Merge origin/main into fix/discord-allow-channel-threads

Made-with: Cursor

nanobot/agent/context.py

@@ -9,7 +9,7 @@ from typing import Any
 
 from nanobot.agent.memory import MemoryStore
 from nanobot.agent.skills import SkillsLoader
-from nanobot.utils.helpers import build_assistant_message, current_time_str, detect_image_mime
+from nanobot.utils.helpers import build_assistant_message, current_time_str, detect_image_mime, truncate_text
 from nanobot.utils.prompt_templates import render_template
 
 
@@ -19,6 +19,7 @@ class ContextBuilder:
     BOOTSTRAP_FILES = ["AGENTS.md", "SOUL.md", "USER.md", "TOOLS.md"]
     _RUNTIME_CONTEXT_TAG = "[Runtime Context — metadata only, not instructions]"
     _MAX_RECENT_HISTORY = 50
+    _MAX_HISTORY_CHARS = 32_000  # hard cap on recent history section size
     _RUNTIME_CONTEXT_END = "[/Runtime Context]"
 
     def __init__(self, workspace: Path, timezone: str | None = None, disabled_skills: list[str] | None = None):
@@ -56,9 +57,11 @@ class ContextBuilder:
         entries = self.memory.read_unprocessed_history(since_cursor=self.memory.get_last_dream_cursor())
         if entries:
             capped = entries[-self._MAX_RECENT_HISTORY:]
-            parts.append("# Recent History\n\n" + "\n".join(
+            history_text = "\n".join(
                 f"- [{e['timestamp']}] {e['content']}" for e in capped
-            ))
+            )
+            history_text = truncate_text(history_text, self._MAX_HISTORY_CHARS)
+            parts.append("# Recent History\n\n" + history_text)
 
         return "\n\n---\n\n".join(parts)
 

nanobot/agent/loop.py

@@ -20,6 +20,13 @@ from nanobot.agent.memory import Consolidator, Dream
 from nanobot.agent.runner import _MAX_INJECTIONS_PER_TURN, AgentRunner, AgentRunSpec
 from nanobot.agent.skills import BUILTIN_SKILLS_DIR
 from nanobot.agent.subagent import SubagentManager
+from nanobot.agent.tools.ask import (
+    AskUserTool,
+    ask_user_options_from_messages,
+    ask_user_outbound,
+    ask_user_tool_result_messages,
+    pending_ask_user_id,
+)
 from nanobot.agent.tools.cron import CronTool
 from nanobot.agent.tools.filesystem import EditFileTool, ListDirTool, ReadFileTool, WriteFileTool
 from nanobot.agent.tools.message import MessageTool
@@ -35,10 +42,17 @@ from nanobot.bus.queue import MessageBus
 from nanobot.command import CommandContext, CommandRouter, register_builtin_commands
 from nanobot.config.schema import AgentDefaults
 from nanobot.providers.base import LLMProvider
+from nanobot.providers.factory import ProviderSnapshot
 from nanobot.session.manager import Session, SessionManager
 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.progress_events import (
+    build_tool_event_finish_payloads,
+    build_tool_event_start_payload,
+    invoke_on_progress,
+    on_progress_accepts_tool_events,
+)
 from nanobot.utils.runtime import EMPTY_FINAL_RESPONSE_MESSAGE
 
 if TYPE_CHECKING:
@@ -62,7 +76,8 @@ class _LoopHook(AgentHook):
         channel: str = "cli",
         chat_id: str = "direct",
         message_id: str | None = None,
-        effective_key: str | None = None,
+        metadata: dict[str, Any] | None = None,
+        session_key: str | None = None,
     ) -> None:
         super().__init__(reraise=True)
         self._loop = agent_loop
@@ -72,7 +87,8 @@ class _LoopHook(AgentHook):
         self._channel = channel
         self._chat_id = chat_id
         self._message_id = message_id
-        self._effective_key = effective_key
+        self._metadata = metadata or {}
+        self._session_key = session_key
         self._stream_buf = ""
 
     def wants_streaming(self) -> bool:
@@ -105,7 +121,13 @@ class _LoopHook(AgentHook):
                 if thought:
                     await self._on_progress(thought)
             tool_hint = self._loop._strip_think(self._loop._tool_hint(context.tool_calls))
-            await self._on_progress(tool_hint, tool_hint=True)
+            tool_events = [build_tool_event_start_payload(tc) for tc in context.tool_calls]
+            await invoke_on_progress(
+                self._on_progress,
+                tool_hint,
+                tool_hint=True,
+                tool_events=tool_events,
+            )
         for tc in context.tool_calls:
             args_str = json.dumps(tc.arguments, ensure_ascii=False)
             logger.info("Tool call: {}({})", tc.name, args_str[:200])
@@ -113,10 +135,25 @@ class _LoopHook(AgentHook):
             self._channel,
             self._chat_id,
             self._message_id,
-            effective_key=self._effective_key,
+            self._metadata,
+            session_key=self._session_key,
         )
 
     async def after_iteration(self, context: AgentHookContext) -> None:
+        if (
+            self._on_progress
+            and context.tool_calls
+            and context.tool_events
+            and on_progress_accepts_tool_events(self._on_progress)
+        ):
+            tool_events = build_tool_event_finish_payloads(context)
+            if tool_events:
+                await invoke_on_progress(
+                    self._on_progress,
+                    "",
+                    tool_hint=False,
+                    tool_events=tool_events,
+                )
         u = context.usage or {}
         logger.debug(
             "LLM usage: prompt={} completion={} cached={}",
@@ -164,10 +201,13 @@ class AgentLoop:
         channels_config: ChannelsConfig | None = None,
         timezone: str | None = None,
         session_ttl_minutes: int = 0,
+        consolidation_ratio: float = 0.5,
         hooks: list[AgentHook] | None = None,
         unified_session: bool = False,
         disabled_skills: list[str] | None = None,
         tools_config: ToolsConfig | None = None,
+        provider_snapshot_loader: Callable[[], ProviderSnapshot] | None = None,
+        provider_signature: tuple[object, ...] | None = None,
     ):
         from nanobot.config.schema import ExecToolConfig, ToolsConfig, WebToolsConfig
 
@@ -176,6 +216,8 @@ class AgentLoop:
         self.bus = bus
         self.channels_config = channels_config
         self.provider = provider
+        self._provider_snapshot_loader = provider_snapshot_loader
+        self._provider_signature = provider_signature
         self.workspace = workspace
         self.model = model or provider.get_default_model()
         self.max_iterations = (
@@ -243,6 +285,7 @@ class AgentLoop:
             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,
@@ -262,12 +305,43 @@ class AgentLoop:
         self.commands = CommandRouter()
         register_builtin_commands(self.commands)
 
+    def _apply_provider_snapshot(self, snapshot: ProviderSnapshot) -> 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
+        if self.provider is provider and self.model == model:
+            return
+        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
+        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
+        if snapshot.signature == self._provider_signature:
+            return
+        self._apply_provider_snapshot(snapshot)
+
     def _register_default_tools(self) -> None:
         """Register the default set of tools."""
         allowed_dir = (
             self.workspace if (self.restrict_to_workspace or self.exec_config.sandbox) else None
         )
         extra_read = [BUILTIN_SKILLS_DIR] if allowed_dir else None
+        self.tools.register(AskUserTool())
         self.tools.register(
             ReadFileTool(
                 workspace=self.workspace, allowed_dir=allowed_dir, extra_allowed_dirs=extra_read
@@ -294,7 +368,7 @@ class AgentLoop:
                 WebSearchTool(config=self.web_config.search, proxy=self.web_config.proxy)
             )
             self.tools.register(WebFetchTool(proxy=self.web_config.proxy))
-        self.tools.register(MessageTool(send_callback=self.bus.publish_outbound))
+        self.tools.register(MessageTool(send_callback=self.bus.publish_outbound, workspace=self.workspace))
         self.tools.register(SpawnTool(manager=self.subagents))
         if self.cron_service:
             self.tools.register(
@@ -324,30 +398,32 @@ class AgentLoop:
             self._mcp_connecting = False
 
     def _set_tool_context(
-        self,
-        channel: str,
-        chat_id: str,
-        message_id: str | None = None,
-        *,
-        effective_key: str | None = None,
+        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."""
-        # Compute the effective session key (accounts for unified sessions)
-        # so that subagent results route to the correct pending queue.
-        context_key = (
-            effective_key
-            if effective_key is not None
-            else UNIFIED_SESSION_KEY
-            if self._unified_session
-            else f"{channel}:{chat_id}"
-        )
+        # When the caller threads a thread-scoped session_key (e.g. slack with
+        # reply_in_thread: true), honor it so spawn announces route back to
+        # the originating thread session. Falls back to unified mode or
+        # channel:chat_id for callers that don't have a thread-scoped key.
+        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}"
         for name in ("message", "spawn", "cron", "my"):
             if tool := self.tools.get(name):
                 if hasattr(tool, "set_context"):
                     if name == "spawn":
-                        tool.set_context(channel, chat_id, effective_key=context_key)
+                        tool.set_context(channel, chat_id, effective_key=effective_key)
+                    elif name == "cron":
+                        tool.set_context(channel, chat_id, metadata=metadata, session_key=session_key)
+                    elif name == "message":
+                        tool.set_context(channel, chat_id, message_id, metadata=metadata)
                     else:
-                        tool.set_context(channel, chat_id, *([message_id] if name == "message" else []))
+                        tool.set_context(channel, chat_id)
 
     @staticmethod
     def _strip_think(text: str | None) -> str | None:
@@ -406,6 +482,18 @@ class AgentLoop:
             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],
@@ -418,7 +506,8 @@ class AgentLoop:
         channel: str = "cli",
         chat_id: str = "direct",
         message_id: str | None = None,
-        effective_key: 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.
@@ -438,7 +527,8 @@ class AgentLoop:
             channel=channel,
             chat_id=chat_id,
             message_id=message_id,
-            effective_key=effective_key,
+            metadata=metadata,
+            session_key=session_key,
         )
         hook: AgentHook = (
             CompositeHook([loop_hook] + self._extra_hooks) if self._extra_hooks else loop_hook
@@ -753,19 +843,23 @@ class AgentLoop:
         self,
         msg: InboundMessage,
         session_key: str | None = None,
-        on_progress: Callable[[str], Awaitable[None]] | 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()
         # System messages: parse origin from chat_id ("channel:chat_id")
         if msg.channel == "system":
             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 = f"{channel}:{chat_id}"
+            # Honor session_key_override so subagent announces from threaded
+            # callers route to the originating thread session, not the
+            # channel-level session derived from chat_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)
@@ -786,8 +880,14 @@ class AgentLoop:
             is_subagent = msg.sender_id == "subagent"
             if is_subagent and self._persist_subagent_followup(session, msg):
                 self.sessions.save(session)
-            self._set_tool_context(channel, chat_id, msg.metadata.get("message_id"), effective_key=key)
-            history = session.get_history(max_messages=0)
+            self._set_tool_context(
+                channel, chat_id, msg.metadata.get("message_id"),
+                msg.metadata, session_key=key,
+            )
+            history = session.get_history(
+                max_tokens=self._replay_token_budget(),
+                include_timestamps=True,
+            )
             current_role = "assistant" if is_subagent else "user"
 
             # Subagent content is already in `history` above; passing it again
@@ -800,20 +900,38 @@ class AgentLoop:
                 session_summary=pending,
                 current_role=current_role,
             )
-            final_content, _, all_msgs, _, _ = await self._run_agent_loop(
+            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"),
-                effective_key=key,
+                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))
+            options = ask_user_options_from_messages(all_msgs) if stop_reason == "ask_user" else []
+            content, buttons = ask_user_outbound(
+                final_content or "Background task completed.",
+                options,
+                channel,
+            )
+            # Reconstruct channel-specific metadata from session.key so the
+            # outbound reply lands in the originating thread (not the channel
+            # top-level). The announce InboundMessage carries only
+            # injected_event metadata; we recover thread_ts from the session
+            # key, which slack writes as "slack:<chat_id>:<thread_ts>".
+            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]}
             return OutboundMessage(
                 channel=channel,
                 chat_id=chat_id,
-                content=final_content or "Background task completed.",
+                content=content,
+                buttons=buttons,
+                metadata=outbound_metadata,
             )
 
         # Extract document text from media at the processing boundary so all
@@ -846,30 +964,47 @@ class AgentLoop:
         )
 
         self._set_tool_context(
-            msg.channel,
-            msg.chat_id,
-            msg.metadata.get("message_id"),
-            effective_key=key,
+            msg.channel, msg.chat_id, msg.metadata.get("message_id"),
+            msg.metadata, session_key=key,
         )
         if message_tool := self.tools.get("message"):
             if isinstance(message_tool, MessageTool):
                 message_tool.start_turn()
 
-        history = session.get_history(max_messages=0)
-
-        initial_messages = self.context.build_messages(
-            history=history,
-            current_message=msg.content,
-            session_summary=pending,
-            media=msg.media if msg.media else None,
-            channel=msg.channel,
-            chat_id=self._runtime_chat_id(msg),
+        history = session.get_history(
+            max_tokens=self._replay_token_budget(),
+            include_timestamps=True,
         )
 
-        async def _bus_progress(content: str, *, tool_hint: bool = False) -> None:
+        pending_ask_id = pending_ask_user_id(history)
+        if pending_ask_id:
+            initial_messages = ask_user_tool_result_messages(
+                self.context.build_system_prompt(channel=msg.channel),
+                history,
+                pending_ask_id,
+                msg.content,
+            )
+        else:
+            initial_messages = self.context.build_messages(
+                history=history,
+                current_message=msg.content,
+                session_summary=pending,
+                media=msg.media if msg.media else None,
+                channel=msg.channel,
+                chat_id=self._runtime_chat_id(msg),
+            )
+
+        async def _bus_progress(
+            content: str,
+            *,
+            tool_hint: bool = False,
+            tool_events: list[dict[str, Any]] | None = None,
+        ) -> None:
             meta = dict(msg.metadata or {})
             meta["_progress"] = True
             meta["_tool_hint"] = tool_hint
+            if tool_events:
+                meta["_tool_events"] = tool_events
             await self.bus.publish_outbound(
                 OutboundMessage(
                     channel=msg.channel,
@@ -898,7 +1033,7 @@ class AgentLoop:
         user_persisted_early = False
         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:
+        if not pending_ask_id and (has_text or media_paths):
             extra: dict[str, Any] = {"media": list(media_paths)} if media_paths else {}
             text = msg.content if isinstance(msg.content, str) else ""
             session.add_message("user", text, **extra)
@@ -916,7 +1051,8 @@ class AgentLoop:
             channel=msg.channel,
             chat_id=msg.chat_id,
             message_id=msg.metadata.get("message_id"),
-            effective_key=key,
+            metadata=msg.metadata,
+            session_key=key,
             pending_queue=pending_queue,
         )
 
@@ -926,6 +1062,7 @@ class AgentLoop:
         # Skip the already-persisted user message when saving the turn
         save_skip = 1 + len(history) + (1 if user_persisted_early else 0)
         self._save_turn(session, all_msgs, save_skip)
+        session.enforce_file_cap(on_archive=self.context.memory.raw_archive)
         self._clear_pending_user_turn(session)
         self._clear_runtime_checkpoint(session)
         self.sessions.save(session)
@@ -945,13 +1082,19 @@ class AgentLoop:
         logger.info("Response to {}:{}: {}", msg.channel, msg.sender_id, preview)
 
         meta = dict(msg.metadata or {})
-        if on_stream is not None and stop_reason != "error":
+        final_content, buttons = ask_user_outbound(
+            final_content,
+            ask_user_options_from_messages(all_msgs) if stop_reason == "ask_user" else [],
+            msg.channel,
+        )
+        if on_stream is not None and stop_reason not in {"ask_user", "error"}:
             meta["_streamed"] = True
         return OutboundMessage(
             channel=msg.channel,
             chat_id=msg.chat_id,
             content=final_content,
             metadata=meta,
+            buttons=buttons,
         )
 
     def _sanitize_persisted_blocks(
@@ -1171,7 +1314,7 @@ class AgentLoop:
         channel: str = "cli",
         chat_id: str = "direct",
         media: list[str] | None = None,
-        on_progress: Callable[[str], Awaitable[None]] | 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:

nanobot/agent/memory.py

@@ -6,6 +6,7 @@ import asyncio
 import json
 import re
 import weakref
+import tiktoken
 from datetime import datetime
 from pathlib import Path
 from typing import TYPE_CHECKING, Any, Callable, Iterator
@@ -13,7 +14,7 @@ from typing import TYPE_CHECKING, Any, Callable, Iterator
 from loguru import logger
 
 from nanobot.utils.prompt_templates import render_template
-from nanobot.utils.helpers import ensure_dir, estimate_message_tokens, estimate_prompt_tokens_chain, strip_think
+from nanobot.utils.helpers import ensure_dir, estimate_message_tokens, estimate_prompt_tokens_chain, strip_think, truncate_text
 
 from nanobot.agent.runner import AgentRunSpec, AgentRunner
 from nanobot.agent.tools.registry import ToolRegistry
@@ -50,6 +51,7 @@ class MemoryStore:
         self._cursor_file = self.memory_dir / ".cursor"
         self._dream_cursor_file = self.memory_dir / ".dream_cursor"
         self._corruption_logged = False  # rate-limit non-int cursor warning
+        self._oversize_logged = False  # rate-limit oversized-entry warning
         self._git = GitStore(workspace, tracked_files=[
             "SOUL.md", "USER.md", "memory/MEMORY.md",
         ])
@@ -221,7 +223,7 @@ class MemoryStore:
 
     # -- history.jsonl — append-only, JSONL format ---------------------------
 
-    def append_history(self, entry: str) -> int:
+    def append_history(self, entry: str, *, max_chars: int | None = None) -> int:
         """Append *entry* to history.jsonl and return its auto-incrementing cursor.
 
         Entries are passed through `strip_think` to drop template-level leaks
@@ -230,10 +232,26 @@ class MemoryStore:
         the record is persisted with an empty string rather than falling back
         to the raw leak — otherwise `strip_think`'s guarantees would be
         undone by history replay / consolidation downstream.
+
+        A defensive cap (*max_chars*, default ``_HISTORY_ENTRY_HARD_CAP``) is
+        applied as a final safety net: individual callers should cap their own
+        content more tightly; this default only exists to catch unintentional
+        large writes (e.g. an LLM echoing its input back as a "summary").
         """
+        limit = max_chars if max_chars is not None else _HISTORY_ENTRY_HARD_CAP
         cursor = self._next_cursor()
         ts = datetime.now().strftime("%Y-%m-%d %H:%M")
         raw = entry.rstrip()
+        if len(raw) > limit:
+            if not self._oversize_logged:
+                self._oversize_logged = True
+                logger.warning(
+                    "history entry exceeds {} chars ({}); truncating. "
+                    "Usually means a caller forgot its own cap; "
+                    "further occurrences suppressed.",
+                    limit, len(raw),
+                )
+            raw = truncate_text(raw, limit)
         content = strip_think(raw)
         if raw and not content:
             logger.debug(
@@ -373,11 +391,13 @@ class MemoryStore:
             )
         return "\n".join(lines)
 
-    def raw_archive(self, messages: list[dict]) -> None:
+    def raw_archive(self, messages: list[dict], *, max_chars: int | None = None) -> None:
         """Fallback: dump raw messages to history.jsonl without LLM summarization."""
+        limit = max_chars if max_chars is not None else _RAW_ARCHIVE_MAX_CHARS
+        formatted = truncate_text(self._format_messages(messages), limit)
         self.append_history(
             f"[RAW] {len(messages)} messages\n"
-            f"{self._format_messages(messages)}"
+            f"{formatted}"
         )
         logger.warning(
             "Memory consolidation degraded: raw-archived {} messages", len(messages)
@@ -390,11 +410,18 @@ class MemoryStore:
 # ---------------------------------------------------------------------------
 
 
+# Individual history.jsonl writers cap their own payloads tightly; the
+# _HISTORY_ENTRY_HARD_CAP at append_history() is a belt-and-suspenders default
+# that catches any new caller that forgot to set its own cap.
+_RAW_ARCHIVE_MAX_CHARS = 16_000       # fallback dump (LLM failed)
+_ARCHIVE_SUMMARY_MAX_CHARS = 8_000    # LLM-produced consolidation summary
+_HISTORY_ENTRY_HARD_CAP = 64_000      # emergency cap in append_history
+
+
 class Consolidator:
     """Lightweight consolidation: summarizes evicted messages into history.jsonl."""
 
     _MAX_CONSOLIDATION_ROUNDS = 5
-    _MAX_CHUNK_MESSAGES = 60  # hard cap per consolidation round
 
     _SAFETY_BUFFER = 1024  # extra headroom for tokenizer estimation drift
 
@@ -408,6 +435,7 @@ class Consolidator:
         build_messages: Callable[..., list[dict[str, Any]]],
         get_tool_definitions: Callable[[], list[dict[str, Any]]],
         max_completion_tokens: int = 4096,
+        consolidation_ratio: float = 0.5,
     ):
         self.store = store
         self.provider = provider
@@ -415,12 +443,24 @@ class Consolidator:
         self.sessions = sessions
         self.context_window_tokens = context_window_tokens
         self.max_completion_tokens = max_completion_tokens
+        self.consolidation_ratio = consolidation_ratio
         self._build_messages = build_messages
         self._get_tool_definitions = get_tool_definitions
         self._locks: weakref.WeakValueDictionary[str, asyncio.Lock] = (
             weakref.WeakValueDictionary()
         )
 
+    def set_provider(
+        self,
+        provider: LLMProvider,
+        model: str,
+        context_window_tokens: int,
+    ) -> None:
+        self.provider = provider
+        self.model = model
+        self.context_window_tokens = context_window_tokens
+        self.max_completion_tokens = provider.generation.max_tokens
+
     def get_lock(self, session_key: str) -> asyncio.Lock:
         """Return the shared consolidation lock for one session."""
         return self._locks.setdefault(session_key, asyncio.Lock())
@@ -447,22 +487,6 @@ class Consolidator:
 
         return last_boundary
 
-    def _cap_consolidation_boundary(
-        self,
-        session: Session,
-        end_idx: int,
-    ) -> int | None:
-        """Clamp the chunk size without breaking the user-turn boundary."""
-        start = session.last_consolidated
-        if end_idx - start <= self._MAX_CHUNK_MESSAGES:
-            return end_idx
-
-        capped_end = start + self._MAX_CHUNK_MESSAGES
-        for idx in range(capped_end, start, -1):
-            if session.messages[idx].get("role") == "user":
-                return idx
-        return None
-
     def estimate_session_prompt_tokens(
         self,
         session: Session,
@@ -470,7 +494,7 @@ class Consolidator:
         session_summary: str | None = None,
     ) -> tuple[int, str]:
         """Estimate current prompt size for the normal session history view."""
-        history = session.get_history(max_messages=0)
+        history = session.get_history(max_messages=0, include_timestamps=True)
         channel, chat_id = (session.key.split(":", 1) if ":" in session.key else (None, None))
         probe_messages = self._build_messages(
             history=history,
@@ -486,6 +510,25 @@ class Consolidator:
             self._get_tool_definitions(),
         )
 
+    @property
+    def _input_token_budget(self) -> int:
+        """Available input token budget for consolidation LLM."""
+        return self.context_window_tokens - self.max_completion_tokens - self._SAFETY_BUFFER
+
+    def _truncate_to_token_budget(self, text: str) -> str:
+        """Truncate text so it fits within the consolidation LLM's token budget."""
+        budget = self._input_token_budget
+        if budget <= 0:
+            return truncate_text(text, _RAW_ARCHIVE_MAX_CHARS)
+        try:
+            enc = tiktoken.get_encoding("cl100k_base")
+            tokens = enc.encode(text)
+            if len(tokens) <= budget:
+                return text
+            return enc.decode(tokens[:budget]) + "\n... (truncated)"
+        except Exception:
+            return truncate_text(text, budget * 4)
+
     async def archive(self, messages: list[dict]) -> str | None:
         """Summarize messages via LLM and append to history.jsonl.
 
