled

    def _effective_session_key(self, msg: InboundMessage) -> str:
        """Return the session key used for task routing and mid-turn injections."""
        if self._unified_session and not msg.session_key_override:
            return UNIFIED_SESSION_KEY
        return msg.session_key

    def _replay_token_budget(self) -> int:
        """Derive a token budget for session history replay from the context window."""
        if self.context_window_tokens <= 0:
            return 0
        max_output = getattr(getattr(self.provider, "generation", None), "max_tokens", 4096)
        try:
            reserved_output = int(max_output)
        except (TypeError, ValueError):
            reserved_output = 4096
        budget = self.context_window_tokens - max(1, reserved_output) - 1024
        return budget if budget > 0 else max(128, self.context_window_tokens // 2)

    async def _run_agent_loop(
        self,
        initial_messages: list[dict],
        on_progress: Callable[..., Awaitable[None]] | None = None,
        on_stream: Callable[[str], Awaitable[None]] | None = None,
        on_stream_end: Callable[..., Awaitable[None]] | None = None,
        on_retry_wait: Callable[[str], Awaitable[None]] | None = None,
        *,
        session: Session | None = None,
        channel: str = "cli",
        chat_id: str = "direct",
        message_id: str | None = None,
        metadata: dict[str, Any] | None = None,
        session_key: str | None = None,
        pending_queue: asyncio.Queue | None = None,
    ) -> tuple[str | None, list[str], list[dict], str, bool]:
        """Run the agent iteration loop.

        *on_stream*: called with each content delta during streaming.
        *on_stream_end(resuming)*: called when a streaming session finishes.
        ``resuming=True`` means tool calls follow (spinner should restart);
        ``resuming=False`` means this is the final response.

        Returns (final_content, tools_used, messages, stop_reason, had_injections).
        """
        self._sync_subagent_runtime_limits()

        loop_hook = AgentProgressHook(
            on_progress=on_progress,
            on_stream=on_stream,
            on_stream_end=on_stream_end,
            channel=channel,
            chat_id=chat_id,
            message_id=message_id,
            metadata=metadata,
            session_key=session_key,
            tool_hint_max_length=self.tool_hint_max_length,
            set_tool_context=self._set_tool_context,
            on_iteration=lambda iteration: setattr(self, "_current_iteration", iteration),
        )
        hook: AgentHook = (
            CompositeHook([loop_hook] + self._extra_hooks) if self._extra_hooks else loop_hook
        )

        async def _checkpoint(payload: dict[str, Any]) -> None:
            if session is None:
                return
            self._set_runtime_checkpoint(session, payload)

        async def _drain_pending(*, limit: int = _MAX_INJECTIONS_PER_TURN) -> list[dict[str, Any]]:
            """Drain follow-up messages from the pending queue.

            When no messages are immediately available but sub-agents
            spawned in this dispatch are still running, blocks until at
            least one result arrives (or timeout).  This keeps the runner
            loop alive so subsequent sub-agent completions are consumed
            in-order rather than dispatched separately.
            """
            if pending_queue is None:
                return []

            def _to_user_message(pending_msg: InboundMessage) -> dict[str, Any]:
                content = pending_msg.content
                media = pending_msg.media if pending_msg.media else None
                if media:
                    content, media = extract_documents(content, media)
                    media = media or None
                user_content = self.context._build_user_content(content, media)
                runtime_ctx = self.context._build_runtime_context(
                    pending_msg.channel,
                    self._runtime_chat_id(pending_msg),
                    self.context.timezone,
                )
                if isinstance(user_content, str):
                    merged: str | list[dict[str, Any]] = f"{runtime_ctx}\n\n{user_content}"
                else:
                    merged = [{"type": "text", "text": runtime_ctx}] + user_content
                return {"role": "user", "content": merged}

            items: list[dict[str, Any]] = []
            while len(items) < limit:
                try:
                    items.append(_to_user_message(pending_queue.get_nowait()))
                except asyncio.QueueEmpty:
                    break

            # Block if nothing drained but sub-agents spawned in this dispatch
            # are still running.  Keeps the runner loop alive so subsequent
            # completions are injected in-order rather than dispatched separately.
            if (not items
                    and session is not None
                    and self.subagents.get_running_count_by_session(session.key) > 0):
                try:
                    msg = await asyncio.wait_for(pending_queue.get(), timeout=300)
                except asyncio.TimeoutError:
                    logger.warning(
                        "Timeout waiting for sub-agent completion in session {}",
                        session.key,
                    )
                    return items
                items.append(_to_user_message(msg))
                while len(items) < limit:
                    try:
                        items.append(_to_user_message(pending_queue.get_nowait()))
                    except asyncio.QueueEmpty:
                        break

            return items

        active_session_key = session.key if session else session_key
        file_state_token = bind_file_states(self._file_state_store.for_session(active_session_key))
        try:
            result = await self.runner.run(AgentRunSpec(
                initial_messages=initial_messages,
                tools=self.tools,
                model=self.model,
                max_iterations=self.max_iterations,
                max_tool_result_chars=self.max_tool_result_chars,
                hook=hook,
                error_message="Sorry, I encountered an error calling the AI model.",
                concurrent_tools=True,
                workspace=self.workspace,
                session_key=session.key if session else None,
                context_window_tokens=self.context_window_tokens,
                context_block_limit=self.context_block_limit,
                provider_retry_mode=self.provider_retry_mode,
                progress_callback=on_progress,
                stream_progress_deltas=on_stream is not None,
                retry_wait_callback=on_retry_wait,
                checkpoint_callback=_checkpoint,
                injection_callback=_drain_pending,
            ))
        finally:
            reset_file_states(file_state_token)
        self._last_usage = result.usage
        if result.stop_reason == "max_iterations":
            logger.warning("Max iterations ({}) reached", self.max_iterations)
            # Push final content through stream so streaming channels (e.g. Feishu)
            # update the card instead of leaving it empty.
            if on_stream and on_stream_end:
                await on_stream(result.final_content or "")
                await on_stream_end(resuming=False)
        elif result.stop_reason == "error":
            logger.error("LLM returned error: {}", (result.final_content or "")[:200])
        return result.final_content, result.tools_used, result.messages, result.stop_reason, result.had_injections

    async def run(self) -> None:
        """Run the agent loop, dispatching messages as tasks to stay responsive to /stop."""
        self._running = True
        await self._connect_mcp()
        logger.info("Agent loop started")

        while self._running:
            try:
                msg = await asyncio.wait_for(self.bus.consume_inbound(), timeout=1.0)
            except asyncio.TimeoutError:
                self.auto_compact.check_expired(
                    self._schedule_background,
                    active_session_keys=self._pending_queues.keys(),
                )
                continue
            except asyncio.CancelledError:
                # Preserve real task cancellation so shutdown can complete cleanly.
                # Only ignore non-task CancelledError signals that may leak from integrations.
                if not self._running or asyncio.current_task().cancelling():
                    raise
                continue
            except Exception as e:
                logger.warning("Error consuming inbound message: {}, continuing...", e)
                continue

            raw = msg.content.strip()
            if self.commands.is_priority(raw):
                await self._dispatch_command_inline(
                    msg, msg.session_key, raw,
                    self.commands.dispatch_priority,
                )
                continue
            effective_key = self._effective_session_key(msg)
            # If this session already has an active pending queue (i.e. a task
            # is processing this session), route the message there for mid-turn
            # injection instead of creating a competing task.
            if effective_key in self._pending_queues:
                # Non-priority commands must not be queued for injection;
                # dispatch them directly (same pattern as priority commands).
                if self.commands.is_dispatchable_command(raw):
                    await self._dispatch_command_inline(
                        msg, effective_key, raw,
                        self.commands.dispatch,
                    )
                    continue
                pending_msg = msg
                if effective_key != msg.session_key:
                    pending_msg = dataclasses.replace(
                        msg,
                        session_key_override=effective_key,
                    )
                try:
                    self._pending_queues[effective_key].put_nowait(pending_msg)
                except asyncio.QueueFull:
                    logger.warning(
                        "Pending queue full for session {}, falling back to queued task",
                        effective_key,
                    )
                else:
                    logger.info(
                        "Routed follow-up message to pending queue for session {}",
                        effective_key,
                    )
                    continue
            # Compute the effective session key before dispatching
            # This ensures /stop command can find tasks correctly when unified session is enabled
            task = asyncio.create_task(self._dispatch(msg))
            self._active_tasks.setdefault(effective_key, []).append(task)
            task.add_done_callback(
                lambda t, k=effective_key: self._active_tasks.get(k, [])
                and self._active_tasks[k].remove(t)
                if t in self._active_tasks.get(k, [])
                else None
            )

    async def _dispatch(self, msg: InboundMessage) -> None:
        """Process a message: per-session serial, cross-session concurrent."""
        session_key = self._effective_session_key(msg)
        if session_key != msg.session_key:
            msg = dataclasses.replace(msg, session_key_override=session_key)
        lock = self._session_locks.setdefault(session_key, asyncio.Lock())
        gate = self._concurrency_gate or nullcontext()