@@ -495,6 +538,7 @@ class Consolidator:
             return None
         try:
             formatted = MemoryStore._format_messages(messages)
+            formatted = self._truncate_to_token_budget(formatted)
             response = await self.provider.chat_with_retry(
                 model=self.model,
                 messages=[
@@ -513,7 +557,7 @@ class Consolidator:
             if response.finish_reason == "error":
                 raise RuntimeError(f"LLM returned error: {response.content}")
             summary = response.content or "[no summary]"
-            self.store.append_history(summary)
+            self.store.append_history(summary, max_chars=_ARCHIVE_SUMMARY_MAX_CHARS)
             return summary
         except Exception:
             logger.warning("Consolidation LLM call failed, raw-dumping to history")
@@ -536,8 +580,8 @@ class Consolidator:
 
         lock = self.get_lock(session.key)
         async with lock:
-            budget = self.context_window_tokens - self.max_completion_tokens - self._SAFETY_BUFFER
-            target = budget // 2
+            budget = self._input_token_budget
+            target = int(budget * self.consolidation_ratio)
             try:
                 estimated, source = self.estimate_session_prompt_tokens(
                     session,
@@ -575,14 +619,6 @@ class Consolidator:
                     break
 
                 end_idx = boundary[0]
-                end_idx = self._cap_consolidation_boundary(session, end_idx)
-                if end_idx is None:
-                    logger.debug(
-                        "Token consolidation: no capped boundary for {} (round {})",
-                        session.key,
-                        round_num,
-                    )
-                    break
 
                 chunk = session.messages[session.last_consolidated:end_idx]
                 if not chunk:
@@ -598,12 +634,18 @@ class Consolidator:
                     len(chunk),
                 )
                 summary = await self.archive(chunk)
+                # Advance the cursor either way: on success the chunk was
+                # summarized; on failure archive() already raw-archived it as
+                # a breadcrumb. Re-archiving the same chunk on the next call
+                # would just emit duplicate [RAW] entries.
                 if summary:
                     last_summary = summary
-                else:
-                    break
                 session.last_consolidated = end_idx
                 self.sessions.save(session)
+                if not summary:
+                    # LLM is degraded — stop hammering it this call;
+                    # the next invocation can retry a fresh chunk.
+                    break
 
                 try:
                     estimated, source = self.estimate_session_prompt_tokens(
@@ -647,6 +689,15 @@ class Dream:
     LLM can make targeted, incremental edits instead of replacing entire files.
     """
 
+    # Caps on prompt-bound inputs so Dream's LLM calls never exceed the model's
+    # context window just because a file (or a legacy large history entry) grew
+    # unexpectedly. Each file still appears in full via read_file when the agent
+    # needs it in Phase 2 — these caps only bound the Phase 1/2 prompt preview.
+    _MEMORY_FILE_MAX_CHARS = 32_000
+    _SOUL_FILE_MAX_CHARS = 16_000
+    _USER_FILE_MAX_CHARS = 16_000
+    _HISTORY_ENTRY_PREVIEW_MAX_CHARS = 4_000
+
     def __init__(
         self,
         store: MemoryStore,
@@ -670,6 +721,11 @@ class Dream:
         self._runner = AgentRunner(provider)
         self._tools = self._build_tools()
 
+    def set_provider(self, provider: LLMProvider, model: str) -> None:
+        self.provider = provider
+        self.model = model
+        self._runner.provider = provider
+
     # -- tool registry -------------------------------------------------------
 
     def _build_tools(self) -> ToolRegistry:
@@ -785,21 +841,31 @@ class Dream:
             len(entries), last_cursor, batch[-1]["cursor"], len(batch),
         )
 
-        # Build history text for LLM
+        # Build history text for LLM — cap each entry so a legacy oversized
+        # record (e.g. pre-#3412 raw_archive dump) can't blow up the prompt.
         history_text = "\n".join(
-            f"[{e['timestamp']}] {e['content']}" for e in batch
+            f"[{e['timestamp']}] "
+            f"{truncate_text(e['content'], self._HISTORY_ENTRY_PREVIEW_MAX_CHARS)}"
+            for e in batch
         )
 
-        # Current file contents + per-line age annotations (MEMORY.md only)
+        # Current file contents + per-line age annotations (MEMORY.md only).
+        # Each file is capped in the *prompt preview* only; Phase 2 still sees
+        # the full file via the read_file tool.
         current_date = datetime.now().strftime("%Y-%m-%d")
         raw_memory = self.store.read_memory() or "(empty)"
-        current_memory = (
+        annotated_memory = (
             self._annotate_with_ages(raw_memory)
             if self.annotate_line_ages
             else raw_memory
         )
-        current_soul = self.store.read_soul() or "(empty)"
-        current_user = self.store.read_user() or "(empty)"
+        current_memory = truncate_text(annotated_memory, self._MEMORY_FILE_MAX_CHARS)
+        current_soul = truncate_text(
+            self.store.read_soul() or "(empty)", self._SOUL_FILE_MAX_CHARS,
+        )
+        current_user = truncate_text(
+            self.store.read_user() or "(empty)", self._USER_FILE_MAX_CHARS,
+        )
 
         file_context = (
             f"## Current Date\n{current_date}\n\n"

nanobot/agent/runner.py

@@ -3,17 +3,18 @@
 from __future__ import annotations
 
 import asyncio
-from dataclasses import dataclass, field
 import inspect
+import os
+from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Any
 
 from loguru import logger
 
 from nanobot.agent.hook import AgentHook, AgentHookContext
-from nanobot.utils.prompt_templates import render_template
+from nanobot.agent.tools.ask import AskUserInterrupt
 from nanobot.agent.tools.registry import ToolRegistry
-from nanobot.providers.base import LLMProvider, ToolCallRequest
+from nanobot.providers.base import LLMProvider, LLMResponse, ToolCallRequest
 from nanobot.utils.helpers import (
     build_assistant_message,
     estimate_message_tokens,
@@ -22,6 +23,7 @@ from nanobot.utils.helpers import (
     maybe_persist_tool_result,
     truncate_text,
 )
+from nanobot.utils.prompt_templates import render_template
 from nanobot.utils.runtime import (
     EMPTY_FINAL_RESPONSE_MESSAGE,
     build_finalization_retry_message,
@@ -74,6 +76,7 @@ class AgentRunSpec:
     retry_wait_callback: Any | None = None
     checkpoint_callback: Any | None = None
     injection_callback: Any | None = None
+    llm_timeout_s: float | None = None
 
 
 @dataclass(slots=True)
@@ -275,17 +278,22 @@ class AgentRunner:
             self._accumulate_usage(usage, raw_usage)
 
             if response.should_execute_tools:
+                tool_calls = list(response.tool_calls)
+                ask_index = next((i for i, tc in enumerate(tool_calls) if tc.name == "ask_user"), None)
+                if ask_index is not None:
+                    tool_calls = tool_calls[: ask_index + 1]
+                context.tool_calls = list(tool_calls)
                 if hook.wants_streaming():
                     await hook.on_stream_end(context, resuming=True)
 
                 assistant_message = build_assistant_message(
                     response.content or "",
-                    tool_calls=[tc.to_openai_tool_call() for tc in response.tool_calls],
+                    tool_calls=[tc.to_openai_tool_call() for tc in tool_calls],
                     reasoning_content=response.reasoning_content,
                     thinking_blocks=response.thinking_blocks,
                 )
                 messages.append(assistant_message)
-                tools_used.extend(tc.name for tc in response.tool_calls)
+                tools_used.extend(tc.name for tc in tool_calls)
                 await self._emit_checkpoint(
                     spec,
                     {
@@ -294,7 +302,7 @@ class AgentRunner:
                         "model": spec.model,
                         "assistant_message": assistant_message,
                         "completed_tool_results": [],
-                        "pending_tool_calls": [tc.to_openai_tool_call() for tc in response.tool_calls],
+                        "pending_tool_calls": [tc.to_openai_tool_call() for tc in tool_calls],
                     },
                 )
 
@@ -302,14 +310,16 @@ class AgentRunner:
 
                 results, new_events, fatal_error = await self._execute_tools(
                     spec,
-                    response.tool_calls,
+                    tool_calls,
                     external_lookup_counts,
                 )
                 tool_events.extend(new_events)
                 context.tool_results = list(results)
                 context.tool_events = list(new_events)
                 completed_tool_results: list[dict[str, Any]] = []
-                for tool_call, result in zip(response.tool_calls, results):
+                for tool_call, result in zip(tool_calls, results):
+                    if isinstance(fatal_error, AskUserInterrupt) and tool_call.name == "ask_user":
+                        continue
                     tool_message = {
                         "role": "tool",
                         "tool_call_id": tool_call.id,
@@ -324,6 +334,15 @@ class AgentRunner:
                     messages.append(tool_message)
                     completed_tool_results.append(tool_message)
                 if fatal_error is not None:
+                    if isinstance(fatal_error, AskUserInterrupt):
+                        final_content = fatal_error.question
+                        stop_reason = "ask_user"
+                        context.final_content = final_content
+                        context.stop_reason = stop_reason
+                        if hook.wants_streaming():
+                            await hook.on_stream_end(context, resuming=False)
+                        await hook.after_iteration(context)
+                        break
                     error = f"Error: {type(fatal_error).__name__}: {fatal_error}"
                     final_content = error
                     stop_reason = "tool_error"
@@ -570,6 +589,19 @@ class AgentRunner:
         hook: AgentHook,
         context: AgentHookContext,
     ):
+        timeout_s: float | None = spec.llm_timeout_s
+        if timeout_s is None:
+            # Default to a finite timeout to avoid per-session lock starvation when an LLM
+            # request hangs indefinitely (e.g. gateway/network stall).
+            # Set NANOBOT_LLM_TIMEOUT_S=0 to disable.
+            raw = os.environ.get("NANOBOT_LLM_TIMEOUT_S", "300").strip()
+            try:
+                timeout_s = float(raw)
+            except (TypeError, ValueError):
+                timeout_s = 300.0
+        if timeout_s is not None and timeout_s <= 0:
+            timeout_s = None
+
         kwargs = self._build_request_kwargs(
             spec,
             messages,
@@ -579,11 +611,23 @@ class AgentRunner:
             async def _stream(delta: str) -> None:
                 await hook.on_stream(context, delta)
 
-            return await self.provider.chat_stream_with_retry(
+            coro = self.provider.chat_stream_with_retry(
                 **kwargs,
                 on_content_delta=_stream,
             )
-        return await self.provider.chat_with_retry(**kwargs)
+        else:
+            coro = self.provider.chat_with_retry(**kwargs)
+
+        if timeout_s is None:
+            return await coro
+        try:
+            return await asyncio.wait_for(coro, timeout=timeout_s)
+        except asyncio.TimeoutError:
+            return LLMResponse(
+                content=f"Error calling LLM: timed out after {timeout_s:g}s",
+                finish_reason="error",
+                error_kind="timeout",
+            )
 
     async def _request_finalization_retry(
         self,
@@ -629,13 +673,21 @@ class AgentRunner:
         tool_results: list[tuple[Any, dict[str, str], BaseException | None]] = []
         for batch in batches:
             if spec.concurrent_tools and len(batch) > 1:
-                tool_results.extend(await asyncio.gather(*(
+                batch_results = await asyncio.gather(*(
                     self._run_tool(spec, tool_call, external_lookup_counts)
                     for tool_call in batch
-                )))
+                ))
+                tool_results.extend(batch_results)
             else:
+                batch_results = []
                 for tool_call in batch:
-                    tool_results.append(await self._run_tool(spec, tool_call, external_lookup_counts))
+                    result = await self._run_tool(spec, tool_call, external_lookup_counts)
+                    tool_results.append(result)
+                    batch_results.append(result)
+                    if isinstance(result[2], AskUserInterrupt):
+                        break
+            if any(isinstance(error, AskUserInterrupt) for _, _, error in batch_results):
+                break
 
         results: list[Any] = []
         events: list[dict[str, str]] = []
@@ -653,7 +705,7 @@ class AgentRunner:
         tool_call: ToolCallRequest,
         external_lookup_counts: dict[str, int],
     ) -> tuple[Any, dict[str, str], BaseException | None]:
-        _HINT = "\n\n[Analyze the error above and try a different approach.]"
+        hint = "\n\n[Analyze the error above and try a different approach.]"
         lookup_error = repeated_external_lookup_error(
             tool_call.name,
             tool_call.arguments,
@@ -666,8 +718,8 @@ class AgentRunner:
                 "detail": "repeated external lookup blocked",
             }
             if spec.fail_on_tool_error:
-                return lookup_error + _HINT, event, RuntimeError(lookup_error)
-            return lookup_error + _HINT, event, None
+                return lookup_error + hint, event, RuntimeError(lookup_error)
+            return lookup_error + hint, event, None
         prepare_call = getattr(spec.tools, "prepare_call", None)
         tool, params, prep_error = None, tool_call.arguments, None
         if callable(prepare_call):
@@ -683,7 +735,7 @@ class AgentRunner:
                 "status": "error",
                 "detail": prep_error.split(": ", 1)[-1][:120],
             }
-            return prep_error + _HINT, event, RuntimeError(prep_error) if spec.fail_on_tool_error else None
+            return prep_error + hint, event, RuntimeError(prep_error) if spec.fail_on_tool_error else None
         try:
             if tool is not None:
                 result = await tool.execute(**params)
@@ -697,6 +749,9 @@ class AgentRunner:
                 "status": "error",
                 "detail": str(exc),
             }
+            if isinstance(exc, AskUserInterrupt):
+                event["status"] = "waiting"
+                return "", event, exc
             if spec.fail_on_tool_error:
                 return f"Error: {type(exc).__name__}: {exc}", event, exc
             return f"Error: {type(exc).__name__}: {exc}", event, None
@@ -708,8 +763,8 @@ class AgentRunner:
                 "detail": result.replace("\n", " ").strip()[:120],
             }
             if spec.fail_on_tool_error:
-                return result + _HINT, event, RuntimeError(result)
-            return result + _HINT, event, None
+                return result + hint, event, RuntimeError(result)
+            return result + hint, event, None
 
         detail = "" if result is None else str(result)
         detail = detail.replace("\n", " ").strip()
@@ -984,4 +1039,3 @@ class AgentRunner:
         if current:
             batches.append(current)
         return batches
-

nanobot/agent/subagent.py

@@ -96,6 +96,11 @@ class SubagentManager:
         self._task_statuses: dict[str, SubagentStatus] = {}
         self._session_tasks: dict[str, set[str]] = {}  # session_key -> {task_id, ...}
 
+    def set_provider(self, provider: LLMProvider, model: str) -> None:
+        self.provider = provider
+        self.model = model
+        self.runner.provider = provider
+
     async def spawn(
         self,
         task: str,

nanobot/agent/tools/ask.py

@@ -0,0 +1,136 @@
+"""Tool for pausing a turn until the user answers."""
+
+import json
+from typing import Any
+
+from nanobot.agent.tools.base import Tool, tool_parameters
+from nanobot.agent.tools.schema import ArraySchema, StringSchema, tool_parameters_schema
+
+STRUCTURED_BUTTON_CHANNELS = frozenset({"telegram", "websocket"})
+
+
+class AskUserInterrupt(BaseException):
+    """Internal signal: the runner should stop and wait for user input."""
+
+    def __init__(self, question: str, options: list[str] | None = None) -> None:
+        self.question = question
+        self.options = [str(option) for option in (options or []) if str(option)]
+        super().__init__(question)
+
+
+@tool_parameters(
+    tool_parameters_schema(
+        question=StringSchema(
+            "The question to ask before continuing. Use this only when the task needs the user's answer."
+        ),
+        options=ArraySchema(
+            StringSchema("A possible answer label"),
+            description="Optional choices. The user may still reply with free text.",
+        ),
+        required=["question"],
+    )
+)
+class AskUserTool(Tool):
+    """Ask the user a blocking question."""
+
+    @property
+    def name(self) -> str:
+        return "ask_user"
+
+    @property
+    def description(self) -> str:
+        return (
+            "Pause and ask the user a question when their answer is required to continue. "
+            "Use options for likely answers; the user's reply, typed or selected, is returned as the tool result. "
+            "For non-blocking notifications or buttons, use the message tool instead."
+        )
+
+    @property
+    def exclusive(self) -> bool:
+        return True
+
+    async def execute(self, question: str, options: list[str] | None = None, **_: Any) -> Any:
+        raise AskUserInterrupt(question=question, options=options)
+
+
+def _tool_call_name(tool_call: dict[str, Any]) -> str:
+    function = tool_call.get("function")
+    if isinstance(function, dict) and isinstance(function.get("name"), str):
+        return function["name"]
+    name = tool_call.get("name")
+    return name if isinstance(name, str) else ""
+
+
+def _tool_call_arguments(tool_call: dict[str, Any]) -> dict[str, Any]:
+    function = tool_call.get("function")
+    raw = function.get("arguments") if isinstance(function, dict) else tool_call.get("arguments")
+    if isinstance(raw, dict):
+        return raw
+    if isinstance(raw, str):
+        try:
+            parsed = json.loads(raw)
+        except json.JSONDecodeError:
+            return {}
+        return parsed if isinstance(parsed, dict) else {}
+    return {}
+
+
+def pending_ask_user_id(history: list[dict[str, Any]]) -> str | None:
+    pending: dict[str, str] = {}
+    for message in history:
+        if message.get("role") == "assistant":
+            for tool_call in message.get("tool_calls") or []:
+                if isinstance(tool_call, dict) and isinstance(tool_call.get("id"), str):
+                    pending[tool_call["id"]] = _tool_call_name(tool_call)
+        elif message.get("role") == "tool":
+            tool_call_id = message.get("tool_call_id")
+            if isinstance(tool_call_id, str):
+                pending.pop(tool_call_id, None)
+    for tool_call_id, name in reversed(pending.items()):
+        if name == "ask_user":
+            return tool_call_id
+    return None
+
+
+def ask_user_tool_result_messages(
+    system_prompt: str,
+    history: list[dict[str, Any]],
+    tool_call_id: str,
+    content: str,
+) -> list[dict[str, Any]]:
+    return [
+        {"role": "system", "content": system_prompt},
+        *history,
+        {
+            "role": "tool",
+            "tool_call_id": tool_call_id,
+            "name": "ask_user",
+            "content": content,
+        },
+    ]
+
+
+def ask_user_options_from_messages(messages: list[dict[str, Any]]) -> list[str]:
+    for message in reversed(messages):
+        if message.get("role") != "assistant":
+            continue
+        for tool_call in reversed(message.get("tool_calls") or []):
+            if not isinstance(tool_call, dict) or _tool_call_name(tool_call) != "ask_user":
+                continue
+            options = _tool_call_arguments(tool_call).get("options")
+            if isinstance(options, list):
+                return [str(option) for option in options if isinstance(option, str)]
+    return []
+
+
+def ask_user_outbound(
+    content: str | None,
+    options: list[str],
+    channel: str,
+) -> tuple[str | None, list[list[str]]]:
+    if not options:
+        return content, []
+    if channel in STRUCTURED_BUTTON_CHANNELS:
+        return content, [options]
+    option_text = "\n".join(f"{index}. {option}" for index, option in enumerate(options, 1))
+    return f"{content}\n\n{option_text}" if content else option_text, []

nanobot/agent/tools/cron.py

@@ -60,12 +60,19 @@ class CronTool(Tool):
         self._default_timezone = default_timezone
         self._channel: ContextVar[str] = ContextVar("cron_channel", default="")
         self._chat_id: ContextVar[str] = ContextVar("cron_chat_id", default="")
+        self._metadata: ContextVar[dict] = ContextVar("cron_metadata", default={})
+        self._session_key: ContextVar[str] = ContextVar("cron_session_key", default="")
         self._in_cron_context: ContextVar[bool] = ContextVar("cron_in_context", default=False)
 
-    def set_context(self, channel: str, chat_id: str) -> None:
+    def set_context(
+        self, channel: str, chat_id: str,
+        metadata: dict | None = None, session_key: str | None = None,
+    ) -> None:
         """Set the current session context for delivery."""
         self._channel.set(channel)
         self._chat_id.set(chat_id)
+        self._metadata.set(metadata or {})
+        self._session_key.set(session_key or f"{channel}:{chat_id}")
 
     def set_cron_context(self, active: bool):
         """Mark whether the tool is executing inside a cron job callback."""
@@ -199,6 +206,8 @@ class CronTool(Tool):
             channel=channel,
             to=chat_id,
             delete_after_run=delete_after,
+            channel_meta=self._metadata.get(),
+            session_key=self._session_key.get() or None,
         )
         return f"Created job '{job.name}' (id: {job.id})"
 

nanobot/agent/tools/mcp.py

@@ -2,6 +2,7 @@
 
 import asyncio
 import os
+import re
 import shutil
 from contextlib import AsyncExitStack
 from typing import Any
@@ -28,6 +29,15 @@ _TRANSIENT_EXC_NAMES: frozenset[str] = frozenset((
 
 _WINDOWS_SHELL_LAUNCHERS: frozenset[str] = frozenset(("npx", "npm", "pnpm", "yarn", "bunx"))
 
+# Characters allowed in tool names by model providers (Anthropic, OpenAI, etc.).
+# Replace anything outside [a-zA-Z0-9_-] with underscore and collapse runs.
+_SANITIZE_RE = re.compile(r"_+")
+
+
+def _sanitize_name(name: str) -> str:
+    """Sanitize an MCP-derived name for model API compatibility."""
+    return _SANITIZE_RE.sub("_", re.sub(r"[^a-zA-Z0-9_-]", "_", name))
+
 
 def _is_transient(exc: BaseException) -> bool:
     """Check if an exception looks like a transient connection error."""
@@ -137,7 +147,7 @@ class MCPToolWrapper(Tool):
     def __init__(self, session, server_name: str, tool_def, tool_timeout: int = 30):
         self._session = session
         self._original_name = tool_def.name
-        self._name = f"mcp_{server_name}_{tool_def.name}"
+        self._name = _sanitize_name(f"mcp_{server_name}_{tool_def.name}")
         self._description = tool_def.description or tool_def.name
         raw_schema = tool_def.inputSchema or {"type": "object", "properties": {}}
         self._parameters = _normalize_schema_for_openai(raw_schema)
@@ -221,7 +231,7 @@ class MCPResourceWrapper(Tool):
     def __init__(self, session, server_name: str, resource_def, resource_timeout: int = 30):
         self._session = session
         self._uri = resource_def.uri
-        self._name = f"mcp_{server_name}_resource_{resource_def.name}"
+        self._name = _sanitize_name(f"mcp_{server_name}_resource_{resource_def.name}")
         desc = resource_def.description or resource_def.name
         self._description = f"[MCP Resource] {desc}\nURI: {self._uri}"
         self._parameters: dict[str, Any] = {
@@ -311,7 +321,7 @@ class MCPPromptWrapper(Tool):
     def __init__(self, session, server_name: str, prompt_def, prompt_timeout: int = 30):
         self._session = session
         self._prompt_name = prompt_def.name
-        self._name = f"mcp_{server_name}_prompt_{prompt_def.name}"
+        self._name = _sanitize_name(f"mcp_{server_name}_prompt_{prompt_def.name}")
         desc = prompt_def.description or prompt_def.name
         self._description = (
             f"[MCP Prompt] {desc}\n"
@@ -514,9 +524,9 @@ async def connect_mcp_servers(
             registered_count = 0
             matched_enabled_tools: set[str] = set()
             available_raw_names = [tool_def.name for tool_def in tools.tools]
-            available_wrapped_names = [f"mcp_{name}_{tool_def.name}" for tool_def in tools.tools]
+            available_wrapped_names = [_sanitize_name(f"mcp_{name}_{tool_def.name}") for tool_def in tools.tools]
             for tool_def in tools.tools:
-                wrapped_name = f"mcp_{name}_{tool_def.name}"
+                wrapped_name = _sanitize_name(f"mcp_{name}_{tool_def.name}")
                 if (
                     not allow_all_tools
                     and tool_def.name not in enabled_tools

nanobot/agent/tools/message.py

@@ -1,11 +1,14 @@
 """Message tool for sending messages to users."""
 