        # Register a pending queue so follow-up messages for this session are
        # routed here (mid-turn injection) instead of spawning a new task.
        pending = asyncio.Queue(maxsize=20)
        self._pending_queues[session_key] = pending

        try:
            async with lock, gate:
                try:
                    on_stream = on_stream_end = None
                    if msg.metadata.get("_wants_stream"):
                        # Split one answer into distinct stream segments.
                        stream_base_id = f"{msg.session_key}:{time.time_ns()}"
                        stream_segment = 0

                        def _current_stream_id() -> str:
                            return f"{stream_base_id}:{stream_segment}"

                        async def on_stream(delta: str) -> None:
                            meta = dict(msg.metadata or {})
                            meta["_stream_delta"] = True
                            meta["_stream_id"] = _current_stream_id()
                            await self.bus.publish_outbound(OutboundMessage(
                                channel=msg.channel, chat_id=msg.chat_id,
                                content=delta,
                                metadata=meta,
                            ))

                        async def on_stream_end(*, resuming: bool = False) -> None:
                            nonlocal stream_segment
                            meta = dict(msg.metadata or {})
                            meta["_stream_end"] = True
                            meta["_resuming"] = resuming
                            meta["_stream_id"] = _current_stream_id()
                            await self.bus.publish_outbound(OutboundMessage(
                                channel=msg.channel, chat_id=msg.chat_id,
                                content="",
                                metadata=meta,
                            ))
                            stream_segment += 1

                    response = await self._process_message(
                        msg, on_stream=on_stream, on_stream_end=on_stream_end,
                        pending_queue=pending,
                    )
                    if response is not None:
                        await self.bus.publish_outbound(response)
                    elif msg.channel == "cli":
                        await self.bus.publish_outbound(OutboundMessage(
                            channel=msg.channel, chat_id=msg.chat_id,
                            content="", metadata=msg.metadata or {},
                        ))
                    if msg.channel == "websocket":
                        # Signal that the turn is fully complete (all tools executed,
                        # final text streamed).  This lets WS clients know when to
                        # definitively stop the loading indicator.
                        await self.bus.publish_outbound(OutboundMessage(
                            channel=msg.channel, chat_id=msg.chat_id,
                            content="", metadata={**msg.metadata, "_turn_end": True},
                        ))
                        if msg.metadata.get("webui") is True:
                            async def _generate_title_and_notify() -> None:
                                generated = await maybe_generate_webui_title_after_turn(
                                    channel=msg.channel,
                                    metadata=msg.metadata,
                                    sessions=self.sessions,
                                    session_key=session_key,
                                    provider=self.provider,
                                    model=self.model,
                                )
                                if generated:
                                    await self.bus.publish_outbound(OutboundMessage(
                                        channel=msg.channel,
                                        chat_id=msg.chat_id,
                                        content="",
                                        metadata={**msg.metadata, "_session_updated": True},
                                    ))

                            self._schedule_background(_generate_title_and_notify())
                except asyncio.CancelledError:
                    logger.info("Task cancelled for session {}", session_key)
                    # Preserve partial context from the interrupted turn so
                    # the user does not lose tool results and assistant
                    # messages accumulated before /stop.  The checkpoint was
                    # already persisted to session metadata by
                    # _emit_checkpoint during tool execution; materializing
                    # it into session history now makes it visible in the
                    # next conversation turn.
                    try:
                        key = self._effective_session_key(msg)
                        session = self.sessions.get_or_create(key)
                        if self._restore_runtime_checkpoint(session):
                            self._clear_pending_user_turn(session)
                            self.sessions.save(session)
                            logger.info(
                                "Restored partial context for cancelled session {}",
                                key,
                            )
                    except Exception:
                        logger.debug(
                            "Could not restore checkpoint for cancelled session {}",
                            session_key,
                            exc_info=True,
                        )
                    raise
                except Exception:
                    logger.exception("Error processing message for session {}", session_key)
                    await self.bus.publish_outbound(OutboundMessage(
                        channel=msg.channel, chat_id=msg.chat_id,
                        content="Sorry, I encountered an error.",
                    ))
        finally:
            # Drain any messages still in the pending queue and re-publish
            # them to the bus so they are processed as fresh inbound messages
            # rather than silently lost.
            queue = self._pending_queues.pop(session_key, None)
            if queue is not None:
                leftover = 0
                while True:
                    try:
                        item = queue.get_nowait()
                    except asyncio.QueueEmpty:
                        break
                    await self.bus.publish_inbound(item)
                    leftover += 1
                if leftover:
                    logger.info(
                        "Re-published {} leftover message(s) to bus for session {}",
                        leftover, session_key,
                    )

    async def close_mcp(self) -> None:
        """Drain pending background archives, then close MCP connections."""
        if self._background_tasks:
            await asyncio.gather(*self._background_tasks, return_exceptions=True)
            self._background_tasks.clear()
        for name, stack in self._mcp_stacks.items():
            try:
                await stack.aclose()
            except (RuntimeError, BaseExceptionGroup):
                logger.debug("MCP server '{}' cleanup error (can be ignored)", name)
        self._mcp_stacks.clear()

    def _schedule_background(self, coro) -> None:
        """Schedule a coroutine as a tracked background task (drained on shutdown)."""
        task = asyncio.create_task(coro)
        self._background_tasks.append(task)
        task.add_done_callback(self._background_tasks.remove)

    def stop(self) -> None:
        """Stop the agent loop."""
        self._running = False
        logger.info("Agent loop stopping")

    async def _process_system_message(
        self,
        msg: InboundMessage,
        session_key: str | None = None,
        on_progress: Callable[..., Awaitable[None]] | None = None,
        on_stream: Callable[[str], Awaitable[None]] | None = None,
        on_stream_end: Callable[..., Awaitable[None]] | None = None,
        pending_queue: asyncio.Queue | None = None,
    ) -> OutboundMessage | None:
        """Process a system inbound message (e.g. subagent announce)."""
        channel, chat_id = (
            msg.chat_id.split(":", 1) if ":" in msg.chat_id else ("cli", msg.chat_id)
        )
        logger.info("Processing system message from {}", msg.sender_id)
        key = msg.session_key_override or f"{channel}:{chat_id}"
        session = self.sessions.get_or_create(key)
        if self._restore_runtime_checkpoint(session):
            self.sessions.save(session)
        if self._restore_pending_user_turn(session):
            self.sessions.save(session)

        session, pending = self.auto_compact.prepare_session(session, key)
        if pending:
            logger.info("Memory compact triggered for session {}", key)

        await self.consolidator.maybe_consolidate_by_tokens(
            session,
            replay_max_messages=self._max_messages,
        )
        is_subagent = msg.sender_id == "subagent"
        if is_subagent and self._persist_subagent_followup(session, msg):
            logger.debug("Subagent result persisted for session {}", key)
            self.sessions.save(session)
        self._set_tool_context(
            channel, chat_id, msg.metadata.get("message_id"),
            msg.metadata, session_key=key,
        )
        _hist_kwargs: dict[str, Any] = {
            "max_messages": self._max_messages,
            "max_tokens": self._replay_token_budget(),
            "include_timestamps": True,
        }
        history = session.get_history(**_hist_kwargs)
        current_role = "assistant" if is_subagent else "user"

        messages = self.context.build_messages(
            history=history,
            current_message="" if is_subagent else msg.content,
            channel=channel,
            chat_id=chat_id,
            current_role=current_role,
            sender_id=msg.sender_id,
            session_summary=pending,
        )
        final_content, _, all_msgs, stop_reason, _ = await self._run_agent_loop(
            messages, session=session, channel=channel, chat_id=chat_id,
            message_id=msg.metadata.get("message_id"),
            metadata=msg.metadata,
            session_key=key,
            pending_queue=pending_queue,
        )
        self._save_turn(session, all_msgs, 1 + len(history))
        session.enforce_file_cap(on_archive=self.context.memory.raw_archive)
        self._clear_runtime_checkpoint(session)
        self.sessions.save(session)
        self._schedule_background(
            self.consolidator.maybe_consolidate_by_tokens(
                session,
                replay_max_messages=self._max_messages,
            )
        )
        content = final_content or "Background task completed."
        outbound_metadata: dict[str, Any] = {}
        if channel == "slack" and key.startswith("slack:") and key.count(":") >= 2:
            outbound_metadata["slack"] = {"thread_ts": key.split(":", 2)[2]}
        if origin_message_id := msg.metadata.get("origin_message_id"):
            outbound_metadata["origin_message_id"] = origin_message_id
        return OutboundMessage(
            channel=channel,
            chat_id=chat_id,
            content=content,
            metadata=outbound_metadata,
        )