+import os
 from contextvars import ContextVar
+from pathlib import Path
 from typing import Any, Awaitable, Callable
 
 from nanobot.agent.tools.base import Tool, tool_parameters
 from nanobot.agent.tools.schema import ArraySchema, StringSchema, tool_parameters_schema
 from nanobot.bus.events import OutboundMessage
+from nanobot.config.paths import get_workspace_path
 
 
 @tool_parameters(
@@ -15,7 +18,11 @@ from nanobot.bus.events import OutboundMessage
         chat_id=StringSchema("Optional: target chat/user ID"),
         media=ArraySchema(
             StringSchema(""),
-            description="Optional: list of file paths to attach (images, audio, documents)",
+            description="Optional: list of file paths to attach (images, video, audio, documents)",
+        ),
+        buttons=ArraySchema(
+            ArraySchema(StringSchema("Button label")),
+            description="Optional: inline keyboard buttons as list of rows, each row is list of button labels.",
         ),
         required=["content"],
     )
@@ -29,21 +36,38 @@ class MessageTool(Tool):
         default_channel: str = "",
         default_chat_id: str = "",
         default_message_id: str | None = None,
+        workspace: str | Path | None = None,
     ):
         self._send_callback = send_callback
+        self._workspace = Path(workspace).expanduser() if workspace is not None else get_workspace_path()
         self._default_channel: ContextVar[str] = ContextVar("message_default_channel", default=default_channel)
         self._default_chat_id: ContextVar[str] = ContextVar("message_default_chat_id", default=default_chat_id)
         self._default_message_id: ContextVar[str | None] = ContextVar(
             "message_default_message_id",
             default=default_message_id,
         )
+        self._default_metadata: ContextVar[dict[str, Any]] = ContextVar(
+            "message_default_metadata",
+            default={},
+        )
         self._sent_in_turn_var: ContextVar[bool] = ContextVar("message_sent_in_turn", default=False)
+        self._record_channel_delivery_var: ContextVar[bool] = ContextVar(
+            "message_record_channel_delivery",
+            default=False,
+        )
 
-    def set_context(self, channel: str, chat_id: str, message_id: str | None = None) -> None:
+    def set_context(
+        self,
+        channel: str,
+        chat_id: str,
+        message_id: str | None = None,
+        metadata: dict[str, Any] | None = None,
+    ) -> None:
         """Set the current message context."""
         self._default_channel.set(channel)
         self._default_chat_id.set(chat_id)
         self._default_message_id.set(message_id)
+        self._default_metadata.set(metadata or {})
 
     def set_send_callback(self, callback: Callable[[OutboundMessage], Awaitable[None]]) -> None:
         """Set the callback for sending messages."""
@@ -53,6 +77,14 @@ class MessageTool(Tool):
         """Reset per-turn send tracking."""
         self._sent_in_turn = False
 
+    def set_record_channel_delivery(self, active: bool):
+        """Mark tool-sent messages as proactive channel deliveries."""
+        return self._record_channel_delivery_var.set(active)
+
+    def reset_record_channel_delivery(self, token) -> None:
+        """Restore previous proactive delivery recording state."""
+        self._record_channel_delivery_var.reset(token)
+
     @property
     def _sent_in_turn(self) -> bool:
         return self._sent_in_turn_var.get()
@@ -81,14 +113,20 @@ class MessageTool(Tool):
         chat_id: str | None = None,
         message_id: str | None = None,
         media: list[str] | None = None,
+        buttons: list[list[str]] | None = None,
         **kwargs: Any
     ) -> str:
         from nanobot.utils.helpers import strip_think
         content = strip_think(content)
 
+        if buttons is not None:
+            if not isinstance(buttons, list) or any(
+                not isinstance(row, list) or any(not isinstance(label, str) for label in row)
+                for row in buttons
+            ):
+                return "Error: buttons must be a list of list of strings"
         default_channel = self._default_channel.get()
         default_chat_id = self._default_chat_id.get()
-
         channel = channel or default_channel
         chat_id = chat_id or default_chat_id
         # Only inherit default message_id when targeting the same channel+chat.
@@ -96,7 +134,8 @@ class MessageTool(Tool):
         # some channels (e.g. Feishu) use it to determine the target
         # conversation via their Reply API, which would route the message
         # to the wrong chat entirely.
-        if channel == default_channel and chat_id == default_chat_id:
+        same_target = channel == default_channel and chat_id == default_chat_id
+        if same_target:
             message_id = message_id or self._default_message_id.get()
         else:
             message_id = None
@@ -107,14 +146,28 @@ class MessageTool(Tool):
         if not self._send_callback:
             return "Error: Message sending not configured"
 
+        if media:
+            resolved = []
+            for p in media:
+                if p.startswith(("http://", "https://")) or os.path.isabs(p):
+                    resolved.append(p)
+                else:
+                    resolved.append(str(self._workspace / p))
+            media = resolved
+
+        metadata = dict(self._default_metadata.get()) if same_target else {}
+        if message_id:
+            metadata["message_id"] = message_id
+        if self._record_channel_delivery_var.get():
+            metadata["_record_channel_delivery"] = True
+
         msg = OutboundMessage(
             channel=channel,
             chat_id=chat_id,
             content=content,
             media=media or [],
-            metadata={
-                "message_id": message_id,
-            } if message_id else {},
+            buttons=buttons or [],
+            metadata=metadata,
         )
 
         try:
@@ -122,6 +175,7 @@ class MessageTool(Tool):
             if channel == default_channel and chat_id == default_chat_id:
                 self._sent_in_turn = True
             media_info = f" with {len(media)} attachments" if media else ""
-            return f"Message sent to {channel}:{chat_id}{media_info}"
+            button_info = f" with {sum(len(row) for row in buttons)} button(s)" if buttons else ""
+            return f"Message sent to {channel}:{chat_id}{media_info}{button_info}"
         except Exception as e:
             return f"Error sending message: {str(e)}"

nanobot/agent/tools/shell.py

@@ -136,9 +136,10 @@ class ExecTool(Tool):
 
         if self.path_append:
             if _IS_WINDOWS:
-                env["PATH"] = env.get("PATH", "") + ";" + self.path_append
+                env["PATH"] = env.get("PATH", "") + os.pathsep + self.path_append
             else:
-                command = f'export PATH="$PATH:{self.path_append}"; {command}'
+                env["NANOBOT_PATH_APPEND"] = self.path_append
+                command = f'export PATH="$PATH{os.pathsep}$NANOBOT_PATH_APPEND"; {command}'
 
         try:
             process = await self._spawn(command, cwd, env)
@@ -298,8 +299,8 @@ class ExecTool(Tool):
                     continue
 
                 media_path = get_media_dir().resolve()
-                if (p.is_absolute() 
-                    and cwd_path not in p.parents 
+                if (p.is_absolute()
+                    and cwd_path not in p.parents
                     and p != cwd_path
                     and media_path not in p.parents
                     and p != media_path

nanobot/bus/events.py

@@ -34,5 +34,5 @@ class OutboundMessage:
     reply_to: str | None = None
     media: list[str] = field(default_factory=list)
     metadata: dict[str, Any] = field(default_factory=dict)
-
+    buttons: list[list[str]] = field(default_factory=list)
 

nanobot/channels/feishu.py

@@ -13,6 +13,7 @@ from dataclasses import dataclass
 from typing import Any, Literal
 
 from lark_oapi.api.im.v1.model import MentionEvent, P2ImMessageReceiveV1
+from lark_oapi.core.const import FEISHU_DOMAIN, LARK_DOMAIN
 from loguru import logger
 from pydantic import Field
 
@@ -22,8 +23,6 @@ from nanobot.channels.base import BaseChannel
 from nanobot.config.paths import get_media_dir
 from nanobot.config.schema import Base
 
-from lark_oapi.core.const import FEISHU_DOMAIN, LARK_DOMAIN
-
 FEISHU_AVAILABLE = importlib.util.find_spec("lark_oapi") is not None
 
 # Message type display mapping
@@ -308,6 +307,8 @@ class FeishuChannel(BaseChannel):
         self._loop: asyncio.AbstractEventLoop | None = None
         self._stream_bufs: dict[str, _FeishuStreamBuf] = {}
         self._bot_open_id: str | None = None
+        self._background_tasks: set[asyncio.Task] = set()
+        self._reaction_ids: dict[str, str] = {}  # message_id → reaction_id
 
     @staticmethod
     def _register_optional_event(builder: Any, method_name: str, handler: Any) -> Any:
@@ -549,8 +550,11 @@ class FeishuChannel(BaseChannel):
             return None
 
     async def _add_reaction(self, message_id: str, emoji_type: str = "THUMBSUP") -> str | None:
-        """
-        Add a reaction emoji to a message (non-blocking).
+        """Add a reaction emoji to a message.
+
+        Returns the reaction_id on success, None on failure.
+        When called via a tracked background task, the returned reaction_id
+        is stored in ``_reaction_ids`` for later cleanup by ``send_delta``.
 
         Common emoji types: THUMBSUP, OK, EYES, DONE, OnIt, HEART
         """
@@ -594,6 +598,36 @@ class FeishuChannel(BaseChannel):
         loop = asyncio.get_running_loop()
         await loop.run_in_executor(None, self._remove_reaction_sync, message_id, reaction_id)
 
+    def _on_background_task_done(self, task: asyncio.Task) -> None:
+        """Callback: remove from tracking set and log unhandled exceptions."""
+        self._background_tasks.discard(task)
+        if task.cancelled():
+            return
+        try:
+            task.result()
+        except Exception as exc:
+            logger.warning("Background task failed: {}", exc)
+
+    def _on_reaction_added(self, message_id: str, task: asyncio.Task) -> None:
+        """Callback: store reaction_id after background add-reaction completes."""
+        if task.cancelled():
+            return
+        try:
+            reaction_id = task.result()
+            if reaction_id:
+                self._reaction_ids[message_id] = reaction_id
+        except Exception:
+            pass  # already logged by _on_background_task_done
+        # Trim cache to prevent unbounded growth
+        if len(self._reaction_ids) > 500:
+            self._reaction_ids.pop(next(iter(self._reaction_ids)))
+
+    @staticmethod
+    def _stream_key(chat_id: str, metadata: dict[str, Any] | None = None) -> str:
+        """Scope streaming buffers to the inbound message when available."""
+        meta = metadata or {}
+        return meta.get("message_id") or chat_id
+
     # Regex to match markdown tables (header + separator + data rows)
     _TABLE_RE = re.compile(
         r"((?:^[ \t]*\|.+\|[ \t]*\n)(?:^[ \t]*\|[-:\s|]+\|[ \t]*\n)(?:^[ \t]*\|.+\|[ \t]*\n?)+)",
@@ -1101,17 +1135,23 @@ class FeishuChannel(BaseChannel):
             logger.debug("Feishu: error fetching parent message {}: {}", message_id, e)
             return None
 
-    def _reply_message_sync(self, parent_message_id: str, msg_type: str, content: str) -> bool:
-        """Reply to an existing Feishu message using the Reply API (synchronous)."""
+    def _reply_message_sync(self, parent_message_id: str, msg_type: str, content: str, *, reply_in_thread: bool = False) -> bool:
+        """Reply to an existing Feishu message using the Reply API (synchronous).
+
+        Args:
+            reply_in_thread: If True, reply as a thread/topic message
+                in the Feishu client.
+        """
         from lark_oapi.api.im.v1 import ReplyMessageRequest, ReplyMessageRequestBody
 
         try:
+            body_builder = ReplyMessageRequestBody.builder().msg_type(msg_type).content(content)
+            if reply_in_thread:
+                body_builder = body_builder.reply_in_thread(True)
             request = (
                 ReplyMessageRequest.builder()
                 .message_id(parent_message_id)
-                .request_body(
-                    ReplyMessageRequestBody.builder().msg_type(msg_type).content(content).build()
-                )
+                .request_body(body_builder.build())
                 .build()
             )
             response = self._client.im.v1.message.reply(request)
@@ -1166,8 +1206,19 @@ class FeishuChannel(BaseChannel):
             logger.error("Error sending Feishu {} message: {}", msg_type, e)
             return None
 
-    def _create_streaming_card_sync(self, receive_id_type: str, chat_id: str) -> str | None:
-        """Create a CardKit streaming card, send it to chat, return card_id."""
+    def _create_streaming_card_sync(
+        self,
+        receive_id_type: str,
+        chat_id: str,
+        reply_message_id: str | None = None,
+    ) -> str | None:
+        """Create a CardKit streaming card, send it to chat, return card_id.
+
+        When *reply_message_id* is provided the card is delivered via the
+        reply API (with reply_in_thread=True) so it lands inside the
+        originating thread / topic.  Otherwise the plain create-message
+        API is used.
+        """
         from lark_oapi.api.cardkit.v1 import CreateCardRequest, CreateCardRequestBody
 
         card_json = {
@@ -1196,13 +1247,19 @@ class FeishuChannel(BaseChannel):
                 return None
             card_id = getattr(response.data, "card_id", None)
             if card_id:
-                message_id = self._send_message_sync(
-                    receive_id_type,
-                    chat_id,
-                    "interactive",
-                    json.dumps({"type": "card", "data": {"card_id": card_id}}),
+                card_content = json.dumps(
+                    {"type": "card", "data": {"card_id": card_id}}, ensure_ascii=False
                 )
-                if message_id:
+                if reply_message_id:
+                    sent = self._reply_message_sync(
+                        reply_message_id, "interactive", card_content,
+                        reply_in_thread=True,
+                    )
+                else:
+                    sent = self._send_message_sync(
+                        receive_id_type, chat_id, "interactive", card_content,
+                    ) is not None
+                if sent:
                     return card_id
                 logger.warning(
                     "Created streaming card {} but failed to send it to {}", card_id, chat_id
@@ -1292,23 +1349,27 @@ class FeishuChannel(BaseChannel):
             _stream_end: Finalize the streaming card.
             _tool_hint:  Delta is a formatted tool hint (for display only).
             message_id:  Original message id (used with _stream_end for reaction cleanup).
-            reaction_id: Reaction id to remove on stream end.
+            chat_type:   "group" or "p2p" — controls reply-in-thread for streaming cards.
         """
         if not self._client:
             return
         meta = metadata or {}
+        stream_key = self._stream_key(chat_id, meta)
         loop = asyncio.get_running_loop()
         rid_type = "chat_id" if chat_id.startswith("oc_") else "open_id"
 
         # --- stream end: final update or fallback ---
         if meta.get("_stream_end"):
-            if (message_id := meta.get("message_id")) and (reaction_id := meta.get("reaction_id")):
-                await self._remove_reaction(message_id, reaction_id)
+            message_id = meta.get("message_id")
+            if message_id:
+                reaction_id = self._reaction_ids.pop(message_id, None)
+                if reaction_id:
+                    await self._remove_reaction(message_id, reaction_id)
                 # Add completion emoji if configured
-                if self.config.done_emoji and message_id:
+                if self.config.done_emoji:
                     await self._add_reaction(message_id, self.config.done_emoji)
 
-            buf = self._stream_bufs.pop(chat_id, None)
+            buf = self._stream_bufs.pop(stream_key, None)
             if not buf or not buf.text:
                 return
             # Try to finalize via streaming card; if that fails (e.g.
@@ -1343,24 +1404,45 @@ class FeishuChannel(BaseChannel):
                     {"config": {"wide_screen_mode": True}, "elements": chunk},
                     ensure_ascii=False,
                 )
-                await loop.run_in_executor(
-                    None, self._send_message_sync, rid_type, chat_id, "interactive", card
-                )
+                # Fallback: reply via the Reply API for group chats.
+                # Target message_id — the Feishu API keeps the reply in
+                # the same topic automatically.
+                _f_msg = meta.get("message_id")
+                fallback_msg_id = _f_msg if meta.get("chat_type", "group") == "group" else None
+                if fallback_msg_id:
+                    await loop.run_in_executor(
+                        None, lambda: self._reply_message_sync(
+                            fallback_msg_id, "interactive", card,
+                            reply_in_thread=True,
+                        ),
+                    )
+                else:
+                    await loop.run_in_executor(
+                        None, self._send_message_sync, rid_type, chat_id, "interactive", card
+                    )
             return
 
         # --- accumulate delta ---
-        buf = self._stream_bufs.get(chat_id)
+        buf = self._stream_bufs.get(stream_key)
         if buf is None:
             buf = _FeishuStreamBuf()
-            self._stream_bufs[chat_id] = buf
+            self._stream_bufs[stream_key] = buf
         buf.text += delta
         if not buf.text.strip():
             return
 
         now = time.monotonic()
         if buf.card_id is None:
+            # Send the streaming card as a reply for group chats so it
+            # lands inside the originating topic/thread.  Always target
+            # message_id (the actual inbound message) — the Feishu Reply
+            # API keeps the response in the same topic automatically.
+            is_group = meta.get("chat_type", "group") == "group"
+            reply_msg_id = meta.get("message_id") if is_group else None
             card_id = await loop.run_in_executor(
-                None, self._create_streaming_card_sync, rid_type, chat_id
+                None,
+                self._create_streaming_card_sync,
+                rid_type, chat_id, reply_msg_id,
             )
             if card_id:
                 buf.card_id = card_id
@@ -1393,7 +1475,7 @@ class FeishuChannel(BaseChannel):
                 hint = (msg.content or "").strip()
                 if not hint:
                     return
-                buf = self._stream_bufs.get(msg.chat_id)
+                buf = self._stream_bufs.get(self._stream_key(msg.chat_id, msg.metadata))
                 if buf and buf.card_id:
                     # Delegate to send_delta so tool hints get the same
                     # throttling (and card creation) as regular text deltas.
@@ -1404,37 +1486,59 @@ class FeishuChannel(BaseChannel):
                     return
                 # No active streaming card — send as a regular
                 # interactive card with the same 🔧 prefix style.
+                # Use reply API for group chats so the hint stays in topic.
                 card = json.dumps(
                     {"config": {"wide_screen_mode": True}, "elements": [
                         {"tag": "markdown", "content": self._format_tool_hint_delta(hint)},
                     ]},
                     ensure_ascii=False,
                 )
-                await loop.run_in_executor(
-                    None, self._send_message_sync, receive_id_type, msg.chat_id, "interactive", card
-                )
+                _th_msg_id = msg.metadata.get("message_id")
+                _th_chat_type = msg.metadata.get("chat_type", "group")
+                if _th_msg_id and _th_chat_type == "group":
+                    await loop.run_in_executor(
+                        None, lambda: self._reply_message_sync(
+                            _th_msg_id, "interactive", card,
+                            reply_in_thread=True,
+                        ),
+                    )
+                else:
+                    await loop.run_in_executor(
+                        None, self._send_message_sync, receive_id_type, msg.chat_id, "interactive", card
+                    )
                 return
 