    async def _process_message(
        self,
        msg: InboundMessage,
        session_key: str | None = None,
        on_progress: Callable[..., Awaitable[None]] | None = None,
        on_stream: Callable[[str], Awaitable[None]] | None = None,
        on_stream_end: Callable[..., Awaitable[None]] | None = None,
        pending_queue: asyncio.Queue | None = None,
    ) -> OutboundMessage | None:
        """Process a single inbound message and return the response."""
        self._refresh_provider_snapshot()

        if msg.channel == "system":
            return await self._process_system_message(
                msg,
                session_key=session_key,
                on_progress=on_progress,
                on_stream=on_stream,
                on_stream_end=on_stream_end,
                pending_queue=pending_queue,
            )

        key = session_key or msg.session_key
        ctx = TurnContext(
            msg=msg,
            session=None,
            session_key=key,
            state=TurnState.RESTORE,
            turn_id=f"{key}:{time.time_ns()}",
            on_progress=on_progress,
            on_stream=on_stream,
            on_stream_end=on_stream_end,
            pending_queue=pending_queue,
        )

        while ctx.state is not TurnState.DONE:
            handler_name = f"_state_{ctx.state.name.lower()}"
            handler = getattr(self, handler_name, None)
            if handler is None:
                raise RuntimeError(f"Missing state handler for {ctx.state}")

            t0 = time.perf_counter()
            try:
                event = await handler(ctx)
            except Exception:
                duration = (time.perf_counter() - t0) * 1000
                ctx.trace.append(
                    StateTraceEntry(
                        state=ctx.state,
                        started_at=t0,
                        duration_ms=duration,
                        event="",
                        error="exception",
                    )
                )
                raise

            duration = (time.perf_counter() - t0) * 1000
            ctx.trace.append(
                StateTraceEntry(
                    state=ctx.state,
                    started_at=t0,
                    duration_ms=duration,
                    event=event,
                )
            )
            logger.debug(
                "[turn {}] State {} took {:.1f}ms -> event {}",
                ctx.turn_id,
                ctx.state.name,
                duration,
                event,
            )

            next_state = self._TRANSITIONS.get((ctx.state, event))
            if next_state is None:
                raise RuntimeError(
                    f"[turn {ctx.turn_id}] No transition from {ctx.state} "
                    f"on event {event!r}"
                )
            ctx.state = next_state

        logger.debug(
            "[turn {}] Turn completed after {} states",
            ctx.turn_id,
            len(ctx.trace),
        )
        return ctx.outbound

    def _assemble_outbound(
        self,
        msg: InboundMessage,
        final_content: str,
        all_msgs: list[dict[str, Any]],
        stop_reason: str,
        had_injections: bool,
        generated_media: list[str],
        on_stream: Callable[[str], Awaitable[None]] | None,
    ) -> OutboundMessage | None:
        """Assemble the final outbound message from turn results."""
        # MessageTool suppression
        if (mt := self.tools.get("message")) and isinstance(mt, MessageTool) and mt._sent_in_turn:
            if not had_injections or stop_reason == "empty_final_response":
                return None

        preview = final_content[:120] + "..." if len(final_content) > 120 else final_content
        logger.info("Response to {}:{}: {}", msg.channel, msg.sender_id, preview)

        meta = dict(msg.metadata or {})
        if on_stream is not None and stop_reason not in {"error", "tool_error"}:
            meta["_streamed"] = True

        return OutboundMessage(
            channel=msg.channel,
            chat_id=msg.chat_id,
            content=final_content,
            media=generated_media,
            metadata=meta,
        )

    async def _state_restore(self, ctx: TurnContext) -> TurnState:
        """Restore checkpoint / pending user turn; extract documents."""
        msg = ctx.msg

        if msg.media:
            new_content, image_only = extract_documents(msg.content, msg.media)
            ctx.msg = dataclasses.replace(msg, content=new_content, media=image_only)
            msg = ctx.msg

        preview = msg.content[:80] + "..." if len(msg.content) > 80 else msg.content
        logger.info("Processing message from {}:{}: {}", msg.channel, msg.sender_id, preview)

        # Session is already fetched by the caller (_process_message) but
        # ensure it exists in case this handler is invoked independently.
        if ctx.session is None:
            ctx.session = self.sessions.get_or_create(ctx.session_key)
        mark_webui_session(ctx.session, msg.metadata)

        if self._restore_runtime_checkpoint(ctx.session):
            self.sessions.save(ctx.session)
        if self._restore_pending_user_turn(ctx.session):
            self.sessions.save(ctx.session)

        return "ok"

    async def _state_compact(self, ctx: TurnContext) -> str:
        ctx.session, pending = self.auto_compact.prepare_session(ctx.session, ctx.session_key)
        ctx.pending_summary = pending
        return "ok"

    async def _state_command(self, ctx: TurnContext) -> str:
        raw = ctx.msg.content.strip()
        cmd_ctx = CommandContext(
            msg=ctx.msg, session=ctx.session, key=ctx.session_key, raw=raw, loop=self
        )
        result = await self.commands.dispatch(cmd_ctx)
        if result is not None:
            ctx.outbound = result
            # Shortcut commands skip BUILD and SAVE, so we must persist the
            # turn here so WebUI history hydration after _turn_end sees the
            # message.  Mark messages with _command so get_history can filter
            # them out of LLM context.  /new is excluded because it
            # intentionally clears the session.
            if raw.lower() != "/new":
                ctx.user_persisted_early = self._persist_user_message_early(
                    ctx.msg, ctx.session, _command=True
                )
                ctx.session.add_message(
                    "assistant", result.content, _command=True
                )
                self.sessions.save(ctx.session)
                self._clear_pending_user_turn(ctx.session)
            return "shortcut"
        return "dispatch"

    async def _state_build(self, ctx: TurnContext) -> str:
        await self.consolidator.maybe_consolidate_by_tokens(
            ctx.session,
            replay_max_messages=self._max_messages,
        )
        self._set_tool_context(
            ctx.msg.channel,
            ctx.msg.chat_id,
            ctx.msg.metadata.get("message_id"),
            ctx.msg.metadata,
            session_key=ctx.session_key,
        )
        if message_tool := self.tools.get("message"):
            if isinstance(message_tool, MessageTool):
                message_tool.start_turn()

        _hist_kwargs: dict[str, Any] = {
            "max_messages": self._max_messages,
            "max_tokens": self._replay_token_budget(),
            "include_timestamps": True,
        }
        ctx.history = ctx.session.get_history(**_hist_kwargs)

        ctx.initial_messages = self._build_initial_messages(
            ctx.msg, ctx.session, ctx.history, ctx.pending_summary
        )
        ctx.user_persisted_early = self._persist_user_message_early(
            ctx.msg, ctx.session
        )

        if ctx.on_progress is None:
            ctx.on_progress = await self._build_bus_progress_callback(ctx.msg)
        if ctx.on_retry_wait is None:
            ctx.on_retry_wait = await self._build_retry_wait_callback(ctx.msg)

        return "ok"

    async def _state_run(self, ctx: TurnContext) -> str:
        result = await self._run_agent_loop(
            ctx.initial_messages,
            on_progress=ctx.on_progress,
            on_stream=ctx.on_stream,
            on_stream_end=ctx.on_stream_end,
            on_retry_wait=ctx.on_retry_wait,
            session=ctx.session,
            channel=ctx.msg.channel,
            chat_id=ctx.msg.chat_id,
            message_id=ctx.msg.metadata.get("message_id"),
            metadata=ctx.msg.metadata,
            session_key=ctx.session_key,
            pending_queue=ctx.pending_queue,
        )
        final_content, tools_used, all_msgs, stop_reason, had_injections = result
        ctx.final_content = final_content
        ctx.tools_used = tools_used
        ctx.all_messages = all_msgs
        ctx.stop_reason = stop_reason
        ctx.had_injections = had_injections
        return "ok"

    async def _state_save(self, ctx: TurnContext) -> str:
        if ctx.final_content is None or not ctx.final_content.strip():
            ctx.final_content = EMPTY_FINAL_RESPONSE_MESSAGE

        ctx.save_skip = 1 + len(ctx.history) + (1 if ctx.user_persisted_early else 0)
        skip_msgs = ctx.all_messages[ctx.save_skip:]
        ctx.generated_media = generated_image_paths_from_messages(skip_msgs)
        last_msg = ctx.all_messages[-1] if ctx.all_messages else None
        if ctx.generated_media and last_msg and last_msg.get("role") == "assistant":
            existing_media = last_msg.get("media")
            media = existing_media if isinstance(existing_media, list) else []
            last_msg["media"] = list(dict.fromkeys([*media, *ctx.generated_media]))