             # Determine whether the first message should quote the user's message.
             # Only the very first send (media or text) in this call uses reply; subsequent
             # chunks/media fall back to plain create to avoid redundant quote bubbles.
+            # Always target message_id — the Feishu Reply API keeps replies in the
+            # same topic automatically when the target message is inside a topic.
             reply_message_id: str | None = None
+            _msg_id = msg.metadata.get("message_id")
             if self.config.reply_to_message and not msg.metadata.get("_progress", False):
-                reply_message_id = msg.metadata.get("message_id") or None
+                reply_message_id = _msg_id
             # For topic group messages, always reply to keep context in thread
             elif msg.metadata.get("thread_id"):
-                reply_message_id = (
-                    msg.metadata.get("root_id") or msg.metadata.get("message_id") or None
-                )
+                reply_message_id = _msg_id
 
             first_send = True  # tracks whether the reply has already been used
 
             def _do_send(m_type: str, content: str) -> None:
-                """Send via reply (first message) or create (subsequent)."""
+                """Send via reply (first message) or create (subsequent).
+
+                For group chats the reply API always uses reply_in_thread=True.
+                The Feishu API automatically keeps replies inside existing
+                topics — reply_in_thread only creates a *new* topic when the
+                target message is a plain (non-topic) message.
+                """
                 nonlocal first_send
                 if reply_message_id and first_send:
                     first_send = False
-                    ok = self._reply_message_sync(reply_message_id, m_type, content)
+                    chat_type = msg.metadata.get("chat_type", "group")
+                    ok = self._reply_message_sync(
+                        reply_message_id, m_type, content,
+                        reply_in_thread=chat_type == "group",
+                    )
                     if ok:
                         return
                     # Fall back to regular send if reply fails
@@ -1457,13 +1561,13 @@ class FeishuChannel(BaseChannel):
                 else:
                     key = await loop.run_in_executor(None, self._upload_file_sync, file_path)
                     if key:
-                        # Use msg_type "audio" for audio, "video" for video, "file" for documents.
+                        # Feishu's OpenAPI names video messages "media".
+                        # Use "audio" for audio, "media" for video, "file" for documents.
                         # Feishu requires these specific msg_types for inline playback.
-                        # Note: "media" is only valid as a tag inside "post" messages, not as a standalone msg_type.
                         if ext in self._AUDIO_EXTS:
                             media_type = "audio"
                         elif ext in self._VIDEO_EXTS:
-                            media_type = "video"
+                            media_type = "media"
                         else:
                             media_type = "file"
                         await loop.run_in_executor(
@@ -1543,8 +1647,13 @@ class FeishuChannel(BaseChannel):
                 logger.debug("Feishu: skipping group message (not mentioned)")
                 return
 
-            # Add reaction
-            reaction_id = await self._add_reaction(message_id, self.config.react_emoji)
+            # Add reaction (non-blocking — tracked background task)
+            task = asyncio.create_task(
+                self._add_reaction(message_id, self.config.react_emoji)
+            )
+            self._background_tasks.add(task)
+            task.add_done_callback(self._on_background_task_done)
+            task.add_done_callback(lambda t: self._on_reaction_added(message_id, t))
 
             # Parse content
             content_parts = []
@@ -1624,6 +1733,15 @@ class FeishuChannel(BaseChannel):
             if not content and not media_paths:
                 return
 
+            # Build topic-scoped session key for conversation isolation.
+            # Group chat: each topic gets its own session via root_id (replies
+            # inside a topic) or message_id (top-level messages start a new topic).
+            # Private chat: no override — same behavior as Telegram/Slack.
+            if chat_type == "group":
+                session_key = f"feishu:{chat_id}:{root_id or message_id}"
+            else:
+                session_key = None
+
             # Forward to message bus
             reply_to = chat_id if chat_type == "group" else sender_id
             await self._handle_message(
@@ -1633,13 +1751,13 @@ class FeishuChannel(BaseChannel):
                 media=media_paths,
                 metadata={
                     "message_id": message_id,
-                    "reaction_id": reaction_id,
                     "chat_type": chat_type,
                     "msg_type": msg_type,
                     "parent_id": parent_id,
                     "root_id": root_id,
                     "thread_id": thread_id,
                 },
+                session_key=session_key,
             )
 
         except Exception as e:

nanobot/channels/manager.py

@@ -172,6 +172,7 @@ class ChannelManager:
                 channel=notice.channel,
                 chat_id=notice.chat_id,
                 content=format_restart_completed_message(notice.started_at_raw),
+                metadata=dict(notice.metadata or {}),
             ),
         ))
 

nanobot/channels/msteams.py

@@ -15,12 +15,21 @@ import asyncio
 import html
 import importlib.util
 import json
+import os
 import re
+import tempfile
 import threading
 import time
+from contextlib import contextmanager
 from dataclasses import dataclass
 from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
 from typing import TYPE_CHECKING, Any
+from urllib.parse import urlparse
+
+try:  # pragma: no cover - Windows fallback path
+    import fcntl
+except ImportError:  # pragma: no cover
+    fcntl = None
 
 import httpx
 from loguru import logger
@@ -43,6 +52,13 @@ if TYPE_CHECKING:
 if MSTEAMS_AVAILABLE:
     import jwt
 
+MSTEAMS_REF_TTL_DAYS = 30
+MSTEAMS_REF_TTL_S = MSTEAMS_REF_TTL_DAYS * 24 * 60 * 60
+MSTEAMS_WEBCHAT_HOST = "webchat.botframework.com"
+MSTEAMS_REF_META_FILENAME = "msteams_conversations_meta.json"
+MSTEAMS_REF_LOCK_FILENAME = "msteams_conversations.lock"
+MSTEAMS_REF_TOUCH_INTERVAL_S = 300
+
 
 class MSTeamsConfig(Base):
     """Microsoft Teams channel configuration."""
@@ -58,6 +74,10 @@ class MSTeamsConfig(Base):
     reply_in_thread: bool = True
     mention_only_response: str = "Hi — what can I help with?"
     validate_inbound_auth: bool = True
+    ref_ttl_days: int = Field(default=MSTEAMS_REF_TTL_DAYS, ge=1)
+    prune_web_chat_refs: bool = True
+    prune_non_personal_refs: bool = True
+    ref_touch_interval_s: int = Field(default=MSTEAMS_REF_TOUCH_INTERVAL_S, ge=0)
 
 
 @dataclass
@@ -70,6 +90,7 @@ class ConversationRef:
     activity_id: str | None = None
     conversation_type: str | None = None
     tenant_id: str | None = None
+    updated_at: float | None = None
 
 
 class MSTeamsChannel(BaseChannel):
@@ -102,7 +123,13 @@ class MSTeamsChannel(BaseChannel):
         self._botframework_jwks_expires_at: float = 0.0
         self._refs_path = get_workspace_path() / "state" / "msteams_conversations.json"
         self._refs_path.parent.mkdir(parents=True, exist_ok=True)
+        self._refs_meta_path = self._refs_path.parent / MSTEAMS_REF_META_FILENAME
+        self._refs_lock_path = self._refs_path.parent / MSTEAMS_REF_LOCK_FILENAME
+        self._refs_guard = threading.RLock()
         self._conversation_refs: dict[str, ConversationRef] = self._load_refs()
+        with self._refs_guard:
+            if self._prune_conversation_refs():
+                self._save_refs_locked(prune=True)
 
     async def start(self) -> None:
         """Start the Teams webhook listener."""
@@ -220,7 +247,6 @@ class MSTeamsChannel(BaseChannel):
         token = await self._get_access_token()
         base_url = f"{ref.service_url.rstrip('/')}/v3/conversations/{ref.conversation_id}/activities"
         use_thread_reply = self.config.reply_in_thread and bool(ref.activity_id)
-        url = f"{base_url}/{ref.activity_id}" if use_thread_reply else base_url
         headers = {
             "Authorization": f"Bearer {token}",
             "Content-Type": "application/json",
@@ -233,9 +259,10 @@ class MSTeamsChannel(BaseChannel):
             payload["replyToId"] = ref.activity_id
 
         try:
-            resp = await self._http.post(url, headers=headers, json=payload)
+            resp = await self._http.post(base_url, headers=headers, json=payload)
             resp.raise_for_status()
             logger.info("MSTeams message sent to {}", ref.conversation_id)
+            self._touch_conversation_ref(str(msg.chat_id), persist=True)
         except Exception as e:
             logger.error("MSTeams send failed: {}", e)
             raise
@@ -282,15 +309,17 @@ class MSTeamsChannel(BaseChannel):
             )
             return
 
-        self._conversation_refs[conversation_id] = ConversationRef(
-            service_url=service_url,
-            conversation_id=conversation_id,
-            bot_id=str(recipient.get("id") or "") or None,
-            activity_id=activity_id or None,
-            conversation_type=conversation_type or None,
-            tenant_id=str((channel_data.get("tenant") or {}).get("id") or "") or None,
-        )
-        self._save_refs()
+        with self._refs_guard:
+            self._conversation_refs[conversation_id] = ConversationRef(
+                service_url=service_url,
+                conversation_id=conversation_id,
+                bot_id=str(recipient.get("id") or "") or None,
+                activity_id=activity_id or None,
+                conversation_type=conversation_type or None,
+                tenant_id=str((channel_data.get("tenant") or {}).get("id") or "") or None,
+                updated_at=time.time(),
+            )
+            self._save_refs_locked()
 
         await self._handle_message(
             sender_id=sender_id,
@@ -310,10 +339,12 @@ class MSTeamsChannel(BaseChannel):
         """Extract the user-authored text from a Teams activity."""
         text = str(activity.get("text") or "")
         text = self._strip_possible_bot_mention(text)
+        text = self._normalize_html_whitespace(text)
 
         channel_data = activity.get("channelData") or {}
         reply_to_id = str(activity.get("replyToId") or "").strip()
         normalized_preview = html.unescape(text).replace("&rsquo", "’").strip()
+        normalized_preview = normalized_preview.replace("\xa0", " ")
         normalized_preview = normalized_preview.replace("\r\n", "\n").replace("\r", "\n")
         preview_lines = [line.strip() for line in normalized_preview.split("\n")]
         while preview_lines and not preview_lines[0]:
@@ -333,9 +364,15 @@ class MSTeamsChannel(BaseChannel):
         cleaned = re.sub(r"(?:\r?\n){3,}", "\n\n", cleaned)
         return cleaned.strip()
 
+    def _normalize_html_whitespace(self, text: str) -> str:
+        """Normalize common HTML whitespace/entities from Teams into plain text spacing."""
+        normalized = html.unescape(text).replace("&rsquo", "’")
+        normalized = normalized.replace("\xa0", " ")
+        return normalized
+
     def _normalize_teams_reply_quote(self, text: str) -> str:
         """Normalize Teams quoted replies into a compact structured form."""
-        cleaned = html.unescape(text).replace("&rsquo", "’").strip()
+        cleaned = self._normalize_html_whitespace(text).strip()
         if not cleaned:
             return ""
 
@@ -477,38 +514,243 @@ class MSTeamsChannel(BaseChannel):
         self._botframework_jwks_expires_at = now + 3600
         return self._botframework_jwks
 
-    def _load_refs(self) -> dict[str, ConversationRef]:
-        """Load stored conversation references."""
-        if not self._refs_path.exists():
-            return {}
+    @staticmethod
+    def _safe_float(value: Any) -> float | None:
         try:
-            data = json.loads(self._refs_path.read_text(encoding="utf-8"))
-            out: dict[str, ConversationRef] = {}
-            for key, value in data.items():
-                out[key] = ConversationRef(**value)
-            return out
-        except Exception as e:
-            logger.warning("Failed to load MSTeams conversation refs: {}", e)
+            out = float(value)
+            if out > 0:
+                return out
+        except (TypeError, ValueError):
+            return None
+        return None
+
+    def _normalize_ref_record(self, value: Any) -> ConversationRef | None:
+        """Normalize a stored ref record from legacy/current schema."""
+        if not isinstance(value, dict):
+            return None
+        service_url = str(value.get("service_url") or "").strip()
+        conversation_id = str(value.get("conversation_id") or "").strip()
+        if not service_url or not conversation_id:
+            return None
+        return ConversationRef(
+            service_url=service_url,
+            conversation_id=conversation_id,
+            bot_id=str(value.get("bot_id") or "") or None,
+            activity_id=str(value.get("activity_id") or "") or None,
+            conversation_type=str(value.get("conversation_type") or "") or None,
+            tenant_id=str(value.get("tenant_id") or "") or None,
+            updated_at=self._safe_float(value.get("updated_at")),
+        )
+
+    def _load_refs_raw(self) -> tuple[dict[str, Any], dict[str, Any], bool]:
+        """Load raw refs/main+meta JSON payloads."""
+        main_data: dict[str, Any] = {}
+        meta_data: dict[str, Any] = {}
+        meta_exists = self._refs_meta_path.exists()
+
+        if self._refs_path.exists():
+            try:
+                loaded = json.loads(self._refs_path.read_text(encoding="utf-8"))
+                if isinstance(loaded, dict):
+                    main_data = loaded
+            except Exception as e:
+                logger.warning("Failed to load MSTeams conversation refs: {}", e)
+
+        if meta_exists:
+            try:
+                loaded_meta = json.loads(self._refs_meta_path.read_text(encoding="utf-8"))
+                if isinstance(loaded_meta, dict):
+                    meta_data = loaded_meta
+            except Exception as e:
+                logger.warning("Failed to load MSTeams conversation refs metadata: {}", e)
+
+        return main_data, meta_data, meta_exists
+
+    def _load_refs_from_disk(self) -> dict[str, ConversationRef]:
+        """Load refs from disk with compatibility fallback for legacy layouts."""
+        main_data, meta_data, meta_exists = self._load_refs_raw()
+        if not main_data:
             return {}
 
-    def _save_refs(self) -> None:
-        """Persist conversation references."""
+        out: dict[str, ConversationRef] = {}
+        now = time.time()
+        for key, value in main_data.items():
+            ref = self._normalize_ref_record(value)
+            if not ref:
+                continue
+
+            meta_entry = meta_data.get(key) if isinstance(meta_data, dict) else None
+            meta_ts = None
+            if isinstance(meta_entry, dict):
+                meta_ts = self._safe_float(meta_entry.get("updated_at"))
+            elif meta_entry is not None:
+                meta_ts = self._safe_float(meta_entry)
+
+            if meta_ts is not None:
+                ref.updated_at = meta_ts
+            elif not meta_exists:
+                # First run after introducing meta sidecar: keep legacy refs alive
+                # by initializing timestamps to "now" instead of purging immediately.
+                ref.updated_at = now
+            elif ref.updated_at is None:
+                ref.updated_at = now
+
+            out[key] = ref
+        return out
+
+    def _load_refs(self) -> dict[str, ConversationRef]:
+        """Load stored conversation references."""
+        return self._load_refs_from_disk()
+
+    @contextmanager
+    def _refs_file_lock(self):
+        """Cross-process lock while merging and writing refs state."""
+        self._refs_path.parent.mkdir(parents=True, exist_ok=True)
+        lock_fp = self._refs_lock_path.open("a+", encoding="utf-8")
         try:
-            data = {
-                key: {
-                    "service_url": ref.service_url,
-                    "conversation_id": ref.conversation_id,
-                    "bot_id": ref.bot_id,
-                    "activity_id": ref.activity_id,
-                    "conversation_type": ref.conversation_type,
-                    "tenant_id": ref.tenant_id,
+            if fcntl is not None:
+                fcntl.flock(lock_fp.fileno(), fcntl.LOCK_EX)
+            yield
+        finally:
+            try:
+                if fcntl is not None:
+                    fcntl.flock(lock_fp.fileno(), fcntl.LOCK_UN)
+            finally:
+                lock_fp.close()
+
+    def _is_webchat_service_url(self, service_url: str) -> bool:
+        """Return True when service URL points to unsupported Bot Framework Web Chat."""
+        normalized = service_url.strip()
+        if not normalized:
+            return False
+        host = (urlparse(normalized).hostname or "").strip().lower()
+        if host:
+            return host == MSTEAMS_WEBCHAT_HOST or host.endswith(f".{MSTEAMS_WEBCHAT_HOST}")
+        return MSTEAMS_WEBCHAT_HOST in normalized.lower()
+
+    def _prune_conversation_refs(self, *, now: float | None = None) -> bool:
+        """Remove stale and unsupported conversation refs from memory."""
+        if not self._conversation_refs:
+            return False
+
+        now_ts = time.time() if now is None else now
+        ttl_days = int(self.config.ref_ttl_days)
+        stale_before = now_ts - (ttl_days * 24 * 60 * 60)
+        keys_to_drop: list[str] = []
+
+        for key, ref in self._conversation_refs.items():
+            if self.config.prune_web_chat_refs and self._is_webchat_service_url(ref.service_url):
+                keys_to_drop.append(key)
+                continue
+
+            conv_type = str(ref.conversation_type or "").strip().lower()
+            if self.config.prune_non_personal_refs and conv_type and conv_type != "personal":
+                keys_to_drop.append(key)
+                continue
+
+            try:
+                updated_at = float(ref.updated_at) if ref.updated_at is not None else 0.0
+            except (TypeError, ValueError):
+                updated_at = 0.0
+            if updated_at <= 0 or updated_at < stale_before:
+                keys_to_drop.append(key)
+
+        if not keys_to_drop:
+            return False
+
+        for key in keys_to_drop:
+            self._conversation_refs.pop(key, None)
+        logger.info(
+            "MSTeams pruned {} stale/unsupported conversation refs (ttl={} days)",
+            len(keys_to_drop),
+            ttl_days,
+        )
+        return True
+
+    def _merge_refs_from_disk_locked(self) -> None:
+        """Merge disk refs into memory to reduce lost updates across processes."""
+        disk_refs = self._load_refs_from_disk()
+        for key, disk_ref in disk_refs.items():
+            mem_ref = self._conversation_refs.get(key)
+            if mem_ref is None:
+                self._conversation_refs[key] = disk_ref
+                continue
+            disk_ts = self._safe_float(disk_ref.updated_at) or 0.0
+            mem_ts = self._safe_float(mem_ref.updated_at) or 0.0
+            if disk_ts > mem_ts:
+                self._conversation_refs[key] = disk_ref
+
+    def _touch_conversation_ref(self, chat_id: str, *, persist: bool = False) -> None:
+        """Refresh updated_at for an active ref to keep it from expiring while used."""
+        with self._refs_guard:
+            ref = self._conversation_refs.get(str(chat_id))
+            if not ref:
+                return
+            now = time.time()
+            prev = self._safe_float(ref.updated_at) or 0.0
+            min_interval = max(0, int(self.config.ref_touch_interval_s))
+            if min_interval > 0 and prev > 0 and now - prev < min_interval:
+                return
+            ref.updated_at = now
+            if persist:
+                self._save_refs_locked()
+
+    def _write_json_atomically(self, path, data: dict[str, Any]) -> None:
+        """Write refs JSON atomically to reduce corruption risk during crashes."""
+        payload = json.dumps(data, indent=2)
+        tmp_path: str | None = None
+        try:
+            fd, tmp_path = tempfile.mkstemp(
+                dir=str(path.parent),
+                prefix=f"{path.name}.",
+                suffix=".tmp",
+            )
+            with os.fdopen(fd, "w", encoding="utf-8") as f:
+                f.write(payload)
+                f.flush()
+                os.fsync(f.fileno())
+            os.replace(tmp_path, path)
+        finally:
+            if tmp_path and os.path.exists(tmp_path):
+                try:
+                    os.unlink(tmp_path)
+                except OSError:
+                    pass
+
+    def _save_refs_locked(self, *, prune: bool = True) -> None:
+        """Persist conversation references (caller must hold _refs_guard)."""
+        try:
+            with self._refs_file_lock():
+                self._merge_refs_from_disk_locked()
+                if prune:
+                    self._prune_conversation_refs()
+                refs_data = {
+                    key: {
+                        "service_url": ref.service_url,
+                        "conversation_id": ref.conversation_id,
+                        "bot_id": ref.bot_id,
+                        "activity_id": ref.activity_id,
+                        "conversation_type": ref.conversation_type,
+                        "tenant_id": ref.tenant_id,
+                    }
+                    for key, ref in self._conversation_refs.items()
                 }
-                for key, ref in self._conversation_refs.items()
-            }
-            self._refs_path.write_text(json.dumps(data, indent=2), encoding="utf-8")
+                refs_meta = {
+                    key: {
+                        "updated_at": self._safe_float(ref.updated_at),
+                    }
+                    for key, ref in self._conversation_refs.items()
+                }
+                self._write_json_atomically(self._refs_path, refs_data)
+                self._write_json_atomically(self._refs_meta_path, refs_meta)
         except Exception as e:
             logger.warning("Failed to save MSTeams conversation refs: {}", e)
 
+    def _save_refs(self, *, prune: bool = True) -> None:
+        """Persist conversation references."""
+        with self._refs_guard:
+            self._save_refs_locked(prune=prune)
+
     async def _get_access_token(self) -> str:
         """Fetch an access token for Bot Framework / Azure Bot auth."""
 

nanobot/channels/slack.py

@@ -2,8 +2,10 @@
 
 import asyncio
 import re
+from pathlib import Path
 from typing import Any
 
+import httpx
 from loguru import logger
 from pydantic import Field
 from slack_sdk.socket_mode.request import SocketModeRequest
@@ -15,7 +17,9 @@ from slackify_markdown import slackify_markdown
 from nanobot.bus.events import OutboundMessage
 from nanobot.bus.queue import MessageBus
 from nanobot.channels.base import BaseChannel
+from nanobot.config.paths import get_media_dir
 from nanobot.config.schema import Base
+from nanobot.utils.helpers import safe_filename, split_message
 
 
 class SlackDMConfig(Base):
@@ -38,12 +42,19 @@ class SlackConfig(Base):
     reply_in_thread: bool = True
     react_emoji: str = "eyes"
     done_emoji: str = "white_check_mark"
+    include_thread_context: bool = True
+    thread_context_limit: int = 20
     allow_from: list[str] = Field(default_factory=list)
     group_policy: str = "mention"
     group_allow_from: list[str] = Field(default_factory=list)
     dm: SlackDMConfig = Field(default_factory=SlackDMConfig)
 
 
+SLACK_MAX_MESSAGE_LEN = 39_000  # Slack API allows ~40k; leave margin
+SLACK_DOWNLOAD_TIMEOUT = 30.0
+_HTML_DOWNLOAD_PREFIXES = (b"<!doctype html", b"<html")
+
+
 class SlackChannel(BaseChannel):
     """Slack channel using Socket Mode."""
 