        self._save_turn(ctx.session, ctx.all_messages, ctx.save_skip)
        ctx.session.enforce_file_cap(on_archive=self.context.memory.raw_archive)
        self._clear_pending_user_turn(ctx.session)
        self._clear_runtime_checkpoint(ctx.session)
        self.sessions.save(ctx.session)
        self._schedule_background(
            self.consolidator.maybe_consolidate_by_tokens(
                ctx.session,
                replay_max_messages=self._max_messages,
            )
        )
        return "ok"

    async def _state_respond(self, ctx: TurnContext) -> str:
        ctx.outbound = self._assemble_outbound(
            ctx.msg,
            ctx.final_content,
            ctx.all_messages,
            ctx.stop_reason,
            ctx.had_injections,
            ctx.generated_media,
            ctx.on_stream,
        )
        return "ok"

    def _sanitize_persisted_blocks(
        self,
        content: list[dict[str, Any]],
        *,
        should_truncate_text: bool = False,
        drop_runtime: bool = False,
    ) -> list[dict[str, Any]]:
        """Strip volatile multimodal payloads before writing session history."""
        filtered: list[dict[str, Any]] = []
        for block in content:
            if not isinstance(block, dict):
                filtered.append(block)
                continue

            if (
                drop_runtime
                and block.get("type") == "text"
                and isinstance(block.get("text"), str)
                and block["text"].startswith(ContextBuilder._RUNTIME_CONTEXT_TAG)
            ):
                continue

            if block.get("type") == "image_url" and block.get("image_url", {}).get(
                "url", ""
            ).startswith("data:image/"):
                path = (block.get("_meta") or {}).get("path", "")
                filtered.append({"type": "text", "text": image_placeholder_text(path)})
                continue

            if block.get("type") == "text" and isinstance(block.get("text"), str):
                text = block["text"]
                if should_truncate_text and len(text) > self.max_tool_result_chars:
                    text = truncate_text_fn(text, self.max_tool_result_chars)
                filtered.append({**block, "text": text})
                continue

            filtered.append(block)

        return filtered

    def _save_turn(self, session: Session, messages: list[dict], skip: int) -> None:
        """Save new-turn messages into session, truncating large tool results."""
        from datetime import datetime

        for m in messages[skip:]:
            entry = dict(m)
            role, content = entry.get("role"), entry.get("content")
            if role == "assistant" and not content and not entry.get("tool_calls"):
                continue  # skip empty assistant messages — they poison session context
            if role == "tool":
                if isinstance(content, str) and len(content) > self.max_tool_result_chars:
                    entry["content"] = truncate_text_fn(content, self.max_tool_result_chars)
                elif isinstance(content, list):
                    filtered = self._sanitize_persisted_blocks(content, should_truncate_text=True)
                    if not filtered:
                        continue
                    entry["content"] = filtered
            elif role == "user":
                if isinstance(content, str) and content.startswith(ContextBuilder._RUNTIME_CONTEXT_TAG):
                    # Strip the entire runtime-context block (including any session summary).
                    # The block is bounded by _RUNTIME_CONTEXT_TAG and _RUNTIME_CONTEXT_END.
                    end_marker = ContextBuilder._RUNTIME_CONTEXT_END
                    end_pos = content.find(end_marker)
                    if end_pos >= 0:
                        after = content[end_pos + len(end_marker):].lstrip("\n")
                        if after:
                            entry["content"] = after
                        else:
                            continue
                    else:
                        # Fallback: no end marker found, strip the tag prefix
                        after_tag = content[len(ContextBuilder._RUNTIME_CONTEXT_TAG):].lstrip("\n")
                        if after_tag.strip():
                            entry["content"] = after_tag
                        else:
                            continue
                if isinstance(content, list):
                    filtered = self._sanitize_persisted_blocks(content, drop_runtime=True)
                    if not filtered:
                        continue
                    entry["content"] = filtered
            entry.setdefault("timestamp", datetime.now().isoformat())
            session.messages.append(entry)
        session.updated_at = datetime.now()

    def _persist_subagent_followup(self, session: Session, msg: InboundMessage) -> bool:
        """Persist subagent follow-ups before prompt assembly so history stays durable.