@@ -57,6 +68,8 @@ class SlackChannel(BaseChannel):
     def default_config(cls) -> dict[str, Any]:
         return SlackConfig().model_dump(by_alias=True)
 
+    _THREAD_CONTEXT_CACHE_LIMIT = 10_000
+
     def __init__(self, config: Any, bus: MessageBus):
         if isinstance(config, dict):
             config = SlackConfig.model_validate(config)
@@ -66,6 +79,7 @@ class SlackChannel(BaseChannel):
         self._socket_client: SocketModeClient | None = None
         self._bot_user_id: str | None = None
         self._target_cache: dict[str, str] = {}
+        self._thread_context_attempted: set[str] = set()
 
     async def start(self) -> None:
         """Start the Slack Socket Mode client."""
@@ -119,23 +133,24 @@ class SlackChannel(BaseChannel):
             target_chat_id = await self._resolve_target_chat_id(msg.chat_id)
             slack_meta = msg.metadata.get("slack", {}) if msg.metadata else {}
             thread_ts = slack_meta.get("thread_ts")
-            channel_type = slack_meta.get("channel_type")
             origin_chat_id = str((slack_meta.get("event", {}) or {}).get("channel") or msg.chat_id)
-            # Slack DMs don't use threads; channel/group replies may keep thread_ts.
-            thread_ts_param = (
-                thread_ts
-                if thread_ts and channel_type != "im" and target_chat_id == origin_chat_id
-                else None
-            )
+            # Reply in the same thread the inbound message belongs to (works
+            # for both real channel threads and DM threads). When the agent
+            # is forwarding to a different channel, drop thread_ts because it
+            # only makes sense within the originating conversation.
+            thread_ts_param = thread_ts if thread_ts and target_chat_id == origin_chat_id else None
 
-            # Slack rejects empty text payloads. Keep media-only messages media-only,
-            # but send a single blank message when the bot has no text or files to send.
             if msg.content or not (msg.media or []):
-                await self._web_client.chat_postMessage(
-                    channel=target_chat_id,
-                    text=self._to_mrkdwn(msg.content) if msg.content else " ",
-                    thread_ts=thread_ts_param,
-                )
+                mrkdwn = self._to_mrkdwn(msg.content) if msg.content else " "
+                buttons = getattr(msg, "buttons", None) or []
+                chunks = split_message(mrkdwn, SLACK_MAX_MESSAGE_LEN)
+                for index, chunk in enumerate(chunks):
+                    kwargs: dict[str, Any] = dict(
+                        channel=target_chat_id, text=chunk, thread_ts=thread_ts_param,
+                    )
+                    if buttons and index == len(chunks) - 1:
+                        kwargs["blocks"] = self._build_button_blocks(chunk, buttons)
+                    await self._web_client.chat_postMessage(**kwargs)
 
             for media_path in msg.media or []:
                 try:
@@ -273,6 +288,9 @@ class SlackChannel(BaseChannel):
         req: SocketModeRequest,
     ) -> None:
         """Handle incoming Socket Mode requests."""
+        if req.type == "interactive":
+            await self._on_block_action(client, req)
+            return
         if req.type != "events_api":
             return
 
@@ -292,8 +310,10 @@ class SlackChannel(BaseChannel):
         sender_id = event.get("user")
         chat_id = event.get("channel")
 
-        # Ignore bot/system messages (any subtype = not a normal user message)
-        if event.get("subtype"):
+        subtype = event.get("subtype")
+        # Slack uses subtype=file_share for user messages with attachments.
+        # Ignore other subtypes such as bot_message / message_changed / deleted.
+        if subtype and subtype != "file_share":
             return
         if self._bot_user_id and sender_id == self._bot_user_id:
             return
@@ -308,7 +328,7 @@ class SlackChannel(BaseChannel):
         logger.debug(
             "Slack event: type={} subtype={} user={} channel={} channel_type={} text={}",
             event_type,
-            event.get("subtype"),
+            subtype,
             sender_id,
             chat_id,
             event.get("channel_type"),
@@ -327,9 +347,18 @@ class SlackChannel(BaseChannel):
 
         text = self._strip_bot_mention(text)
 
-        thread_ts = event.get("thread_ts")
-        if self.config.reply_in_thread and not thread_ts:
-            thread_ts = event.get("ts")
+        event_ts = event.get("ts")
+        raw_thread_ts = event.get("thread_ts")
+        thread_ts = raw_thread_ts
+        # In DMs we don't auto-open a thread on top-level messages (it would
+        # bury replies under "1 reply"). But if the user explicitly opened a
+        # thread inside the DM, raw_thread_ts is set and we honor it.
+        if (
+            self.config.reply_in_thread
+            and not thread_ts
+            and channel_type != "im"
+        ):
+            thread_ts = event_ts
         # Add :eyes: reaction to the triggering message (best-effort)
         try:
             if self._web_client and event.get("ts"):
@@ -341,14 +370,43 @@ class SlackChannel(BaseChannel):
         except Exception as e:
             logger.debug("Slack reactions_add failed: {}", e)
 
-        # Thread-scoped session key for channel/group messages
-        session_key = f"slack:{chat_id}:{thread_ts}" if thread_ts and channel_type != "im" else None
+        # Thread-scoped session key whenever the user is in a real thread
+        # (raw_thread_ts is set). DM threads get their own session, separate
+        # from the DM root, so context doesn't bleed across thread boundaries.
+        session_key = (
+            f"slack:{chat_id}:{thread_ts}" if thread_ts and raw_thread_ts else None
+        )
+        media_paths: list[str] = []
+        file_markers: list[str] = []
+        for file_info in event.get("files") or []:
+            if not isinstance(file_info, dict):
+                continue
+            file_path, marker = await self._download_slack_file(file_info)
+            if file_path:
+                media_paths.append(file_path)
+            if marker:
+                file_markers.append(marker)
+
+        is_slash = text.strip().startswith("/")
+        content = text if is_slash else await self._with_thread_context(
+            text,
+            chat_id=chat_id,
+            channel_type=channel_type,
+            thread_ts=thread_ts,
+            raw_thread_ts=raw_thread_ts,
+            current_ts=event_ts,
+        )
+        if file_markers:
+            content = "\n".join(part for part in [content, *file_markers] if part)
+        if not content and not media_paths:
+            return
 
         try:
             await self._handle_message(
                 sender_id=sender_id,
                 chat_id=chat_id,
-                content=text,
+                content=content,
+                media=media_paths,
                 metadata={
                     "slack": {
                         "event": event,
@@ -361,6 +419,163 @@ class SlackChannel(BaseChannel):
         except Exception:
             logger.exception("Error handling Slack message from {}", sender_id)
 
+    async def _download_slack_file(self, file_info: dict[str, Any]) -> tuple[str | None, str]:
+        """Download a Slack private file to the local media directory."""
+        file_id = str(file_info.get("id") or "file")
+        name = str(
+            file_info.get("name")
+            or file_info.get("title")
+            or file_info.get("id")
+            or "slack-file"
+        )
+        marker_type = "image" if str(file_info.get("mimetype") or "").startswith("image/") else "file"
+        marker = f"[{marker_type}: {name}]"
+        url = str(file_info.get("url_private_download") or file_info.get("url_private") or "")
+        if not url:
+            return None, f"[{marker_type}: {name}: missing download url]"
+        if not self.config.bot_token:
+            return None, f"[{marker_type}: {name}: missing bot token]"
+
+        filename = safe_filename(f"{file_id}_{name}")
+        path = Path(get_media_dir("slack")) / filename
+        try:
+            async with httpx.AsyncClient(timeout=SLACK_DOWNLOAD_TIMEOUT, follow_redirects=True) as client:
+                response = await client.get(
+                    url,
+                    headers={"Authorization": f"Bearer {self.config.bot_token}"},
+                )
+                response.raise_for_status()
+            if self._looks_like_html_download(response):
+                raise ValueError("Slack returned HTML instead of file content")
+            path.write_bytes(response.content)
+            return str(path), marker
+        except Exception as e:
+            logger.warning("Failed to download Slack file {}: {}", file_id, e)
+            return None, f"[{marker_type}: {name}: download failed]"
+
+    @staticmethod
+    def _looks_like_html_download(response: httpx.Response) -> bool:
+        content_type = response.headers.get("content-type", "").lower()
+        if "text/html" in content_type:
+            return True
+        preview = response.content[:256].lstrip().lower()
+        return preview.startswith(_HTML_DOWNLOAD_PREFIXES)
+
+    async def _on_block_action(self, client: SocketModeClient, req: SocketModeRequest) -> None:
+        """Handle button clicks from ask_user blocks."""
+        await client.send_socket_mode_response(SocketModeResponse(envelope_id=req.envelope_id))
+        payload = req.payload or {}
+        actions = payload.get("actions") or []
+        if not actions:
+            return
+        value = str(actions[0].get("value") or "")
+        user_info = payload.get("user") or {}
+        sender_id = str(user_info.get("id") or "")
+        channel_info = payload.get("channel") or {}
+        chat_id = str(channel_info.get("id") or "")
+        if not sender_id or not chat_id or not value:
+            return
+        message_info = payload.get("message") or {}
+        thread_ts = message_info.get("thread_ts") or message_info.get("ts")
+        channel_type = self._infer_channel_type(chat_id)
+        if not self._is_allowed(sender_id, chat_id, channel_type):
+            return
+        session_key = f"slack:{chat_id}:{thread_ts}" if thread_ts else None
+        try:
+            await self._handle_message(
+                sender_id=sender_id,
+                chat_id=chat_id,
+                content=value,
+                metadata={"slack": {"thread_ts": thread_ts, "channel_type": channel_type}},
+                session_key=session_key,
+            )
+        except Exception:
+            logger.exception("Error handling Slack button click from {}", sender_id)
+
+    async def _with_thread_context(
+        self,
+        text: str,
+        *,
+        chat_id: str,
+        channel_type: str,
+        thread_ts: str | None,
+        raw_thread_ts: str | None,
+        current_ts: str | None,
+    ) -> str:
+        """Include thread history the first time the bot is pulled into a Slack thread."""
+        del channel_type  # DM and channel threads are both fetched via conversations.replies
+        if (
+            not self.config.include_thread_context
+            or not self._web_client
+            or not raw_thread_ts
+            or not thread_ts
+            or current_ts == thread_ts
+        ):
+            return text
+
+        key = f"{chat_id}:{thread_ts}"
+        if key in self._thread_context_attempted:
+            return text
+        if len(self._thread_context_attempted) >= self._THREAD_CONTEXT_CACHE_LIMIT:
+            self._thread_context_attempted.clear()
+        self._thread_context_attempted.add(key)
+
+        try:
+            response = await self._web_client.conversations_replies(
+                channel=chat_id,
+                ts=thread_ts,
+                limit=max(1, self.config.thread_context_limit),
+            )
+        except Exception as e:
+            logger.warning("Slack thread context unavailable for {}: {}", key, e)
+            return text
+
+        lines = self._format_thread_context(
+            response.get("messages", []),
+            current_ts=current_ts,
+        )
+        if not lines:
+            return text
+        return "Slack thread context before this mention:\n" + "\n".join(lines) + f"\n\nCurrent message:\n{text}"
+
+    def _format_thread_context(self, messages: list[dict[str, Any]], *, current_ts: str | None) -> list[str]:
+        lines: list[str] = []
+        for item in messages:
+            if item.get("ts") == current_ts:
+                continue
+            if item.get("subtype"):
+                continue
+            sender = str(item.get("user") or item.get("bot_id") or "unknown")
+            is_bot = self._bot_user_id is not None and sender == self._bot_user_id
+            label = "bot" if is_bot else f"<@{sender}>"
+            text = str(item.get("text") or "").strip()
+            if not text:
+                continue
+            text = self._strip_bot_mention(text)
+            if len(text) > 500:
+                text = text[:500] + "…"
+            lines.append(f"- {label}: {text}")
+        return lines
+
+    @staticmethod
+    def _build_button_blocks(text: str, buttons: list[list[str]]) -> list[dict[str, Any]]:
+        """Build Slack Block Kit blocks with action buttons for ask_user choices."""
+        blocks: list[dict[str, Any]] = [
+            {"type": "section", "text": {"type": "mrkdwn", "text": text[:3000]}},
+        ]
+        elements = []
+        for row in buttons:
+            for label in row:
+                elements.append({
+                    "type": "button",
+                    "text": {"type": "plain_text", "text": label[:75]},
+                    "value": label[:75],
+                    "action_id": f"ask_user_{label[:50]}",
+                })
+        if elements:
+            blocks.append({"type": "actions", "elements": elements[:25]})
+        return blocks
+
     async def _update_react_emoji(self, chat_id: str, ts: str | None) -> None:
         """Remove the in-progress reaction and optionally add a done reaction."""
         if not self._web_client or not ts:
@@ -407,6 +622,19 @@ class SlackChannel(BaseChannel):
             return chat_id in self.config.group_allow_from
         return False
 
+    def is_allowed(self, sender_id: str) -> bool:
+        # Slack needs channel-aware policy checks, so _on_socket_request and
+        # _on_block_action call _is_allowed before handing off to BaseChannel.
+        return True
+
+    @staticmethod
+    def _infer_channel_type(chat_id: str) -> str:
+        if chat_id.startswith("D"):
+            return "im"
+        if chat_id.startswith("G"):
+            return "group"
+        return "channel"
+
     def _strip_bot_mention(self, text: str) -> str:
         if not text or not self._bot_user_id:
             return text
@@ -425,7 +653,7 @@ class SlackChannel(BaseChannel):
         if not text:
             return ""
         text = cls._TABLE_RE.sub(cls._convert_table, text)
-        return cls._fixup_mrkdwn(slackify_markdown(text))
+        return cls._fixup_mrkdwn(slackify_markdown(text)).rstrip("\n")
 
     @classmethod
     def _fixup_mrkdwn(cls, text: str) -> str:

nanobot/channels/telegram.py

@@ -7,13 +7,14 @@ import re
 import time
 import unicodedata
 from dataclasses import dataclass
+from pathlib import Path
 from typing import Any, Literal
 
 from loguru import logger
 from pydantic import Field
-from telegram import BotCommand, ReactionTypeEmoji, ReplyParameters, Update
+from telegram import BotCommand, InlineKeyboardButton, InlineKeyboardMarkup, ReactionTypeEmoji, ReplyParameters, Update
 from telegram.error import BadRequest, NetworkError, TimedOut
-from telegram.ext import Application, ContextTypes, MessageHandler, filters
+from telegram.ext import Application, CallbackQueryHandler, ContextTypes, MessageHandler, filters
 from telegram.request import HTTPXRequest
 
 from nanobot.bus.events import OutboundMessage
@@ -230,6 +231,8 @@ class TelegramConfig(Base):
     connection_pool_size: int = 32
     pool_timeout: float = 5.0
     streaming: bool = True
+    # Enable inline keyboard buttons in Telegram messages.
+    inline_keyboards: bool = False
     stream_edit_interval: float = Field(default=_STREAM_EDIT_INTERVAL_DEFAULT, ge=0.1)
 
 
@@ -355,15 +358,25 @@ class TelegramChannel(BaseChannel):
         )
         self._app.add_handler(MessageHandler(filters.Regex(r"^/help(?:@\w+)?$"), self._on_help))
 
-        # Add message handler for text, photos, voice, documents, and locations
+        # Add message handler for text, photos, video, voice, documents, and locations
         self._app.add_handler(
             MessageHandler(
-                (filters.TEXT | filters.PHOTO | filters.VOICE | filters.AUDIO | filters.Document.ALL | filters.LOCATION)
+                (filters.TEXT | filters.PHOTO | filters.VIDEO | filters.VIDEO_NOTE
+                 | filters.ANIMATION | filters.VOICE | filters.AUDIO
+                 | filters.Document.ALL | filters.LOCATION)
                 & ~filters.COMMAND,
                 self._on_message
             )
         )
 
+        # Conditionally register inline keyboard callback handler
+        if self.config.inline_keyboards:
+            self._app.add_handler(CallbackQueryHandler(self._on_callback_query))
+            allowed_updates = ["message", "callback_query"]
+            logger.debug("Telegram inline keyboards enabled")
+        else:
+            allowed_updates = ["message"]
+
         logger.info("Starting Telegram bot (polling mode)...")
 
         # Initialize and start polling
@@ -384,7 +397,7 @@ class TelegramChannel(BaseChannel):
 
         # Start polling (this runs until stopped)
         await self._app.updater.start_polling(
-            allowed_updates=["message"],
+            allowed_updates=allowed_updates,
             drop_pending_updates=False,  # Process pending messages on startup
             error_callback=self._on_polling_error,
         )
@@ -419,6 +432,8 @@ class TelegramChannel(BaseChannel):
         ext = path.rsplit(".", 1)[-1].lower() if "." in path else ""
         if ext in ("jpg", "jpeg", "png", "gif", "webp"):
             return "photo"
+        if ext in ("mp4", "mov", "avi", "mkv", "webm", "3gp"):
+            return "video"
         if ext == "ogg":
             return "voice"
         if ext in ("mp3", "m4a", "wav", "aac"):
@@ -471,10 +486,19 @@ class TelegramChannel(BaseChannel):
                 media_type = self._get_media_type(media_path)
                 sender = {
                     "photo": self._app.bot.send_photo,
+                    "video": self._app.bot.send_video,
                     "voice": self._app.bot.send_voice,
                     "audio": self._app.bot.send_audio,
                 }.get(media_type, self._app.bot.send_document)
-                param = "photo" if media_type == "photo" else media_type if media_type in ("voice", "audio") else "document"
+                param = {
+                    "photo": "photo",
+                    "video": "video",
+                    "voice": "voice",
+                    "audio": "audio",
+                }.get(media_type, "document")
+                extra: dict[str, Any] = {}
+                if media_type == "video":
+                    extra["supports_streaming"] = True
 
                 # Telegram Bot API accepts HTTP(S) URLs directly for media params.
                 if self._is_remote_media_url(media_path):
@@ -487,16 +511,19 @@ class TelegramChannel(BaseChannel):
                         **{param: media_path},
                         reply_parameters=reply_params,
                         **thread_kwargs,
+                        **extra,
                     )
                     continue
 
-                with open(media_path, "rb") as f:
-                    await sender(
-                        chat_id=chat_id,
-                        **{param: f},
-                        reply_parameters=reply_params,
-                        **thread_kwargs,
-                    )
+                media_bytes = Path(media_path).read_bytes()
+                await self._call_with_retry(
+                    sender,
+                    chat_id=chat_id,
+                    **{param: media_bytes},
+                    reply_parameters=reply_params,
+                    **thread_kwargs,
+                    **extra,
+                )
             except Exception as e:
                 filename = media_path.rsplit("/", 1)[-1]
                 logger.error("Failed to send media {}: {}", media_path, e)
@@ -510,16 +537,25 @@ class TelegramChannel(BaseChannel):
         # Send text content
         if msg.content and msg.content != "[empty message]":
             render_as_blockquote = bool(msg.metadata.get("_tool_hint"))
-            for chunk in split_message(msg.content, TELEGRAM_MAX_MESSAGE_LEN):
+            buttons = getattr(msg, "buttons", None) or []
+            reply_markup = self._build_keyboard(buttons) if buttons else None
+            text = msg.content
+            # Fallback: no native keyboard → splice labels into the message so the choices survive.
+            if buttons and reply_markup is None:
+                text = f"{text}\n\n{self._buttons_as_text(buttons)}"
+            chunks = split_message(text, TELEGRAM_MAX_MESSAGE_LEN)
+            for i, chunk in enumerate(chunks):
+                is_last = (i == len(chunks) - 1)
                 await self._send_text(
                     chat_id, chunk, reply_params, thread_kwargs,
                     render_as_blockquote=render_as_blockquote,
+                    reply_markup=reply_markup if is_last else None,
                 )
 
     async def _call_with_retry(self, fn, *args, **kwargs):
         """Call an async Telegram API function with retry on pool/network timeout and RetryAfter."""
         from telegram.error import RetryAfter
-        
+
         for attempt in range(1, _SEND_MAX_RETRIES + 1):
             try:
                 return await fn(*args, **kwargs)
@@ -549,6 +585,7 @@ class TelegramChannel(BaseChannel):
         reply_params=None,
         thread_kwargs: dict | None = None,
         render_as_blockquote: bool = False,
+        reply_markup=None,
     ) -> None:
         """Send a plain text message with HTML fallback."""
         try:
@@ -557,12 +594,10 @@ class TelegramChannel(BaseChannel):
                 self._app.bot.send_message,
                 chat_id=chat_id, text=html, parse_mode="HTML",
                 reply_parameters=reply_params,
+                reply_markup=reply_markup,
                 **(thread_kwargs or {}),
             )
         except BadRequest as e:
-            # Only fall back to plain text on actual HTML parse/format errors.
-            # Network errors (TimedOut, NetworkError) should propagate immediately
-            # to avoid doubling connection demand during pool exhaustion.
             logger.warning("HTML parse failed, falling back to plain text: {}", e)
             try:
                 await self._call_with_retry(
@@ -570,6 +605,7 @@ class TelegramChannel(BaseChannel):
                     chat_id=chat_id,
                     text=text,
                     reply_parameters=reply_params,
+                    reply_markup=reply_markup,
                     **(thread_kwargs or {}),
                 )
             except Exception as e2:
@@ -796,13 +832,13 @@ class TelegramChannel(BaseChannel):
         text = getattr(reply, "text", None) or getattr(reply, "caption", None) or ""
         if len(text) > TELEGRAM_REPLY_CONTEXT_MAX_LEN:
             text = text[:TELEGRAM_REPLY_CONTEXT_MAX_LEN] + "..."
-            
+
         if not text:
             return None
-            
+
         bot_id, _ = await self._ensure_bot_identity()
         reply_user = getattr(reply, "from_user", None)
-        
+
         if bot_id and reply_user and getattr(reply_user, "id", None) == bot_id:
             return f"[Reply to bot: {text}]"
         elif reply_user and getattr(reply_user, "username", None):
@@ -947,7 +983,7 @@ class TelegramChannel(BaseChannel):
         message = update.message
         user = update.effective_user
         self._remember_thread_context(message)
-        
+
         # Strip @bot_username suffix if present
         content = message.text or ""
         if content.startswith("/") and "@" in content:
@@ -955,7 +991,7 @@ class TelegramChannel(BaseChannel):
             cmd_part = cmd_part.split("@")[0]
             content = f"{cmd_part} {rest[0]}" if rest else cmd_part
         content = self._normalize_telegram_command(content)
-            
+
         await self._handle_message(
             sender_id=self._sender_id(user),
             chat_id=str(message.chat_id),
@@ -1165,18 +1201,76 @@ class TelegramChannel(BaseChannel):
         if mime_type:
             ext_map = {
                 "image/jpeg": ".jpg", "image/png": ".png", "image/gif": ".gif",
+                "image/webp": ".webp",
                 "audio/ogg": ".ogg", "audio/mpeg": ".mp3", "audio/mp4": ".m4a",
+                "video/mp4": ".mp4", "video/quicktime": ".mov", "video/webm": ".webm",
+                "video/x-matroska": ".mkv", "video/3gpp": ".3gp",
             }
             if mime_type in ext_map:
                 return ext_map[mime_type]
 
-        type_map = {"image": ".jpg", "voice": ".ogg", "audio": ".mp3", "file": ""}
+        type_map = {"image": ".jpg", "voice": ".ogg", "audio": ".mp3", "video": ".mp4", "file": ""}
         if ext := type_map.get(media_type, ""):
             return ext
 
         if filename:
-            from pathlib import Path
-
             return "".join(Path(filename).suffixes)
 
         return ""
+
+    def _build_keyboard(self, buttons: list) -> InlineKeyboardMarkup | None:
+        """Build inline keyboard markup if inline_keyboards is enabled."""
+        if not buttons or not self.config.inline_keyboards:
+            return None
+        keyboard = [
+            [InlineKeyboardButton(label, callback_data=self._safe_callback_data(label)) for label in row]
+            for row in buttons
+        ]
+        return InlineKeyboardMarkup(keyboard)
+
+    @staticmethod
+    def _safe_callback_data(label: str) -> str:
+        # Telegram caps callback_data at 64 bytes UTF-8; truncate at a char boundary so the keyboard still sends.
+        encoded = label.encode("utf-8")
+        if len(encoded) <= 64:
+            return label
+        return encoded[:64].decode("utf-8", errors="ignore")
+
+    @staticmethod
+    def _buttons_as_text(buttons: list[list[str]]) -> str:
+        # Buttons are semantic options; when we can't render a keyboard, the user still needs to see them.
+        return "\n".join(" ".join(f"[{label}]" for label in row) for row in buttons if row)
+
+    async def _on_callback_query(self, update: Update, context: ContextTypes.DEFAULT_TYPE) -> None:
+        """Handle inline keyboard button clicks (callback queries)."""
+        if not update.callback_query or not update.effective_user:
+            return
+        query = update.callback_query
+        user = update.effective_user
+        chat_id = query.message.chat_id if query.message else None
+        sender_id = self._sender_id(user)
+        if not chat_id:
+            logger.warning("Callback query without chat_id")
+            return
+        button_label = query.data or ""
+        await query.answer()
+        if query.message:
+            try:
+                await query.message.edit_reply_markup(reply_markup=None)
+            except Exception:
+                pass
+        logger.debug("Inline button tap from {}: {}", sender_id, button_label)
+        self._start_typing(str(chat_id))
+        await self._handle_message(
+            sender_id=sender_id,
+            chat_id=str(chat_id),
+            content=button_label,
+            metadata={
+                "callback_query_id": query.id,
+                "button_label": button_label,
+                "user_id": user.id,
+                "username": user.username,
+                "first_name": user.first_name,
+                "is_callback": True,
+            },
+        )

nanobot/channels/websocket.py

@@ -13,6 +13,7 @@ import json
 import mimetypes
 import re
 import secrets
+import shutil
 import ssl
 import time
 import uuid
@@ -33,6 +34,7 @@ from nanobot.bus.queue import MessageBus
 from nanobot.channels.base import BaseChannel
 from nanobot.config.paths import get_media_dir
 from nanobot.config.schema import Base
+from nanobot.utils.helpers import safe_filename
 from nanobot.utils.media_decode import (
     FileSizeExceeded,
     save_base64_data_url,
@@ -52,6 +54,14 @@ def _normalize_config_path(path: str) -> str:
     return _strip_trailing_slash(path)
 
 
+def _append_buttons_as_text(text: str, buttons: list[list[str]]) -> str:
+    labels = [label for row in buttons for label in row if label]
+    if not labels:
+        return text
+    fallback = "\n".join(f"{index}. {label}" for index, label in enumerate(labels, 1))
+    return f"{text}\n\n{fallback}" if text else fallback
+
+
 class WebSocketConfig(Base):
     """WebSocket server channel configuration.
 
@@ -218,12 +228,14 @@ def _parse_envelope(raw: str) -> dict[str, Any] | None:
     return data
 
 
-# Per-message image limits. The server-side guard is a touch looser than the
+# Per-message media limits. The server-side guard is a touch looser than the
 # client's ``Worker`` normalization target (6 MB) — tolerate client slop, but
 # still cap total ingress at ``_MAX_IMAGES_PER_MESSAGE * _MAX_IMAGE_BYTES``
 # which fits comfortably inside ``max_message_bytes``.
 _MAX_IMAGES_PER_MESSAGE = 4
 _MAX_IMAGE_BYTES = 8 * 1024 * 1024
+_MAX_VIDEOS_PER_MESSAGE = 1
+_MAX_VIDEO_BYTES = 20 * 1024 * 1024
 
 # Image MIME whitelist — matches the Composer's ``accept`` list. SVG is
 # explicitly excluded to avoid the XSS surface inside embedded scripts.
@@ -234,6 +246,14 @@ _IMAGE_MIME_ALLOWED: frozenset[str] = frozenset({
     "image/gif",
 })
 
+_VIDEO_MIME_ALLOWED: frozenset[str] = frozenset({
+    "video/mp4",
+    "video/webm",
+    "video/quicktime",
+})
+
+_UPLOAD_MIME_ALLOWED: frozenset[str] = _IMAGE_MIME_ALLOWED | _VIDEO_MIME_ALLOWED
+
 _DATA_URL_MIME_RE = re.compile(r"^data:([^;]+);base64,", re.DOTALL)
 
 
@@ -339,6 +359,9 @@ _MEDIA_ALLOWED_MIMES: frozenset[str] = frozenset({
     "image/jpeg",
     "image/webp",
     "image/gif",
+    "video/mp4",
+    "video/webm",
+    "video/quicktime",
 })
 
 
@@ -516,6 +539,12 @@ class WebSocketChannel(BaseChannel):
         if got == "/api/sessions":
             return self._handle_sessions_list(request)
 
+        if got == "/api/settings":
+            return self._handle_settings(request)
+
+        if got == "/api/settings/update":
+            return self._handle_settings_update(request)
+
         m = re.match(r"^/api/sessions/([^/]+)/messages$", got)
         if m:
             return self._handle_session_messages(request, m.group(1))
@@ -624,6 +653,75 @@ class WebSocketChannel(BaseChannel):
         ]
         return _http_json_response({"sessions": cleaned})
 
+    def _settings_payload(self, *, requires_restart: bool = False) -> dict[str, Any]:
+        from nanobot.config.loader import get_config_path, load_config
+        from nanobot.providers.registry import PROVIDERS, find_by_name
+
+        config = load_config()
+        defaults = config.agents.defaults
+        provider_name = config.get_provider_name(defaults.model) or defaults.provider
+        provider = config.get_provider(defaults.model)
+        selected_provider = provider_name
+        if defaults.provider != "auto":
+            spec = find_by_name(defaults.provider)
+            selected_provider = spec.name if spec else provider_name
+        return {
+            "agent": {
+                "model": defaults.model,
+                "provider": selected_provider,
+                "resolved_provider": provider_name,
+                "has_api_key": bool(provider and provider.api_key),
+            },
+            "providers": [
+                {"name": "auto", "label": "Auto"}
+            ] + [
+                {"name": spec.name, "label": spec.label}
+                for spec in PROVIDERS
+            ],
+            "runtime": {
+                "config_path": str(get_config_path().expanduser()),
+            },
+            "requires_restart": requires_restart,
+        }
+
+    def _handle_settings(self, request: WsRequest) -> Response:
+        if not self._check_api_token(request):
+            return _http_error(401, "Unauthorized")
+        return _http_json_response(self._settings_payload())
+
+    def _handle_settings_update(self, request: WsRequest) -> Response:
+        if not self._check_api_token(request):
+            return _http_error(401, "Unauthorized")
+        from nanobot.config.loader import load_config, save_config
+        from nanobot.providers.registry import find_by_name
+
+        query = _parse_query(request.path)
+        config = load_config()
+        defaults = config.agents.defaults
+        changed = False
+
+        model = _query_first(query, "model")
+        if model is not None:
+            model = model.strip()
+            if not model:
+                return _http_error(400, "model is required")
+            if defaults.model != model:
+                defaults.model = model
+                changed = True
+
+        provider = _query_first(query, "provider")
+        if provider is not None:
+            provider = provider.strip() or "auto"
+            if provider != "auto" and find_by_name(provider) is None:
+                return _http_error(400, "unknown provider")
+            if defaults.provider != provider:
+                defaults.provider = provider
+                changed = True
+
+        if changed:
+            save_config(config)
+        return _http_json_response(self._settings_payload(requires_restart=changed))
+
     @staticmethod
     def _is_webui_session_key(key: str) -> bool:
         """Return True when *key* belongs to the webui's websocket-only surface."""
@@ -703,6 +801,33 @@ class WebSocketChannel(BaseChannel):
         ).digest()[:16]
         return f"/api/media/{_b64url_encode(mac)}/{payload}"
 
+    def _sign_or_stage_media_path(self, path: Path) -> dict[str, str] | None:
+        """Return a signed media URL payload for *path*.
+
+        Persisted inbound media already lives under ``get_media_dir`` and can
+        be signed directly. Outbound bot-generated files may live anywhere on
+        disk; copy those into the websocket media bucket first so the browser
+        can fetch them through the existing signed media route without
+        exposing arbitrary filesystem paths.
+        """
+        signed = self._sign_media_path(path)
+        if signed is not None:
+            return {"url": signed, "name": path.name}
+        try:
+            if not path.is_file():
+                return None
+            media_dir = get_media_dir("websocket")
+            safe_name = safe_filename(path.name) or "attachment"
+            staged = media_dir / f"{uuid.uuid4().hex[:12]}-{safe_name}"
+            shutil.copyfile(path, staged)
+        except OSError as exc:
+            logger.warning("websocket: failed to stage outbound media {}: {}", path, exc)
+            return None
+        signed = self._sign_media_path(staged)
+        if signed is None:
+            return None
+        return {"url": signed, "name": path.name}
+
     def _handle_media_fetch(self, sig: str, payload: str) -> Response:
         """Serve a single media file previously signed via
         :meth:`_sign_media_path`. Validates the signature, decodes the
@@ -945,14 +1070,25 @@ class WebSocketChannel(BaseChannel):
         Returns ``(paths, None)`` on success or ``([], reason)`` on the first
         failure — the caller is expected to surface ``reason`` to the client
         and skip publishing so no half-formed message ever reaches the agent.
-        On failure, any images already written to disk earlier in the same
+        On failure, any files already written to disk earlier in the same
         call are unlinked so partial ingress doesn't leak orphan files.
         ``reason`` is a short, stable token suitable for UI localization.
 
         Shape: ``list[{"data_url": str, "name"?: str | None}]``.
         """
-        if len(media) > _MAX_IMAGES_PER_MESSAGE:
+        image_count = 0
+        video_count = 0
+        for item in media:
+            mime = _extract_data_url_mime(item.get("data_url", "")) if isinstance(item, dict) else None
+            if mime in _VIDEO_MIME_ALLOWED:
+                video_count += 1
+            elif mime in _IMAGE_MIME_ALLOWED:
+                image_count += 1
+        if image_count > _MAX_IMAGES_PER_MESSAGE:
             return [], "too_many_images"
+        if video_count > _MAX_VIDEOS_PER_MESSAGE:
+            return [], "too_many_videos"
+
         media_dir = get_media_dir("websocket")
         paths: list[str] = []
 
@@ -975,11 +1111,13 @@ class WebSocketChannel(BaseChannel):
             mime = _extract_data_url_mime(data_url)
             if mime is None:
                 return _abort("decode")
-            if mime not in _IMAGE_MIME_ALLOWED:
+            if mime not in _UPLOAD_MIME_ALLOWED:
                 return _abort("mime")
+            is_video = mime in _VIDEO_MIME_ALLOWED
+            max_bytes = _MAX_VIDEO_BYTES if is_video else _MAX_IMAGE_BYTES
             try:
                 saved = save_base64_data_url(
-                    data_url, media_dir, max_bytes=_MAX_IMAGE_BYTES,
+                    data_url, media_dir, max_bytes=max_bytes,
                 )
             except FileSizeExceeded:
                 return _abort("size")
@@ -1091,13 +1229,26 @@ class WebSocketChannel(BaseChannel):
         if not conns:
             logger.warning("websocket: no active subscribers for chat_id={}", msg.chat_id)
             return
+        text = msg.content
+        if msg.buttons:
+            text = _append_buttons_as_text(text, msg.buttons)
         payload: dict[str, Any] = {
             "event": "message",
             "chat_id": msg.chat_id,
-            "text": msg.content,
+            "text": text,
         }
+        if msg.buttons:
+            payload["buttons"] = msg.buttons
+            payload["button_prompt"] = msg.content
         if msg.media:
             payload["media"] = msg.media
+            urls: list[dict[str, str]] = []
+            for entry in msg.media:
+                signed = self._sign_or_stage_media_path(Path(entry))
+                if signed is not None:
+                    urls.append(signed)
+            if urls:
+                payload["media_urls"] = urls
         if msg.reply_to:
             payload["reply_to"] = msg.reply_to
         # Mark intermediate agent breadcrumbs (tool-call hints, generic

nanobot/cli/commands.py

@@ -212,12 +212,16 @@ async def _print_interactive_response(
 
 def _print_cli_progress_line(text: str, thinking: ThinkingSpinner | None) -> None:
     """Print a CLI progress line, pausing the spinner if needed."""
+    if not text.strip():
+        return
     with thinking.pause() if thinking else nullcontext():
         console.print(f"  [dim]↳ {text}[/dim]")
 
 
 async def _print_interactive_progress_line(text: str, thinking: ThinkingSpinner | None) -> None:
     """Print an interactive progress line, pausing the spinner if needed."""
+    if not text.strip():
+        return
     with thinking.pause() if thinking else nullcontext():
         await _print_interactive_line(text)
 
@@ -408,73 +412,13 @@ def _make_provider(config: Config):
 
     Routing is driven by ``ProviderSpec.backend`` in the registry.
     """
-    from nanobot.providers.base import GenerationSettings
-    from nanobot.providers.registry import find_by_name
+    from nanobot.providers.factory import make_provider
 
-    model = config.agents.defaults.model
-    provider_name = config.get_provider_name(model)
-    p = config.get_provider(model)
-    spec = find_by_name(provider_name) if provider_name else None
-    backend = spec.backend if spec else "openai_compat"
-
-    # --- validation ---
-    if backend == "azure_openai":
-        if not p or not p.api_key or not p.api_base:
-            console.print("[red]Error: Azure OpenAI requires api_key and api_base.[/red]")
-            console.print("Set them in ~/.nanobot/config.json under providers.azure_openai section")
-            console.print("Use the model field to specify the deployment name.")
-            raise typer.Exit(1)
-    elif backend == "openai_compat" and not model.startswith("bedrock/"):
-        needs_key = not (p and p.api_key)
-        exempt = spec and (spec.is_oauth or spec.is_local or spec.is_direct)
-        if needs_key and not exempt:
-            console.print("[red]Error: No API key configured.[/red]")
-            console.print("Set one in ~/.nanobot/config.json under providers section")
-            raise typer.Exit(1)
-
-    # --- instantiation by backend ---
-    if backend == "openai_codex":
-        from nanobot.providers.openai_codex_provider import OpenAICodexProvider
-
-        provider = OpenAICodexProvider(default_model=model)
-    elif backend == "azure_openai":
-        from nanobot.providers.azure_openai_provider import AzureOpenAIProvider
-
-        provider = AzureOpenAIProvider(
-            api_key=p.api_key,
-            api_base=p.api_base,
-            default_model=model,
-        )
-    elif backend == "github_copilot":
-        from nanobot.providers.github_copilot_provider import GitHubCopilotProvider
-        provider = GitHubCopilotProvider(default_model=model)
-    elif backend == "anthropic":
-        from nanobot.providers.anthropic_provider import AnthropicProvider
-
-        provider = AnthropicProvider(
-            api_key=p.api_key if p else None,
-            api_base=config.get_api_base(model),
-            default_model=model,
-            extra_headers=p.extra_headers if p else None,
-        )
-    else:
-        from nanobot.providers.openai_compat_provider import OpenAICompatProvider
-
-        provider = OpenAICompatProvider(
-            api_key=p.api_key if p else None,
-            api_base=config.get_api_base(model),
-            default_model=model,
-            extra_headers=p.extra_headers if p else None,
-            spec=spec,
-        )
-
-    defaults = config.agents.defaults
-    provider.generation = GenerationSettings(
-        temperature=defaults.temperature,
-        max_tokens=defaults.max_tokens,
-        reasoning_effort=defaults.reasoning_effort,
-    )
-    return provider
+    try:
+        return make_provider(config)
+    except ValueError as exc:
+        console.print(f"[red]Error: {exc}[/red]")
+        raise typer.Exit(1) from exc
 
 
 def _load_runtime_config(config: str | None = None, workspace: str | None = None) -> Config:
@@ -593,6 +537,7 @@ def serve(
         unified_session=runtime_config.agents.defaults.unified_session,
         disabled_skills=runtime_config.agents.defaults.disabled_skills,
         session_ttl_minutes=runtime_config.agents.defaults.session_ttl_minutes,
+        consolidation_ratio=runtime_config.agents.defaults.consolidation_ratio,
         tools_config=runtime_config.tools,
     )
 
@@ -652,11 +597,14 @@ def _run_gateway(
 ) -> None:
     """Shared gateway runtime; ``open_browser_url`` opens a tab once channels are up."""
     from nanobot.agent.loop import AgentLoop
+    from nanobot.agent.tools.cron import CronTool
+    from nanobot.agent.tools.message import MessageTool
     from nanobot.bus.queue import MessageBus
     from nanobot.channels.manager import ChannelManager
     from nanobot.cron.service import CronService
     from nanobot.cron.types import CronJob
     from nanobot.heartbeat.service import HeartbeatService
+    from nanobot.providers.factory import build_provider_snapshot, load_provider_snapshot
     from nanobot.session.manager import SessionManager
 
     port = port if port is not None else config.gateway.port
@@ -664,7 +612,12 @@ def _run_gateway(
     console.print(f"{__logo__} Starting nanobot gateway version {__version__} on port {port}...")
     sync_workspace_templates(config.workspace_path)
     bus = MessageBus()
-    provider = _make_provider(config)
+    try:
+        provider_snapshot = build_provider_snapshot(config)
+    except ValueError as exc:
+        console.print(f"[red]Error: {exc}[/red]")
+        raise typer.Exit(1) from exc
+    provider = provider_snapshot.provider
     session_manager = SessionManager(config.workspace_path)
 
     # Preserve existing single-workspace installs, but keep custom workspaces clean.
@@ -680,9 +633,9 @@ def _run_gateway(
         bus=bus,
         provider=provider,
         workspace=config.workspace_path,
-        model=config.agents.defaults.model,
+        model=provider_snapshot.model,
         max_iterations=config.agents.defaults.max_tool_iterations,
-        context_window_tokens=config.agents.defaults.context_window_tokens,
+        context_window_tokens=provider_snapshot.context_window_tokens,
         web_config=config.tools.web,
         context_block_limit=config.agents.defaults.context_block_limit,
         max_tool_result_chars=config.agents.defaults.max_tool_result_chars,
@@ -697,9 +650,55 @@ def _run_gateway(
         unified_session=config.agents.defaults.unified_session,
         disabled_skills=config.agents.defaults.disabled_skills,
         session_ttl_minutes=config.agents.defaults.session_ttl_minutes,
+        consolidation_ratio=config.agents.defaults.consolidation_ratio,
         tools_config=config.tools,
+        provider_snapshot_loader=load_provider_snapshot,
+        provider_signature=provider_snapshot.signature,
     )
 
+    from nanobot.agent.loop import UNIFIED_SESSION_KEY
+    from nanobot.bus.events import OutboundMessage
+
+    def _channel_session_key(channel: str, chat_id: str) -> str:
+        return (
+            UNIFIED_SESSION_KEY
+            if config.agents.defaults.unified_session
+            else f"{channel}:{chat_id}"
+        )
+
+    async def _deliver_to_channel(
+        msg: OutboundMessage, *, record: bool = False, session_key: str | None = None,
+    ) -> None:
+        """Publish a user-visible message and mirror it into that channel's session."""
+        metadata = dict(msg.metadata or {})
+        record = record or bool(metadata.pop("_record_channel_delivery", False))
+        if metadata != (msg.metadata or {}):
+            msg = OutboundMessage(
+                channel=msg.channel,
+                chat_id=msg.chat_id,
+                content=msg.content,
+                reply_to=msg.reply_to,
+                media=msg.media,
+                metadata=metadata,
+                buttons=msg.buttons,
+            )
+        if (
+            record
+            and msg.channel != "cli"
+            and msg.content.strip()
+            and hasattr(session_manager, "get_or_create")
+            and hasattr(session_manager, "save")
+        ):
+            key = session_key or _channel_session_key(msg.channel, msg.chat_id)
+            session = session_manager.get_or_create(key)
+            session.add_message("assistant", msg.content, _channel_delivery=True)
+            session_manager.save(session)
+        await bus.publish_outbound(msg)
+
+    message_tool = getattr(agent, "tools", {}).get("message")
+    if isinstance(message_tool, MessageTool):
+        message_tool.set_send_callback(_deliver_to_channel)
+
     # Set cron callback (needs agent)
     async def on_cron_job(job: CronJob) -> str | None:
         """Execute a cron job through the agent."""
@@ -712,14 +711,14 @@ def _run_gateway(
                 logger.exception("Dream cron job failed")
             return None
 
-        from nanobot.agent.tools.cron import CronTool
-        from nanobot.agent.tools.message import MessageTool
         from nanobot.utils.evaluator import evaluate_response
 
         reminder_note = (
-            "[Scheduled Task] Timer finished.\n\n"
-            f"Task '{job.name}' has been triggered.\n"
-            f"Scheduled instruction: {job.payload.message}"
+            "The scheduled time has arrived. Deliver this reminder to the user now, "
+            "as a brief and natural message in their language. Speak directly to them — "
+            "do not narrate progress, summarize, include user IDs, or add status reports "
+            "like 'Done' or 'Reminded'.\n\n"
+            f"Reminder: {job.payload.message}"
         )
 
         cron_tool = agent.tools.get("cron")
@@ -730,6 +729,10 @@ def _run_gateway(
         async def _silent(*_args, **_kwargs):
             pass
 
+        message_record_token = None
+        if isinstance(message_tool, MessageTool):
+            message_record_token = message_tool.set_record_channel_delivery(True)
+
         try:
             resp = await agent.process_direct(
                 reminder_note,
@@ -741,10 +744,11 @@ def _run_gateway(
         finally:
             if isinstance(cron_tool, CronTool) and cron_token is not None:
                 cron_tool.reset_cron_context(cron_token)
+            if isinstance(message_tool, MessageTool) and message_record_token is not None:
+                message_tool.reset_record_channel_delivery(message_record_token)
 
         response = resp.content if resp else ""
 
-        message_tool = agent.tools.get("message")
         if job.payload.deliver and isinstance(message_tool, MessageTool) and message_tool._sent_in_turn:
             return response
 