        Returns True if a new entry was appended; False if the follow-up was
        deduped (same ``subagent_task_id`` already in session) or carries no
        content worth persisting.
        """
        if not msg.content:
            return False
        task_id = msg.metadata.get("subagent_task_id") if isinstance(msg.metadata, dict) else None
        if task_id and any(
            m.get("injected_event") == "subagent_result" and m.get("subagent_task_id") == task_id
            for m in session.messages
        ):
            return False
        session.add_message(
            "assistant",
            msg.content,
            sender_id=msg.sender_id,
            injected_event="subagent_result",
            subagent_task_id=task_id,
        )
        return True

    def _set_runtime_checkpoint(self, session: Session, payload: dict[str, Any]) -> None:
        """Persist the latest in-flight turn state into session metadata."""
        session.metadata[self._RUNTIME_CHECKPOINT_KEY] = payload
        self.sessions.save(session)

    def _mark_pending_user_turn(self, session: Session) -> None:
        session.metadata[self._PENDING_USER_TURN_KEY] = True

    def _clear_pending_user_turn(self, session: Session) -> None:
        session.metadata.pop(self._PENDING_USER_TURN_KEY, None)

    def _clear_runtime_checkpoint(self, session: Session) -> None:
        if self._RUNTIME_CHECKPOINT_KEY in session.metadata:
            session.metadata.pop(self._RUNTIME_CHECKPOINT_KEY, None)

    @staticmethod
    def _checkpoint_message_key(message: dict[str, Any]) -> tuple[Any, ...]:
        return (
            message.get("role"),
            message.get("content"),
            message.get("tool_call_id"),
            message.get("name"),
            message.get("tool_calls"),
            message.get("reasoning_content"),
            message.get("thinking_blocks"),
        )

    def _restore_runtime_checkpoint(self, session: Session) -> bool:
        """Materialize an unfinished turn into session history before a new request."""
        from datetime import datetime

        checkpoint = session.metadata.get(self._RUNTIME_CHECKPOINT_KEY)
        if not isinstance(checkpoint, dict):
            return False

        assistant_message = checkpoint.get("assistant_message")
        completed_tool_results = checkpoint.get("completed_tool_results") or []
        pending_tool_calls = checkpoint.get("pending_tool_calls") or []

        restored_messages: list[dict[str, Any]] = []
        if isinstance(assistant_message, dict):
            restored = dict(assistant_message)
            restored.setdefault("timestamp", datetime.now().isoformat())
            restored_messages.append(restored)
        for message in completed_tool_results:
            if isinstance(message, dict):
                restored = dict(message)
                restored.setdefault("timestamp", datetime.now().isoformat())
                restored_messages.append(restored)
        for tool_call in pending_tool_calls:
            if not isinstance(tool_call, dict):
                continue
            tool_id = tool_call.get("id")
            name = ((tool_call.get("function") or {}).get("name")) or "tool"
            restored_messages.append(
                {
                    "role": "tool",
                    "tool_call_id": tool_id,
                    "name": name,
                    "content": "Error: Task interrupted before this tool finished.",
                    "timestamp": datetime.now().isoformat(),
                }
            )

        overlap = 0
        max_overlap = min(len(session.messages), len(restored_messages))
        for size in range(max_overlap, 0, -1):
            existing = session.messages[-size:]
            restored = restored_messages[:size]
            if all(
                self._checkpoint_message_key(left) == self._checkpoint_message_key(right)
                for left, right in zip(existing, restored)
            ):
                overlap = size
                break
        session.messages.extend(restored_messages[overlap:])

        self._clear_pending_user_turn(session)
        self._clear_runtime_checkpoint(session)
        return True

    def _restore_pending_user_turn(self, session: Session) -> bool:
        """Close a turn that only persisted the user message before crashing."""
        from datetime import datetime

        if not session.metadata.get(self._PENDING_USER_TURN_KEY):
            return False

        if session.messages and session.messages[-1].get("role") == "user":
            session.messages.append(
                {
                    "role": "assistant",
                    "content": "Error: Task interrupted before a response was generated.",
                    "timestamp": datetime.now().isoformat(),
                }
            )
            session.updated_at = datetime.now()

        self._clear_pending_user_turn(session)
        return True

    async def process_direct(
        self,
        content: str,
        session_key: str = "cli:direct",
        channel: str = "cli",
        chat_id: str = "direct",
        media: list[str] | None = None,
        on_progress: Callable[..., Awaitable[None]] | None = None,
        on_stream: Callable[[str], Awaitable[None]] | None = None,
        on_stream_end: Callable[..., Awaitable[None]] | None = None,
    ) -> OutboundMessage | None:
        """Process a message directly and return the outbound payload."""
        await self._connect_mcp()
        msg = InboundMessage(
            channel=channel, sender_id="user", chat_id=chat_id,
            content=content, media=media or [],
        )
        return await self._process_message(
            msg,
            session_key=session_key,
            on_progress=on_progress,
            on_stream=on_stream,
            on_stream_end=on_stream_end,
        )