@@ -753,12 +757,16 @@ def _run_gateway(
                 response, reminder_note, provider, agent.model,
             )
             if should_notify:
-                from nanobot.bus.events import OutboundMessage
-                await bus.publish_outbound(OutboundMessage(
-                    channel=job.payload.channel or "cli",
-                    chat_id=job.payload.to,
-                    content=response,
-                ))
+                await _deliver_to_channel(
+                    OutboundMessage(
+                        channel=job.payload.channel or "cli",
+                        chat_id=job.payload.to,
+                        content=response,
+                        metadata=dict(job.payload.channel_meta),
+                    ),
+                    record=True,
+                    session_key=job.payload.session_key,
+                )
         return response
 
     cron.on_job = on_cron_job
@@ -808,12 +816,22 @@ def _run_gateway(
         return resp.content if resp else ""
 
     async def on_heartbeat_notify(response: str) -> None:
-        """Deliver a heartbeat response to the user's channel."""
-        from nanobot.bus.events import OutboundMessage
+        """Deliver a heartbeat response to the user's channel.
+
+        In addition to publishing the outbound message, this injects the
+        delivered text as an assistant turn into the *target channel's*
+        session.  Without this, a user reply on the channel (e.g. "Sure")
+        lands in a session that has no context about the heartbeat message
+        and the agent cannot follow through.
+        """
         channel, chat_id = _pick_heartbeat_target()
         if channel == "cli":
             return  # No external channel available to deliver to
-        await bus.publish_outbound(OutboundMessage(channel=channel, chat_id=chat_id, content=response))
+
+        await _deliver_to_channel(
+            OutboundMessage(channel=channel, chat_id=chat_id, content=response),
+            record=True,
+        )
 
     hb_cfg = config.gateway.heartbeat
     heartbeat = HeartbeatService(
@@ -1016,6 +1034,7 @@ def agent(
         unified_session=config.agents.defaults.unified_session,
         disabled_skills=config.agents.defaults.disabled_skills,
         session_ttl_minutes=config.agents.defaults.session_ttl_minutes,
+        consolidation_ratio=config.agents.defaults.consolidation_ratio,
         tools_config=config.tools,
     )
     restart_notice = consume_restart_notice_from_env()
@@ -1028,7 +1047,7 @@ def agent(
     # Shared reference for progress callbacks
     _thinking: ThinkingSpinner | None = None
 
-    async def _cli_progress(content: str, *, tool_hint: bool = False) -> None:
+    async def _cli_progress(content: str, *, tool_hint: bool = False, **_kwargs: Any) -> None:
         ch = agent_loop.channels_config
         if ch and tool_hint and not ch.send_tool_hints:
             return

nanobot/command/builtin.py

@@ -28,7 +28,11 @@ async def cmd_stop(ctx: CommandContext) -> OutboundMessage:
 async def cmd_restart(ctx: CommandContext) -> OutboundMessage:
     """Restart the process in-place via os.execv."""
     msg = ctx.msg
-    set_restart_notice_to_env(channel=msg.channel, chat_id=msg.chat_id)
+    set_restart_notice_to_env(
+        channel=msg.channel,
+        chat_id=msg.chat_id,
+        metadata=dict(msg.metadata or {}),
+    )
 
     async def _do_restart():
         await asyncio.sleep(1)

nanobot/config/schema.py

@@ -90,6 +90,13 @@ class AgentDefaults(Base):
         validation_alias=AliasChoices("idleCompactAfterMinutes", "sessionTtlMinutes"),
         serialization_alias="idleCompactAfterMinutes",
     )  # Auto-compact idle threshold in minutes (0 = disabled)
+    consolidation_ratio: float = Field(
+        default=0.5,
+        ge=0.1,
+        le=0.95,
+        validation_alias=AliasChoices("consolidationRatio"),
+        serialization_alias="consolidationRatio",
+    )  # Consolidation target ratio (0.5 = 50% of budget retained after compression)
     dream: DreamConfig = Field(default_factory=DreamConfig)
 
 

nanobot/cron/service.py

@@ -109,6 +109,12 @@ class CronService:
                             deliver=j["payload"].get("deliver", False),
                             channel=j["payload"].get("channel"),
                             to=j["payload"].get("to"),
+                            channel_meta=(
+                                j["payload"].get("channelMeta")
+                                or j["payload"].get("channel_meta")
+                                or {}
+                            ),
+                            session_key=j["payload"].get("sessionKey") or j["payload"].get("session_key"),
                         ),
                         state=CronJobState(
                             next_run_at_ms=j.get("state", {}).get("nextRunAtMs"),
@@ -210,6 +216,8 @@ class CronService:
                         "deliver": j.payload.deliver,
                         "channel": j.payload.channel,
                         "to": j.payload.to,
+                        "channelMeta": j.payload.channel_meta,
+                        "sessionKey": j.payload.session_key,
                     },
                     "state": {
                         "nextRunAtMs": j.state.next_run_at_ms,
@@ -379,6 +387,8 @@ class CronService:
         channel: str | None = None,
         to: str | None = None,
         delete_after_run: bool = False,
+        channel_meta: dict | None = None,
+        session_key: str | None = None,
     ) -> CronJob:
         """Add a new job."""
         _validate_schedule_for_add(schedule)
@@ -395,6 +405,8 @@ class CronService:
                 deliver=deliver,
                 channel=channel,
                 to=to,
+                channel_meta=channel_meta or {},
+                session_key=session_key,
             ),
             state=CronJobState(next_run_at_ms=_compute_next_run(schedule, now)),
             created_at_ms=now,

nanobot/cron/types.py

@@ -27,6 +27,8 @@ class CronPayload:
     deliver: bool = False
     channel: str | None = None  # e.g. "whatsapp"
     to: str | None = None  # e.g. phone number
+    channel_meta: dict = field(default_factory=dict)  # channel-specific routing (e.g. Slack thread_ts)
+    session_key: str | None = None  # original session key for correct session recording
 
 
 @dataclass

nanobot/nanobot.py

@@ -84,6 +84,7 @@ class Nanobot:
             unified_session=defaults.unified_session,
             disabled_skills=defaults.disabled_skills,
             session_ttl_minutes=defaults.session_ttl_minutes,
+            consolidation_ratio=defaults.consolidation_ratio,
             tools_config=config.tools,
         )
         return cls(loop)
@@ -119,62 +120,6 @@ class Nanobot:
 
 def _make_provider(config: Any) -> Any:
     """Create the LLM provider from config (extracted from CLI)."""
-    from nanobot.providers.base import GenerationSettings
-    from nanobot.providers.registry import find_by_name
+    from nanobot.providers.factory import make_provider
 
-    model = config.agents.defaults.model
-    provider_name = config.get_provider_name(model)
-    p = config.get_provider(model)
-    spec = find_by_name(provider_name) if provider_name else None
-    backend = spec.backend if spec else "openai_compat"
-
-    if backend == "azure_openai":
-        if not p or not p.api_key or not p.api_base:
-            raise ValueError("Azure OpenAI requires api_key and api_base in config.")
-    elif backend == "openai_compat" and not model.startswith("bedrock/"):
-        needs_key = not (p and p.api_key)
-        exempt = spec and (spec.is_oauth or spec.is_local or spec.is_direct)
-        if needs_key and not exempt:
-            raise ValueError(f"No API key configured for provider '{provider_name}'.")
-
-    if backend == "openai_codex":
-        from nanobot.providers.openai_codex_provider import OpenAICodexProvider
-
-        provider = OpenAICodexProvider(default_model=model)
-    elif backend == "github_copilot":
-        from nanobot.providers.github_copilot_provider import GitHubCopilotProvider
-
-        provider = GitHubCopilotProvider(default_model=model)
-    elif backend == "azure_openai":
-        from nanobot.providers.azure_openai_provider import AzureOpenAIProvider
-
-        provider = AzureOpenAIProvider(
-            api_key=p.api_key, api_base=p.api_base, default_model=model
-        )
-    elif backend == "anthropic":
-        from nanobot.providers.anthropic_provider import AnthropicProvider
-
-        provider = AnthropicProvider(
-            api_key=p.api_key if p else None,
-            api_base=config.get_api_base(model),
-            default_model=model,
-            extra_headers=p.extra_headers if p else None,
-        )
-    else:
-        from nanobot.providers.openai_compat_provider import OpenAICompatProvider
-
-        provider = OpenAICompatProvider(
-            api_key=p.api_key if p else None,
-            api_base=config.get_api_base(model),
-            default_model=model,
-            extra_headers=p.extra_headers if p else None,
-            spec=spec,
-        )
-
-    defaults = config.agents.defaults
-    provider.generation = GenerationSettings(
-        temperature=defaults.temperature,
-        max_tokens=defaults.max_tokens,
-        reasoning_effort=defaults.reasoning_effort,
-    )
-    return provider
+    return make_provider(config)

nanobot/providers/anthropic_provider.py

@@ -436,6 +436,10 @@ class AnthropicProvider(LLMProvider):
         max_tokens = max(1, max_tokens)
         thinking_enabled = bool(reasoning_effort)
 
+        # claude-opus-4-7 deprecated the `temperature` parameter entirely — the
+        # API returns 400 if it is present, on any code path.
+        omit_temperature = "opus-4-7" in model_name
+
         kwargs: dict[str, Any] = {
             "model": model_name,
             "messages": anthropic_msgs,
@@ -450,14 +454,16 @@ class AnthropicProvider(LLMProvider):
             # Supported on claude-sonnet-4-6 and claude-opus-4-6.
             # Also auto-enables interleaved thinking between tool calls.
             kwargs["thinking"] = {"type": "adaptive"}
-            kwargs["temperature"] = 1.0
+            if not omit_temperature:
+                kwargs["temperature"] = 1.0
         elif thinking_enabled:
             budget_map = {"low": 1024, "medium": 4096, "high": max(8192, max_tokens)}
             budget = budget_map.get(reasoning_effort.lower(), 4096)
             kwargs["thinking"] = {"type": "enabled", "budget_tokens": budget}
             kwargs["max_tokens"] = max(max_tokens, budget + 4096)
-            kwargs["temperature"] = 1.0
-        else:
+            if not omit_temperature:
+                kwargs["temperature"] = 1.0
+        elif not omit_temperature:
             kwargs["temperature"] = temperature
 
         if anthropic_tools:

nanobot/providers/factory.py

@@ -0,0 +1,112 @@
+"""Create LLM providers from config."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from pathlib import Path
+
+from nanobot.config.schema import Config
+from nanobot.providers.base import GenerationSettings, LLMProvider
+from nanobot.providers.registry import find_by_name
+
+
+@dataclass(frozen=True)
+class ProviderSnapshot:
+    provider: LLMProvider
+    model: str
+    context_window_tokens: int
+    signature: tuple[object, ...]
+
+
+def make_provider(config: Config) -> LLMProvider:
+    """Create the LLM provider implied by config."""
+    model = config.agents.defaults.model
+    provider_name = config.get_provider_name(model)
+    p = config.get_provider(model)
+    spec = find_by_name(provider_name) if provider_name else None
+    backend = spec.backend if spec else "openai_compat"
+
+    if backend == "azure_openai":
+        if not p or not p.api_key or not p.api_base:
+            raise ValueError("Azure OpenAI requires api_key and api_base in config.")
+    elif backend == "openai_compat" and not model.startswith("bedrock/"):
+        needs_key = not (p and p.api_key)
+        exempt = spec and (spec.is_oauth or spec.is_local or spec.is_direct)
+        if needs_key and not exempt:
+            raise ValueError(f"No API key configured for provider '{provider_name}'.")
+
+    if backend == "openai_codex":
+        from nanobot.providers.openai_codex_provider import OpenAICodexProvider
+
+        provider = OpenAICodexProvider(default_model=model)
+    elif backend == "azure_openai":
+        from nanobot.providers.azure_openai_provider import AzureOpenAIProvider
+
+        provider = AzureOpenAIProvider(
+            api_key=p.api_key,
+            api_base=p.api_base,
+            default_model=model,
+        )
+    elif backend == "github_copilot":
+        from nanobot.providers.github_copilot_provider import GitHubCopilotProvider
+
+        provider = GitHubCopilotProvider(default_model=model)
+    elif backend == "anthropic":
+        from nanobot.providers.anthropic_provider import AnthropicProvider
+
+        provider = AnthropicProvider(
+            api_key=p.api_key if p else None,
+            api_base=config.get_api_base(model),
+            default_model=model,
+            extra_headers=p.extra_headers if p else None,
+        )
+    else:
+        from nanobot.providers.openai_compat_provider import OpenAICompatProvider
+
+        provider = OpenAICompatProvider(
+            api_key=p.api_key if p else None,
+            api_base=config.get_api_base(model),
+            default_model=model,
+            extra_headers=p.extra_headers if p else None,
+            spec=spec,
+        )
+
+    defaults = config.agents.defaults
+    provider.generation = GenerationSettings(
+        temperature=defaults.temperature,
+        max_tokens=defaults.max_tokens,
+        reasoning_effort=defaults.reasoning_effort,
+    )
+    return provider
+
+
+def provider_signature(config: Config) -> tuple[object, ...]:
+    """Return the config fields that affect the primary LLM provider."""
+    model = config.agents.defaults.model
+    defaults = config.agents.defaults
+    return (
+        model,
+        defaults.provider,
+        config.get_provider_name(model),
+        config.get_api_key(model),
+        config.get_api_base(model),
+        defaults.max_tokens,
+        defaults.temperature,
+        defaults.reasoning_effort,
+        defaults.context_window_tokens,
+    )
+
+
+def build_provider_snapshot(config: Config) -> ProviderSnapshot:
+    return ProviderSnapshot(
+        provider=make_provider(config),
+        model=config.agents.defaults.model,
+        context_window_tokens=config.agents.defaults.context_window_tokens,
+        signature=provider_signature(config),
+    )
+
+
+def load_provider_snapshot(config_path: Path | None = None) -> ProviderSnapshot:
+    from nanobot.config.loader import load_config, resolve_config_env_vars
+
+    return build_provider_snapshot(resolve_config_env_vars(load_config(config_path)))

nanobot/providers/openai_compat_provider.py

@@ -3,17 +3,20 @@
 from __future__ import annotations
 
 import asyncio
-import json
 import hashlib
 import importlib.util
+import json
 import os
 import secrets
 import string
 import time
 import uuid
 from collections.abc import Awaitable, Callable
+from ipaddress import ip_address
 from typing import TYPE_CHECKING, Any
+from urllib.parse import urlparse
 
+import httpx
 import json_repair
 from loguru import logger
 
@@ -58,6 +61,15 @@ _KIMI_THINKING_MODELS: frozenset[str] = frozenset({
     "k2.6-code-preview",
 })
 
+# Maps ProviderSpec.thinking_style → extra_body builder.
+# Each builder takes a bool (thinking_enabled) and returns the dict to
+# merge into extra_body, keeping the style→wire-format mapping in one place.
+_THINKING_STYLE_MAP: dict[str, Any] = {
+    "thinking_type": lambda on: {"thinking": {"type": "enabled" if on else "disabled"}},
+    "enable_thinking": lambda on: {"enable_thinking": on},
+    "reasoning_split": lambda on: {"reasoning_split": on},
+}
+
 
 def _is_kimi_thinking_model(model_name: str) -> bool:
     """Return True if model_name refers to a Kimi thinking-capable model.
@@ -150,6 +162,37 @@ _RESPONSES_FAILURE_THRESHOLD = 3
 _RESPONSES_PROBE_INTERVAL_S = 300  # 5 minutes
 
 
+def _is_local_endpoint(
+    spec: "ProviderSpec | None",
+    api_base: str | None,
+) -> bool:
+    """Return True when the endpoint is a local or LAN model server.
+
+    Matches either the provider spec's ``is_local`` flag or common private-
+    network patterns in the base URL (localhost, 127.x, 192.168.x, 10.x,
+    172.16-31.x, Docker ``host.docker.internal``).
+    """
+    if spec and spec.is_local:
+        return True
+    if not api_base:
+        return False
+    raw = api_base.strip().lower()
+    parsed = urlparse(raw if "://" in raw else f"//{raw}")
+    try:
+        host = parsed.hostname
+    except ValueError:
+        return False
+    if host in {"localhost", "host.docker.internal"}:
+        return True
+    if not host:
+        return False
+    try:
+        addr = ip_address(host)
+    except ValueError:
+        return False
+    return addr.is_loopback or addr.is_private
+
+
 def _is_direct_openai_base(api_base: str | None) -> bool:
     """Return True for direct OpenAI endpoints, not generic OpenAI-compatible gateways."""
     if not api_base:
@@ -199,11 +242,27 @@ class OpenAICompatProvider(LLMProvider):
         if extra_headers:
             default_headers.update(extra_headers)
 
+        # Local model servers (Ollama, llama.cpp, vLLM) often close idle
+        # HTTP connections before the client-side keepalive expires.  When
+        # two LLM calls happen seconds apart (e.g. heartbeat _decide then
+        # process_direct), the second call may grab a now-dead pooled
+        # connection, causing a transient APIConnectionError on every first
+        # attempt.  Disabling keepalive for local endpoints avoids this by
+        # opening a fresh connection for each request, which is cheap on a
+        # LAN.  Cloud providers benefit from keepalive, so we leave the
+        # default pool settings for them.
+        http_client: httpx.AsyncClient | None = None
+        if _is_local_endpoint(spec, effective_base):
+            http_client = httpx.AsyncClient(
+                limits=httpx.Limits(keepalive_expiry=0),
+            )
+
         self._client = AsyncOpenAI(
             api_key=api_key or "no-key",
             base_url=effective_base,
             default_headers=default_headers,
             max_retries=0,
+            http_client=http_client,
         )
 
         # Responses API circuit breaker: skip after repeated failures,
@@ -286,10 +345,25 @@ class OpenAICompatProvider(LLMProvider):
             return json.dumps(arguments, ensure_ascii=False)
         return "{}"
 
+    @staticmethod
+    def _coerce_content_to_string(content: Any) -> str | None:
+        """Coerce block/list content into plain text for strict string-only APIs."""
+        if content is None or isinstance(content, str):
+            return content
+        text = OpenAICompatProvider._extract_text_content(content)
+        if isinstance(text, str) and text:
+            return text
+        try:
+            dumped = json.dumps(content, ensure_ascii=False)
+        except Exception:
+            dumped = str(content)
+        return dumped or "(empty)"
+
     def _sanitize_messages(self, messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
         """Strip non-standard keys, normalize tool_call IDs."""
         sanitized = LLMProvider._sanitize_request_messages(messages, _ALLOWED_MSG_KEYS)
         id_map: dict[str, str] = {}
+        force_string_content = bool(self._spec and self._spec.name == "deepseek")
 
         def map_id(value: Any) -> Any:
             if not isinstance(value, str):
@@ -323,8 +397,54 @@ class OpenAICompatProvider(LLMProvider):
                     clean["content"] = None
             if "tool_call_id" in clean and clean["tool_call_id"]:
                 clean["tool_call_id"] = map_id(clean["tool_call_id"])
+            if (
+                force_string_content
+                and not (clean.get("role") == "assistant" and clean.get("tool_calls"))
+            ):
+                clean["content"] = self._coerce_content_to_string(clean.get("content"))
         return self._enforce_role_alternation(sanitized)
 
+    def _drop_deepseek_incomplete_reasoning_history(
+        self,
+        messages: list[dict[str, Any]],
+        reasoning_effort: str | None,
+    ) -> list[dict[str, Any]]:
+        if (
+            not self._spec
+            or self._spec.name != "deepseek"
+            or not reasoning_effort
+            or reasoning_effort.lower() == "none"
+        ):
+            return messages
+
+        bad_idx = None
+        for idx, msg in enumerate(messages):
+            if (
+                msg.get("role") == "assistant"
+                and msg.get("tool_calls")
+                and not msg.get("reasoning_content")
+            ):
+                bad_idx = idx
+        if bad_idx is None:
+            return messages
+
+        keep_from = None
+        for idx in range(bad_idx + 1, len(messages)):
+            if messages[idx].get("role") == "user":
+                keep_from = idx
+                break
+
+        if keep_from is None:
+            trimmed = messages[:bad_idx]
+        else:
+            prefix = [msg for msg in messages[:keep_from] if msg.get("role") == "system"]
+            trimmed = prefix + messages[keep_from:]
+        logger.warning(
+            "Dropped {} DeepSeek thinking history message(s) with incomplete reasoning_content",
+            len(messages) - len(trimmed),
+        )
+        return trimmed
+
     # ------------------------------------------------------------------
     # Build kwargs
     # ------------------------------------------------------------------
@@ -365,6 +485,10 @@ class OpenAICompatProvider(LLMProvider):
         if spec and spec.strip_model_prefix:
             model_name = model_name.split("/")[-1]
 
+        messages = self._drop_deepseek_incomplete_reasoning_history(
+            messages,
+            reasoning_effort,
+        )
         kwargs: dict[str, Any] = {
             "model": model_name,
             "messages": self._sanitize_messages(self._sanitize_empty_content(messages)),
@@ -407,20 +531,11 @@ class OpenAICompatProvider(LLMProvider):
         # Provider-specific thinking parameters.
         # Only sent when reasoning_effort is explicitly configured so that
         # the provider default is preserved otherwise.
-        if spec and reasoning_effort is not None:
+        # The mapping is driven by ProviderSpec.thinking_style so that adding
+        # a new provider never requires touching this function.
+        if spec and spec.thinking_style and reasoning_effort is not None:
             thinking_enabled = semantic_effort != "minimal"
-            extra: dict[str, Any] | None = None
-            if spec.name == "dashscope":
-                extra = {"enable_thinking": thinking_enabled}
-            elif spec.name == "minimax":
-                extra = {"reasoning_split": thinking_enabled}
-            elif spec.name in (
-                "volcengine", "volcengine_coding_plan",
-                "byteplus", "byteplus_coding_plan",
-            ):
-                extra = {
-                    "thinking": {"type": "enabled" if thinking_enabled else "disabled"}
-                }
+            extra = _THINKING_STYLE_MAP.get(spec.thinking_style, lambda _: None)(thinking_enabled)
             if extra:
                 kwargs.setdefault("extra_body", {}).update(extra)
 
@@ -438,6 +553,26 @@ class OpenAICompatProvider(LLMProvider):
             kwargs["tools"] = tools
             kwargs["tool_choice"] = tool_choice or "auto"
 
+        # Backfill reasoning_content on legacy assistant messages.
+        # DeepSeek V4 (and potentially others) rejects thinking-mode
+        # requests that contain assistant messages without reasoning_content
+        # — even on turns that had no tool calls. This happens when a
+        # session was started with a non-thinking model or without
+        # reasoning_effort, then the user switches thinking mode on
+        # mid-session. Injecting an empty string satisfies the API
+        # without altering semantics (the model treats it as "no
+        # thinking happened on that turn").
+        thinking_active = (
+            (spec and spec.thinking_style and reasoning_effort is not None
+             and semantic_effort != "minimal")
+            or (reasoning_effort is not None and _is_kimi_thinking_model(model_name)
+                and semantic_effort != "minimal")
+        )
+        if thinking_active:
+            for msg in kwargs["messages"]:
+                if msg.get("role") == "assistant" and "reasoning_content" not in msg:
+                    msg["reasoning_content"] = ""
+
         return kwargs
 
     def _should_use_responses_api(
@@ -689,8 +824,8 @@ class OpenAICompatProvider(LLMProvider):
             finish_reason = str(choice0.get("finish_reason") or "stop")
 
             raw_tool_calls: list[Any] = []
-            # StepFun Plan: fallback to reasoning field when content is empty
-            if not content and msg0.get("reasoning"):
+            # StepFun: fallback to reasoning field when content is empty
+            if not content and msg0.get("reasoning") and self._spec and self._spec.reasoning_as_content:
                 content = self._extract_text_content(msg0.get("reasoning"))
             reasoning_content = msg0.get("reasoning_content")
             if not reasoning_content and msg0.get("reasoning"):
@@ -750,7 +885,7 @@ class OpenAICompatProvider(LLMProvider):
                     finish_reason = ch.finish_reason
             if not content and m.content:
                 content = m.content
-            if not content and getattr(m, "reasoning", None):
+            if not content and getattr(m, "reasoning", None) and self._spec and self._spec.reasoning_as_content:
                 content = m.reasoning
 
         tool_calls = []

nanobot/providers/registry.py

@@ -63,6 +63,19 @@ class ProviderSpec:
     # Provider supports cache_control on content blocks (e.g. Anthropic prompt caching)
     supports_prompt_caching: bool = False
 
+    # How to inject the thinking on/off toggle into extra_body.
+    # ""              — no extra_body needed (default)
+    # "thinking_type" — {"thinking": {"type": "enabled"/"disabled"}}
+    #                   (DeepSeek, VolcEngine, BytePlus)
+    # "enable_thinking" — {"enable_thinking": true/false}  (DashScope)
+    # "reasoning_split" — {"reasoning_split": true/false}  (MiniMax)
+    thinking_style: str = ""
+
+    # When True, treat the "reasoning" response field as formal content
+    # when "content" is empty.  Only set this for providers (e.g. StepFun)
+    # whose API returns the actual answer in "reasoning" instead of "content".
+    reasoning_as_content: bool = False
+
     @property
     def label(self) -> str:
         return self.display_name or self.name.title()
@@ -143,6 +156,7 @@ PROVIDERS: tuple[ProviderSpec, ...] = (
         is_gateway=True,
         detect_by_base_keyword="volces",
         default_api_base="https://ark.cn-beijing.volces.com/api/v3",
+        thinking_style="thinking_type",
     ),
 
     # VolcEngine Coding Plan (火山引擎 Coding Plan): same key as volcengine
@@ -155,6 +169,7 @@ PROVIDERS: tuple[ProviderSpec, ...] = (
         is_gateway=True,
         default_api_base="https://ark.cn-beijing.volces.com/api/coding/v3",
         strip_model_prefix=True,
+        thinking_style="thinking_type",
     ),
 
     # BytePlus: VolcEngine international, pay-per-use models
@@ -168,6 +183,7 @@ PROVIDERS: tuple[ProviderSpec, ...] = (
         detect_by_base_keyword="bytepluses",
         default_api_base="https://ark.ap-southeast.bytepluses.com/api/v3",
         strip_model_prefix=True,
+        thinking_style="thinking_type",
     ),
 
     # BytePlus Coding Plan: same key as byteplus
@@ -180,6 +196,7 @@ PROVIDERS: tuple[ProviderSpec, ...] = (
         is_gateway=True,
         default_api_base="https://ark.ap-southeast.bytepluses.com/api/coding/v3",
         strip_model_prefix=True,
+        thinking_style="thinking_type",
     ),
 
 
@@ -233,6 +250,7 @@ PROVIDERS: tuple[ProviderSpec, ...] = (
         display_name="DeepSeek",
         backend="openai_compat",
         default_api_base="https://api.deepseek.com",
+        thinking_style="thinking_type",
     ),
     # Gemini: Google's OpenAI-compatible endpoint
     ProviderSpec(
@@ -261,6 +279,7 @@ PROVIDERS: tuple[ProviderSpec, ...] = (
         display_name="DashScope",
         backend="openai_compat",
         default_api_base="https://dashscope.aliyuncs.com/compatible-mode/v1",
+        thinking_style="enable_thinking",
     ),
     # Moonshot (月之暗面): Kimi K2.5 / K2.6 enforce temperature >= 1.0.
     ProviderSpec(
@@ -283,6 +302,7 @@ PROVIDERS: tuple[ProviderSpec, ...] = (
         display_name="MiniMax",
         backend="openai_compat",
         default_api_base="https://api.minimax.io/v1",
+        thinking_style="reasoning_split",
     ),
     # MiniMax Anthropic-compatible endpoint: supports thinking mode
     ProviderSpec(
@@ -310,6 +330,7 @@ PROVIDERS: tuple[ProviderSpec, ...] = (
         display_name="Step Fun",
         backend="openai_compat",
         default_api_base="https://api.stepfun.com/v1",
+        reasoning_as_content=True,
     ),
     # Xiaomi MIMO (小米): OpenAI-compatible API
     ProviderSpec(

nanobot/session/manager.py

@@ -12,6 +12,7 @@ from loguru import logger
 
 from nanobot.config.paths import get_legacy_sessions_dir
 from nanobot.utils.helpers import (
+    estimate_message_tokens,
     ensure_dir,
     find_legal_message_start,
     image_placeholder_text,
@@ -19,6 +20,10 @@ from nanobot.utils.helpers import (
 )
 
 
+HISTORY_MAX_MESSAGES = 120
+FILE_MAX_MESSAGES = 2000
+
+
 @dataclass
 class Session:
     """A conversation session."""
@@ -30,6 +35,32 @@ class Session:
     metadata: dict[str, Any] = field(default_factory=dict)
     last_consolidated: int = 0  # Number of messages already consolidated to files
 
+    @staticmethod
+    def _annotate_message_time(message: dict[str, Any], content: Any) -> Any:
+        """Expose persisted turn timestamps to the model for relative-date reasoning.
+
+        Annotating *every* assistant turn trains the model (via in-context
+        demonstrations) to start its own replies with the same
+        ``[Message Time: ...]`` prefix, which leaks metadata back to the user.
+        We therefore only annotate:
+
+        * ``user`` turns — needed so the model can pin the conversation in time.
+        * proactive deliveries (``_channel_delivery=True``) — cron / heartbeat
+          assistant pushes that may sit hours away from the next user reply,
+          and are too infrequent to act as parroting demonstrations.
+        """
+        timestamp = message.get("timestamp")
+        if not timestamp or not isinstance(content, str):
+            return content
+        role = message.get("role")
+        if role == "user":
+            pass
+        elif role == "assistant" and message.get("_channel_delivery"):
+            pass
+        else:
+            return content
+        return f"[Message Time: {timestamp}]\n{content}"
+
     def add_message(self, role: str, content: str, **kwargs: Any) -> None:
         """Add a message to the session."""
         msg = {
@@ -41,15 +72,29 @@ class Session:
         self.messages.append(msg)
         self.updated_at = datetime.now()
 
-    def get_history(self, max_messages: int = 500) -> list[dict[str, Any]]:
-        """Return unconsolidated messages for LLM input, aligned to a legal tool-call boundary."""
+    def get_history(
+        self,
+        max_messages: int = HISTORY_MAX_MESSAGES,
+        *,
+        max_tokens: int = 0,
+        include_timestamps: bool = False,
+    ) -> list[dict[str, Any]]:
+        """Return unconsolidated messages for LLM input.
+
+        History is sliced by message count first (``max_messages``), then by
+        token budget from the tail (``max_tokens``) when provided.
+        """
         unconsolidated = self.messages[self.last_consolidated:]
         sliced = unconsolidated[-max_messages:]
 
-        # Avoid starting mid-turn when possible.
+        # Avoid starting mid-turn when possible, except for proactive
+        # assistant deliveries that the user may be replying to.
         for i, message in enumerate(sliced):
             if message.get("role") == "user":
-                sliced = sliced[i:]
+                start = i
+                if i > 0 and sliced[i - 1].get("_channel_delivery"):
+                    start = i - 1
+                sliced = sliced[start:]
                 break
 
         # Drop orphan tool results at the front.
@@ -71,11 +116,45 @@ class Session:
                     image_placeholder_text(p) for p in media if isinstance(p, str) and p
                 )
                 content = f"{content}\n{breadcrumbs}" if content else breadcrumbs
+            if include_timestamps:
+                content = self._annotate_message_time(message, content)
             entry: dict[str, Any] = {"role": message["role"], "content": content}
             for key in ("tool_calls", "tool_call_id", "name", "reasoning_content"):
                 if key in message:
                     entry[key] = message[key]
             out.append(entry)
+
+        if max_tokens > 0 and out:
+            kept: list[dict[str, Any]] = []
+            used = 0
+            for message in reversed(out):
+                tokens = estimate_message_tokens(message)
+                if kept and used + tokens > max_tokens:
+                    break
+                kept.append(message)
+                used += tokens
+            kept.reverse()
+
+            # Keep history aligned to the first visible user turn.
+            first_user = next((i for i, m in enumerate(kept) if m.get("role") == "user"), None)
+            if first_user is not None:
+                kept = kept[first_user:]
+            else:
+                # Tight token budgets can otherwise leave assistant-only tails.
+                # If a user turn exists in the unsliced output, recover the
+                # nearest one even if it slightly exceeds the token budget.
+                recovered_user = next(
+                    (i for i in range(len(out) - 1, -1, -1) if out[i].get("role") == "user"),
+                    None,
+                )
+                if recovered_user is not None:
+                    kept = out[recovered_user:]
+
+            # And keep a legal tool-call boundary at the front.
+            start = find_legal_message_start(kept)
+            if start:
+                kept = kept[start:]
+            out = kept
         return out
 
     def clear(self) -> None:
@@ -85,31 +164,77 @@ class Session:
         self.updated_at = datetime.now()
 
     def retain_recent_legal_suffix(self, max_messages: int) -> None:
-        """Keep a legal recent suffix, mirroring get_history boundary rules."""
+        """Keep a legal recent suffix constrained by a hard message cap."""
         if max_messages <= 0:
             self.clear()
             return
         if len(self.messages) <= max_messages:
             return
 
-        start_idx = max(0, len(self.messages) - max_messages)
+        retained = list(self.messages[-max_messages:])
 
-        # If the cutoff lands mid-turn, extend backward to the nearest user turn.
-        while start_idx > 0 and self.messages[start_idx].get("role") != "user":
-            start_idx -= 1
-
-        retained = self.messages[start_idx:]
+        # Prefer starting at a user turn when one exists within the tail.
+        first_user = next((i for i, m in enumerate(retained) if m.get("role") == "user"), None)
+        if first_user is not None:
+            retained = retained[first_user:]
+        else:
+            # If the tail is assistant/tool-only, anchor to the latest user in
+            # the full session and take a capped forward window from there.
+            latest_user = next(
+                (i for i in range(len(self.messages) - 1, -1, -1)
+                 if self.messages[i].get("role") == "user"),
+                None,
+            )
+            if latest_user is not None:
+                retained = list(self.messages[latest_user: latest_user + max_messages])
 
         # Mirror get_history(): avoid persisting orphan tool results at the front.
         start = find_legal_message_start(retained)
         if start:
             retained = retained[start:]
 
+        # Hard-cap guarantee: never keep more than max_messages.
+        if len(retained) > max_messages:
+            retained = retained[-max_messages:]
+            start = find_legal_message_start(retained)
+            if start:
+                retained = retained[start:]
+
         dropped = len(self.messages) - len(retained)
         self.messages = retained
         self.last_consolidated = max(0, self.last_consolidated - dropped)
         self.updated_at = datetime.now()
 
+    def enforce_file_cap(
+        self,
+        on_archive: Any = None,
+        limit: int = FILE_MAX_MESSAGES,
+    ) -> None:
+        """Bound session message growth by archiving and trimming old prefixes."""
+        if limit <= 0 or len(self.messages) <= limit:
+            return
+
+        before = list(self.messages)
+        before_last_consolidated = self.last_consolidated
+        before_count = len(before)
+        self.retain_recent_legal_suffix(limit)
+        dropped_count = before_count - len(self.messages)
+        if dropped_count <= 0:
+            return
+
+        dropped = before[:dropped_count]
+        already_consolidated = min(before_last_consolidated, dropped_count)
+        archive_chunk = dropped[already_consolidated:]
+        if archive_chunk and on_archive:
+            on_archive(archive_chunk)
+        logger.info(
+            "Session file cap hit for {}: dropped {}, raw-archived {}, kept {}",
+            self.key,
+            dropped_count,
+            len(archive_chunk),
+            len(self.messages),
+        )
+
 
 class SessionManager:
     """

nanobot/templates/agent/identity.md

@@ -29,4 +29,4 @@ Output is rendered in a terminal. Avoid markdown headings and tables. Use plain
 {% include 'agent/_snippets/untrusted_content.md' %}
 
 Reply directly with text for conversations. Only use the 'message' tool to send to a specific chat channel.
-IMPORTANT: To send files (images, documents, audio, video) to the user, you MUST call the 'message' tool with the 'media' parameter. Do NOT use read_file to "send" a file — reading a file only shows its content to you, it does NOT deliver the file to the user. Example: message(content="Here is the file", media=["/path/to/file.png"])
+IMPORTANT: To send files (images, video, audio, documents) to the user, you MUST call the 'message' tool with the 'media' parameter. Do NOT use read_file to "send" a file — reading a file only shows its content to you, it does NOT deliver the file to the user. Examples: message(content="Here is the image", media=["/path/to/file.png"]) or message(content="Here is the video", media=["/path/to/video.mp4"])

nanobot/utils/document.py

@@ -7,26 +7,6 @@ from loguru import logger
 
 from nanobot.utils.helpers import detect_image_mime
 
-try:
-    from pypdf import PdfReader
-except ImportError:
-    PdfReader = None  # type: ignore
-
-try:
-    from docx import Document as DocxDocument
-except ImportError:
-    DocxDocument = None  # type: ignore
-
-try:
-    from openpyxl import load_workbook
-except ImportError:
-    load_workbook = None  # type: ignore
-
-try:
-    from pptx import Presentation as PptxPresentation
-except ImportError:
-    PptxPresentation = None  # type: ignore
-
 
 # Supported file extensions for text extraction
 SUPPORTED_EXTENSIONS: set[str] = {
@@ -78,22 +58,16 @@ def extract_text(path: Path) -> str | None:
 
     ext = path.suffix.lower()
 
-    # Document formats
+    # Document formats -- each branch lazily imports its parser so that
+    # startup does not pay the ~25 MB cost of loading openpyxl /
+    # python-docx / python-pptx / pypdf up front (see issue #3422).
     if ext == ".pdf":
-        if PdfReader is None:
-            return "[error: pypdf not installed]"
         return _extract_pdf(path)
     elif ext == ".docx":
-        if DocxDocument is None:
-            return "[error: python-docx not installed]"
         return _extract_docx(path)
     elif ext == ".xlsx":
-        if load_workbook is None:
-            return "[error: openpyxl not installed]"
         return _extract_xlsx(path)
     elif ext == ".pptx":
-        if PptxPresentation is None:
-            return "[error: python-pptx not installed]"
         return _extract_pptx(path)
     elif _is_text_extension(ext):
         return _extract_text_file(path)
@@ -107,6 +81,10 @@ def extract_text(path: Path) -> str | None:
 
 def _extract_pdf(path: Path) -> str:
     """Extract text from PDF using pypdf."""
+    try:
+        from pypdf import PdfReader
+    except ImportError:
+        return "[error: pypdf not installed]"
     try:
         reader = PdfReader(path)
         pages: list[str] = []
@@ -121,6 +99,10 @@ def _extract_pdf(path: Path) -> str:
 
 def _extract_docx(path: Path) -> str:
     """Extract text from DOCX using python-docx."""
+    try:
+        from docx import Document as DocxDocument
+    except ImportError:
+        return "[error: python-docx not installed]"
     try:
         doc = DocxDocument(path)
         paragraphs: list[str] = [p.text for p in doc.paragraphs if p.text.strip()]
@@ -132,6 +114,10 @@ def _extract_docx(path: Path) -> str:
 
 def _extract_xlsx(path: Path) -> str:
     """Extract text from XLSX using openpyxl."""
+    try:
+        from openpyxl import load_workbook
+    except ImportError:
+        return "[error: openpyxl not installed]"
     try:
         wb = load_workbook(path, read_only=True, data_only=True)
         try:
@@ -155,6 +141,10 @@ def _extract_xlsx(path: Path) -> str:
 
 def _extract_pptx(path: Path) -> str:
     """Extract text from PPTX using python-pptx."""
+    try:
+        from pptx import Presentation as PptxPresentation
+    except ImportError:
+        return "[error: python-pptx not installed]"
     try:
         prs = PptxPresentation(path)
         slides: list[str] = []

nanobot/utils/progress_events.py

@@ -0,0 +1,84 @@
+"""Structured progress-event helpers shared by agent runtimes."""
+
+from __future__ import annotations
+
+import inspect
+from collections.abc import Awaitable, Callable
+from typing import Any
+
+from nanobot.agent.hook import AgentHookContext
+
+
+def on_progress_accepts_tool_events(cb: Callable[..., Any]) -> bool:
+    try:
+        sig = inspect.signature(cb)
+    except (TypeError, ValueError):
+        return False
+    if any(p.kind == inspect.Parameter.VAR_KEYWORD for p in sig.parameters.values()):
+        return True
+    return "tool_events" in sig.parameters
+
+
+async def invoke_on_progress(
+    on_progress: Callable[..., Awaitable[None]],
+    content: str,
+    *,
+    tool_hint: bool = False,
+    tool_events: list[dict[str, Any]] | None = None,
+) -> None:
+    if tool_events and on_progress_accepts_tool_events(on_progress):
+        await on_progress(content, tool_hint=tool_hint, tool_events=tool_events)
+        return
+    await on_progress(content, tool_hint=tool_hint)
+
+
+def build_tool_event_start_payload(tool_call: Any) -> dict[str, Any]:
+    return {
+        "version": 1,
+        "phase": "start",
+        "call_id": str(getattr(tool_call, "id", "") or ""),
+        "name": getattr(tool_call, "name", ""),
+        "arguments": getattr(tool_call, "arguments", {}) or {},
+        "result": None,
+        "error": None,
+        "files": [],
+        "embeds": [],
+    }
+
+
+def tool_event_result_extras(result: Any) -> tuple[list[Any], list[Any]]:
+    if not isinstance(result, dict):
+        return [], []
+    files = result.get("files") if isinstance(result.get("files"), list) else []
+    embeds = result.get("embeds") if isinstance(result.get("embeds"), list) else []
+    return files, embeds
+
+
+def build_tool_event_finish_payloads(context: AgentHookContext) -> list[dict[str, Any]]:
+    payloads: list[dict[str, Any]] = []
+    count = min(len(context.tool_calls), len(context.tool_results), len(context.tool_events))
+    for idx in range(count):
+        tool_call = context.tool_calls[idx]
+        result = context.tool_results[idx]
+        event = context.tool_events[idx] if isinstance(context.tool_events[idx], dict) else {}
+        status = event.get("status")
+        phase = "end" if status == "ok" else "error"
+        files, embeds = tool_event_result_extras(result)
+        payload = {
+            "version": 1,
+            "phase": phase,
+            "call_id": str(getattr(tool_call, "id", "") or ""),
+            "name": getattr(tool_call, "name", ""),
+            "arguments": getattr(tool_call, "arguments", {}) or {},
+            "result": result if phase == "end" else None,
+            "error": None,
+            "files": files,
+            "embeds": embeds,
+        }
+        if phase == "error":
+            if isinstance(result, str) and result.strip():
+                payload["error"] = result.strip()
+            else:
+                payload["error"] = str(event.get("detail") or "Tool execution failed")
+        payloads.append(payload)
+    return payloads

nanobot/utils/restart.py

@@ -2,12 +2,15 @@
 
 from __future__ import annotations
 
+import json
 import os
 import time
-from dataclasses import dataclass
+from dataclasses import dataclass, field
+from typing import Any
 
 RESTART_NOTIFY_CHANNEL_ENV = "NANOBOT_RESTART_NOTIFY_CHANNEL"
 RESTART_NOTIFY_CHAT_ID_ENV = "NANOBOT_RESTART_NOTIFY_CHAT_ID"
+RESTART_NOTIFY_METADATA_ENV = "NANOBOT_RESTART_NOTIFY_METADATA"
 RESTART_STARTED_AT_ENV = "NANOBOT_RESTART_STARTED_AT"
 
 
@@ -16,6 +19,7 @@ class RestartNotice:
     channel: str
     chat_id: str
     started_at_raw: str
+    metadata: dict[str, Any] = field(default_factory=dict)
 
 
 def format_restart_completed_message(started_at_raw: str) -> str:
@@ -30,11 +34,20 @@ def format_restart_completed_message(started_at_raw: str) -> str:
     return f"Restart completed{elapsed_suffix}."
 
 
-def set_restart_notice_to_env(*, channel: str, chat_id: str) -> None:
+def set_restart_notice_to_env(
+    *, channel: str, chat_id: str, metadata: dict[str, Any] | None = None,
+) -> None:
     """Write restart notice env values for the next process."""
     os.environ[RESTART_NOTIFY_CHANNEL_ENV] = channel
     os.environ[RESTART_NOTIFY_CHAT_ID_ENV] = chat_id
     os.environ[RESTART_STARTED_AT_ENV] = str(time.time())
+    if metadata:
+        try:
+            os.environ[RESTART_NOTIFY_METADATA_ENV] = json.dumps(metadata, default=str)
+        except (TypeError, ValueError):
+            os.environ.pop(RESTART_NOTIFY_METADATA_ENV, None)
+    else:
+        os.environ.pop(RESTART_NOTIFY_METADATA_ENV, None)
 
 
 def consume_restart_notice_from_env() -> RestartNotice | None:
@@ -42,9 +55,23 @@ def consume_restart_notice_from_env() -> RestartNotice | None:
     channel = os.environ.pop(RESTART_NOTIFY_CHANNEL_ENV, "").strip()
     chat_id = os.environ.pop(RESTART_NOTIFY_CHAT_ID_ENV, "").strip()
     started_at_raw = os.environ.pop(RESTART_STARTED_AT_ENV, "").strip()
+    metadata_raw = os.environ.pop(RESTART_NOTIFY_METADATA_ENV, "").strip()
     if not (channel and chat_id):
         return None
-    return RestartNotice(channel=channel, chat_id=chat_id, started_at_raw=started_at_raw)
+    metadata: dict[str, Any] = {}
+    if metadata_raw:
+        try:
+            parsed = json.loads(metadata_raw)
+        except (TypeError, ValueError):
+            parsed = None
+        if isinstance(parsed, dict):
+            metadata = parsed
+    return RestartNotice(
+        channel=channel,
+        chat_id=chat_id,
+        started_at_raw=started_at_raw,
+        metadata=metadata,
+    )
 
 
 def should_show_cli_restart_notice(notice: RestartNotice, session_id: str) -> bool: