Merge commit '29c25b4ce5' into dev

refactor(memory): streamline memory extraction

surfsense_backend/.env.example

@@ -282,6 +282,9 @@ LANGSMITH_PROJECT=surfsense
 # =============================================================================
 # OPTIONAL: New-chat agent feature flags
 # =============================================================================
+# Multi-agent orchestrator switch for authenticated chat streaming.
+# MULTI_AGENT_CHAT_ENABLED=false
+
 # Master kill-switch — when true, every flag below is forced OFF.
 # SURFSENSE_DISABLE_NEW_AGENT_STACK=false
 

surfsense_backend/app/agents/multi_agent_chat/__init__.py

@@ -0,0 +1,7 @@
+"""Deepagents-backed routes: ``subagents/``; main-agent graph under ``main_agent/`` (SRP subpackages)."""
+
+from __future__ import annotations
+
+from .main_agent import create_surfsense_deep_agent
+
+__all__ = ["create_surfsense_deep_agent"]

surfsense_backend/app/agents/multi_agent_chat/constants.py

@@ -0,0 +1,43 @@
+"""Connector-type to subagent name; subagent name to availability tokens for build_subagents."""
+
+from __future__ import annotations
+
+CONNECTOR_TYPE_TO_CONNECTOR_AGENT_MAPS: dict[str, str] = {
+    "GOOGLE_GMAIL_CONNECTOR": "gmail",
+    "COMPOSIO_GMAIL_CONNECTOR": "gmail",
+    "GOOGLE_CALENDAR_CONNECTOR": "calendar",
+    "COMPOSIO_GOOGLE_CALENDAR_CONNECTOR": "calendar",
+    "DISCORD_CONNECTOR": "discord",
+    "TEAMS_CONNECTOR": "teams",
+    "LUMA_CONNECTOR": "luma",
+    "LINEAR_CONNECTOR": "linear",
+    "JIRA_CONNECTOR": "jira",
+    "CLICKUP_CONNECTOR": "clickup",
+    "SLACK_CONNECTOR": "slack",
+    "AIRTABLE_CONNECTOR": "airtable",
+    "NOTION_CONNECTOR": "notion",
+    "CONFLUENCE_CONNECTOR": "confluence",
+    "GOOGLE_DRIVE_CONNECTOR": "google_drive",
+    "COMPOSIO_GOOGLE_DRIVE_CONNECTOR": "google_drive",
+    "DROPBOX_CONNECTOR": "dropbox",
+    "ONEDRIVE_CONNECTOR": "onedrive",
+}
+
+SUBAGENT_TO_REQUIRED_CONNECTOR_MAP: dict[str, frozenset[str]] = {
+    "deliverables": frozenset(),
+    "airtable": frozenset({"AIRTABLE_CONNECTOR"}),
+    "calendar": frozenset({"GOOGLE_CALENDAR_CONNECTOR"}),
+    "clickup": frozenset({"CLICKUP_CONNECTOR"}),
+    "confluence": frozenset({"CONFLUENCE_CONNECTOR"}),
+    "discord": frozenset({"DISCORD_CONNECTOR"}),
+    "dropbox": frozenset({"DROPBOX_FILE"}),
+    "gmail": frozenset({"GOOGLE_GMAIL_CONNECTOR"}),
+    "google_drive": frozenset({"GOOGLE_DRIVE_FILE"}),
+    "jira": frozenset({"JIRA_CONNECTOR"}),
+    "linear": frozenset({"LINEAR_CONNECTOR"}),
+    "luma": frozenset({"LUMA_CONNECTOR"}),
+    "notion": frozenset({"NOTION_CONNECTOR"}),
+    "onedrive": frozenset({"ONEDRIVE_FILE"}),
+    "slack": frozenset({"SLACK_CONNECTOR"}),
+    "teams": frozenset({"TEAMS_CONNECTOR"}),
+}

surfsense_backend/app/agents/multi_agent_chat/main_agent/__init__.py

@@ -0,0 +1,7 @@
+"""Main-agent deep agent: ``runtime/`` (factory), ``graph/`` (compile), ``system_prompt/``, etc."""
+
+from __future__ import annotations
+
+from .runtime import create_surfsense_deep_agent
+
+__all__ = ["create_surfsense_deep_agent"]

surfsense_backend/app/agents/multi_agent_chat/main_agent/context_prune/__init__.py

@@ -0,0 +1,7 @@
+"""Tool-name pruning for context editing (exclude lists without dropping protected tools)."""
+
+from __future__ import annotations
+
+from .prune_tool_names import PRUNE_PROTECTED_TOOL_NAMES, safe_exclude_tools
+
+__all__ = ["PRUNE_PROTECTED_TOOL_NAMES", "safe_exclude_tools"]

surfsense_backend/app/agents/multi_agent_chat/main_agent/context_prune/prune_tool_names.py

@@ -0,0 +1,26 @@
+"""Tool names excluded from context-editing prune when bound."""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+
+from langchain_core.tools import BaseTool
+
+PRUNE_PROTECTED_TOOL_NAMES: frozenset[str] = frozenset(
+    {
+        "generate_report",
+        "generate_resume",
+        "generate_podcast",
+        "generate_video_presentation",
+        "generate_image",
+        "read_email",
+        "search_emails",
+        "invalid",
+    },
+)
+
+
+def safe_exclude_tools(tools: Sequence[BaseTool]) -> tuple[str, ...]:
+    """Names from ``PRUNE_PROTECTED_TOOL_NAMES`` that appear in ``tools``."""
+    enabled = {t.name for t in tools}
+    return tuple(n for n in PRUNE_PROTECTED_TOOL_NAMES if n in enabled)

surfsense_backend/app/agents/multi_agent_chat/main_agent/graph/__init__.py

@@ -0,0 +1,7 @@
+"""Sync compile of the main-agent LangGraph graph (middleware + ``create_agent``)."""
+
+from __future__ import annotations
+
+from .compile_graph_sync import build_compiled_agent_graph_sync
+
+__all__ = ["build_compiled_agent_graph_sync"]

surfsense_backend/app/agents/multi_agent_chat/main_agent/graph/compile_graph_sync.py

@@ -0,0 +1,85 @@
+"""Synchronous graph compile (middleware + ``create_agent``)."""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+from typing import Any
+
+from deepagents import __version__ as deepagents_version
+from langchain.agents import create_agent
+from langchain_core.language_models import BaseChatModel
+from langchain_core.tools import BaseTool
+from langgraph.types import Checkpointer
+
+from app.agents.multi_agent_chat.subagents.shared.permissions import (
+    ToolsPermissions,
+)
+from app.agents.new_chat.context import SurfSenseContextSchema
+from app.agents.new_chat.feature_flags import AgentFeatureFlags
+from app.agents.new_chat.filesystem_selection import FilesystemMode
+from app.db import ChatVisibility
+
+from .middleware import build_main_agent_deepagent_middleware
+
+
+def build_compiled_agent_graph_sync(
+    *,
+    llm: BaseChatModel,
+    tools: Sequence[BaseTool],
+    final_system_prompt: str,
+    backend_resolver: Any,
+    filesystem_mode: FilesystemMode,
+    search_space_id: int,
+    user_id: str | None,
+    thread_id: int | None,
+    visibility: ChatVisibility,
+    anon_session_id: str | None,
+    available_connectors: list[str] | None,
+    available_document_types: list[str] | None,
+    mentioned_document_ids: list[int] | None,
+    max_input_tokens: int | None,
+    flags: AgentFeatureFlags,
+    checkpointer: Checkpointer,
+    subagent_dependencies: dict[str, Any],
+    mcp_tools_by_agent: dict[str, ToolsPermissions] | None = None,
+    disabled_tools: list[str] | None = None,
+):
+    """Sync compile: middleware + ``create_agent`` (run via ``asyncio.to_thread``)."""
+    main_agent_middleware = build_main_agent_deepagent_middleware(
+        llm=llm,
+        tools=tools,
+        backend_resolver=backend_resolver,
+        filesystem_mode=filesystem_mode,
+        search_space_id=search_space_id,
+        user_id=user_id,
+        thread_id=thread_id,
+        visibility=visibility,
+        anon_session_id=anon_session_id,
+        available_connectors=available_connectors,
+        available_document_types=available_document_types,
+        mentioned_document_ids=mentioned_document_ids,
+        max_input_tokens=max_input_tokens,
+        flags=flags,
+        subagent_dependencies=subagent_dependencies,
+        checkpointer=checkpointer,
+        mcp_tools_by_agent=mcp_tools_by_agent,
+        disabled_tools=disabled_tools,
+    )
+
+    agent = create_agent(
+        llm,
+        system_prompt=final_system_prompt,
+        tools=list(tools),
+        middleware=main_agent_middleware,
+        context_schema=SurfSenseContextSchema,
+        checkpointer=checkpointer,
+    )
+    return agent.with_config(
+        {
+            "recursion_limit": 10_000,
+            "metadata": {
+                "ls_integration": "deepagents",
+                "versions": {"deepagents": deepagents_version},
+            },
+        }
+    )

surfsense_backend/app/agents/multi_agent_chat/main_agent/graph/middleware/__init__.py

@@ -0,0 +1,7 @@
+"""Main-agent graph middleware assembly (SurfSense + LangChain + deepagents)."""
+
+from __future__ import annotations
+
+from .deepagent_stack import build_main_agent_deepagent_middleware
+
+__all__ = ["build_main_agent_deepagent_middleware"]

surfsense_backend/app/agents/multi_agent_chat/main_agent/graph/middleware/checkpointed_subagent_middleware/__init__.py

@@ -0,0 +1,26 @@
+"""SubAgent ``task`` tool wiring required for HITL inside subagents.
+
+Replaces upstream ``SubAgentMiddleware`` to:
+
+- share the parent's checkpointer with each subagent,
+- forward ``runtime.config`` (thread_id, recursion_limit, …) into nested invokes,
+- bridge ``Command(resume=...)`` from the parent into the subagent via the
+  ``config["configurable"]["surfsense_resume_value"]`` side-channel,
+- target the resume at the captured interrupt id so a follow-up
+  ``HumanInTheLoopMiddleware.after_model`` does not consume the same payload,
+- re-raise any new subagent interrupt at the parent so the SSE stream surfaces it.
+
+Module layout
+-------------
+
+- ``constants``   — shared keys / limits.
+- ``config``      — RunnableConfig + side-channel resume read.
+- ``resume``      — pending-interrupt detection, fan-out, ``Command(resume=...)`` builder.
+- ``propagation`` — re-raise pending subagent interrupts at the parent.
+- ``task_tool``   — the ``task`` tool factory (sync + async).
+- ``middleware``  — :class:`SurfSenseCheckpointedSubAgentMiddleware` itself.
+"""
+
+from .middleware import SurfSenseCheckpointedSubAgentMiddleware
+
+__all__ = ["SurfSenseCheckpointedSubAgentMiddleware"]

surfsense_backend/app/agents/multi_agent_chat/main_agent/graph/middleware/checkpointed_subagent_middleware/config.py

@@ -0,0 +1,44 @@
+"""RunnableConfig wiring for nested subagent invocations.
+
+Forwards the parent's ``runtime.config`` (thread_id, …) into the subagent and
+exposes the side-channel ``stream_resume_chat`` uses to ferry resume payloads.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from langchain.tools import ToolRuntime
+
+from .constants import DEFAULT_SUBAGENT_RECURSION_LIMIT
+
+
+def subagent_invoke_config(runtime: ToolRuntime) -> dict[str, Any]:
+    """RunnableConfig for the nested invoke; raises ``recursion_limit`` to the parent's budget."""
+    merged: dict[str, Any] = dict(runtime.config) if runtime.config else {}
+    current_limit = merged.get("recursion_limit")
+    try:
+        current_int = int(current_limit) if current_limit is not None else 0
+    except (TypeError, ValueError):
+        current_int = 0
+    if current_int < DEFAULT_SUBAGENT_RECURSION_LIMIT:
+        merged["recursion_limit"] = DEFAULT_SUBAGENT_RECURSION_LIMIT
+    return merged
+
+
+def consume_surfsense_resume(runtime: ToolRuntime) -> Any:
+    """Pop the resume payload; siblings share ``configurable`` by reference."""
+    cfg = runtime.config or {}
+    configurable = cfg.get("configurable") if isinstance(cfg, dict) else None
+    if not isinstance(configurable, dict):
+        return None
+    return configurable.pop("surfsense_resume_value", None)
+
+
+def has_surfsense_resume(runtime: ToolRuntime) -> bool:
+    """True iff a resume payload is queued on this runtime (non-destructive)."""
+    cfg = runtime.config or {}
+    configurable = cfg.get("configurable") if isinstance(cfg, dict) else None
+    if not isinstance(configurable, dict):
+        return False
+    return "surfsense_resume_value" in configurable

surfsense_backend/app/agents/multi_agent_chat/main_agent/graph/middleware/checkpointed_subagent_middleware/constants.py

@@ -0,0 +1,18 @@
+"""Constants shared by the checkpointed subagent middleware."""
+
+from __future__ import annotations
+
+# Mirror of deepagents.middleware.subagents._EXCLUDED_STATE_KEYS.
+EXCLUDED_STATE_KEYS = frozenset(
+    {
+        "messages",
+        "todos",
+        "structured_response",
+        "skills_metadata",
+        "memory_contents",
+    }
+)
+
+# Match the parent graph's budget; the LangGraph default of 25 trips on
+# multi-step subagent runs.
+DEFAULT_SUBAGENT_RECURSION_LIMIT = 10_000

surfsense_backend/app/agents/multi_agent_chat/main_agent/graph/middleware/checkpointed_subagent_middleware/middleware.py

@@ -0,0 +1,103 @@
+"""SubAgent middleware that compiles each subagent against the parent checkpointer."""
+
+from __future__ import annotations
+
+from typing import Any, cast
+
+from deepagents.backends.protocol import BackendFactory, BackendProtocol
+from deepagents.middleware.subagents import (
+    TASK_SYSTEM_PROMPT,
+    CompiledSubAgent,
+    SubAgent,
+    SubAgentMiddleware,
+)
+from langchain.agents import create_agent
+from langchain.agents.middleware import HumanInTheLoopMiddleware
+from langchain.chat_models import init_chat_model
+from langgraph.types import Checkpointer
+
+from .task_tool import build_task_tool_with_parent_config
+
+
+class SurfSenseCheckpointedSubAgentMiddleware(SubAgentMiddleware):
+    """``SubAgentMiddleware`` variant that compiles each subagent against the parent checkpointer."""
+
+    def __init__(
+        self,
+        *,
+        checkpointer: Checkpointer,
+        backend: BackendProtocol | BackendFactory,
+        subagents: list[SubAgent | CompiledSubAgent],
+        system_prompt: str | None = TASK_SYSTEM_PROMPT,
+        task_description: str | None = None,
+    ) -> None:
+        self._surf_checkpointer = checkpointer
+        super(SubAgentMiddleware, self).__init__()
+        if not subagents:
+            raise ValueError(
+                "At least one subagent must be specified when using the new API"
+            )
+        self._backend = backend
+        self._subagents = subagents
+        subagent_specs = self._surf_compile_subagent_graphs()
+        task_tool = build_task_tool_with_parent_config(subagent_specs, task_description)
+        if system_prompt and subagent_specs:
+            agents_desc = "\n".join(
+                f"- {s['name']}: {s['description']}" for s in subagent_specs
+            )
+            self.system_prompt = (
+                system_prompt + "\n\nAvailable subagent types:\n" + agents_desc
+            )
+        else:
+            self.system_prompt = system_prompt
+        self.tools = [task_tool]
+
+    def _surf_compile_subagent_graphs(self) -> list[dict[str, Any]]:
+        """Mirror of ``SubAgentMiddleware._get_subagents`` that threads the parent checkpointer."""
+        specs: list[dict[str, Any]] = []
+
+        for spec in self._subagents:
+            if "runnable" in spec:
+                compiled = cast(CompiledSubAgent, spec)
+                specs.append(
+                    {
+                        "name": compiled["name"],
+                        "description": compiled["description"],
+                        "runnable": compiled["runnable"],
+                    }
+                )
+                continue
+
+            if "model" not in spec:
+                msg = f"SubAgent '{spec['name']}' must specify 'model'"
+                raise ValueError(msg)
+            if "tools" not in spec:
+                msg = f"SubAgent '{spec['name']}' must specify 'tools'"
+                raise ValueError(msg)
+
+            model = spec["model"]
+            if isinstance(model, str):
+                model = init_chat_model(model)
+
+            middleware: list[Any] = list(spec.get("middleware", []))
+
+            interrupt_on = spec.get("interrupt_on")
+            if interrupt_on:
+                middleware.append(HumanInTheLoopMiddleware(interrupt_on=interrupt_on))
+
+            specs.append(
+                {
+                    "name": spec["name"],
+                    "description": spec["description"],
+                    "runnable": create_agent(
+                        model,
+                        system_prompt=spec["system_prompt"],
+                        tools=spec["tools"],
+                        middleware=middleware,
+                        name=spec["name"],
+                        checkpointer=self._surf_checkpointer,
+                    ),
+                }
+            )
+
+        return specs

surfsense_backend/app/agents/multi_agent_chat/main_agent/graph/middleware/checkpointed_subagent_middleware/propagation.py

@@ -0,0 +1,74 @@
+"""Re-raise still-pending subagent interrupts at the parent graph level.
+
+After ``subagent.[a]invoke(Command(resume=...))`` returns, the subagent may
+still hold a pending interrupt (e.g. the LLM produced a follow-up tool call
+that fired a fresh ``interrupt()``). The parent's pregel cannot see that
+interrupt because it lives in a separate compiled graph; we re-raise it here
+so the parent's SSE stream surfaces it as the next approval card.
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import Any
+
+from langchain_core.runnables import Runnable
+from langgraph.types import interrupt as _lg_interrupt
+
+from .resume import get_first_pending_subagent_interrupt
+
+logger = logging.getLogger(__name__)
+
+
+def maybe_propagate_subagent_interrupt(
+    subagent: Runnable,
+    sub_config: dict[str, Any],
+    subagent_type: str,
+) -> None:
+    """Re-raise a still-pending subagent interrupt at the parent so the SSE stream surfaces it."""
+    get_state_sync = getattr(subagent, "get_state", None)
+    if not callable(get_state_sync):
+        return
+    try:
+        snapshot = get_state_sync(sub_config)
+    except Exception:  # pragma: no cover - defensive
+        logger.debug(
+            "Subagent get_state failed during re-interrupt check",
+            exc_info=True,
+        )
+        return
+    _pending_id, pending_value = get_first_pending_subagent_interrupt(snapshot)
+    if pending_value is None:
+        return
+    logger.info(
+        "Re-raising subagent %r interrupt to parent (multi-step HITL)",
+        subagent_type,
+    )
+    _lg_interrupt(pending_value)
+
+
+async def amaybe_propagate_subagent_interrupt(
+    subagent: Runnable,
+    sub_config: dict[str, Any],
+    subagent_type: str,
+) -> None:
+    """Async counterpart of :func:`maybe_propagate_subagent_interrupt`."""
+    aget_state = getattr(subagent, "aget_state", None)
+    if not callable(aget_state):
+        return
+    try:
+        snapshot = await aget_state(sub_config)
+    except Exception:  # pragma: no cover - defensive
+        logger.debug(
+            "Subagent aget_state failed during re-interrupt check",
+            exc_info=True,
+        )
+        return
+    _pending_id, pending_value = get_first_pending_subagent_interrupt(snapshot)
+    if pending_value is None:
+        return
+    logger.info(
+        "Re-raising subagent %r interrupt to parent (multi-step HITL)",
+        subagent_type,
+    )
+    _lg_interrupt(pending_value)

surfsense_backend/app/agents/multi_agent_chat/main_agent/graph/middleware/checkpointed_subagent_middleware/resume.py

@@ -0,0 +1,76 @@
+"""Resume-payload shaping and pending-interrupt detection for subagents.
+
+Splits the work of "given a state snapshot and a parent-stashed resume value,
+produce the right ``Command(resume=...)`` for the subagent" into pure helpers.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from langgraph.types import Command
+
+
+def hitlrequest_action_count(pending_value: Any) -> int:
+    """Bundle size for a LangChain ``HITLRequest`` payload; ``0`` for non-bundle interrupts."""
+    if not isinstance(pending_value, dict):
+        return 0
+    actions = pending_value.get("action_requests")
+    if isinstance(actions, list):
+        return len(actions)
+    return 0
+
+
+def fan_out_decisions_to_match(resume_value: Any, expected_count: int) -> Any:
+    """Legacy fallback: pad a 1-decision resume to N for an ``action_requests=N`` bundle.
+
+    Modern frontend submits N decisions per bundle (one per action_request) so
+    this is a no-op; kept for backwards compatibility with old in-flight
+    threads or non-bundle clients that send a single decision.
+    """
+    if expected_count <= 1:
+        return resume_value
+    if not isinstance(resume_value, dict):
+        return resume_value
+    decisions = resume_value.get("decisions")
+    if not isinstance(decisions, list) or len(decisions) >= expected_count:
+        return resume_value
+    if not decisions:
+        return resume_value
+    padded = list(decisions) + [decisions[-1]] * (expected_count - len(decisions))
+    return {**resume_value, "decisions": padded}
+
+
+def get_first_pending_subagent_interrupt(state: Any) -> tuple[str | None, Any]:
+    """First pending ``(interrupt_id, value)``; ``(None, None)`` if no interrupt.
+
+    Assumes at most one pending interrupt per snapshot (sequential tool nodes).
+    Parallel tool nodes would need an id-aware lookup instead of first-wins.
+    """
+    if state is None:
+        return None, None
+    for it in getattr(state, "interrupts", None) or ():
+        value = getattr(it, "value", None)
+        interrupt_id = getattr(it, "id", None)
+        if value is not None:
+            return (
+                interrupt_id if isinstance(interrupt_id, str) else None,
+                value,
+            )
+    for sub_task in getattr(state, "tasks", None) or ():
+        for it in getattr(sub_task, "interrupts", None) or ():
+            value = getattr(it, "value", None)
+            interrupt_id = getattr(it, "id", None)
+            if value is not None:
+                return (
+                    interrupt_id if isinstance(interrupt_id, str) else None,
+                    value,
+                )
+    return None, None
+
+
+def build_resume_command(resume_value: Any, pending_id: str | None) -> Command:
+    """``Command(resume={id: value})`` when ``id`` is known, else fall back to scalar."""
+    if pending_id is None:
+        return Command(resume=resume_value)
+    return Command(resume={pending_id: resume_value})

surfsense_backend/app/agents/multi_agent_chat/main_agent/graph/middleware/checkpointed_subagent_middleware/task_tool.py

@@ -0,0 +1,231 @@
+"""Build the ``task`` tool that invokes subagents with HITL bridging.
+
+The tool's body is the only place where the parent and the subagent meet at
+runtime: it reads the parent's stashed resume value, decides whether to send
+fresh state or a targeted ``Command(resume=...)`` to the subagent, then
+re-raises any new pending interrupt back to the parent.
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import Annotated, Any
+
+from deepagents.middleware.subagents import TASK_TOOL_DESCRIPTION
+from langchain.tools import BaseTool, ToolRuntime
+from langchain_core.messages import HumanMessage, ToolMessage
+from langchain_core.runnables import Runnable
+from langchain_core.tools import StructuredTool
+from langgraph.types import Command
+
+from .config import (
+    consume_surfsense_resume,
+    has_surfsense_resume,
+    subagent_invoke_config,
+)
+from .constants import EXCLUDED_STATE_KEYS
+from .propagation import (
+    amaybe_propagate_subagent_interrupt,
+    maybe_propagate_subagent_interrupt,
+)
+from .resume import (
+    build_resume_command,
+    fan_out_decisions_to_match,
+    get_first_pending_subagent_interrupt,
+    hitlrequest_action_count,
+)
+
+logger = logging.getLogger(__name__)
+
+
+def build_task_tool_with_parent_config(
+    subagents: list[dict[str, Any]],
+    task_description: str | None = None,
+) -> BaseTool:
+    """Upstream ``_build_task_tool`` + parent ``runtime.config`` propagation + resume bridging."""
+    subagent_graphs: dict[str, Runnable] = {
+        spec["name"]: spec["runnable"] for spec in subagents
+    }
+    subagent_description_str = "\n".join(
+        f"- {s['name']}: {s['description']}" for s in subagents
+    )
+
+    if task_description is None:
+        description = TASK_TOOL_DESCRIPTION.format(
+            available_agents=subagent_description_str
+        )
+    elif "{available_agents}" in task_description:
+        description = task_description.format(available_agents=subagent_description_str)
+    else:
+        description = task_description
+
+    def _return_command_with_state_update(result: dict, tool_call_id: str) -> Command:
+        if "messages" not in result:
+            msg = (
+                "CompiledSubAgent must return a state containing a 'messages' key. "
+                "Custom StateGraphs used with CompiledSubAgent should include 'messages' "
+                "in their state schema to communicate results back to the main agent."
+            )
+            raise ValueError(msg)
+
+        state_update = {k: v for k, v in result.items() if k not in EXCLUDED_STATE_KEYS}
+        message_text = (
+            result["messages"][-1].text.rstrip() if result["messages"][-1].text else ""
+        )
+        return Command(
+            update={
+                **state_update,
+                "messages": [ToolMessage(message_text, tool_call_id=tool_call_id)],
+            }
+        )
+
+    def _validate_and_prepare_state(
+        subagent_type: str, description: str, runtime: ToolRuntime
+    ) -> tuple[Runnable, dict]:
+        subagent = subagent_graphs[subagent_type]
+        subagent_state = {
+            k: v for k, v in runtime.state.items() if k not in EXCLUDED_STATE_KEYS
+        }
+        subagent_state["messages"] = [HumanMessage(content=description)]
+        return subagent, subagent_state
+
+    def task(
+        description: Annotated[
+            str,
+            "A detailed description of the task for the subagent to perform autonomously. Include all necessary context and specify the expected output format.",
+        ],
+        subagent_type: Annotated[
+            str,
+            "The type of subagent to use. Must be one of the available agent types listed in the tool description.",
+        ],
+        runtime: ToolRuntime,
+    ) -> str | Command:
+        if subagent_type not in subagent_graphs:
+            allowed_types = ", ".join([f"`{k}`" for k in subagent_graphs])
+            return (
+                f"We cannot invoke subagent {subagent_type} because it does not exist, "
+                f"the only allowed types are {allowed_types}"
+            )
+        if not runtime.tool_call_id:
+            raise ValueError("Tool call ID is required for subagent invocation")
+        subagent, subagent_state = _validate_and_prepare_state(
+            subagent_type, description, runtime
+        )
+        sub_config = subagent_invoke_config(runtime)
+
+        # Resume bridge: forward the parent's stashed decision into the
+        # subagent's pending ``interrupt()``, targeted by id.
+        pending_id: str | None = None
+        pending_value: Any = None
+        get_state = getattr(subagent, "get_state", None)
+        if callable(get_state):
+            try:
+                snapshot = get_state(sub_config)
+                pending_id, pending_value = get_first_pending_subagent_interrupt(
+                    snapshot
+                )
+            except Exception:
+                # Fail loud if a resume is queued: silent fallback would
+                # replay the original interrupt to the user.
+                if has_surfsense_resume(runtime):
+                    logger.exception(
+                        "Subagent %r get_state raised with resume queued; re-raising.",
+                        subagent_type,
+                    )
+                    raise
+                logger.debug(
+                    "Subagent get_state failed; falling back to fresh invoke",
+                    exc_info=True,
+                )
+
+        if pending_value is not None:
+            resume_value = consume_surfsense_resume(runtime)
+            if resume_value is None:
+                # Bridge invariant: a queued resume must accompany any pending
+                # subagent interrupt. Fall-through replay would silently re-prompt
+                # the user; raise so the streaming layer surfaces a clear error.
+                raise RuntimeError(
+                    f"Subagent {subagent_type!r} has a pending interrupt but no "
+                    "surfsense_resume_value on config; resume bridge is broken."
+                )
+            expected = hitlrequest_action_count(pending_value)
+            resume_value = fan_out_decisions_to_match(resume_value, expected)
+            result = subagent.invoke(
+                build_resume_command(resume_value, pending_id),
+                config=sub_config,
+            )
+        else:
+            result = subagent.invoke(subagent_state, config=sub_config)
+        maybe_propagate_subagent_interrupt(subagent, sub_config, subagent_type)
+        return _return_command_with_state_update(result, runtime.tool_call_id)
+
+    async def atask(
+        description: Annotated[
+            str,
+            "A detailed description of the task for the subagent to perform autonomously. Include all necessary context and specify the expected output format.",
+        ],
+        subagent_type: Annotated[
+            str,
+            "The type of subagent to use. Must be one of the available agent types listed in the tool description.",
+        ],
+        runtime: ToolRuntime,
+    ) -> str | Command:
+        if subagent_type not in subagent_graphs:
+            allowed_types = ", ".join([f"`{k}`" for k in subagent_graphs])
+            return (
+                f"We cannot invoke subagent {subagent_type} because it does not exist, "
+                f"the only allowed types are {allowed_types}"
+            )
+        if not runtime.tool_call_id:
+            raise ValueError("Tool call ID is required for subagent invocation")
+        subagent, subagent_state = _validate_and_prepare_state(
+            subagent_type, description, runtime
+        )
+        sub_config = subagent_invoke_config(runtime)
+
+        # Resume bridge — see ``task`` above.
+        pending_id: str | None = None
+        pending_value: Any = None
+        aget_state = getattr(subagent, "aget_state", None)
+        if callable(aget_state):
+            try:
+                snapshot = await aget_state(sub_config)
+                pending_id, pending_value = get_first_pending_subagent_interrupt(
+                    snapshot
+                )
+            except Exception:
+                if has_surfsense_resume(runtime):
+                    logger.exception(
+                        "Subagent %r aget_state raised with resume queued; re-raising.",
+                        subagent_type,
+                    )
+                    raise
+                logger.debug(
+                    "Subagent aget_state failed; falling back to fresh ainvoke",
+                    exc_info=True,
+                )
+
+        if pending_value is not None:
+            resume_value = consume_surfsense_resume(runtime)
+            if resume_value is None:
+                raise RuntimeError(
+                    f"Subagent {subagent_type!r} has a pending interrupt but no "
+                    "surfsense_resume_value on config; resume bridge is broken."
+                )
+            expected = hitlrequest_action_count(pending_value)
+            resume_value = fan_out_decisions_to_match(resume_value, expected)
+            result = await subagent.ainvoke(
+                build_resume_command(resume_value, pending_id),
+                config=sub_config,
+            )
+        else:
+            result = await subagent.ainvoke(subagent_state, config=sub_config)
+        await amaybe_propagate_subagent_interrupt(subagent, sub_config, subagent_type)
+        return _return_command_with_state_update(result, runtime.tool_call_id)
+
+    return StructuredTool.from_function(
+        name="task",
+        func=task,
+        coroutine=atask,
+        description=description,
+    )

surfsense_backend/app/agents/multi_agent_chat/main_agent/graph/middleware/deepagent_stack.py

@@ -0,0 +1,506 @@
+"""Assemble the main-agent deep-agent middleware list (LangChain + SurfSense + deepagents)."""
+
+from __future__ import annotations
+
+import logging
+from collections.abc import Sequence
+from typing import Any
+
+from deepagents import SubAgent
+from deepagents.backends import StateBackend
+from deepagents.middleware.patch_tool_calls import PatchToolCallsMiddleware
+from deepagents.middleware.skills import SkillsMiddleware
+from deepagents.middleware.subagents import GENERAL_PURPOSE_SUBAGENT
+from langchain.agents.middleware import (
+    LLMToolSelectorMiddleware,
+    ModelCallLimitMiddleware,
+    ModelFallbackMiddleware,
+    TodoListMiddleware,
+    ToolCallLimitMiddleware,
+)
+from langchain_anthropic.middleware import AnthropicPromptCachingMiddleware
+from langchain_core.language_models import BaseChatModel
+from langchain_core.tools import BaseTool
+from langgraph.types import Checkpointer
+
+from app.agents.multi_agent_chat.subagents import (
+    build_subagents,
+    get_subagents_to_exclude,
+)
+from app.agents.multi_agent_chat.subagents.shared.permissions import (
+    ToolsPermissions,
+)
+from app.agents.new_chat.feature_flags import AgentFeatureFlags
+from app.agents.new_chat.filesystem_selection import FilesystemMode
+from app.agents.new_chat.middleware import (
+    ActionLogMiddleware,
+    AnonymousDocumentMiddleware,
+    BusyMutexMiddleware,
+    ClearToolUsesEdit,
+    DedupHITLToolCallsMiddleware,
+    DoomLoopMiddleware,
+    FileIntentMiddleware,
+    KnowledgeBasePersistenceMiddleware,
+    KnowledgePriorityMiddleware,
+    KnowledgeTreeMiddleware,
+    MemoryInjectionMiddleware,
+    NoopInjectionMiddleware,
+    OtelSpanMiddleware,
+    PermissionMiddleware,
+    RetryAfterMiddleware,
+    SpillingContextEditingMiddleware,
+    SpillToBackendEdit,
+    SurfSenseFilesystemMiddleware,
+    ToolCallNameRepairMiddleware,
+    build_skills_backend_factory,
+    create_surfsense_compaction_middleware,
+    default_skills_sources,
+)
+from app.agents.new_chat.permissions import Rule, Ruleset
+from app.agents.new_chat.plugin_loader import (
+    PluginContext,
+    load_allowed_plugin_names_from_env,
+    load_plugin_middlewares,
+)
+from app.agents.new_chat.tools.registry import BUILTIN_TOOLS
+from app.db import ChatVisibility
+
+from ...context_prune.prune_tool_names import safe_exclude_tools
+from .checkpointed_subagent_middleware import SurfSenseCheckpointedSubAgentMiddleware
+
+
+def build_main_agent_deepagent_middleware(
+    *,
+    llm: BaseChatModel,
+    tools: Sequence[BaseTool],
+    backend_resolver: Any,
+    filesystem_mode: FilesystemMode,
+    search_space_id: int,
+    user_id: str | None,
+    thread_id: int | None,
+    visibility: ChatVisibility,
+    anon_session_id: str | None,
+    available_connectors: list[str] | None,
+    available_document_types: list[str] | None,
+    mentioned_document_ids: list[int] | None,
+    max_input_tokens: int | None,
+    flags: AgentFeatureFlags,
+    subagent_dependencies: dict[str, Any],
+    checkpointer: Checkpointer,
+    mcp_tools_by_agent: dict[str, ToolsPermissions] | None = None,
+    disabled_tools: list[str] | None = None,
+) -> list[Any]:
+    """Build ordered middleware for ``create_agent`` (Nones already stripped)."""
+    _memory_middleware = MemoryInjectionMiddleware(
+        user_id=user_id,
+        search_space_id=search_space_id,
+        thread_visibility=visibility,
+    )
+
+    gp_middleware = [
+        TodoListMiddleware(),
+        _memory_middleware,
+        FileIntentMiddleware(llm=llm),
+        SurfSenseFilesystemMiddleware(
+            backend=backend_resolver,
+            filesystem_mode=filesystem_mode,
+            search_space_id=search_space_id,
+            created_by_id=user_id,
+            thread_id=thread_id,
+        ),
+        create_surfsense_compaction_middleware(llm, StateBackend),
+        PatchToolCallsMiddleware(),
+        AnthropicPromptCachingMiddleware(unsupported_model_behavior="ignore"),
+    ]
+
+    # Build permission rulesets up front so the GP subagent can mirror ``ask``
+    # rules into ``interrupt_on``: tool calls emitted from within ``task`` runs
+    # never reach the parent's ``PermissionMiddleware``.
+    is_desktop_fs = filesystem_mode == FilesystemMode.DESKTOP_LOCAL_FOLDER
+    permission_enabled = flags.enable_permission and not flags.disable_new_agent_stack
+    permission_rulesets: list[Ruleset] = []
+    if permission_enabled or is_desktop_fs:
+        permission_rulesets.append(
+            Ruleset(
+                rules=[Rule(permission="*", pattern="*", action="allow")],
+                origin="surfsense_defaults",
+            )
+        )
+        if is_desktop_fs:
+            permission_rulesets.append(
+                Ruleset(
+                    rules=[
+                        Rule(permission="rm", pattern="*", action="ask"),
+                        Rule(permission="rmdir", pattern="*", action="ask"),
+                        Rule(permission="move_file", pattern="*", action="ask"),
+                        Rule(permission="edit_file", pattern="*", action="ask"),
+                        Rule(permission="write_file", pattern="*", action="ask"),
+                    ],
+                    origin="desktop_safety",
+                )
+            )
+
+    # Tools that self-prompt via ``request_approval`` must not also appear
+    # as ``ask`` rules — that would double-prompt the user for one call.
+    _tool_names_in_use = {t.name for t in tools}
+
+    # Deny parent-bound tools whose ``required_connector`` is missing.
+    # No-op today (connector subagents are pruned upstream); guards future
+    # additions to the parent's tool list.
+    if permission_enabled:
+        _available_set = set(available_connectors or [])
+        _synthesized: list[Rule] = []
+        for tool_def in BUILTIN_TOOLS:
+            if tool_def.name not in _tool_names_in_use:
+                continue
+            rc = tool_def.required_connector
+            if rc and rc not in _available_set:
+                _synthesized.append(
+                    Rule(permission=tool_def.name, pattern="*", action="deny")
+                )
+        if _synthesized:
+            permission_rulesets.append(
+                Ruleset(rules=_synthesized, origin="connector_synthesized")
+            )
+    gp_interrupt_on: dict[str, bool] = {
+        rule.permission: True
+        for rs in permission_rulesets
+        for rule in rs.rules
+        if rule.action == "ask" and rule.permission in _tool_names_in_use
+    }
+
+    general_purpose_spec: SubAgent = {  # type: ignore[typeddict-unknown-key]
+        **GENERAL_PURPOSE_SUBAGENT,
+        "model": llm,
+        "tools": tools,
+        "middleware": gp_middleware,
+    }
+    if gp_interrupt_on:
+        general_purpose_spec["interrupt_on"] = gp_interrupt_on
+
+    # Deny-only on subagents: ``task`` runs bypass the parent's
+    # PermissionMiddleware, while bucket-based ask gates own the ask path.
+    subagent_deny_rulesets: list[Ruleset] = [
+        Ruleset(
+            rules=[r for r in rs.rules if r.action == "deny"],
+            origin=rs.origin,
+        )
+        for rs in permission_rulesets
+    ]
+    subagent_deny_rulesets = [rs for rs in subagent_deny_rulesets if rs.rules]
+
+    subagent_deny_permission_mw: PermissionMiddleware | None = (
+        PermissionMiddleware(rulesets=subagent_deny_rulesets)
+        if subagent_deny_rulesets
+        else None
+    )
+
+    if subagent_deny_permission_mw is not None:
+        # Run deny check on already-repaired tool calls; insert before
+        # PatchToolCallsMiddleware (append if the slot moves).
+        _patch_idx = next(
+            (
+                i
+                for i, m in enumerate(gp_middleware)
+                if isinstance(m, PatchToolCallsMiddleware)
+            ),
+            len(gp_middleware),
+        )
+        gp_middleware.insert(_patch_idx, subagent_deny_permission_mw)
+
+    registry_subagents: list[SubAgent] = []
+    try:
+        subagent_extra_middleware: list[Any] = [
+            TodoListMiddleware(),
+            SurfSenseFilesystemMiddleware(
+                backend=backend_resolver,
+                filesystem_mode=filesystem_mode,
+                search_space_id=search_space_id,
+                created_by_id=user_id,
+                thread_id=thread_id,
+            ),
+        ]
+        if subagent_deny_permission_mw is not None:
+            subagent_extra_middleware.append(subagent_deny_permission_mw)
+        registry_subagents = build_subagents(
+            dependencies=subagent_dependencies,
+            model=llm,
+            extra_middleware=subagent_extra_middleware,
+            mcp_tools_by_agent=mcp_tools_by_agent or {},
+            exclude=get_subagents_to_exclude(available_connectors),
+            disabled_tools=disabled_tools,
+        )
+        logging.info(
+            "Registry subagents: %s",
+            [s["name"] for s in registry_subagents],
+        )
+    except Exception:
+        logging.exception("Registry subagent build failed")
+        raise
+
+    subagent_specs: list[SubAgent] = [general_purpose_spec, *registry_subagents]
+
+    summarization_mw = create_surfsense_compaction_middleware(llm, StateBackend)
+
+    context_edit_mw = None
+    if (
+        flags.enable_context_editing
+        and not flags.disable_new_agent_stack
+        and max_input_tokens
+    ):
+        spill_edit = SpillToBackendEdit(
+            trigger=int(max_input_tokens * 0.55),
+            clear_at_least=int(max_input_tokens * 0.15),
+            keep=5,
+            exclude_tools=safe_exclude_tools(tools),
+            clear_tool_inputs=True,
+        )
+        clear_edit = ClearToolUsesEdit(
+            trigger=int(max_input_tokens * 0.55),
+            clear_at_least=int(max_input_tokens * 0.15),
+            keep=5,
+            exclude_tools=safe_exclude_tools(tools),
+            clear_tool_inputs=True,
+            placeholder="[cleared - older tool output trimmed for context]",
+        )
+        context_edit_mw = SpillingContextEditingMiddleware(
+            edits=[spill_edit, clear_edit],
+            backend_resolver=backend_resolver,
+        )
+
+    retry_mw = (
+        RetryAfterMiddleware(max_retries=3)
+        if flags.enable_retry_after and not flags.disable_new_agent_stack
+        else None
+    )
+    fallback_mw: ModelFallbackMiddleware | None = None
+    if flags.enable_model_fallback and not flags.disable_new_agent_stack:
+        try:
+            fallback_mw = ModelFallbackMiddleware(
+                "openai:gpt-4o-mini",
+                "anthropic:claude-3-5-haiku-20241022",
+            )
+        except Exception:
+            logging.warning("ModelFallbackMiddleware init failed; skipping.")
+            fallback_mw = None
+    model_call_limit_mw = (
+        ModelCallLimitMiddleware(
+            thread_limit=120,
+            run_limit=80,
+            exit_behavior="end",
+        )
+        if flags.enable_model_call_limit and not flags.disable_new_agent_stack
+        else None
+    )
+    tool_call_limit_mw = (
+        ToolCallLimitMiddleware(
+            thread_limit=300, run_limit=80, exit_behavior="continue"
+        )
+        if flags.enable_tool_call_limit and not flags.disable_new_agent_stack
+        else None
+    )
+
+    noop_mw = (
+        NoopInjectionMiddleware()
+        if flags.enable_compaction_v2 and not flags.disable_new_agent_stack
+        else None
+    )
+
+    repair_mw = None
+    if flags.enable_tool_call_repair and not flags.disable_new_agent_stack:
+        registered_names: set[str] = {t.name for t in tools}
+        registered_names |= {
+            "write_todos",
+            "ls",
+            "read_file",
+            "write_file",
+            "edit_file",
+            "glob",
+            "grep",
+            "execute",
+            "task",
+            "mkdir",
+            "cd",
+            "pwd",
+            "move_file",
+            "rm",
+            "rmdir",
+            "list_tree",
+            "execute_code",
+        }
+        repair_mw = ToolCallNameRepairMiddleware(
+            registered_tool_names=registered_names,
+            fuzzy_match_threshold=None,
+        )
+
+    doom_loop_mw = (
+        DoomLoopMiddleware(threshold=3)
+        if flags.enable_doom_loop and not flags.disable_new_agent_stack
+        else None
+    )
+
+    permission_mw: PermissionMiddleware | None = (
+        PermissionMiddleware(rulesets=permission_rulesets)
+        if permission_rulesets
+        else None
+    )
+
+    action_log_mw: ActionLogMiddleware | None = None
+    if (
+        flags.enable_action_log
+        and not flags.disable_new_agent_stack
+        and thread_id is not None
+    ):
+        try:
+            tool_defs_by_name = {td.name: td for td in BUILTIN_TOOLS}
+            action_log_mw = ActionLogMiddleware(
+                thread_id=thread_id,
+                search_space_id=search_space_id,
+                user_id=user_id,
+                tool_definitions=tool_defs_by_name,
+            )
+        except Exception:  # pragma: no cover - defensive
+            logging.warning(
+                "ActionLogMiddleware init failed; running without it.",
+                exc_info=True,
+            )
+            action_log_mw = None
+
+    busy_mutex_mw: BusyMutexMiddleware | None = (
+        BusyMutexMiddleware()
+        if flags.enable_busy_mutex and not flags.disable_new_agent_stack
+        else None
+    )
+
+    otel_mw: OtelSpanMiddleware | None = (
+        OtelSpanMiddleware()
+        if flags.enable_otel and not flags.disable_new_agent_stack
+        else None
+    )
+
+    plugin_middlewares: list[Any] = []
+    if flags.enable_plugin_loader and not flags.disable_new_agent_stack:
+        try:
+            allowed_names = load_allowed_plugin_names_from_env()
+            if allowed_names:
+                plugin_middlewares = load_plugin_middlewares(
+                    PluginContext.build(
+                        search_space_id=search_space_id,
+                        user_id=user_id,
+                        thread_visibility=visibility,
+                        llm=llm,
+                    ),
+                    allowed_plugin_names=allowed_names,
+                )
+        except Exception:  # pragma: no cover - defensive
+            logging.warning(
+                "Plugin loader failed; continuing without plugins.",
+                exc_info=True,
+            )
+            plugin_middlewares = []
+
+    skills_mw: SkillsMiddleware | None = None
+    if flags.enable_skills and not flags.disable_new_agent_stack:
+        try:
+            skills_factory = build_skills_backend_factory(
+                search_space_id=search_space_id
+                if filesystem_mode == FilesystemMode.CLOUD
+                else None,
+            )
+            skills_mw = SkillsMiddleware(
+                backend=skills_factory,
+                sources=default_skills_sources(),
+            )
+        except Exception as exc:  # pragma: no cover - defensive
+            logging.warning("SkillsMiddleware init failed; skipping: %s", exc)
+            skills_mw = None
+
+    selector_mw: LLMToolSelectorMiddleware | None = None
+    if (
+        flags.enable_llm_tool_selector
+        and not flags.disable_new_agent_stack
+        and len(tools) > 30
+    ):
+        try:
+            selector_mw = LLMToolSelectorMiddleware(
+                model="openai:gpt-4o-mini",
+                max_tools=12,
+                always_include=[
+                    name
+                    for name in (
+                        "update_memory",
+                        "get_connected_accounts",
+                        "scrape_webpage",
+                    )
+                    if name in {t.name for t in tools}
+                ],
+            )
+        except Exception:
+            logging.warning("LLMToolSelectorMiddleware init failed; skipping.")
+            selector_mw = None
+
+    deepagent_middleware = [
+        busy_mutex_mw,
+        otel_mw,
+        TodoListMiddleware(),
+        _memory_middleware,
+        AnonymousDocumentMiddleware(
+            anon_session_id=anon_session_id,
+        )
+        if filesystem_mode == FilesystemMode.CLOUD
+        else None,
+        KnowledgeTreeMiddleware(
+            search_space_id=search_space_id,
+            filesystem_mode=filesystem_mode,
+            llm=llm,
+        )
+        if filesystem_mode == FilesystemMode.CLOUD
+        else None,
+        KnowledgePriorityMiddleware(
+            llm=llm,
+            search_space_id=search_space_id,
+            filesystem_mode=filesystem_mode,
+            available_connectors=available_connectors,
+            available_document_types=available_document_types,
+            mentioned_document_ids=mentioned_document_ids,
+        ),
+        FileIntentMiddleware(llm=llm),
+        SurfSenseFilesystemMiddleware(
+            backend=backend_resolver,
+            filesystem_mode=filesystem_mode,
+            search_space_id=search_space_id,
+            created_by_id=user_id,
+            thread_id=thread_id,
+        ),
+        KnowledgeBasePersistenceMiddleware(
+            search_space_id=search_space_id,
+            created_by_id=user_id,
+            filesystem_mode=filesystem_mode,
+            thread_id=thread_id,
+        )
+        if filesystem_mode == FilesystemMode.CLOUD
+        else None,
+        skills_mw,
+        SurfSenseCheckpointedSubAgentMiddleware(
+            checkpointer=checkpointer,
+            backend=StateBackend,
+            subagents=subagent_specs,
+        ),
+        selector_mw,
+        model_call_limit_mw,
+        tool_call_limit_mw,
+        context_edit_mw,
+        summarization_mw,
+        noop_mw,
+        retry_mw,
+        fallback_mw,
+        repair_mw,
+        permission_mw,
+        doom_loop_mw,
+        action_log_mw,
+        PatchToolCallsMiddleware(),
+        DedupHITLToolCallsMiddleware(agent_tools=list(tools)),
+        *plugin_middlewares,
+        AnthropicPromptCachingMiddleware(unsupported_model_behavior="ignore"),
+    ]
+    return [m for m in deepagent_middleware if m is not None]

surfsense_backend/app/agents/multi_agent_chat/main_agent/runtime/__init__.py

@@ -0,0 +1,7 @@
+"""Async factory: wiring tools, prompts, MCP buckets, then graph compile."""
+
+from __future__ import annotations
+
+from .factory import create_surfsense_deep_agent
+
+__all__ = ["create_surfsense_deep_agent"]

surfsense_backend/app/agents/multi_agent_chat/main_agent/runtime/factory.py

@@ -0,0 +1,230 @@
+"""Async factory: tools, system prompt, MCP buckets for subagents, then sync graph compile."""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import time
+from collections.abc import Sequence
+from typing import Any
+
+from deepagents.graph import BASE_AGENT_PROMPT
+from langchain_core.language_models import BaseChatModel
+from langchain_core.tools import BaseTool
+from langgraph.types import Checkpointer
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.agents.multi_agent_chat.subagents import (
+    get_subagents_to_exclude,
+    main_prompt_registry_subagent_lines,
+)
+from app.agents.multi_agent_chat.subagents.mcp_tools.index import (
+    load_mcp_tools_by_connector,
+)
+from app.agents.new_chat.chat_deepagent import _map_connectors_to_searchable_types
+from app.agents.new_chat.feature_flags import AgentFeatureFlags, get_flags
+from app.agents.new_chat.filesystem_backends import build_backend_resolver
+from app.agents.new_chat.filesystem_selection import FilesystemMode, FilesystemSelection
+from app.agents.new_chat.llm_config import AgentConfig
+from app.agents.new_chat.tools.invalid_tool import INVALID_TOOL_NAME, invalid_tool
+from app.agents.new_chat.tools.registry import build_tools_async
+from app.db import ChatVisibility
+from app.services.connector_service import ConnectorService
+from app.utils.perf import get_perf_logger
+
+from ..graph.compile_graph_sync import build_compiled_agent_graph_sync
+from ..system_prompt import build_main_agent_system_prompt
+from ..tools import (
+    MAIN_AGENT_SURFSENSE_TOOL_NAMES,
+    MAIN_AGENT_SURFSENSE_TOOL_NAMES_ORDERED,
+)
+
+_perf_log = get_perf_logger()
+
+
+async def create_surfsense_deep_agent(
+    llm: BaseChatModel,
+    search_space_id: int,
+    db_session: AsyncSession,
+    connector_service: ConnectorService,
+    checkpointer: Checkpointer,
+    user_id: str | None = None,
+    thread_id: int | None = None,
+    agent_config: AgentConfig | None = None,
+    enabled_tools: list[str] | None = None,
+    disabled_tools: list[str] | None = None,
+    additional_tools: Sequence[BaseTool] | None = None,
+    firecrawl_api_key: str | None = None,
+    thread_visibility: ChatVisibility | None = None,
+    mentioned_document_ids: list[int] | None = None,
+    anon_session_id: str | None = None,
+    filesystem_selection: FilesystemSelection | None = None,
+):
+    """Deep agent with SurfSense tools/middleware; registry route subagents behind ``task`` when enabled."""
+    _t_agent_total = time.perf_counter()
+    filesystem_selection = filesystem_selection or FilesystemSelection()
+    backend_resolver = build_backend_resolver(
+        filesystem_selection,
+        search_space_id=search_space_id
+        if filesystem_selection.mode == FilesystemMode.CLOUD
+        else None,
+    )
+
+    available_connectors: list[str] | None = None
+    available_document_types: list[str] | None = None
+
+    _t0 = time.perf_counter()
+    try:
+        connector_types = await connector_service.get_available_connectors(
+            search_space_id
+        )
+        available_connectors = _map_connectors_to_searchable_types(connector_types)
+
+        available_document_types = await connector_service.get_available_document_types(
+            search_space_id
+        )
+
+    except Exception as e:
+        logging.warning("Failed to discover available connectors/document types: %s", e)
+    _perf_log.info(
+        "[create_agent] Connector/doc-type discovery in %.3fs",
+        time.perf_counter() - _t0,
+    )
+
+    visibility = thread_visibility or ChatVisibility.PRIVATE
+
+    _model_profile = getattr(llm, "profile", None)
+    _max_input_tokens: int | None = (
+        _model_profile.get("max_input_tokens")
+        if isinstance(_model_profile, dict)
+        else None
+    )
+
+    dependencies: dict[str, Any] = {
+        "search_space_id": search_space_id,
+        "db_session": db_session,
+        "connector_service": connector_service,
+        "firecrawl_api_key": firecrawl_api_key,
+        "user_id": user_id,
+        "thread_id": thread_id,
+        "thread_visibility": visibility,
+        "available_connectors": available_connectors,
+        "available_document_types": available_document_types,
+        "max_input_tokens": _max_input_tokens,
+        "llm": llm,
+    }
+
+    _t0 = time.perf_counter()
+    mcp_tools_by_agent = await load_mcp_tools_by_connector(db_session, search_space_id)
+    _perf_log.info(
+        "[create_agent] load_mcp_tools_by_connector in %.3fs (%d buckets)",
+        time.perf_counter() - _t0,
+        len(mcp_tools_by_agent),
+    )
+
+    modified_disabled_tools = list(disabled_tools) if disabled_tools else []
+
+    if "search_knowledge_base" not in modified_disabled_tools:
+        modified_disabled_tools.append("search_knowledge_base")
+
+    if enabled_tools is not None:
+        main_agent_enabled_tools = [
+            n for n in enabled_tools if n in MAIN_AGENT_SURFSENSE_TOOL_NAMES
+        ]
+    else:
+        main_agent_enabled_tools = list(MAIN_AGENT_SURFSENSE_TOOL_NAMES_ORDERED)
+
+    _t0 = time.perf_counter()
+    tools = await build_tools_async(
+        dependencies=dependencies,
+        enabled_tools=main_agent_enabled_tools,
+        disabled_tools=modified_disabled_tools,
+        additional_tools=list(additional_tools) if additional_tools else None,
+        include_mcp_tools=False,
+    )
+
+    _flags: AgentFeatureFlags = get_flags()
+    if _flags.enable_tool_call_repair and INVALID_TOOL_NAME not in {
+        t.name for t in tools
+    }:
+        tools = [*list(tools), invalid_tool]
+    _perf_log.info(
+        "[create_agent] build_tools_async in %.3fs (%d tools)",
+        time.perf_counter() - _t0,
+        len(tools),
+    )
+
+    _t0 = time.perf_counter()
+    _enabled_tool_names = {t.name for t in tools}
+    _user_disabled_tool_names = set(disabled_tools) if disabled_tools else set()
+
+    _model_name: str | None = None
+    prof = getattr(llm, "model_name", None) or getattr(llm, "model", None)
+    if isinstance(prof, str):
+        _model_name = prof
+
+    _connector_exclude = get_subagents_to_exclude(available_connectors)
+    _registry_subagent_prompt_lines = main_prompt_registry_subagent_lines(
+        _connector_exclude
+    )
+
+    if agent_config is not None:
+        system_prompt = build_main_agent_system_prompt(
+            today=None,
+            thread_visibility=thread_visibility,
+            enabled_tool_names=_enabled_tool_names,
+            disabled_tool_names=_user_disabled_tool_names,
+            custom_system_instructions=agent_config.system_instructions,
+            use_default_system_instructions=agent_config.use_default_system_instructions,
+            citations_enabled=agent_config.citations_enabled,
+            model_name=_model_name or getattr(agent_config, "model_name", None),
+            registry_subagent_prompt_lines=_registry_subagent_prompt_lines,
+        )
+    else:
+        system_prompt = build_main_agent_system_prompt(
+            thread_visibility=thread_visibility,
+            enabled_tool_names=_enabled_tool_names,
+            disabled_tool_names=_user_disabled_tool_names,
+            citations_enabled=True,
+            model_name=_model_name,
+            registry_subagent_prompt_lines=_registry_subagent_prompt_lines,
+        )
+    _perf_log.info(
+        "[create_agent] System prompt built in %.3fs", time.perf_counter() - _t0
+    )
+
+    final_system_prompt = system_prompt + "\n\n" + BASE_AGENT_PROMPT
+
+    _t0 = time.perf_counter()
+    agent = await asyncio.to_thread(
+        build_compiled_agent_graph_sync,
+        llm=llm,
+        tools=tools,
+        final_system_prompt=final_system_prompt,
+        backend_resolver=backend_resolver,
+        filesystem_mode=filesystem_selection.mode,
+        search_space_id=search_space_id,
+        user_id=user_id,
+        thread_id=thread_id,
+        visibility=visibility,
+        anon_session_id=anon_session_id,
+        available_connectors=available_connectors,
+        available_document_types=available_document_types,
+        mentioned_document_ids=mentioned_document_ids,
+        max_input_tokens=_max_input_tokens,
+        flags=_flags,
+        checkpointer=checkpointer,
+        subagent_dependencies=dependencies,
+        mcp_tools_by_agent=mcp_tools_by_agent,
+        disabled_tools=disabled_tools,
+    )
+    _perf_log.info(
+        "[create_agent] Middleware stack + graph compiled in %.3fs",
+        time.perf_counter() - _t0,
+    )
+
+    _perf_log.info(
+        "[create_agent] Total agent creation in %.3fs",
+        time.perf_counter() - _t_agent_total,
+    )
+    return agent

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/__init__.py

@@ -0,0 +1,7 @@
+"""Main-agent system prompt — not shared verbatim with single-agent ``new_chat``."""
+
+from __future__ import annotations
+
+from .builder import build_main_agent_system_prompt
+
+__all__ = ["build_main_agent_system_prompt"]

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/builder/__init__.py

@@ -0,0 +1,7 @@
+"""Assemble the main-agent system prompt from ``markdown/*.md`` fragments."""
+
+from __future__ import annotations
+
+from .compose import build_main_agent_system_prompt
+
+__all__ = ["build_main_agent_system_prompt"]

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/builder/compose.py

@@ -0,0 +1,53 @@
+"""Assemble the **main-agent** deep-agent system string only.
+
+Sections (order matters): core instructions → provider → citations → dynamic
+``<registry_subagents>`` → SurfSense ``<tools>``.
+"""
+
+from __future__ import annotations
+
+from datetime import UTC, datetime
+
+from app.db import ChatVisibility
+
+from .sections.citations import build_citations_section
+from .sections.provider import build_provider_section
+from .sections.registry_subagents import build_registry_subagents_section
+from .sections.system_instruction import build_default_system_instruction_xml
+from .sections.tools import build_tools_section
+
+
+def build_main_agent_system_prompt(
+    *,
+    today: datetime | None = None,
+    thread_visibility: ChatVisibility | None = None,
+    enabled_tool_names: set[str] | None = None,
+    disabled_tool_names: set[str] | None = None,
+    custom_system_instructions: str | None = None,
+    use_default_system_instructions: bool = True,
+    citations_enabled: bool = True,
+    model_name: str | None = None,
+    registry_subagent_prompt_lines: list[tuple[str, str]] | None = None,
+) -> str:
+    resolved_today = (today or datetime.now(UTC)).astimezone(UTC).date().isoformat()
+    visibility = thread_visibility or ChatVisibility.PRIVATE
+
+    if custom_system_instructions and custom_system_instructions.strip():
+        system_block = custom_system_instructions.format(resolved_today=resolved_today)
+    elif use_default_system_instructions:
+        system_block = build_default_system_instruction_xml(
+            visibility=visibility,
+            resolved_today=resolved_today,
+        )
+    else:
+        system_block = ""
+
+    system_block += build_provider_section(model_name=model_name)
+    system_block += build_citations_section(citations_enabled=citations_enabled)
+    system_block += build_registry_subagents_section(registry_subagent_prompt_lines)
+    system_block += build_tools_section(
+        visibility=visibility,
+        enabled_tool_names=enabled_tool_names,
+        disabled_tool_names=disabled_tool_names,
+    )
+    return system_block

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/builder/load_md.py

@@ -0,0 +1,16 @@
+"""Load main-agent-only markdown from ``system_prompt/markdown/`` (``importlib.resources``)."""
+
+from __future__ import annotations
+
+from importlib import resources
+
+_PROMPTS_PACKAGE = "app.agents.multi_agent_chat.main_agent.system_prompt.markdown"
+
+
+def read_prompt_md(filename: str) -> str:
+    """Load ``markdown/{filename}`` (e.g. ``agent_private.md`` or ``tools/_preamble.md``)."""
+    ref = resources.files(_PROMPTS_PACKAGE).joinpath(filename)
+    if not ref.is_file():
+        return ""
+    text = ref.read_text(encoding="utf-8")
+    return text[:-1] if text.endswith("\n") else text

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/builder/provider_hints.py

@@ -0,0 +1,50 @@
+"""Provider-specific style hints from ``markdown/providers/`` (main agent only)."""
+
+from __future__ import annotations
+
+import re
+
+from .load_md import read_prompt_md
+
+ProviderVariant = str
+
+_OPENAI_CODEX_RE = re.compile(
+    r"\b(gpt-codex|codex-mini|gpt-[\d.]+-codex)\b", re.IGNORECASE
+)
+_OPENAI_REASONING_RE = re.compile(r"\b(gpt-5|o\d|o-)", re.IGNORECASE)
+_OPENAI_CLASSIC_RE = re.compile(r"\bgpt-4", re.IGNORECASE)
+_ANTHROPIC_RE = re.compile(r"\bclaude\b", re.IGNORECASE)
+_GOOGLE_RE = re.compile(r"\bgemini\b", re.IGNORECASE)
+_KIMI_RE = re.compile(r"\b(kimi[-\d.]*|moonshot)\b", re.IGNORECASE)
+_GROK_RE = re.compile(r"\bgrok\b", re.IGNORECASE)
+_DEEPSEEK_RE = re.compile(r"\bdeepseek\b", re.IGNORECASE)
+
+
+def detect_provider_variant(model_name: str | None) -> ProviderVariant:
+    if not model_name:
+        return "default"
+    name = model_name.strip()
+    if _OPENAI_CODEX_RE.search(name):
+        return "openai_codex"
+    if _OPENAI_REASONING_RE.search(name):
+        return "openai_reasoning"
+    if _OPENAI_CLASSIC_RE.search(name):
+        return "openai_classic"
+    if _ANTHROPIC_RE.search(name):
+        return "anthropic"
+    if _GOOGLE_RE.search(name):
+        return "google"
+    if _KIMI_RE.search(name):
+        return "kimi"
+    if _GROK_RE.search(name):
+        return "grok"
+    if _DEEPSEEK_RE.search(name):
+        return "deepseek"
+    return "default"
+
+
+def build_provider_hint_block(provider_variant: ProviderVariant) -> str:
+    if not provider_variant or provider_variant == "default":
+        return ""
+    text = read_prompt_md(f"providers/{provider_variant}.md")
+    return f"\n{text}\n" if text else ""

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/builder/sections/__init__.py

@@ -0,0 +1 @@
+"""Rendered slices of the main-agent system prompt."""

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/builder/sections/citations.py

@@ -0,0 +1,11 @@
+"""Citation fragment for the main agent (chunk-tagged context only)."""
+
+from __future__ import annotations
+
+from ..load_md import read_prompt_md
+
+
+def build_citations_section(*, citations_enabled: bool) -> str:
+    name = "citations_on.md" if citations_enabled else "citations_off.md"
+    fragment = read_prompt_md(name)
+    return f"\n{fragment}\n" if fragment else ""

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/builder/sections/provider.py

@@ -0,0 +1,9 @@
+"""Provider-specific style hints."""
+
+from __future__ import annotations
+
+from ..provider_hints import build_provider_hint_block, detect_provider_variant
+
+
+def build_provider_section(*, model_name: str | None) -> str:
+    return build_provider_hint_block(detect_provider_variant(model_name))

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/builder/sections/registry_subagents.py

@@ -0,0 +1,27 @@
+"""Dynamic ``<registry_subagents>`` block: **task** specialists actually built for this workspace."""
+
+from __future__ import annotations
+
+
+def build_registry_subagents_section(
+    registry_subagent_lines: list[tuple[str, str]] | None,
+) -> str:
+    if registry_subagent_lines is None:
+        return ""
+    if not registry_subagent_lines:
+        return (
+            "\n<registry_subagents>\n"
+            "No registry specialists are listed for **task** in this workspace.\n"
+            "</registry_subagents>\n"
+        )
+    bullets = "\n".join(
+        f"- **{name}** — {desc}" for name, desc in registry_subagent_lines
+    )
+    return (
+        "\n<registry_subagents>\n"
+        "These specialists are registered for **task** (routes without a matching connector are omitted).\n"
+        f"{bullets}\n"
+        "The runtime may also offer a general-purpose **task** helper with your tools in a separate context.\n"
+        "Pick the specialist by **name**. Put full instructions in the task prompt; they do not see this thread.\n"
+        "</registry_subagents>\n"
+    )

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/builder/sections/system_instruction.py

@@ -0,0 +1,35 @@
+"""Default ``<system_instruction>`` block for the main agent only."""
+
+from __future__ import annotations
+
+from app.db import ChatVisibility
+
+from ..load_md import read_prompt_md
+
+_PRIVATE_ORDER = (
+    "agent_private.md",
+    "kb_only_policy_private.md",
+    "main_agent_tool_routing.md",
+    "parameter_resolution.md",
+    "memory_protocol_private.md",
+)
+_TEAM_ORDER = (
+    "agent_team.md",
+    "kb_only_policy_team.md",
+    "main_agent_tool_routing.md",
+    "parameter_resolution.md",
+    "memory_protocol_team.md",
+)
+
+
+def build_default_system_instruction_xml(
+    *,
+    visibility: ChatVisibility,
+    resolved_today: str,
+) -> str:
+    order = _TEAM_ORDER if visibility == ChatVisibility.SEARCH_SPACE else _PRIVATE_ORDER
+    parts = [read_prompt_md(name) for name in order]
+    body = "\n\n".join(p for p in parts if p)
+    return f"\n<system_instruction>\n{body}\n\n</system_instruction>\n".format(
+        resolved_today=resolved_today,
+    )

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/builder/sections/tools.py

@@ -0,0 +1,20 @@
+"""Main-agent ``<tools>`` block (memory + research builtins only; see ``main_agent.tools``)."""
+
+from __future__ import annotations
+
+from app.db import ChatVisibility
+
+from ..tool_instruction_block import build_tools_instruction_block
+
+
+def build_tools_section(
+    *,
+    visibility: ChatVisibility,
+    enabled_tool_names: set[str] | None,
+    disabled_tool_names: set[str] | None,
+) -> str:
+    return build_tools_instruction_block(
+        visibility=visibility,
+        enabled_tool_names=enabled_tool_names,
+        disabled_tool_names=disabled_tool_names,
+    )

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/builder/tool_instruction_block.py

@@ -0,0 +1,86 @@
+"""``<tools>`` + ``<tool_call_examples>`` from ``system_prompt/markdown/{tools,examples}/``.
+
+Only documents tools the main agent actually binds — not full ``new_chat``.
+"""
+
+from __future__ import annotations
+
+from app.db import ChatVisibility
+
+from ...tools import MAIN_AGENT_SURFSENSE_TOOL_NAMES_ORDERED
+from .load_md import read_prompt_md
+
+_MEMORY_VARIANT_TOOLS: frozenset[str] = frozenset({"update_memory"})
+
+
+def _tool_fragment_path(tool_name: str, variant: str) -> str:
+    if tool_name in _MEMORY_VARIANT_TOOLS:
+        return f"tools/{tool_name}_{variant}.md"
+    return f"tools/{tool_name}.md"
+
+
+def _example_fragment_path(tool_name: str, variant: str) -> str:
+    if tool_name in _MEMORY_VARIANT_TOOLS:
+        return f"examples/{tool_name}_{variant}.md"
+    return f"examples/{tool_name}.md"
+
+
+def _format_tool_label(tool_name: str) -> str:
+    return tool_name.replace("_", " ").title()
+
+
+def build_tools_instruction_block(
+    *,
+    visibility: ChatVisibility,
+    enabled_tool_names: set[str] | None,
+    disabled_tool_names: set[str] | None,
+) -> str:
+    variant = "team" if visibility == ChatVisibility.SEARCH_SPACE else "private"
+
+    parts: list[str] = []
+    preamble = read_prompt_md("tools/_preamble.md")
+    if preamble:
+        parts.append(preamble + "\n")
+
+    examples: list[str] = []
+
+    for tool_name in MAIN_AGENT_SURFSENSE_TOOL_NAMES_ORDERED:
+        if enabled_tool_names is not None and tool_name not in enabled_tool_names:
+            continue
+
+        instruction = read_prompt_md(_tool_fragment_path(tool_name, variant))
+        if instruction:
+            parts.append(instruction + "\n")
+
+        example = read_prompt_md(_example_fragment_path(tool_name, variant))
+        if example:
+            examples.append(example + "\n")
+
+    known_disabled = (
+        set(disabled_tool_names) & set(MAIN_AGENT_SURFSENSE_TOOL_NAMES_ORDERED)
+        if disabled_tool_names
+        else set()
+    )
+    if known_disabled:
+        disabled_list = ", ".join(
+            _format_tool_label(n)
+            for n in MAIN_AGENT_SURFSENSE_TOOL_NAMES_ORDERED
+            if n in known_disabled
+        )
+        parts.append(
+            "\n"
+            "DISABLED TOOLS (by user, main-agent scope):\n"
+            f"These SurfSense tools were disabled on the main agent for this session: {disabled_list}.\n"
+            "You do NOT have access to them and MUST NOT claim you can use them.\n"
+            "If the user still needs that capability, delegate with **task** if a subagent covers it,\n"
+            "otherwise explain it is disabled on the main agent for this session.\n"
+        )
+
+    parts.append("\n</tools>\n")
+
+    if examples:
+        parts.append("<tool_call_examples>")
+        parts.extend(examples)
+        parts.append("</tool_call_examples>\n")
+
+    return "".join(parts)

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/__init__.py

@@ -0,0 +1 @@
+"""Markdown fragments for the **main-agent** system prompt only (`importlib.resources`)."""

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/agent_private.md

@@ -0,0 +1,9 @@
+You are SurfSense’s **main agent**: you answer using the user’s knowledge context,
+lightweight research tools, and memory — and you **delegate** integrations and
+specialized work via **task** (see `<tool_routing>` in this prompt).
+
+Today's date (UTC): {resolved_today}
+
+When writing mathematical formulas or equations, ALWAYS use LaTeX notation. NEVER use backtick code spans or Unicode symbols for math.
+
+NEVER expose internal tool parameter names, backend IDs, or implementation details to the user. Always use natural, user-friendly language instead.

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/agent_team.md

@@ -0,0 +1,11 @@
+You are SurfSense’s **main agent** for this team space: you answer using shared
+knowledge context, lightweight research tools, and memory — and you **delegate**
+integrations and specialized work via **task** (see `<tool_routing>` in this prompt).
+
+In this team thread, each message is prefixed with **[DisplayName of the author]**. Use this to attribute and reference the author of anything in the discussion (who asked a question, made a suggestion, or contributed an idea) and to cite who said what in your answers.
+
+Today's date (UTC): {resolved_today}
+
+When writing mathematical formulas or equations, ALWAYS use LaTeX notation. NEVER use backtick code spans or Unicode symbols for math.
+
+NEVER expose internal tool parameter names, backend IDs, or implementation details to the user. Always use natural, user-friendly language instead.

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/citations_off.md

@@ -0,0 +1,15 @@
+<citation_instructions>
+IMPORTANT: Citations are DISABLED for this configuration.
+
+DO NOT include `[citation:…]` markers anywhere — even if tool descriptions or examples
+mention them. Ignore citation-format reminders elsewhere in this prompt when they conflict
+with this block.
+
+Instead:
+1. Answer in plain prose; optional markdown links to public URLs when sources are URLs.
+2. Do NOT expose raw chunk IDs, document IDs, or internal IDs to the user.
+3. Present indexed or doc-search facts naturally without attribution markers.
+
+When answering from workspace or docs context: integrate facts cleanly without claiming
+“this comes from chunk X”.
+</citation_instructions>

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/citations_on.md

@@ -0,0 +1,15 @@
+<citation_instructions>
+This block appears **before** `<tools>` so it wins over any tool-example wording below.
+
+Apply chunk citations **only** when the runtime injects `<document>` / `<chunk id='…'>` blocks
+(e.g. from SurfSense docs search or priority documents).
+
+1. For each factual statement taken from those chunks, add `[citation:chunk_id]` using the **exact** `chunk_id` string from `<chunk id='…'>`.
+2. Multiple chunks → `[citation:id1], [citation:id2]` (comma-separated).
+3. Never invent or normalize ids; if unsure, omit the citation.
+4. Plain brackets only — no markdown links, no `([citation:…](url))`, no footnote numbering.
+
+Chunk ids may be numeric, prefixed (e.g. `doc-45`), or URLs when the source is web-shaped — copy verbatim.
+
+If no chunk-tagged documents appear in context this turn, do not fabricate citations.
+</citation_instructions>

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/examples/__init__.py

@@ -0,0 +1 @@
+

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/examples/scrape_webpage.md

@@ -0,0 +1,13 @@
+
+- User: "Check out https://dev.to/some-article"
+  - Call: `scrape_webpage(url="https://dev.to/some-article")`
+  - Respond with a structured analysis — key points, takeaways.
+- User: "Read this article and summarize it for me: https://example.com/blog/ai-trends"
+  - Call: `scrape_webpage(url="https://example.com/blog/ai-trends")`
+  - Respond with a thorough summary using headings and bullet points.
+- User: (after discussing https://example.com/stats) "Can you get the live data from that page?"
+  - Call: `scrape_webpage(url="https://example.com/stats")`
+  - IMPORTANT: Always attempt scraping first. Never refuse before trying the tool.
+- User: "https://example.com/blog/weekend-recipes"
+  - Call: `scrape_webpage(url="https://example.com/blog/weekend-recipes")`
+  - When a user sends just a URL with no instructions, scrape it and provide a concise summary of the content.

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/examples/search_surfsense_docs.md

@@ -0,0 +1,9 @@
+
+- User: "How do I install SurfSense?"
+  - Call: `search_surfsense_docs(query="installation setup")`
+- User: "What connectors does SurfSense support?"
+  - Call: `search_surfsense_docs(query="available connectors integrations")`
+- User: "How do I set up the Notion connector?"
+  - Call: `search_surfsense_docs(query="Notion connector setup configuration")` (how-to docs). Changing data inside Notion itself → **task**.
+- User: "How do I use Docker to run SurfSense?"
+  - Call: `search_surfsense_docs(query="Docker installation setup")`

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/examples/update_memory_private.md

@@ -0,0 +1,16 @@
+
+- <user_name>Alex</user_name>, <user_memory> is empty. User: "I'm a space enthusiast, explain astrophage to me"
+  - The user casually shared a durable fact. Use their first name in the entry, short neutral heading:
+    update_memory(updated_memory="## Interests & background\n- (2025-03-15) [fact] Alex is a space enthusiast\n")
+- User: "Remember that I prefer concise answers over detailed explanations"
+  - Durable preference. Merge with existing memory, add a new heading:
+    update_memory(updated_memory="## Interests & background\n- (2025-03-15) [fact] Alex is a space enthusiast\n\n## Response style\n- (2025-03-15) [pref] Alex prefers concise answers over detailed explanations\n")
+- User: "I actually moved to Tokyo last month"
+  - Updated fact, date prefix reflects when recorded:
+    update_memory(updated_memory="## Interests & background\n...\n\n## Personal context\n- (2025-03-15) [fact] Alex lives in Tokyo (previously London)\n...")
+- User: "I'm a freelance photographer working on a nature documentary"
+  - Durable background info under a fitting heading:
+    update_memory(updated_memory="...\n\n## Current focus\n- (2025-03-15) [fact] Alex is a freelance photographer\n- (2025-03-15) [fact] Alex is working on a nature documentary\n")
+- User: "Always respond in bullet points"
+  - Standing instruction:
+    update_memory(updated_memory="...\n\n## Response style\n- (2025-03-15) [instr] Always respond to Alex in bullet points\n")

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/examples/update_memory_team.md

@@ -0,0 +1,7 @@
+
+- User: "Let's remember that we decided to do weekly standup meetings on Mondays"
+  - Durable team decision:
+    update_memory(updated_memory="- (2025-03-15) [fact] Weekly standup meetings on Mondays\n...")
+- User: "Our office is in downtown Seattle, 5th floor"
+  - Durable team fact:
+    update_memory(updated_memory="- (2025-03-15) [fact] Office location: downtown Seattle, 5th floor\n...")

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/examples/web_search.md

@@ -0,0 +1,8 @@
+
+- User: "What's the current USD to INR exchange rate?"
+  - Call: `web_search(query="current USD to INR exchange rate")`
+  - Answer from returned snippets or scrape a top URL if needed; use markdown links to sources.
+- User: "What's the latest news about AI?"
+  - Call: `web_search(query="latest AI news today")`
+- User: "What's the weather in New York?"
+  - Call: `web_search(query="weather New York today")`

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/kb_only_policy_private.md

@@ -0,0 +1,19 @@
+<knowledge_base_only_policy>
+CRITICAL RULE — KNOWLEDGE BASE FIRST, NEVER DEFAULT TO GENERAL KNOWLEDGE:
+- Ground factual answers in what you actually receive this turn: injected workspace
+  documents (when present), **search_surfsense_docs**, **web_search**, **scrape_webpage**,
+  or substantive results summarized from a **task** subagent you invoked.
+- Do NOT answer factual or informational questions from general knowledge unless the user
+  explicitly grants permission after you say you did not find enough in those sources.
+- If indexed/docs search returns nothing relevant AND **web_search** / **scrape_webpage**
+  (and **task**, if already tried appropriately) still do not supply an answer, you MUST:
+  1. Say you could not find enough in their workspace/docs/tools output.
+  2. Ask: "Would you like me to answer from my general knowledge instead?"
+  3. ONLY then answer from general knowledge after they clearly say yes.
+- This policy does NOT apply to:
+  * Casual conversation, greetings, or meta-questions about SurfSense (e.g. "what can you do?")
+  * Formatting or analysis of content already in the chat
+  * Clear rewrite/edit instructions ("bullet-point this paragraph")
+  * Lightweight research with **web_search** / **scrape_webpage**
+  * Work that belongs on a specialist — use **task**; see `<tool_routing>`
+</knowledge_base_only_policy>

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/kb_only_policy_team.md

@@ -0,0 +1,19 @@
+<knowledge_base_only_policy>
+CRITICAL RULE — KNOWLEDGE BASE FIRST, NEVER DEFAULT TO GENERAL KNOWLEDGE:
+- Ground factual answers in what you actually receive this turn: injected shared
+  workspace documents (when present), **search_surfsense_docs**, **web_search**,
+  **scrape_webpage**, or substantive results summarized from a **task** subagent you invoked.
+- Do NOT answer factual questions from general knowledge unless a team member explicitly
+  grants permission after you say you did not find enough in those sources.
+- If indexed/docs search returns nothing relevant AND **web_search** / **scrape_webpage**
+  (and **task**, if already tried appropriately) still do not supply an answer, you MUST:
+  1. Say you could not find enough in shared docs/tools output.
+  2. Ask: "Would you like me to answer from my general knowledge instead?"
+  3. ONLY then answer from general knowledge after they clearly say yes.
+- This policy does NOT apply to:
+  * Casual conversation, greetings, or meta-questions about SurfSense
+  * Formatting or analysis of content already in the chat
+  * Clear rewrite/edit instructions
+  * Lightweight research with **web_search** / **scrape_webpage**
+  * Work that belongs on a specialist — use **task**; see `<tool_routing>`
+</knowledge_base_only_policy>

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/main_agent_tool_routing.md

@@ -0,0 +1,27 @@
+<tool_routing>
+Use **task** for anything beyond your direct SurfSense tools: calendar, mail,
+chat, tickets, documents in third-party systems, connector-specific discovery,
+deliverables (reports, podcasts, images, etc.), and other specialized routes.
+The live list of specialists you may target with **task** for this workspace is in
+`<registry_subagents>` (later in this prompt).
+
+Your **direct** SurfSense tools are only: **update_memory**, **web_search**,
+**scrape_webpage**, and **search_surfsense_docs**. The runtime may also attach
+deep-agent helpers (e.g. todos, filesystem, **task** itself). Use **task** whenever
+the user needs capabilities **not** listed in the `<tools>` section (that section appears
+later in this system prompt, after citation rules).
+
+Do not treat live third-party state as if it were already in the indexed knowledge
+base; reach it via **task**.
+
+Never emit more than one **task** tool call in the same turn. Bundle related work
+for the same specialist into a single **task** invocation (the subagent itself can
+call its own tools in parallel inside that one run). Parallel **task** calls would
+fan out into multiple concurrent subagent runs whose human-approval interrupts
+cannot be coordinated; one **task** at a time is required.
+</tool_routing>
+
+<!-- TODO: lift the single-task constraint once the runtime supports parallel task
+interrupts end-to-end (multi-interrupt SSE + interrupt-id-keyed Command(resume)
++ keyed surfsense_resume_value side-channel). Until then this nudge is the only
+guard; the parent graph's resume cannot address multiple pending interrupts. -->

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/memory_protocol_private.md

@@ -0,0 +1,6 @@
+<memory_protocol>
+IMPORTANT — After understanding each user message, ALWAYS check: does this message
+reveal durable facts about the user (role, interests, preferences, projects,
+background, or standing instructions)? If yes, you MUST call update_memory
+alongside your normal response — do not defer this to a later turn.
+</memory_protocol>

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/memory_protocol_team.md

@@ -0,0 +1,6 @@
+<memory_protocol>
+IMPORTANT — After understanding each user message, ALWAYS check: does this message
+reveal durable facts about the team (decisions, conventions, architecture, processes,
+or key facts)? If yes, you MUST call update_memory alongside your normal response —
+do not defer this to a later turn.
+</memory_protocol>

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/parameter_resolution.md

@@ -0,0 +1,15 @@
+<parameter_resolution>
+You do **not** call connector-specific discovery tools yourself (accounts, channels,
+Jira cloud IDs, Airtable bases, Slack channels, etc.). Those tools exist only on
+**task** subagents.
+
+When the user needs work inside a connected product, delegate with **task** and a
+clear goal. If several Slack channels, Jira projects, calendar calendars, etc. could
+match and only the integration can list them, **you must not** ask the human for
+internal IDs (UUIDs, cloud IDs, opaque keys). The **task** subagent uses connector
+tools to list candidates and either picks the only sensible match or asks the user
+to choose using **normal labels** (e.g. channel display name, project title), not raw IDs.
+
+If you already have plain-language choices from the user or from prior tool output,
+you may pass them through to **task** without re-discovery.
+</parameter_resolution>

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/providers/__init__.py

@@ -0,0 +1 @@
+

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/providers/anthropic.md

@@ -0,0 +1,16 @@
+<provider_hints>
+You are running on an Anthropic Claude model (SurfSense **main agent**).
+
+Structured reasoning:
+- For non-trivial work, `<thinking>` / short `<plan>` before tool calls is fine.
+
+Professional objectivity:
+- Accuracy over flattery; verify with **search_surfsense_docs**, **web_search**, **scrape_webpage**, or **task** when unsure — don’t invent connector access.
+
+Task management:
+- For 3+ steps, use todo tooling; update statuses promptly.
+
+Tool calls:
+- Parallelise independent calls; sequence only when outputs chain.
+- Never pretend you can run connector-specific tools directly — route through **task** when needed.
+</provider_hints>

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/providers/deepseek.md

@@ -0,0 +1,18 @@
+<provider_hints>
+You are running on a DeepSeek model (SurfSense **main agent**).
+
+Reasoning hygiene (R1-aware):
+- Keep internal scratch separate from the user-facing answer; don’t leak chain-of-thought into tool arguments.
+
+Output style:
+- Concise; lead with the answer or the next action; avoid sycophantic openers.
+
+Attribution:
+- When citations are **enabled** and facts come from chunk-tagged context, follow the citation block above.
+- When citations are **disabled**, do not use `[citation:…]`.
+
+Tool calls:
+- Parallelise independent calls.
+- Prefer **search_surfsense_docs** for SurfSense docs/product questions before **web_search** when that fits the ask.
+- Don’t invent paths, chunk ids, or URLs — only values from tools or the user.
+</provider_hints>

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/providers/default.md

@@ -0,0 +1 @@
+

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/providers/google.md

@@ -0,0 +1,18 @@
+<provider_hints>
+You are running on a Google Gemini model (SurfSense **main agent**).
+
+Output style:
+- Concise & direct. Fewer than ~3 lines of prose when the task allows (excluding tool output and code).
+- No filler openers/closers — move straight to the answer or the tool call.
+- GitHub-flavoured Markdown; monospace-friendly.
+
+Workflow (Understand → Plan → Act → Verify):
+1. **Understand:** parse the ask; use **search_surfsense_docs** / injected workspace context before guessing.
+2. **Plan:** for multi-step work, a short plan first.
+3. **Act:** only with tools you actually have on this agent (see `<tools>` and `<tool_routing>`). Connector work → **task**.
+4. **Verify:** re-read or re-search only when it materially reduces risk.
+
+Discipline:
+- Do not imply access to connectors, MCP tools, or deliverable generators except via **task**.
+- Path arguments for filesystem tools must be exact strings from tool results — never invent paths.
+</provider_hints>

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/providers/grok.md

@@ -0,0 +1,16 @@
+<provider_hints>
+You are running on an xAI Grok model (SurfSense **main agent**).
+
+Maximum terseness:
+- Fewer than 4 lines unless detail is requested; skip preamble/postamble.
+
+Tool discipline:
+- Typically one investigative tool per turn unless several independent read-only queries are clearly needed; don’t repeat identical calls.
+
+Attribution:
+- When citations are **enabled** (see citation block above) and you answer from chunk-tagged documents, use `[citation:chunk_id]` exactly as specified there.
+- When citations are **disabled**, never emit `[citation:…]` — plain prose and links per tool guidance.
+
+Style:
+- No emojis unless asked; flat lists for short answers.
+</provider_hints>

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/providers/kimi.md

@@ -0,0 +1,21 @@
+<provider_hints>
+You are running on a Moonshot Kimi model (Kimi-K1.5 / Kimi-K2 / Kimi-K2.5+), SurfSense **main agent**.
+
+Action bias:
+- Default to taking action with tools rather than describing solutions in prose. If a tool can answer the question, call the tool.
+- Don't narrate routine reads, searches, or obvious next steps. Combine related progress into one short status line.
+- Be thorough in actions (test what you build, verify what you change). Be brief in explanations.
+
+Tool calls:
+- Output multiple non-interfering tool calls in a SINGLE response — parallelism is a major efficiency win on this model.
+- When the `task` tool is available, delegate focused subtasks to a subagent with full context (subagents don't inherit yours).
+- Don't apologise or pre-announce tool calls. The tool call itself is self-explanatory.
+
+Language:
+- Respond in the SAME language as the user's most recent turn unless explicitly instructed otherwise.
+
+Discipline:
+- Stay on track. Never give the user more than what they asked for.
+- Fact-check with tools; don’t fabricate chunk ids or connector outcomes.
+- Keep it stupidly simple. Don't overcomplicate.
+</provider_hints>

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/providers/openai_classic.md

@@ -0,0 +1,20 @@
+<provider_hints>
+You are running on a classic OpenAI chat model (GPT-4 family), SurfSense **main agent**.
+
+Persistence:
+- Finish the user’s request in the same turn when tools allow — don’t stop at intent only.
+- If a tool errors, fix arguments and retry once before giving up.
+
+Planning:
+- For 3+ steps, use the todo / planning tool; mark `in_progress` / `completed` promptly.
+- One short sentence before non-trivial tool use is fine.
+
+Output style:
+- Conversational but professional; bullets for findings; fenced code with language tags when needed.
+- Summarize tool output — don’t paste walls of text.
+
+Tool calls:
+- Parallelise independent calls in one turn.
+- Prefer **search_surfsense_docs** for SurfSense-product questions, **web_search** / **scrape_webpage**
+  for fresh public facts; integrations and heavy workflows → **task**.
+</provider_hints>

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/providers/openai_codex.md

@@ -0,0 +1,13 @@
+<provider_hints>
+You are running on an OpenAI Codex-class model (SurfSense **main agent**).
+
+Output style:
+- Concise; don’t paste huge fetch blobs — summarize.
+- When citations are **enabled** and you rely on chunk-tagged docs, references may use `[citation:chunk_id]` per the citation block above; when **disabled**, use prose and URLs only.
+- Numbered lists work well when the user should reply with a single option index.
+- No emojis; single-level bullets.
+
+Tool calls:
+- Parallelise independent calls; chain only when required.
+- Don’t ask permission for obvious safe defaults — state what you did.
+</provider_hints>

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/providers/openai_reasoning.md

@@ -0,0 +1,22 @@
+<provider_hints>
+You are running on an OpenAI reasoning model (GPT-5+ / o-series), SurfSense **main agent**.
+
+Output style:
+- Be terse and direct. Don't restate the user's request before answering.
+- Don't begin with conversational openers ("Done!", "Got it", "Great question", "Sure thing"). Get to the answer or the action.
+- Match response complexity to the task: simple questions → one-line answer; substantial work → lead with the outcome, then context, then any next steps.
+- No nested bullets — keep lists flat (single level). For options the user can pick by replying with a number, use `1.` `2.` `3.`.
+- Use inline backticks for paths/commands/identifiers; fenced code blocks (with language tags) for multi-line snippets.
+
+Channels (for clients that support them):
+- `commentary` — short progress updates only when they add genuinely new information (a discovery, a tradeoff, a blocker, the start of a non-trivial step). Don't narrate routine reads or obvious next steps.
+- `final` — the completed response. Keep it self-contained; no "see above" / "see below" cross-references.
+
+Tool calls:
+- Parallelise independent tool calls in a single response (`multi_tool_use.parallel` where supported). Only sequence when a later call needs an earlier one's output.
+- Connector or integration execution belongs in **task**, not invented main-agent tools.
+- Don't ask permission ("Should I proceed?", "Do you want me to…?"). Pick the most reasonable default, do it, and state what you did.
+
+Autonomy:
+- Persist until the task is fully resolved within the current turn whenever feasible — within tools you actually have; delegate the rest via **task**.
+</provider_hints>

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/tools/__init__.py

@@ -0,0 +1 @@
+

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/tools/_preamble.md

@@ -0,0 +1,9 @@
+<tools>
+You have access to the following **SurfSense** tools (main-agent scope only):
+
+IMPORTANT: You can ONLY use the tools listed below. Anything else — connectors,
+deliverables, or multi-step integration work — goes through **task**, not as a
+tool in this list.
+
+Do NOT claim you can use a capability if it is not listed here.
+

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/tools/scrape_webpage.md

@@ -0,0 +1,10 @@
+
+- scrape_webpage: Fetch and extract readable content from a single HTTP(S) URL.
+  - Use when the user wants the *actual page body* (article, table, dashboard snapshot), not just search snippets.
+  - Try the tool when a URL is given or referenced; don’t refuse without attempting unless the URL is clearly unsafe/invalid.
+  - Args:
+    - url: Page to fetch
+    - max_length: Cap on returned characters (default: 50000)
+  - Returns: Title, metadata, and markdown-ish body.
+  - Summarize clearly afterward; link back with `[label](url)`.
+  - If indexed workspace material is insufficient and the user points at a public URL, scraping is appropriate — still not a substitute for **task** on private connectors.

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/tools/search_surfsense_docs.md

@@ -0,0 +1,9 @@
+
+- search_surfsense_docs: Search official SurfSense documentation (product help).
+  - Use when the user asks how SurfSense works, setup, connectors at a high level, configuration, etc.
+  - Not a substitute for **task** when they need actions inside Gmail/Slack/Jira/etc.
+  - Args:
+    - query: What to look up in SurfSense docs
+    - top_k: Number of chunks to retrieve (default: 10)
+  - Returns: Doc excerpts; chunk ids may appear for attribution — follow the **citation**
+    instructions block above when citations are enabled; otherwise summarize without `[citation:…]`.

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/tools/update_memory_private.md

@@ -0,0 +1,12 @@
+
+- update_memory: Curate the **personal** long-term memory document for this user.
+  - Current memory (if any) appears in `<user_memory>` with usage vs limit.
+  - Call when the user asks to remember/forget, or shares durable facts/preferences/instructions.
+  - Use the first name from `<user_name>` when writing entries — write “Alex prefers…” not “The user prefers…”.
+    Do not store the name alone as a memory entry.
+  - Skip ephemeral chat noise (one-off q/a, greetings, session logistics).
+  - Args:
+    - updated_memory: FULL replacement markdown (merge and curate — don’t only append).
+  - Formatting rules:
+    - Bullets: `- (YYYY-MM-DD) [marker] text` with markers `[fact]`, `[pref]`, `[instr]` (priority when trimming: instr > pref > fact).
+    - Each bullet under a short `##` heading; keep total size under the limit shown in `<user_memory>`.

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/tools/update_memory_team.md

@@ -0,0 +1,26 @@
+
+- update_memory: Update the team's shared memory document for this search space.
+  - Your current team memory is already in <team_memory> in your context.  The `chars`
+    and `limit` attributes show current usage and the maximum allowed size.
+  - This is the team's curated long-term memory — decisions, conventions, key facts.
+  - NEVER store personal memory in team memory (e.g. personal bio, individual
+    preferences, or user-only standing instructions).
+  - Call update_memory when:
+    * A team member explicitly asks to remember or forget something
+    * The conversation surfaces durable team decisions, conventions, or facts
+      that will matter in future conversations
+  - Do not store short-lived or ephemeral info: one-off questions, greetings,
+    session logistics, or things that only matter for the current task.
+  - Args:
+    - updated_memory: The FULL updated markdown document (not a diff).
+      Merge new facts with existing ones, update contradictions, remove outdated entries.
+      Treat every update as a curation pass — consolidate, don't just append.
+  - Every bullet MUST use this format: - (YYYY-MM-DD) [fact] text
+    Team memory uses ONLY the [fact] marker. Never use [pref] or [instr] in team memory.
+  - Keep it concise and well under the character limit shown in <team_memory>.
+  - Every entry MUST be under a `##` heading. Keep heading names short (2-3 words) and
+    natural. Organize by context — e.g. what the team decided, current architecture,
+    active processes. Create, split, or merge headings freely as the memory grows.
+  - Each entry MUST be a single bullet point. Be descriptive but concise — include relevant
+    details and context rather than just a few words.
+  - During consolidation, prioritize keeping: decisions/conventions > key facts > current priorities.

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/tools/web_search.md

@@ -0,0 +1,10 @@
+
+- web_search: Live public-web search (whatever search backends the workspace configured).
+  - Use for current events, prices, weather, news, or anything needing fresh public web data.
+  - For those queries, call this tool rather than guessing from memory or claiming you lack network access.
+  - If results are thin, say so and offer to refine the query.
+  - Args:
+    - query: Specific search terms
+    - top_k: Max hits (default: 10, max: 50)
+  - If snippets are too shallow, follow up with **scrape_webpage** on the best URL.
+  - Present sources with readable markdown links `[label](url)` — never bare URLs.

surfsense_backend/app/agents/multi_agent_chat/main_agent/tools/__init__.py

@@ -0,0 +1,10 @@
+"""Main-agent SurfSense tool allowlist."""
+
+from __future__ import annotations
+
+from .index import (
+    MAIN_AGENT_SURFSENSE_TOOL_NAMES,
+    MAIN_AGENT_SURFSENSE_TOOL_NAMES_ORDERED,
+)
+
+__all__ = ["MAIN_AGENT_SURFSENSE_TOOL_NAMES", "MAIN_AGENT_SURFSENSE_TOOL_NAMES_ORDERED"]

surfsense_backend/app/agents/multi_agent_chat/main_agent/tools/index.py

@@ -0,0 +1,17 @@
+"""Main-agent SurfSense builtin tool names (not full ``new_chat``).
+
+Connector integrations, MCP, deliverables, etc. are delegated via ``task`` subagents.
+"""
+
+from __future__ import annotations
+
+MAIN_AGENT_SURFSENSE_TOOL_NAMES_ORDERED: tuple[str, ...] = (
+    "search_surfsense_docs",
+    "web_search",
+    "scrape_webpage",
+    "update_memory",
+)
+
+MAIN_AGENT_SURFSENSE_TOOL_NAMES: frozenset[str] = frozenset(
+    MAIN_AGENT_SURFSENSE_TOOL_NAMES_ORDERED,
+)

surfsense_backend/app/agents/multi_agent_chat/subagents/__init__.py

@@ -0,0 +1,19 @@
+"""Registry-backed subagent builders and helpers."""
+
+from __future__ import annotations
+
+from .registry import (
+    SUBAGENT_BUILDERS_BY_NAME,
+    SubagentBuilder,
+    build_subagents,
+    get_subagents_to_exclude,
+    main_prompt_registry_subagent_lines,
+)
+
+__all__ = [
+    "SUBAGENT_BUILDERS_BY_NAME",
+    "SubagentBuilder",
+    "build_subagents",
+    "get_subagents_to_exclude",
+    "main_prompt_registry_subagent_lines",
+]

surfsense_backend/app/agents/multi_agent_chat/subagents/builtins/deliverables/agent.py

@@ -0,0 +1,55 @@
+"""`deliverables` route: ``SubAgent`` spec for deepagents."""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+from typing import Any
+
+from deepagents import SubAgent
+from langchain_core.language_models import BaseChatModel
+
+from app.agents.multi_agent_chat.subagents.shared.md_file_reader import (
+    read_md_file,
+)
+from app.agents.multi_agent_chat.subagents.shared.permissions import (
+    ToolsPermissions,
+    merge_tools_permissions,
+    middleware_gated_interrupt_on,
+)
+from app.agents.multi_agent_chat.subagents.shared.subagent_builder import (
+    pack_subagent,
+)
+
+from .tools.index import load_tools
+
+NAME = "deliverables"
+
+
+def build_subagent(
+    *,
+    dependencies: dict[str, Any],
+    model: BaseChatModel | None = None,
+    extra_middleware: Sequence[Any] | None = None,
+    extra_tools_bucket: ToolsPermissions | None = None,
+) -> SubAgent:
+    buckets = load_tools(dependencies=dependencies)
+    merged_tools_bucket = merge_tools_permissions(buckets, extra_tools_bucket)
+    tools = [
+        row["tool"]
+        for row in (*merged_tools_bucket["allow"], *merged_tools_bucket["ask"])
+        if row.get("tool") is not None
+    ]
+    interrupt_on = middleware_gated_interrupt_on(merged_tools_bucket)
+    description = read_md_file(__package__, "description").strip()
+    if not description:
+        description = "Handles deliverables tasks for this workspace."
+    system_prompt = read_md_file(__package__, "system_prompt").strip()
+    return pack_subagent(
+        name=NAME,
+        description=description,
+        system_prompt=system_prompt,
+        tools=tools,
+        interrupt_on=interrupt_on,
+        model=model,
+        extra_middleware=extra_middleware,
+    )

surfsense_backend/app/agents/multi_agent_chat/subagents/builtins/deliverables/description.md

@@ -0,0 +1 @@
+Use for deliverables and shareable artifacts: generated reports, podcasts, video presentations, resumes, and images—not for routine lookups or single small edits elsewhere.

surfsense_backend/app/agents/multi_agent_chat/subagents/builtins/deliverables/system_prompt.md

@@ -0,0 +1,55 @@
+You are the SurfSense deliverables operations sub-agent.
+You receive delegated instructions from a supervisor agent and return structured results for supervisor synthesis.
+
+<goal>
+Produce **deliverables**: shareable **artifacts** the user keeps (reports, slide-style video presentations, podcasts, resumes, images). Use explicit constraints and reliable proof of what was generated.
+</goal>
+
+<available_tools>
+- `generate_report`
+- `generate_podcast`
+- `generate_video_presentation`
+- `generate_resume`
+- `generate_image`
+</available_tools>
+
+<tool_policy>
+- Use only tools in `<available_tools>`.
+- Require essential generation constraints (audience, format, tone, core content).
+- If critical constraints are missing, return `status=blocked` with `missing_fields`.
+- Never claim artifact generation success without tool confirmation.
+</tool_policy>
+
+<out_of_scope>
+- Do not perform connector data mutations unrelated to artifact generation.
+</out_of_scope>
+
+<safety>
+- Avoid generating artifacts with missing critical constraints.
+- Prefer one complete artifact over partial multi-artifact output.
+</safety>
+
+<failure_policy>
+- On generation failure, return `status=error` with best retry guidance.
+- On missing constraints, return `status=blocked` with required fields.
+</failure_policy>
+
+<output_contract>
+Return **only** one JSON object (no markdown/prose):
+{
+  "status": "success" | "partial" | "blocked" | "error",
+  "action_summary": string,
+  "evidence": {
+    "artifact_type": "report" | "podcast" | "video_presentation" | "resume" | "image" | null,
+    "artifact_id": string | null,
+    "artifact_location": string | null
+  },
+  "next_step": string | null,
+  "missing_fields": string[] | null,
+  "assumptions": string[] | null
+}
+Rules:
+- `status=success` -> `next_step=null`, `missing_fields=null`.
+- `status=partial|blocked|error` -> `next_step` must be non-null.
+- `status=blocked` due to missing required inputs -> `missing_fields` must be non-null.
+</output_contract>

surfsense_backend/app/agents/multi_agent_chat/subagents/builtins/deliverables/tools/__init__.py

@@ -0,0 +1,15 @@
+"""Deliverable generators: reports, podcasts, video decks, resumes, images."""
+
+from .generate_image import create_generate_image_tool
+from .podcast import create_generate_podcast_tool
+from .report import create_generate_report_tool
+from .resume import create_generate_resume_tool
+from .video_presentation import create_generate_video_presentation_tool
+
+__all__ = [
+    "create_generate_image_tool",
+    "create_generate_podcast_tool",
+    "create_generate_report_tool",
+    "create_generate_resume_tool",
+    "create_generate_video_presentation_tool",
+]

surfsense_backend/app/agents/multi_agent_chat/subagents/builtins/deliverables/tools/generate_image.py

@@ -0,0 +1,247 @@
+"""Image generation via litellm; resolves model config from the search space and returns UI-ready payloads."""
+
+import hashlib
+import logging
+from typing import Any
+
+from langchain_core.tools import tool
+from litellm import aimage_generation
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.config import config
+from app.db import (
+    ImageGeneration,
+    ImageGenerationConfig,
+    SearchSpace,
+    shielded_async_session,
+)
+from app.services.image_gen_router_service import (
+    IMAGE_GEN_AUTO_MODE_ID,
+    ImageGenRouterService,
+    is_image_gen_auto_mode,
+)
+from app.utils.signed_image_urls import generate_image_token
+
+logger = logging.getLogger(__name__)
+
+# Provider mapping (same as routes)
+_PROVIDER_MAP = {
+    "OPENAI": "openai",
+    "AZURE_OPENAI": "azure",
+    "GOOGLE": "gemini",
+    "VERTEX_AI": "vertex_ai",
+    "BEDROCK": "bedrock",
+    "RECRAFT": "recraft",
+    "OPENROUTER": "openrouter",
+    "XINFERENCE": "xinference",
+    "NSCALE": "nscale",
+}
+
+
+def _build_model_string(
+    provider: str, model_name: str, custom_provider: str | None
+) -> str:
+    if custom_provider:
+        return f"{custom_provider}/{model_name}"
+    prefix = _PROVIDER_MAP.get(provider.upper(), provider.lower())
+    return f"{prefix}/{model_name}"
+
+
+def _get_global_image_gen_config(config_id: int) -> dict | None:
+    """Get a global image gen config by negative ID."""
+    for cfg in config.GLOBAL_IMAGE_GEN_CONFIGS:
+        if cfg.get("id") == config_id:
+            return cfg
+    return None
+
+
+def create_generate_image_tool(
+    search_space_id: int,
+    db_session: AsyncSession,
+):
+    """Create ``generate_image`` with bound search space; DB work uses a per-call session."""
+    del db_session  # use a fresh per-call session, see below
+
+    @tool
+    async def generate_image(
+        prompt: str,
+        n: int = 1,
+    ) -> dict[str, Any]:
+        """
+        Generate an image from a text description using AI image models.
+
+        Use this tool when the user asks you to create, generate, draw, or make an image.
+        The generated image will be displayed directly in the chat.
+
+        Args:
+            prompt: A detailed text description of the image to generate.
+                    Be specific about subject, style, colors, composition, and mood.
+            n: Number of images to generate (1-4). Default: 1
+
+        Returns:
+            A dictionary containing the generated image(s) for display in the chat.
+        """
+        try:
+            # Use a per-call session so concurrent tool calls don't share an
+            # AsyncSession (which is not concurrency-safe). The streaming
+            # task's session is shared across every tool; without isolation,
+            # autoflushes from a concurrent writer poison this tool too.
+            async with shielded_async_session() as session:
+                result = await session.execute(
+                    select(SearchSpace).filter(SearchSpace.id == search_space_id)
+                )
+                search_space = result.scalars().first()
+                if not search_space:
+                    return {"error": "Search space not found"}
+
+                config_id = (
+                    search_space.image_generation_config_id or IMAGE_GEN_AUTO_MODE_ID
+                )
+
+                # Build generation kwargs
+                # NOTE: size, quality, and style are intentionally NOT passed.
+                # Different models support different values for these params
+                # (e.g. DALL-E 3 wants "hd"/"standard" for quality while
+                # gpt-image-1 wants "high"/"medium"/"low"; size options also
+                # differ). Letting the model use its own defaults avoids errors.
+                gen_kwargs: dict[str, Any] = {}
+                if n is not None and n > 1:
+                    gen_kwargs["n"] = n
+
+                # Call litellm based on config type
+                if is_image_gen_auto_mode(config_id):
+                    if not ImageGenRouterService.is_initialized():
+                        return {
+                            "error": "No image generation models configured. "
+                            "Please add an image model in Settings > Image Models."
+                        }
+                    response = await ImageGenRouterService.aimage_generation(
+                        prompt=prompt, model="auto", **gen_kwargs
+                    )
+                elif config_id < 0:
+                    cfg = _get_global_image_gen_config(config_id)
+                    if not cfg:
+                        return {
+                            "error": f"Image generation config {config_id} not found"
+                        }
+
+                    model_string = _build_model_string(
+                        cfg.get("provider", ""),
+                        cfg["model_name"],
+                        cfg.get("custom_provider"),
+                    )
+                    gen_kwargs["api_key"] = cfg.get("api_key")
+                    if cfg.get("api_base"):
+                        gen_kwargs["api_base"] = cfg["api_base"]
+                    if cfg.get("api_version"):
+                        gen_kwargs["api_version"] = cfg["api_version"]
+                    if cfg.get("litellm_params"):
+                        gen_kwargs.update(cfg["litellm_params"])
+
+                    response = await aimage_generation(
+                        prompt=prompt, model=model_string, **gen_kwargs
+                    )
+                else:
+                    # Positive ID = user-created ImageGenerationConfig
+                    cfg_result = await session.execute(
+                        select(ImageGenerationConfig).filter(
+                            ImageGenerationConfig.id == config_id
+                        )
+                    )
+                    db_cfg = cfg_result.scalars().first()
+                    if not db_cfg:
+                        return {
+                            "error": f"Image generation config {config_id} not found"
+                        }
+
+                    model_string = _build_model_string(
+                        db_cfg.provider.value,
+                        db_cfg.model_name,
+                        db_cfg.custom_provider,
+                    )
+                    gen_kwargs["api_key"] = db_cfg.api_key
+                    if db_cfg.api_base:
+                        gen_kwargs["api_base"] = db_cfg.api_base
+                    if db_cfg.api_version:
+                        gen_kwargs["api_version"] = db_cfg.api_version
+                    if db_cfg.litellm_params:
+                        gen_kwargs.update(db_cfg.litellm_params)
+
+                    response = await aimage_generation(
+                        prompt=prompt, model=model_string, **gen_kwargs
+                    )
+
+                # Parse the response and store in DB
+                response_dict = (
+                    response.model_dump()
+                    if hasattr(response, "model_dump")
+                    else dict(response)
+                )
+
+                # Generate a random access token for this image
+                access_token = generate_image_token()
+
+                # Save to image_generations table for history
+                db_image_gen = ImageGeneration(
+                    prompt=prompt,
+                    model=getattr(response, "_hidden_params", {}).get("model"),
+                    n=n,
+                    image_generation_config_id=config_id,
+                    response_data=response_dict,
+                    search_space_id=search_space_id,
+                    access_token=access_token,
+                )
+                session.add(db_image_gen)
+                await session.commit()
+                await session.refresh(db_image_gen)
+                db_image_gen_id = db_image_gen.id
+
+            # Extract image URLs from response
+            images = response_dict.get("data", [])
+            if not images:
+                return {"error": "No images were generated"}
+
+            first_image = images[0]
+            revised_prompt = first_image.get("revised_prompt", prompt)
+
+            # Resolve image URL:
+            # - If the API returned a URL, use it directly.
+            # - If the API returned b64_json (e.g. gpt-image-1), serve the
+            #   image through our backend endpoint to avoid bloating the
+            #   LLM context with megabytes of base64 data.
+            if first_image.get("url"):
+                image_url = first_image["url"]
+            elif first_image.get("b64_json"):
+                backend_url = config.BACKEND_URL or "http://localhost:8000"
+                image_url = (
+                    f"{backend_url}/api/v1/image-generations/"
+                    f"{db_image_gen_id}/image?token={access_token}"
+                )
+            else:
+                return {"error": "No displayable image data in the response"}
+
+            image_id = f"image-{hashlib.md5(image_url.encode()).hexdigest()[:12]}"
+
+            return {
+                "id": image_id,
+                "assetId": image_url,
+                "src": image_url,
+                "alt": revised_prompt or prompt,
+                "title": "Generated Image",
+                "description": revised_prompt if revised_prompt != prompt else None,
+                "domain": "ai-generated",
+                "ratio": "auto",
+                "generated": True,
+                "prompt": prompt,
+                "image_count": len(images),
+            }
+
+        except Exception as e:
+            logger.exception("Image generation failed in tool")
+            return {
+                "error": f"Image generation failed: {e!s}",
+                "prompt": prompt,
+            }
+
+    return generate_image

surfsense_backend/app/agents/multi_agent_chat/subagents/builtins/deliverables/tools/index.py

@@ -0,0 +1,54 @@
+from __future__ import annotations
+
+from typing import Any
+
+from app.agents.multi_agent_chat.subagents.shared.permissions import (
+    ToolsPermissions,
+)
+
+from .generate_image import create_generate_image_tool
+from .podcast import create_generate_podcast_tool
+from .report import create_generate_report_tool
+from .resume import create_generate_resume_tool
+from .video_presentation import create_generate_video_presentation_tool
+
+
+def load_tools(
+    *, dependencies: dict[str, Any] | None = None, **kwargs: Any
+) -> ToolsPermissions:
+    resolved_dependencies = {**(dependencies or {}), **kwargs}
+    podcast = create_generate_podcast_tool(
+        search_space_id=resolved_dependencies["search_space_id"],
+        db_session=resolved_dependencies["db_session"],
+        thread_id=resolved_dependencies["thread_id"],
+    )
+    video = create_generate_video_presentation_tool(
+        search_space_id=resolved_dependencies["search_space_id"],
+        db_session=resolved_dependencies["db_session"],
+        thread_id=resolved_dependencies["thread_id"],
+    )
+    report = create_generate_report_tool(
+        search_space_id=resolved_dependencies["search_space_id"],
+        thread_id=resolved_dependencies["thread_id"],
+        connector_service=resolved_dependencies.get("connector_service"),
+        available_connectors=resolved_dependencies.get("available_connectors"),
+        available_document_types=resolved_dependencies.get("available_document_types"),
+    )
+    resume = create_generate_resume_tool(
+        search_space_id=resolved_dependencies["search_space_id"],
+        thread_id=resolved_dependencies["thread_id"],
+    )
+    image = create_generate_image_tool(
+        search_space_id=resolved_dependencies["search_space_id"],
+        db_session=resolved_dependencies["db_session"],
+    )
+    return {
+        "allow": [
+            {"name": getattr(podcast, "name", "") or "", "tool": podcast},
+            {"name": getattr(video, "name", "") or "", "tool": video},
+            {"name": getattr(report, "name", "") or "", "tool": report},
+            {"name": getattr(resume, "name", "") or "", "tool": resume},
+            {"name": getattr(image, "name", "") or "", "tool": image},
+        ],
+        "ask": [],
+    }

surfsense_backend/app/agents/multi_agent_chat/subagents/builtins/deliverables/tools/podcast.py

@@ -0,0 +1,92 @@
+"""Factory for a podcast-generation tool that queues background work and returns an ID for polling."""
+
+from typing import Any
+
+from langchain_core.tools import tool
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.db import Podcast, PodcastStatus, shielded_async_session
+
+
+def create_generate_podcast_tool(
+    search_space_id: int,
+    db_session: AsyncSession,
+    thread_id: int | None = None,
+):
+    """Create ``generate_podcast`` with bound search space and thread; DB writes use a tool-local session."""
+    del db_session  # writes use a fresh tool-local session, see below
+
+    @tool
+    async def generate_podcast(
+        source_content: str,
+        podcast_title: str = "SurfSense Podcast",
+        user_prompt: str | None = None,
+    ) -> dict[str, Any]:
+        """
+        Generate a podcast from the provided content.
+
+        Use this tool when the user asks to create, generate, or make a podcast.
+        Common triggers include phrases like:
+        - "Give me a podcast about this"
+        - "Create a podcast from this conversation"
+        - "Generate a podcast summary"
+        - "Make a podcast about..."
+        - "Turn this into a podcast"
+
+        Args:
+            source_content: The text content to convert into a podcast.
+            podcast_title: Title for the podcast (default: "SurfSense Podcast")
+            user_prompt: Optional instructions for podcast style, tone, or format.
+
+        Returns:
+            A dictionary containing:
+            - status: PodcastStatus value (pending, generating, or failed)
+            - podcast_id: The podcast ID for polling (when status is pending or generating)
+            - title: The podcast title
+            - message: Status message (or "error" field if status is failed)
+        """
+        try:
+            # One DB session per tool call so parallel invocations never share an AsyncSession.
+            async with shielded_async_session() as session:
+                podcast = Podcast(
+                    title=podcast_title,
+                    status=PodcastStatus.PENDING,
+                    search_space_id=search_space_id,
+                    thread_id=thread_id,
+                )
+                session.add(podcast)
+                await session.commit()
+                await session.refresh(podcast)
+                podcast_id = podcast.id
+
+            from app.tasks.celery_tasks.podcast_tasks import (
+                generate_content_podcast_task,
+            )
+
+            task = generate_content_podcast_task.delay(
+                podcast_id=podcast_id,
+                source_content=source_content,
+                search_space_id=search_space_id,
+                user_prompt=user_prompt,
+            )
+
+            print(f"[generate_podcast] Created podcast {podcast_id}, task: {task.id}")
+
+            return {
+                "status": PodcastStatus.PENDING.value,
+                "podcast_id": podcast_id,
+                "title": podcast_title,
+                "message": "Podcast generation started. This may take a few minutes.",
+            }
+
+        except Exception as e:
+            error_message = str(e)
+            print(f"[generate_podcast] Error: {error_message}")
+            return {
+                "status": PodcastStatus.FAILED.value,
+                "error": error_message,
+                "title": podcast_title,
+                "podcast_id": None,
+            }
+
+    return generate_podcast

surfsense_backend/app/agents/multi_agent_chat/subagents/builtins/deliverables/tools/resume.py

@@ -0,0 +1,799 @@
+"""Resume as Typst: LLM fills the body; backend prepends a template from ``_TEMPLATES`` and compiles."""
+
+import io
+import logging
+import re
+from datetime import UTC, datetime
+from typing import Any
+
+import pypdf
+import typst
+from langchain_core.callbacks import dispatch_custom_event
+from langchain_core.messages import HumanMessage
+from langchain_core.tools import tool
+
+from app.db import Report, shielded_async_session
+from app.services.llm_service import get_document_summary_llm
+
+logger = logging.getLogger(__name__)
+
+
+# ─── Template Registry ───────────────────────────────────────────────────────
+# Each template defines:
+#   header              - Typst import + show rule with {name}, {year}, {month}, {day} placeholders
+#   component_reference - component docs shown to the LLM
+#   rules               - generation rules for the LLM
+
+_TEMPLATES: dict[str, dict[str, str]] = {
+    "classic": {
+        "header": """\
+#import "@preview/rendercv:0.3.0": *
+
+#show: rendercv.with(
+  name: "{name}",
+  title: "{name} - Resume",
+  footer: context {{ [#emph[{name} -- #str(here().page())\\/#str(counter(page).final().first())]] }},
+  top-note: [ #emph[Last updated in {month_name} {year}] ],
+  locale-catalog-language: "en",
+  text-direction: ltr,
+  page-size: "us-letter",
+  page-top-margin: 0.7in,
+  page-bottom-margin: 0.7in,
+  page-left-margin: 0.7in,
+  page-right-margin: 0.7in,
+  page-show-footer: false,
+  page-show-top-note: true,
+  colors-body: rgb(0, 0, 0),
+  colors-name: rgb(0, 0, 0),
+  colors-headline: rgb(0, 0, 0),
+  colors-connections: rgb(0, 0, 0),
+  colors-section-titles: rgb(0, 0, 0),
+  colors-links: rgb(0, 0, 0),
+  colors-footer: rgb(128, 128, 128),
+  colors-top-note: rgb(128, 128, 128),
+  typography-line-spacing: 0.6em,
+  typography-alignment: "justified",
+  typography-date-and-location-column-alignment: right,
+  typography-font-family-body: "XCharter",
+  typography-font-family-name: "XCharter",
+  typography-font-family-headline: "XCharter",
+  typography-font-family-connections: "XCharter",
+  typography-font-family-section-titles: "XCharter",
+  typography-font-size-body: 10pt,
+  typography-font-size-name: 25pt,
+  typography-font-size-headline: 10pt,
+  typography-font-size-connections: 10pt,
+  typography-font-size-section-titles: 1.2em,
+  typography-small-caps-name: false,
+  typography-small-caps-headline: false,
+  typography-small-caps-connections: false,
+  typography-small-caps-section-titles: false,
+  typography-bold-name: false,
+  typography-bold-headline: false,
+  typography-bold-connections: false,
+  typography-bold-section-titles: true,
+  links-underline: true,
+  links-show-external-link-icon: false,
+  header-alignment: center,
+  header-photo-width: 3.5cm,
+  header-space-below-name: 0.7cm,
+  header-space-below-headline: 0.7cm,
+  header-space-below-connections: 0.7cm,
+  header-connections-hyperlink: true,
+  header-connections-show-icons: false,
+  header-connections-display-urls-instead-of-usernames: true,
+  header-connections-separator: "|",
+  header-connections-space-between-connections: 0.5cm,
+  section-titles-type: "with_full_line",
+  section-titles-line-thickness: 0.5pt,
+  section-titles-space-above: 0.5cm,
+  section-titles-space-below: 0.3cm,
+  sections-allow-page-break: true,
+  sections-space-between-text-based-entries: 0.15cm,
+  sections-space-between-regular-entries: 0.42cm,
+  entries-date-and-location-width: 4.15cm,
+  entries-side-space: 0cm,
+  entries-space-between-columns: 0.1cm,
+  entries-allow-page-break: false,
+  entries-short-second-row: false,
+  entries-degree-width: 1cm,
+  entries-summary-space-left: 0cm,
+  entries-summary-space-above: 0.08cm,
+  entries-highlights-bullet: text(13pt, [\\u{2022}], baseline: -0.6pt),
+  entries-highlights-nested-bullet: text(13pt, [\\u{2022}], baseline: -0.6pt),
+  entries-highlights-space-left: 0cm,
+  entries-highlights-space-above: 0.08cm,
+  entries-highlights-space-between-items: 0.02cm,
+  entries-highlights-space-between-bullet-and-text: 0.3em,
+  date: datetime(
+    year: {year},
+    month: {month},
+    day: {day},
+  ),
+)
+
+""",
+        "component_reference": """\
+Available components (use ONLY these):
+
+= Full Name                              // Top-level heading — person's full name
+
+#connections(                            // Contact info row (pipe-separated)
+  [City, Country],
+  [#link("mailto:email@example.com", icon: false, if-underline: false, if-color: false)[email\\@example.com]],
+  [#link("https://linkedin.com/in/user", icon: false, if-underline: false, if-color: false)[linkedin.com\\/in\\/user]],
+  [#link("https://github.com/user", icon: false, if-underline: false, if-color: false)[github.com\\/user]],
+)
+
+== Section Title                         // Section heading (arbitrary name)
+
+#regular-entry(                          // Work experience, projects, publications, etc.
+  [
+    #strong[Role/Title], Company Name -- Location
+  ],
+  [
+    Start -- End
+  ],
+  main-column-second-row: [
+    - Achievement or responsibility
+    - Another bullet point
+  ],
+)
+
+#education-entry(                        // Education entries
+  [
+    #strong[Institution], Degree in Field -- Location
+  ],
+  [
+    Start -- End
+  ],
+  main-column-second-row: [
+    - GPA, honours, relevant coursework
+  ],
+)
+
+#summary([Short paragraph summary])     // Optional summary inside an entry
+#content-area([Free-form content])       // Freeform text block
+
+For skills sections, use one bullet per category label:
+- #strong[Category:] item1, item2, item3
+
+For simple list sections (e.g. Honors), use plain bullet points:
+- Item one
+- Item two
+""",
+        "rules": """\
+RULES:
+- Do NOT include any #import or #show lines. Start directly with = Full Name.
+- Output ONLY valid Typst content. No explanatory text before or after.
+- Do NOT wrap output in ```typst code fences.
+- The = heading MUST use the person's COMPLETE full name exactly as provided. NEVER shorten or abbreviate.
+- Escape @ symbols inside link labels with a backslash: email\\@example.com
+- Escape forward slashes in link display text: linkedin.com\\/in\\/user
+- Every section MUST use == heading.
+- Use #regular-entry() for experience, projects, publications, certifications, and similar entries.
+- Use #education-entry() for education.
+- For skills sections, use one bullet line per category with a bold label.
+- Keep content professional, concise, and achievement-oriented.
+- Use action verbs for bullet points (Led, Built, Designed, Reduced, etc.).
+- This template works for ALL professions — adapt sections to the user's field.
+- Default behavior should prioritize concise one-page content.
+""",
+    },
+}
+
+DEFAULT_TEMPLATE = "classic"
+MIN_RESUME_PAGES = 1
+MAX_RESUME_PAGES = 5
+MAX_COMPRESSION_ATTEMPTS = 2
+
+
+# ─── Template Helpers ─────────────────────────────────────────────────────────
+
+
+def _get_template(template_id: str | None = None) -> dict[str, str]:
+    """Get a template by ID, falling back to default."""
+    return _TEMPLATES.get(template_id or DEFAULT_TEMPLATE, _TEMPLATES[DEFAULT_TEMPLATE])
+
+
+_MONTH_NAMES = [
+    "",
+    "Jan",
+    "Feb",
+    "Mar",
+    "Apr",
+    "May",
+    "Jun",
+    "Jul",
+    "Aug",
+    "Sep",
+    "Oct",
+    "Nov",
+    "Dec",
+]
+
+
+def _build_header(template: dict[str, str], name: str) -> str:
+    """Build the template header with the person's name and current date."""
+    now = datetime.now(tz=UTC)
+    return (
+        template["header"]
+        .replace("{name}", name)
+        .replace("{year}", str(now.year))
+        .replace("{month}", str(now.month))
+        .replace("{day}", str(now.day))
+        .replace("{month_name}", _MONTH_NAMES[now.month])
+    )
+
+
+def _strip_header(full_source: str) -> str:
+    """Strip the import + show rule from stored source to get the body only.
+
+    Finds the closing parenthesis of the rendercv.with(...) block by tracking
+    nesting depth, then returns everything after it.
+    """
+    show_match = re.search(r"#show:\s*rendercv\.with\(", full_source)
+    if not show_match:
+        return full_source
+
+    start = show_match.end()
+    depth = 1
+    i = start
+    while i < len(full_source) and depth > 0:
+        if full_source[i] == "(":
+            depth += 1
+        elif full_source[i] == ")":
+            depth -= 1
+        i += 1
+
+    return full_source[i:].lstrip("\n")
+
+
+def _extract_name(body: str) -> str | None:
+    """Extract the person's full name from the = heading in the body."""
+    match = re.search(r"^=\s+(.+)$", body, re.MULTILINE)
+    return match.group(1).strip() if match else None
+
+
+def _strip_imports(body: str) -> str:
+    """Remove any #import or #show lines the LLM might accidentally include."""
+    lines = body.split("\n")
+    cleaned: list[str] = []
+    skip_show = False
+    depth = 0
+
+    for line in lines:
+        stripped = line.strip()
+
+        if stripped.startswith("#import"):
+            continue
+
+        if skip_show:
+            depth += stripped.count("(") - stripped.count(")")
+            if depth <= 0:
+                skip_show = False
+            continue
+
+        if stripped.startswith("#show:") and "rendercv" in stripped:
+            depth = stripped.count("(") - stripped.count(")")
+            if depth > 0:
+                skip_show = True
+            continue
+
+        cleaned.append(line)
+
+    result = "\n".join(cleaned).strip()
+    return result
+
+
+def _build_llm_reference(template: dict[str, str]) -> str:
+    """Build the LLM prompt reference from a template."""
+    return f"""\
+You MUST output valid Typst content for a resume.
+Do NOT include any #import or #show lines — those are handled automatically.
+Start directly with the = Full Name heading.
+
+{template["component_reference"]}
+
+{template["rules"]}"""
+
+
+# ─── Prompts ─────────────────────────────────────────────────────────────────
+
+_RESUME_PROMPT = """\
+You are an expert resume writer. Generate professional resume content as Typst markup.
+
+{llm_reference}
+
+**User Information:**
+{user_info}
+
+**Target Maximum Pages:** {max_pages}
+
+{user_instructions_section}
+
+Generate the resume content now (starting with = Full Name):
+"""
+
+_REVISION_PROMPT = """\
+You are an expert resume editor. Modify the existing resume according to the instructions.
+Apply ONLY the requested changes — do NOT rewrite sections that are not affected.
+
+{llm_reference}
+
+**Target Maximum Pages:** {max_pages}
+
+**Modification Instructions:** {user_instructions}
+
+**EXISTING RESUME CONTENT:**
+
+{previous_content}
+
+---
+
+Output the complete, updated resume content with the changes applied (starting with = Full Name):
+"""
+
+_FIX_COMPILE_PROMPT = """\
+The resume content you generated failed to compile. Fix the error while preserving all content.
+
+{llm_reference}
+
+**Compilation Error:**
+{error}
+
+**Full Typst Source (for context — error line numbers refer to this):**
+{full_source}
+
+**Your content starts after the template header. Output ONLY the content portion \
+(starting with = Full Name), NOT the #import or #show rule:**
+"""
+
+_COMPRESS_TO_PAGE_LIMIT_PROMPT = """\
+The resume compiles, but it exceeds the maximum allowed page count.
+Compress the resume while preserving high-impact accomplishments and role relevance.
+
+{llm_reference}
+
+**Target Maximum Pages:** {max_pages}
+**Current Page Count:** {actual_pages}
+**Compression Attempt:** {attempt_number}
+
+Compression priorities (in this order):
+1) Keep recent, high-impact, role-relevant bullets.
+2) Remove low-impact or redundant bullets.
+3) Shorten verbose wording while preserving meaning.
+4) Trim older or less relevant details before recent ones.
+
+Return the complete updated Typst content (starting with = Full Name), and keep it at or below the target pages.
+
+**EXISTING RESUME CONTENT:**
+{previous_content}
+"""
+
+
+# ─── Helpers ─────────────────────────────────────────────────────────────────
+
+
+def _strip_typst_fences(text: str) -> str:
+    """Remove wrapping ```typst ... ``` fences that LLMs sometimes add."""
+    stripped = text.strip()
+    m = re.match(r"^(`{3,})(?:typst|typ)?\s*\n", stripped)
+    if m:
+        fence = m.group(1)
+        if stripped.endswith(fence):
+            stripped = stripped[m.end() :]
+            stripped = stripped[: -len(fence)].rstrip()
+    return stripped
+
+
+def _compile_typst(source: str) -> bytes:
+    """Compile Typst source to PDF bytes. Raises on failure."""
+    return typst.compile(source.encode("utf-8"))
+
+
+def _count_pdf_pages(pdf_bytes: bytes) -> int:
+    """Count the number of pages in compiled PDF bytes."""
+    with io.BytesIO(pdf_bytes) as pdf_stream:
+        reader = pypdf.PdfReader(pdf_stream)
+        return len(reader.pages)
+
+
+def _validate_max_pages(max_pages: int) -> int:
+    """Validate and normalize max_pages input."""
+    if MIN_RESUME_PAGES <= max_pages <= MAX_RESUME_PAGES:
+        return max_pages
+    msg = (
+        f"max_pages must be between {MIN_RESUME_PAGES} and "
+        f"{MAX_RESUME_PAGES}. Received: {max_pages}"
+    )
+    raise ValueError(msg)
+
+
+# ─── Tool Factory ───────────────────────────────────────────────────────────
+
+
+def create_generate_resume_tool(
+    search_space_id: int,
+    thread_id: int | None = None,
+):
+    """
+    Factory function to create the generate_resume tool.
+
+    Generates a Typst-based resume, validates it via compilation,
+    and stores the source in the Report table with content_type='typst'.
+    The LLM generates only the content body; the template header is
+    prepended by the backend.
+    """
+
+    @tool
+    async def generate_resume(
+        user_info: str,
+        user_instructions: str | None = None,
+        parent_report_id: int | None = None,
+        max_pages: int = 1,
+    ) -> dict[str, Any]:
+        """
+        Generate a professional resume as a Typst document.
+
+        Use this tool when the user asks to create, build, generate, write,
+        or draft a resume or CV. Also use it when the user wants to modify,
+        update, or revise an existing resume generated in this conversation.
+
+        Trigger phrases include:
+        - "build me a resume", "create my resume", "generate a CV"
+        - "update my resume", "change my title", "add my new job"
+        - "make my resume more concise", "reformat my resume"
+
+        Do NOT use this tool for:
+        - General questions about resumes or career advice
+        - Reviewing or critiquing a resume without changes
+        - Cover letters (use generate_report instead)
+
+        VERSIONING — parent_report_id:
+        - Set parent_report_id when the user wants to MODIFY an existing
+          resume that was already generated in this conversation.
+        - Leave as None for new resumes.
+
+        Args:
+            user_info: The user's resume content — work experience,
+                education, skills, contact info, etc. Can be structured
+                or unstructured text.
+            user_instructions: Optional style or content preferences
+                (e.g. "emphasize leadership", "keep it to one page",
+                "use a modern style"). For revisions, describe what to change.
+            parent_report_id: ID of a previous resume to revise (creates
+                new version in the same version group).
+            max_pages: Maximum number of pages for the generated resume.
+                Defaults to 1. Allowed range: 1-5.
+
+        Returns:
+            Dict with status, report_id, title, and content_type.
+        """
+        report_group_id: int | None = None
+        parent_content: str | None = None
+
+        template = _get_template()
+        llm_reference = _build_llm_reference(template)
+
+        async def _save_failed_report(error_msg: str) -> int | None:
+            try:
+                async with shielded_async_session() as session:
+                    failed = Report(
+                        title="Resume",
+                        content=None,
+                        content_type="typst",
+                        report_metadata={
+                            "status": "failed",
+                            "error_message": error_msg,
+                        },
+                        report_style="resume",
+                        search_space_id=search_space_id,
+                        thread_id=thread_id,
+                        report_group_id=report_group_id,
+                    )
+                    session.add(failed)
+                    await session.commit()
+                    await session.refresh(failed)
+                    if not failed.report_group_id:
+                        failed.report_group_id = failed.id
+                        await session.commit()
+                    logger.info(
+                        f"[generate_resume] Saved failed report {failed.id}: {error_msg}"
+                    )
+                    return failed.id
+            except Exception:
+                logger.exception(
+                    "[generate_resume] Could not persist failed report row"
+                )
+                return None
+
+        try:
+            try:
+                validated_max_pages = _validate_max_pages(max_pages)
+            except ValueError as e:
+                error_msg = str(e)
+                report_id = await _save_failed_report(error_msg)
+                return {
+                    "status": "failed",
+                    "error": error_msg,
+                    "report_id": report_id,
+                    "title": "Resume",
+                    "content_type": "typst",
+                }
+
+            # ── Phase 1: READ ─────────────────────────────────────────────
+            async with shielded_async_session() as read_session:
+                if parent_report_id:
+                    parent_report = await read_session.get(Report, parent_report_id)
+                    if parent_report:
+                        report_group_id = parent_report.report_group_id
+                        parent_content = parent_report.content
+                        logger.info(
+                            f"[generate_resume] Revising from parent {parent_report_id} "
+                            f"(group {report_group_id})"
+                        )
+
+                llm = await get_document_summary_llm(read_session, search_space_id)
+
+            if not llm:
+                error_msg = (
+                    "No LLM configured. Please configure a language model in Settings."
+                )
+                report_id = await _save_failed_report(error_msg)
+                return {
+                    "status": "failed",
+                    "error": error_msg,
+                    "report_id": report_id,
+                    "title": "Resume",
+                    "content_type": "typst",
+                }
+
+            # ── Phase 2: LLM GENERATION ───────────────────────────────────
+
+            user_instructions_section = ""
+            if user_instructions:
+                user_instructions_section = (
+                    f"**Additional Instructions:** {user_instructions}"
+                )
+
+            if parent_content:
+                dispatch_custom_event(
+                    "report_progress",
+                    {"phase": "writing", "message": "Updating your resume"},
+                )
+                parent_body = _strip_header(parent_content)
+                prompt = _REVISION_PROMPT.format(
+                    llm_reference=llm_reference,
+                    max_pages=validated_max_pages,
+                    user_instructions=user_instructions
+                    or "Improve and refine the resume.",
+                    previous_content=parent_body,
+                )
+            else:
+                dispatch_custom_event(
+                    "report_progress",
+                    {"phase": "writing", "message": "Building your resume"},
+                )
+                prompt = _RESUME_PROMPT.format(
+                    llm_reference=llm_reference,
+                    user_info=user_info,
+                    max_pages=validated_max_pages,
+                    user_instructions_section=user_instructions_section,
+                )
+
+            response = await llm.ainvoke([HumanMessage(content=prompt)])
+            body = response.content
+
+            if not body or not isinstance(body, str):
+                error_msg = "LLM returned empty or invalid content"
+                report_id = await _save_failed_report(error_msg)
+                return {
+                    "status": "failed",
+                    "error": error_msg,
+                    "report_id": report_id,
+                    "title": "Resume",
+                    "content_type": "typst",
+                }
+
+            body = _strip_typst_fences(body)
+            body = _strip_imports(body)
+
+            # ── Phase 3: ASSEMBLE + COMPILE ───────────────────────────────
+            dispatch_custom_event(
+                "report_progress",
+                {"phase": "compiling", "message": "Compiling resume..."},
+            )
+
+            name = _extract_name(body) or "Resume"
+            typst_source = ""
+            actual_pages = 0
+            compression_attempts = 0
+            target_page_met = False
+
+            for compression_round in range(MAX_COMPRESSION_ATTEMPTS + 1):
+                header = _build_header(template, name)
+                typst_source = header + body
+                compile_error: str | None = None
+                pdf_bytes: bytes | None = None
+
+                for compile_attempt in range(2):
+                    try:
+                        pdf_bytes = _compile_typst(typst_source)
+                        compile_error = None
+                        break
+                    except Exception as e:
+                        compile_error = str(e)
+                        logger.warning(
+                            "[generate_resume] Compile attempt %s failed: %s",
+                            compile_attempt + 1,
+                            compile_error,
+                        )
+
+                        if compile_attempt == 0:
+                            dispatch_custom_event(
+                                "report_progress",
+                                {
+                                    "phase": "fixing",
+                                    "message": "Fixing compilation issue...",
+                                },
+                            )
+                            fix_prompt = _FIX_COMPILE_PROMPT.format(
+                                llm_reference=llm_reference,
+                                error=compile_error,
+                                full_source=typst_source,
+                            )
+                            fix_response = await llm.ainvoke(
+                                [HumanMessage(content=fix_prompt)]
+                            )
+                            if fix_response.content and isinstance(
+                                fix_response.content, str
+                            ):
+                                body = _strip_typst_fences(fix_response.content)
+                                body = _strip_imports(body)
+                                name = _extract_name(body) or name
+                                header = _build_header(template, name)
+                                typst_source = header + body
+
+                if compile_error or not pdf_bytes:
+                    error_msg = (
+                        "Typst compilation failed after 2 attempts: "
+                        f"{compile_error or 'Unknown compile error'}"
+                    )
+                    report_id = await _save_failed_report(error_msg)
+                    return {
+                        "status": "failed",
+                        "error": error_msg,
+                        "report_id": report_id,
+                        "title": "Resume",
+                        "content_type": "typst",
+                    }
+
+                actual_pages = _count_pdf_pages(pdf_bytes)
+                if actual_pages <= validated_max_pages:
+                    target_page_met = True
+                    break
+
+                if compression_round >= MAX_COMPRESSION_ATTEMPTS:
+                    break
+
+                compression_attempts += 1
+                dispatch_custom_event(
+                    "report_progress",
+                    {
+                        "phase": "compressing",
+                        "message": f"Condensing resume to {validated_max_pages} page(s)...",
+                    },
+                )
+                compress_prompt = _COMPRESS_TO_PAGE_LIMIT_PROMPT.format(
+                    llm_reference=llm_reference,
+                    max_pages=validated_max_pages,
+                    actual_pages=actual_pages,
+                    attempt_number=compression_attempts,
+                    previous_content=body,
+                )
+                compress_response = await llm.ainvoke(
+                    [HumanMessage(content=compress_prompt)]
+                )
+                if not compress_response.content or not isinstance(
+                    compress_response.content, str
+                ):
+                    error_msg = "LLM returned empty content while compressing resume"
+                    report_id = await _save_failed_report(error_msg)
+                    return {
+                        "status": "failed",
+                        "error": error_msg,
+                        "report_id": report_id,
+                        "title": "Resume",
+                        "content_type": "typst",
+                    }
+
+                body = _strip_typst_fences(compress_response.content)
+                body = _strip_imports(body)
+                name = _extract_name(body) or name
+
+            if actual_pages > MAX_RESUME_PAGES:
+                error_msg = (
+                    "Resume exceeds hard page limit after compression retries. "
+                    f"Hard limit: <= {MAX_RESUME_PAGES} page(s), actual: {actual_pages}."
+                )
+                report_id = await _save_failed_report(error_msg)
+                return {
+                    "status": "failed",
+                    "error": error_msg,
+                    "report_id": report_id,
+                    "title": "Resume",
+                    "content_type": "typst",
+                }
+
+            # ── Phase 4: SAVE ─────────────────────────────────────────────
+            dispatch_custom_event(
+                "report_progress",
+                {"phase": "saving", "message": "Saving your resume"},
+            )
+
+            resume_title = f"{name} - Resume" if name != "Resume" else "Resume"
+
+            metadata: dict[str, Any] = {
+                "status": "ready",
+                "word_count": len(typst_source.split()),
+                "char_count": len(typst_source),
+                "target_max_pages": validated_max_pages,
+                "actual_page_count": actual_pages,
+                "page_limit_enforced": True,
+                "compression_attempts": compression_attempts,
+                "target_page_met": target_page_met,
+            }
+
+            async with shielded_async_session() as write_session:
+                report = Report(
+                    title=resume_title,
+                    content=typst_source,
+                    content_type="typst",
+                    report_metadata=metadata,
+                    report_style="resume",
+                    search_space_id=search_space_id,
+                    thread_id=thread_id,
+                    report_group_id=report_group_id,
+                )
+                write_session.add(report)
+                await write_session.commit()
+                await write_session.refresh(report)
+
+                if not report.report_group_id:
+                    report.report_group_id = report.id
+                    await write_session.commit()
+
+                saved_id = report.id
+
+            logger.info(f"[generate_resume] Created resume {saved_id}: {resume_title}")
+
+            return {
+                "status": "ready",
+                "report_id": saved_id,
+                "title": resume_title,
+                "content_type": "typst",
+                "is_revision": bool(parent_content),
+                "message": (
+                    f"Resume generated successfully: {resume_title}"
+                    if target_page_met
+                    else (
+                        f"Resume generated, but could not fit the target of <= {validated_max_pages} "
+                        f"page(s). Final length: {actual_pages} page(s)."
+                    )
+                ),
+            }
+
+        except Exception as e:
+            error_message = str(e)
+            logger.exception(f"[generate_resume] Error: {error_message}")
+            report_id = await _save_failed_report(error_message)
+            return {
+                "status": "failed",
+                "error": error_message,
+                "report_id": report_id,
+                "title": "Resume",
+                "content_type": "typst",
+            }
+
+    return generate_resume

surfsense_backend/app/agents/multi_agent_chat/subagents/builtins/deliverables/tools/video_presentation.py

@@ -0,0 +1,80 @@
+"""Factory for a video-presentation tool that queues background work and returns an ID for polling."""
+
+from typing import Any
+
+from langchain_core.tools import tool
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.db import VideoPresentation, VideoPresentationStatus, shielded_async_session
+
+
+def create_generate_video_presentation_tool(
+    search_space_id: int,
+    db_session: AsyncSession,
+    thread_id: int | None = None,
+):
+    """Create ``generate_video_presentation`` with bound search space and thread; writes use a tool-local session."""
+    del db_session  # writes use a fresh tool-local session, see below
+
+    @tool
+    async def generate_video_presentation(
+        source_content: str,
+        video_title: str = "SurfSense Presentation",
+        user_prompt: str | None = None,
+    ) -> dict[str, Any]:
+        """Generate a video presentation from the provided content.
+
+        Use this tool when the user asks to create a video, presentation, slides, or slide deck.
+
+        Args:
+            source_content: The text content to turn into a presentation.
+            video_title: Title for the presentation (default: "SurfSense Presentation")
+            user_prompt: Optional style/tone instructions.
+        """
+        try:
+            # One DB session per tool call so parallel invocations never share an AsyncSession.
+            async with shielded_async_session() as session:
+                video_pres = VideoPresentation(
+                    title=video_title,
+                    status=VideoPresentationStatus.PENDING,
+                    search_space_id=search_space_id,
+                    thread_id=thread_id,
+                )
+                session.add(video_pres)
+                await session.commit()
+                await session.refresh(video_pres)
+                video_pres_id = video_pres.id
+
+            from app.tasks.celery_tasks.video_presentation_tasks import (
+                generate_video_presentation_task,
+            )
+
+            task = generate_video_presentation_task.delay(
+                video_presentation_id=video_pres_id,
+                source_content=source_content,
+                search_space_id=search_space_id,
+                user_prompt=user_prompt,
+            )
+
+            print(
+                f"[generate_video_presentation] Created video presentation {video_pres_id}, task: {task.id}"
+            )
+
+            return {
+                "status": VideoPresentationStatus.PENDING.value,
+                "video_presentation_id": video_pres_id,
+                "title": video_title,
+                "message": "Video presentation generation started. This may take a few minutes.",
+            }
+
+        except Exception as e:
+            error_message = str(e)
+            print(f"[generate_video_presentation] Error: {error_message}")
+            return {
+                "status": VideoPresentationStatus.FAILED.value,
+                "error": error_message,
+                "title": video_title,
+                "video_presentation_id": None,
+            }
+
+    return generate_video_presentation

surfsense_backend/app/agents/multi_agent_chat/subagents/builtins/memory/agent.py

@@ -0,0 +1,55 @@
+"""`memory` route: ``SubAgent`` spec for deepagents."""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+from typing import Any
+
+from deepagents import SubAgent
+from langchain_core.language_models import BaseChatModel
+
+from app.agents.multi_agent_chat.subagents.shared.md_file_reader import (
+    read_md_file,
+)
+from app.agents.multi_agent_chat.subagents.shared.permissions import (
+    ToolsPermissions,
+    merge_tools_permissions,
+    middleware_gated_interrupt_on,
+)
+from app.agents.multi_agent_chat.subagents.shared.subagent_builder import (
+    pack_subagent,
+)
+
+from .tools.index import load_tools
+
+NAME = "memory"
+
+
+def build_subagent(
+    *,
+    dependencies: dict[str, Any],
+    model: BaseChatModel | None = None,
+    extra_middleware: Sequence[Any] | None = None,
+    extra_tools_bucket: ToolsPermissions | None = None,
+) -> SubAgent:
+    buckets = load_tools(dependencies=dependencies)
+    merged_tools_bucket = merge_tools_permissions(buckets, extra_tools_bucket)
+    tools = [
+        row["tool"]
+        for row in (*merged_tools_bucket["allow"], *merged_tools_bucket["ask"])
+        if row.get("tool") is not None
+    ]
+    interrupt_on = middleware_gated_interrupt_on(merged_tools_bucket)
+    description = read_md_file(__package__, "description").strip()
+    if not description:
+        description = "Handles memory tasks for this workspace."
+    system_prompt = read_md_file(__package__, "system_prompt").strip()
+    return pack_subagent(
+        name=NAME,
+        description=description,
+        system_prompt=system_prompt,
+        tools=tools,
+        interrupt_on=interrupt_on,
+        model=model,
+        extra_middleware=extra_middleware,
+    )

surfsense_backend/app/agents/multi_agent_chat/subagents/builtins/memory/description.md

@@ -0,0 +1 @@
+Use for storing durable user memory (private team variant selected at runtime).

surfsense_backend/app/agents/multi_agent_chat/subagents/builtins/memory/system_prompt.md

@@ -0,0 +1,56 @@
+You are the SurfSense memory operations sub-agent.
+You receive delegated instructions from a supervisor agent and return structured results for supervisor synthesis.
+
+<goal>
+Persist durable preferences/facts/instructions with `update_memory` while avoiding transient or unsafe storage.
+</goal>
+
+<visibility_scope>
+{{MEMORY_VISIBILITY_POLICY}}
+</visibility_scope>
+
+<available_tools>
+- `update_memory`
+</available_tools>
+
+<tool_policy>
+- Save only durable information with future value.
+- Do not store transient chatter.
+- Do not store secrets unless explicitly instructed.
+- If memory intent is unclear, return `status=blocked` with the missing intent signal.
+</tool_policy>
+
+<out_of_scope>
+- Do not execute non-memory tool actions.
+- Do not store irrelevant, transient, or speculative information.
+</out_of_scope>
+
+<safety>
+- Prefer minimal-memory writes over over-collection.
+- Never claim memory was updated unless `update_memory` succeeded.
+</safety>
+
+<failure_policy>
+- On tool failure, return `status=error` with concise recovery steps.
+- When intent is ambiguous, return `status=blocked` with required disambiguation fields.
+</failure_policy>
+
+<output_contract>
+Return **only** one JSON object (no markdown/prose):
+{
+  "status": "success" | "partial" | "blocked" | "error",
+  "action_summary": string,
+  "evidence": {
+    "memory_updated": boolean,
+    "memory_category": "preference" | "fact" | "instruction" | null,
+    "stored_summary": string | null
+  },
+  "next_step": string | null,
+  "missing_fields": string[] | null,
+  "assumptions": string[] | null
+}
+Rules:
+- `status=success` -> `next_step=null`, `missing_fields=null`.
+- `status=partial|blocked|error` -> `next_step` must be non-null.
+- `status=blocked` due to missing required inputs -> `missing_fields` must be non-null.
+</output_contract>

surfsense_backend/app/agents/multi_agent_chat/subagents/builtins/memory/tools/__init__.py

@@ -0,0 +1,8 @@
+"""Memory tools: persist user or team markdown memory for later turns."""
+
+from .update_memory import create_update_memory_tool, create_update_team_memory_tool
+
+__all__ = [
+    "create_update_memory_tool",
+    "create_update_team_memory_tool",
+]

surfsense_backend/app/agents/multi_agent_chat/subagents/builtins/memory/tools/index.py

@@ -0,0 +1,32 @@
+from __future__ import annotations
+
+from typing import Any
+
+from app.agents.multi_agent_chat.subagents.shared.permissions import (
+    ToolsPermissions,
+)
+from app.db import ChatVisibility
+
+from .update_memory import create_update_memory_tool, create_update_team_memory_tool
+
+
+def load_tools(
+    *, dependencies: dict[str, Any] | None = None, **kwargs: Any
+) -> ToolsPermissions:
+    resolved_dependencies = {**(dependencies or {}), **kwargs}
+    if resolved_dependencies.get("thread_visibility") == ChatVisibility.SEARCH_SPACE:
+        mem = create_update_team_memory_tool(
+            search_space_id=resolved_dependencies["search_space_id"],
+            db_session=resolved_dependencies["db_session"],
+            llm=resolved_dependencies.get("llm"),
+        )
+        return {
+            "allow": [{"name": getattr(mem, "name", "") or "", "tool": mem}],
+            "ask": [],
+        }
+    mem = create_update_memory_tool(
+        user_id=resolved_dependencies["user_id"],
+        db_session=resolved_dependencies["db_session"],
+        llm=resolved_dependencies.get("llm"),
+    )
+    return {"allow": [{"name": getattr(mem, "name", "") or "", "tool": mem}], "ask": []}

surfsense_backend/app/agents/multi_agent_chat/subagents/builtins/memory/tools/update_memory.py

@@ -0,0 +1,375 @@
+"""Overwrite one markdown memory document per user or team, with size and shrink guards."""
+
+from __future__ import annotations
+
+import logging
+import re
+from typing import Any, Literal
+from uuid import UUID
+
+from langchain_core.messages import HumanMessage
+from langchain_core.tools import tool
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.db import SearchSpace, User
+
+logger = logging.getLogger(__name__)
+
+MEMORY_SOFT_LIMIT = 18_000
+MEMORY_HARD_LIMIT = 25_000
+
+_SECTION_HEADING_RE = re.compile(r"^##\s+(.+)$", re.MULTILINE)
+_HEADING_NORMALIZE_RE = re.compile(r"\s+")
+
+_MARKER_RE = re.compile(r"\[(fact|pref|instr)\]")
+_BULLET_FORMAT_RE = re.compile(r"^- \(\d{4}-\d{2}-\d{2}\) \[(fact|pref|instr)\] .+$")
+_PERSONAL_ONLY_MARKERS = {"pref", "instr"}
+
+
+# ---------------------------------------------------------------------------
+# Diff validation
+# ---------------------------------------------------------------------------
+
+
+def _extract_headings(memory: str) -> set[str]:
+    """Return all ``## …`` heading texts (without the ``## `` prefix)."""
+    return set(_SECTION_HEADING_RE.findall(memory))
+
+
+def _normalize_heading(heading: str) -> str:
+    """Normalize heading text for robust scope checks."""
+    return _HEADING_NORMALIZE_RE.sub(" ", heading.strip().lower())
+
+
+def _validate_memory_scope(
+    content: str, scope: Literal["user", "team"]
+) -> dict[str, Any] | None:
+    """Reject personal-only markers ([pref], [instr]) in team memory."""
+    if scope != "team":
+        return None
+
+    markers = set(_MARKER_RE.findall(content))
+    leaked = sorted(markers & _PERSONAL_ONLY_MARKERS)
+    if leaked:
+        tags = ", ".join(f"[{m}]" for m in leaked)
+        return {
+            "status": "error",
+            "message": (
+                f"Team memory cannot include personal markers: {tags}. "
+                "Use [fact] only in team memory."
+            ),
+        }
+    return None
+
+
+def _validate_bullet_format(content: str) -> list[str]:
+    """Return warnings for bullet lines that don't match the required format.
+
+    Expected: ``- (YYYY-MM-DD) [fact|pref|instr] text``
+    """
+    warnings: list[str] = []
+    for line in content.splitlines():
+        stripped = line.strip()
+        if not stripped.startswith("- "):
+            continue
+        if not _BULLET_FORMAT_RE.match(stripped):
+            short = stripped[:80] + ("..." if len(stripped) > 80 else "")
+            warnings.append(f"Malformed bullet: {short}")
+    return warnings
+
+
+def _validate_diff(old_memory: str | None, new_memory: str) -> list[str]:
+    """Return a list of warning strings about suspicious changes."""
+    if not old_memory:
+        return []
+
+    warnings: list[str] = []
+    old_headings = _extract_headings(old_memory)
+    new_headings = _extract_headings(new_memory)
+    dropped = old_headings - new_headings
+    if dropped:
+        names = ", ".join(sorted(dropped))
+        warnings.append(
+            f"Sections removed: {names}. "
+            "If unintentional, the user can restore from the settings page."
+        )
+
+    old_len = len(old_memory)
+    new_len = len(new_memory)
+    if old_len > 0 and new_len < old_len * 0.4:
+        warnings.append(
+            f"Memory shrank significantly ({old_len:,} -> {new_len:,} chars). "
+            "Possible data loss."
+        )
+    return warnings
+
+
+# ---------------------------------------------------------------------------
+# Size validation & soft warning
+# ---------------------------------------------------------------------------
+
+
+def _validate_memory_size(content: str) -> dict[str, Any] | None:
+    """Return an error/warning dict if *content* is too large, else None."""
+    length = len(content)
+    if length > MEMORY_HARD_LIMIT:
+        return {
+            "status": "error",
+            "message": (
+                f"Memory exceeds {MEMORY_HARD_LIMIT:,} character limit "
+                f"({length:,} chars). Consolidate by merging related items, "
+                "removing outdated entries, and shortening descriptions. "
+                "Then call update_memory again."
+            ),
+        }
+    return None
+
+
+def _soft_warning(content: str) -> str | None:
+    """Return a warning string if content exceeds the soft limit."""
+    length = len(content)
+    if length > MEMORY_SOFT_LIMIT:
+        return (
+            f"Memory is at {length:,}/{MEMORY_HARD_LIMIT:,} characters. "
+            "Consolidate by merging related items and removing less important "
+            "entries on your next update."
+        )
+    return None
+
+
+# ---------------------------------------------------------------------------
+# Forced rewrite when memory exceeds the hard limit
+# ---------------------------------------------------------------------------
+
+_FORCED_REWRITE_PROMPT = """\
+You are a memory curator. The following memory document exceeds the character \
+limit and must be shortened.
+
+RULES:
+1. Rewrite the document to be under {target} characters.
+2. Preserve existing ## headings. Every entry must remain under a heading. You may merge
+   or rename headings to consolidate, but keep names personal and descriptive.
+3. Priority for keeping content: [instr] > [pref] > [fact].
+4. Merge duplicate entries, remove outdated entries, shorten verbose descriptions.
+5. Every bullet MUST have format: - (YYYY-MM-DD) [fact|pref|instr] text
+6. Preserve the user's first name in entries — do not replace it with "the user".
+7. Output ONLY the consolidated markdown — no explanations, no wrapping.
+
+<memory_document>
+{content}
+</memory_document>"""
+
+
+async def _forced_rewrite(content: str, llm: Any) -> str | None:
+    """Use a focused LLM call to compress *content* under the hard limit.
+
+    Returns the rewritten string, or ``None`` if the call fails.
+    """
+    try:
+        prompt = _FORCED_REWRITE_PROMPT.format(
+            target=MEMORY_HARD_LIMIT, content=content
+        )
+        response = await llm.ainvoke(
+            [HumanMessage(content=prompt)],
+            config={"tags": ["surfsense:internal"]},
+        )
+        text = (
+            response.content
+            if isinstance(response.content, str)
+            else str(response.content)
+        )
+        return text.strip()
+    except Exception:
+        logger.exception("Forced rewrite LLM call failed")
+        return None
+
+
+# ---------------------------------------------------------------------------
+# Shared save-and-respond logic
+# ---------------------------------------------------------------------------
+
+
+async def _save_memory(
+    *,
+    updated_memory: str,
+    old_memory: str | None,
+    llm: Any | None,
+    apply_fn,
+    commit_fn,
+    rollback_fn,
+    label: str,
+    scope: Literal["user", "team"],
+) -> dict[str, Any]:
+    """Validate, optionally force-rewrite if over the hard limit, save, and
+    return a response dict.
+
+    Parameters
+    ----------
+    updated_memory : str
+        The new document the agent submitted.
+    old_memory : str | None
+        The previously persisted document (for diff checks).
+    llm : Any | None
+        LLM instance for forced rewrite (may be ``None``).
+    apply_fn : callable(str) -> None
+        Callback that sets the new memory on the ORM object.
+    commit_fn : coroutine
+        ``session.commit``.
+    rollback_fn : coroutine
+        ``session.rollback``.
+    label : str
+        Human label for log messages (e.g. "user memory", "team memory").
+    """
+    content = updated_memory
+
+    # --- forced rewrite if over the hard limit ---
+    if len(content) > MEMORY_HARD_LIMIT and llm is not None:
+        rewritten = await _forced_rewrite(content, llm)
+        if rewritten is not None and len(rewritten) < len(content):
+            content = rewritten
+
+    # --- hard-limit gate (reject if still too large after rewrite) ---
+    size_err = _validate_memory_size(content)
+    if size_err:
+        return size_err
+
+    scope_err = _validate_memory_scope(content, scope)
+    if scope_err:
+        return scope_err
+
+    # --- persist ---
+    try:
+        apply_fn(content)
+        await commit_fn()
+    except Exception as e:
+        logger.exception("Failed to update %s: %s", label, e)
+        await rollback_fn()
+        return {"status": "error", "message": f"Failed to update {label}: {e}"}
+
+    # --- build response ---
+    resp: dict[str, Any] = {
+        "status": "saved",
+        "message": f"{label.capitalize()} updated.",
+    }
+
+    if content is not updated_memory:
+        resp["notice"] = "Memory was automatically rewritten to fit within limits."
+
+    diff_warnings = _validate_diff(old_memory, content)
+    if diff_warnings:
+        resp["diff_warnings"] = diff_warnings
+
+    format_warnings = _validate_bullet_format(content)
+    if format_warnings:
+        resp["format_warnings"] = format_warnings
+
+    warning = _soft_warning(content)
+    if warning:
+        resp["warning"] = warning
+
+    return resp
+
+
+# ---------------------------------------------------------------------------
+# Tool factories
+# ---------------------------------------------------------------------------
+
+
+def create_update_memory_tool(
+    user_id: str | UUID,
+    db_session: AsyncSession,
+    llm: Any | None = None,
+):
+    uid = UUID(user_id) if isinstance(user_id, str) else user_id
+
+    @tool
+    async def update_memory(updated_memory: str) -> dict[str, Any]:
+        """Update the user's personal memory document.
+
+        Your current memory is shown in <user_memory> in the system prompt.
+        When the user shares important long-term information (preferences,
+        facts, instructions, context), rewrite the memory document to include
+        the new information.  Merge new facts with existing ones, update
+        contradictions, remove outdated entries, and keep it concise.
+
+        Args:
+            updated_memory: The FULL updated markdown document (not a diff).
+        """
+        try:
+            result = await db_session.execute(select(User).where(User.id == uid))
+            user = result.scalars().first()
+            if not user:
+                return {"status": "error", "message": "User not found."}
+
+            old_memory = user.memory_md
+
+            return await _save_memory(
+                updated_memory=updated_memory,
+                old_memory=old_memory,
+                llm=llm,
+                apply_fn=lambda content: setattr(user, "memory_md", content),
+                commit_fn=db_session.commit,
+                rollback_fn=db_session.rollback,
+                label="memory",
+                scope="user",
+            )
+        except Exception as e:
+            logger.exception("Failed to update user memory: %s", e)
+            await db_session.rollback()
+            return {
+                "status": "error",
+                "message": f"Failed to update memory: {e}",
+            }
+
+    return update_memory
+
+
+def create_update_team_memory_tool(
+    search_space_id: int,
+    db_session: AsyncSession,
+    llm: Any | None = None,
+):
+    @tool
+    async def update_memory(updated_memory: str) -> dict[str, Any]:
+        """Update the team's shared memory document for this search space.
+
+        Your current team memory is shown in <team_memory> in the system
+        prompt.  When the team shares important long-term information
+        (decisions, conventions, key facts, priorities), rewrite the memory
+        document to include the new information.  Merge new facts with
+        existing ones, update contradictions, remove outdated entries, and
+        keep it concise.
+
+        Args:
+            updated_memory: The FULL updated markdown document (not a diff).
+        """
+        try:
+            result = await db_session.execute(
+                select(SearchSpace).where(SearchSpace.id == search_space_id)
+            )
+            space = result.scalars().first()
+            if not space:
+                return {"status": "error", "message": "Search space not found."}
+
+            old_memory = space.shared_memory_md
+
+            return await _save_memory(
+                updated_memory=updated_memory,
+                old_memory=old_memory,
+                llm=llm,
+                apply_fn=lambda content: setattr(space, "shared_memory_md", content),
+                commit_fn=db_session.commit,
+                rollback_fn=db_session.rollback,
+                label="team memory",
+                scope="team",
+            )
+        except Exception as e:
+            logger.exception("Failed to update team memory: %s", e)
+            await db_session.rollback()
+            return {
+                "status": "error",
+                "message": f"Failed to update team memory: {e}",
+            }
+
+    return update_memory

surfsense_backend/app/agents/multi_agent_chat/subagents/builtins/research/agent.py

@@ -0,0 +1,55 @@
+"""`research` route: ``SubAgent`` spec for deepagents."""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+from typing import Any
+
+from deepagents import SubAgent
+from langchain_core.language_models import BaseChatModel
+
+from app.agents.multi_agent_chat.subagents.shared.md_file_reader import (
+    read_md_file,
+)
+from app.agents.multi_agent_chat.subagents.shared.permissions import (
+    ToolsPermissions,
+    merge_tools_permissions,
+    middleware_gated_interrupt_on,
+)
+from app.agents.multi_agent_chat.subagents.shared.subagent_builder import (
+    pack_subagent,
+)
+
+from .tools.index import load_tools
+
+NAME = "research"
+
+
+def build_subagent(
+    *,
+    dependencies: dict[str, Any],
+    model: BaseChatModel | None = None,
+    extra_middleware: Sequence[Any] | None = None,
+    extra_tools_bucket: ToolsPermissions | None = None,
+) -> SubAgent:
+    buckets = load_tools(dependencies=dependencies)
+    merged_tools_bucket = merge_tools_permissions(buckets, extra_tools_bucket)
+    tools = [
+        row["tool"]
+        for row in (*merged_tools_bucket["allow"], *merged_tools_bucket["ask"])
+        if row.get("tool") is not None
+    ]
+    interrupt_on = middleware_gated_interrupt_on(merged_tools_bucket)
+    description = read_md_file(__package__, "description").strip()
+    if not description:
+        description = "Handles research tasks for this workspace."
+    system_prompt = read_md_file(__package__, "system_prompt").strip()
+    return pack_subagent(
+        name=NAME,
+        description=description,
+        system_prompt=system_prompt,
+        tools=tools,
+        interrupt_on=interrupt_on,
+        model=model,
+        extra_middleware=extra_middleware,
+    )

surfsense_backend/app/agents/multi_agent_chat/subagents/builtins/research/description.md

@@ -0,0 +1 @@
+Use for external research: find sources on the web, extract evidence, and answer documentation questions.

surfsense_backend/app/agents/multi_agent_chat/subagents/builtins/research/system_prompt.md

@@ -0,0 +1,53 @@
+You are the SurfSense research operations sub-agent.
+You receive delegated instructions from a supervisor agent and return structured results for supervisor synthesis.
+
+<goal>
+Gather and synthesize evidence using SurfSense research tools with clear citations and uncertainty reporting.
+</goal>
+
+<available_tools>
+- `web_search`
+- `scrape_webpage`
+- `search_surfsense_docs`
+</available_tools>
+
+<tool_policy>
+- Use only tools in `<available_tools>`.
+- Prefer primary and recent sources when recency matters.
+- If the delegated request is underspecified, return `status=blocked` with the missing research constraints.
+- Never fabricate facts, citations, URLs, or quote text.
+</tool_policy>
+
+<out_of_scope>
+- Do not execute connector mutations (email/calendar/docs/chat writes) or deliverable generation.
+</out_of_scope>
+
+<safety>
+- Report uncertainty explicitly when evidence is incomplete or conflicting.
+- Never present unverified claims as facts.
+</safety>
+
+<failure_policy>
+- On tool failure, return `status=error` with a concise recovery `next_step`.
+- On no useful evidence, return `status=blocked` with recommended narrower filters.
+</failure_policy>
+
+<output_contract>
+Return **only** one JSON object (no markdown/prose):
+{
+  "status": "success" | "partial" | "blocked" | "error",
+  "action_summary": string,
+  "evidence": {
+    "findings": string[],
+    "sources": string[],
+    "confidence": "high" | "medium" | "low"
+  },
+  "next_step": string | null,
+  "missing_fields": string[] | null,
+  "assumptions": string[] | null
+}
+Rules:
+- `status=success` -> `next_step=null`, `missing_fields=null`.
+- `status=partial|blocked|error` -> `next_step` must be non-null.
+- `status=blocked` due to missing required inputs -> `missing_fields` must be non-null.
+</output_contract>

surfsense_backend/app/agents/multi_agent_chat/subagents/builtins/research/tools/__init__.py

@@ -0,0 +1,11 @@
+"""Research-stage tools: web search, scrape, and in-product doc search."""
+
+from .scrape_webpage import create_scrape_webpage_tool
+from .search_surfsense_docs import create_search_surfsense_docs_tool
+from .web_search import create_web_search_tool
+
+__all__ = [
+    "create_scrape_webpage_tool",
+    "create_search_surfsense_docs_tool",
+    "create_web_search_tool",
+]

surfsense_backend/app/agents/multi_agent_chat/subagents/builtins/research/tools/index.py

@@ -0,0 +1,35 @@
+from __future__ import annotations
+
+from typing import Any
+
+from app.agents.multi_agent_chat.subagents.shared.permissions import (
+    ToolsPermissions,
+)
+
+from .scrape_webpage import create_scrape_webpage_tool
+from .search_surfsense_docs import create_search_surfsense_docs_tool
+from .web_search import create_web_search_tool
+
+
+def load_tools(
+    *, dependencies: dict[str, Any] | None = None, **kwargs: Any
+) -> ToolsPermissions:
+    resolved_dependencies = {**(dependencies or {}), **kwargs}
+    web = create_web_search_tool(
+        search_space_id=resolved_dependencies.get("search_space_id"),
+        available_connectors=resolved_dependencies.get("available_connectors"),
+    )
+    scrape = create_scrape_webpage_tool(
+        firecrawl_api_key=resolved_dependencies.get("firecrawl_api_key")
+    )
+    docs = create_search_surfsense_docs_tool(
+        db_session=resolved_dependencies["db_session"]
+    )
+    return {
+        "allow": [
+            {"name": getattr(web, "name", "") or "", "tool": web},
+            {"name": getattr(scrape, "name", "") or "", "tool": scrape},
+            {"name": getattr(docs, "name", "") or "", "tool": docs},
+        ],
+        "ask": [],
+    }

surfsense_backend/app/agents/multi_agent_chat/subagents/builtins/research/tools/scrape_webpage.py

@@ -0,0 +1,300 @@
+"""Scrape pages via WebCrawlerConnector; YouTube URLs use the transcript API instead of HTML crawl."""
+
+import hashlib
+import logging
+from typing import Any
+from urllib.parse import urlparse
+
+import aiohttp
+from fake_useragent import UserAgent
+from langchain_core.tools import tool
+from requests import Session
+from youtube_transcript_api import YouTubeTranscriptApi
+
+from app.connectors.webcrawler_connector import WebCrawlerConnector
+from app.tasks.document_processors.youtube_processor import get_youtube_video_id
+from app.utils.proxy_config import get_requests_proxies
+
+logger = logging.getLogger(__name__)
+
+
+def extract_domain(url: str) -> str:
+    """Extract the domain from a URL."""
+    try:
+        parsed = urlparse(url)
+        domain = parsed.netloc
+        # Remove 'www.' prefix if present
+        if domain.startswith("www."):
+            domain = domain[4:]
+        return domain
+    except Exception:
+        return ""
+
+
+def generate_scrape_id(url: str) -> str:
+    """Generate a unique ID for a scraped webpage."""
+    hash_val = hashlib.md5(url.encode()).hexdigest()[:12]
+    return f"scrape-{hash_val}"
+
+
+def truncate_content(content: str, max_length: int = 50000) -> tuple[str, bool]:
+    """
+    Truncate content to a maximum length.
+
+    Returns:
+        Tuple of (truncated_content, was_truncated)
+    """
+    if len(content) <= max_length:
+        return content, False
+
+    # Try to truncate at a sentence boundary
+    truncated = content[:max_length]
+    last_period = truncated.rfind(".")
+    last_newline = truncated.rfind("\n\n")
+
+    # Use the later of the two boundaries, or just truncate
+    boundary = max(last_period, last_newline)
+    if boundary > max_length * 0.8:  # Only use boundary if it's not too far back
+        truncated = content[: boundary + 1]
+
+    return truncated + "\n\n[Content truncated...]", True
+
+
+async def _scrape_youtube_video(
+    url: str, video_id: str, max_length: int
+) -> dict[str, Any]:
+    """
+    Fetch YouTube video metadata and transcript via the YouTubeTranscriptApi.
+
+    Returns a result dict in the same shape as the regular scrape_webpage output.
+    """
+    scrape_id = generate_scrape_id(url)
+    domain = "youtube.com"
+
+    # --- Video metadata via oEmbed ---
+    residential_proxies = get_requests_proxies()
+
+    params = {
+        "format": "json",
+        "url": f"https://www.youtube.com/watch?v={video_id}",
+    }
+    oembed_url = "https://www.youtube.com/oembed"
+
+    try:
+        async with (
+            aiohttp.ClientSession() as http_session,
+            http_session.get(
+                oembed_url,
+                params=params,
+                proxy=residential_proxies["http"] if residential_proxies else None,
+            ) as response,
+        ):
+            video_data = await response.json()
+    except Exception:
+        video_data = {}
+
+    title = video_data.get("title", "YouTube Video")
+    author = video_data.get("author_name", "Unknown")
+
+    # --- Transcript via YouTubeTranscriptApi ---
+    try:
+        ua = UserAgent()
+        http_client = Session()
+        http_client.headers.update({"User-Agent": ua.random})
+        if residential_proxies:
+            http_client.proxies.update(residential_proxies)
+        ytt_api = YouTubeTranscriptApi(http_client=http_client)
+
+        # List all available transcripts and pick the first one
+        # (the video's primary language) instead of defaulting to English
+        transcript_list = ytt_api.list(video_id)
+        transcript = next(iter(transcript_list))
+        captions = transcript.fetch()
+
+        logger.info(
+            f"[scrape_webpage] Fetched transcript for {video_id} "
+            f"in {transcript.language} ({transcript.language_code})"
+        )
+
+        transcript_segments = []
+        for line in captions:
+            start_time = line.start
+            duration = line.duration
+            text = line.text
+            timestamp = f"[{start_time:.2f}s-{start_time + duration:.2f}s]"
+            transcript_segments.append(f"{timestamp} {text}")
+        transcript_text = "\n".join(transcript_segments)
+    except Exception as e:
+        logger.warning(f"[scrape_webpage] No transcript for video {video_id}: {e}")
+        transcript_text = f"No captions available for this video. Error: {e!s}"
+
+    # Build combined content
+    content = f"# {title}\n\n**Author:** {author}\n**Video ID:** {video_id}\n\n## Transcript\n\n{transcript_text}"
+
+    # Truncate if needed
+    content, was_truncated = truncate_content(content, max_length)
+    word_count = len(content.split())
+
+    description = f"YouTube video by {author}"
+
+    return {
+        "id": scrape_id,
+        "assetId": url,
+        "kind": "article",
+        "href": url,
+        "title": title,
+        "description": description,
+        "content": content,
+        "domain": domain,
+        "word_count": word_count,
+        "was_truncated": was_truncated,
+        "crawler_type": "youtube_transcript",
+        "author": author,
+    }
+
+
+def create_scrape_webpage_tool(firecrawl_api_key: str | None = None):
+    """
+    Factory function to create the scrape_webpage tool.
+
+    Args:
+        firecrawl_api_key: Optional Firecrawl API key for premium web scraping.
+                          Falls back to Chromium/Trafilatura if not provided.
+
+    Returns:
+        A configured tool function for scraping webpages.
+    """
+
+    @tool
+    async def scrape_webpage(
+        url: str,
+        max_length: int = 50000,
+    ) -> dict[str, Any]:
+        """
+        Scrape and extract the main content from a webpage.
+
+        Use this tool when the user wants you to read, summarize, or answer
+        questions about a specific webpage's content. This tool actually
+        fetches and reads the full page content. For YouTube video URLs it
+        fetches the transcript directly instead of crawling the page.
+
+        Common triggers:
+        - "Read this article and summarize it"
+        - "What does this page say about X?"
+        - "Summarize this blog post for me"
+        - "Tell me the key points from this article"
+        - "What's in this webpage?"
+
+        Args:
+            url: The URL of the webpage to scrape (must be HTTP/HTTPS)
+            max_length: Maximum content length to return (default: 50000 chars)
+
+        Returns:
+            A dictionary containing:
+            - id: Unique identifier for this scrape
+            - assetId: The URL (for deduplication)
+            - kind: "article" (type of content)
+            - href: The URL to open when clicked
+            - title: Page title
+            - description: Brief description or excerpt
+            - content: The extracted main content (markdown format)
+            - domain: The domain name
+            - word_count: Approximate word count
+            - was_truncated: Whether content was truncated
+            - error: Error message (if scraping failed)
+        """
+        scrape_id = generate_scrape_id(url)
+        domain = extract_domain(url)
+
+        # Validate and normalize URL
+        if not url.startswith(("http://", "https://")):
+            url = f"https://{url}"
+
+        try:
+            # Check if this is a YouTube URL and use transcript API instead
+            video_id = get_youtube_video_id(url)
+            if video_id:
+                return await _scrape_youtube_video(url, video_id, max_length)
+
+            # Create webcrawler connector
+            connector = WebCrawlerConnector(firecrawl_api_key=firecrawl_api_key)
+
+            # Crawl the URL
+            result, error = await connector.crawl_url(url, formats=["markdown"])
+
+            if error:
+                return {
+                    "id": scrape_id,
+                    "assetId": url,
+                    "kind": "article",
+                    "href": url,
+                    "title": domain or "Webpage",
+                    "domain": domain,
+                    "error": error,
+                }
+
+            if not result:
+                return {
+                    "id": scrape_id,
+                    "assetId": url,
+                    "kind": "article",
+                    "href": url,
+                    "title": domain or "Webpage",
+                    "domain": domain,
+                    "error": "No content returned from crawler",
+                }
+
+            # Extract content and metadata
+            content = result.get("content", "")
+            metadata = result.get("metadata", {})
+
+            # Get title from metadata
+            title = metadata.get("title", "")
+            if not title:
+                title = domain or url.split("/")[-1] or "Webpage"
+
+            # Get description from metadata
+            description = metadata.get("description", "")
+            if not description and content:
+                # Use first paragraph as description
+                first_para = content.split("\n\n")[0] if content else ""
+                description = (
+                    first_para[:300] + "..." if len(first_para) > 300 else first_para
+                )
+
+            # Truncate content if needed
+            content, was_truncated = truncate_content(content, max_length)
+
+            # Calculate word count
+            word_count = len(content.split())
+
+            return {
+                "id": scrape_id,
+                "assetId": url,
+                "kind": "article",
+                "href": url,
+                "title": title,
+                "description": description,
+                "content": content,
+                "domain": domain,
+                "word_count": word_count,
+                "was_truncated": was_truncated,
+                "crawler_type": result.get("crawler_type", "unknown"),
+                "author": metadata.get("author"),
+                "date": metadata.get("date"),
+            }
+
+        except Exception as e:
+            error_message = str(e)
+            logger.error(f"[scrape_webpage] Error scraping {url}: {error_message}")
+            return {
+                "id": scrape_id,
+                "assetId": url,
+                "kind": "article",
+                "href": url,
+                "title": domain or "Webpage",
+                "domain": domain,
+                "error": f"Failed to scrape: {error_message[:100]}",
+            }
+
+    return scrape_webpage

surfsense_backend/app/agents/multi_agent_chat/subagents/builtins/research/tools/search_surfsense_docs.py

@@ -0,0 +1,143 @@
+"""Semantic search over pre-indexed in-app documentation chunks for user how-to questions."""
+
+import asyncio
+import json
+
+from langchain_core.tools import tool
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.db import SurfsenseDocsChunk, SurfsenseDocsDocument
+from app.utils.document_converters import embed_text
+
+
+def format_surfsense_docs_results(results: list[tuple]) -> str:
+    """Format (chunk, document) rows as XML with ``doc-`` chunk IDs for citations and UI routing."""
+    if not results:
+        return "No relevant Surfsense documentation found for your query."
+
+    # Group chunks by document
+    grouped: dict[int, dict] = {}
+    for chunk, doc in results:
+        if doc.id not in grouped:
+            grouped[doc.id] = {
+                "document_id": f"doc-{doc.id}",
+                "document_type": "SURFSENSE_DOCS",
+                "title": doc.title,
+                "url": doc.source,
+                "metadata": {"source": doc.source},
+                "chunks": [],
+            }
+        grouped[doc.id]["chunks"].append(
+            {
+                "chunk_id": f"doc-{chunk.id}",
+                "content": chunk.content,
+            }
+        )
+
+    # Render XML matching format_documents_for_context structure
+    parts: list[str] = []
+    for g in grouped.values():
+        metadata_json = json.dumps(g["metadata"], ensure_ascii=False)
+
+        parts.append("<document>")
+        parts.append("<document_metadata>")
+        parts.append(f"  <document_id>{g['document_id']}</document_id>")
+        parts.append(f"  <document_type>{g['document_type']}</document_type>")
+        parts.append(f"  <title><![CDATA[{g['title']}]]></title>")
+        parts.append(f"  <url><![CDATA[{g['url']}]]></url>")
+        parts.append(f"  <metadata_json><![CDATA[{metadata_json}]]></metadata_json>")
+        parts.append("</document_metadata>")
+        parts.append("")
+        parts.append("<document_content>")
+
+        for ch in g["chunks"]:
+            parts.append(
+                f"  <chunk id='{ch['chunk_id']}'><![CDATA[{ch['content']}]]></chunk>"
+            )
+
+        parts.append("</document_content>")
+        parts.append("</document>")
+        parts.append("")
+
+    return "\n".join(parts).strip()
+
+
+async def search_surfsense_docs_async(
+    query: str,
+    db_session: AsyncSession,
+    top_k: int = 10,
+) -> str:
+    """
+    Search Surfsense documentation using vector similarity.
+
+    Args:
+        query: The search query about Surfsense usage
+        db_session: Database session for executing queries
+        top_k: Number of results to return
+
+    Returns:
+        Formatted string with relevant documentation content
+    """
+    # Get embedding for the query
+    query_embedding = await asyncio.to_thread(embed_text, query)
+
+    # Vector similarity search on chunks, joining with documents
+    stmt = (
+        select(SurfsenseDocsChunk, SurfsenseDocsDocument)
+        .join(
+            SurfsenseDocsDocument,
+            SurfsenseDocsChunk.document_id == SurfsenseDocsDocument.id,
+        )
+        .order_by(SurfsenseDocsChunk.embedding.op("<=>")(query_embedding))
+        .limit(top_k)
+    )
+
+    result = await db_session.execute(stmt)
+    rows = result.all()
+
+    return format_surfsense_docs_results(rows)
+
+
+def create_search_surfsense_docs_tool(db_session: AsyncSession):
+    """
+    Factory function to create the search_surfsense_docs tool.
+
+    Args:
+        db_session: Database session for executing queries
+
+    Returns:
+        A configured tool function for searching Surfsense documentation
+    """
+
+    @tool
+    async def search_surfsense_docs(query: str, top_k: int = 10) -> str:
+        """
+        Search Surfsense documentation for help with using the application.
+
+        Use this tool when the user asks questions about:
+        - How to use Surfsense features
+        - Installation and setup instructions
+        - Configuration options and settings
+        - Troubleshooting common issues
+        - Available connectors and integrations
+        - Browser extension usage
+        - API documentation
+
+        This searches the official Surfsense documentation that was indexed
+        at deployment time. It does NOT search the user's personal knowledge base.
+
+        Args:
+            query: The search query about Surfsense usage or features
+            top_k: Number of documentation chunks to retrieve (default: 10)
+
+        Returns:
+            Relevant documentation content formatted with chunk IDs for citations
+        """
+        return await search_surfsense_docs_async(
+            query=query,
+            db_session=db_session,
+            top_k=top_k,
+        )
+
+    return search_surfsense_docs

surfsense_backend/app/agents/multi_agent_chat/subagents/builtins/research/tools/web_search.py

@@ -0,0 +1,241 @@
+"""Real-time web search: SearXNG plus configured live-search connectors (Tavily, Linkup, Baidu, etc.)."""
+
+import asyncio
+import json
+import time
+from typing import Any
+
+from langchain_core.tools import StructuredTool
+from pydantic import BaseModel, Field
+
+from app.db import shielded_async_session
+from app.services.connector_service import ConnectorService
+from app.utils.perf import get_perf_logger
+
+_LIVE_SEARCH_CONNECTORS: set[str] = {
+    "TAVILY_API",
+    "LINKUP_API",
+    "BAIDU_SEARCH_API",
+}
+
+_LIVE_CONNECTOR_SPECS: dict[str, tuple[str, bool, bool, dict[str, Any]]] = {
+    "TAVILY_API": ("search_tavily", False, True, {}),
+    "LINKUP_API": ("search_linkup", False, False, {"mode": "standard"}),
+    "BAIDU_SEARCH_API": ("search_baidu", False, True, {}),
+}
+
+_CONNECTOR_LABELS: dict[str, str] = {
+    "TAVILY_API": "Tavily",
+    "LINKUP_API": "Linkup",
+    "BAIDU_SEARCH_API": "Baidu",
+}
+
+
+class WebSearchInput(BaseModel):
+    """Input schema for the web_search tool."""
+
+    query: str = Field(
+        description="The search query to look up on the web. Use specific, descriptive terms.",
+    )
+    top_k: int = Field(
+        default=10,
+        description="Number of results to retrieve (default: 10, max: 50).",
+    )
+
+
+def _format_web_results(
+    documents: list[dict[str, Any]],
+    *,
+    max_chars: int = 50_000,
+) -> str:
+    """Format web search results into XML suitable for the LLM context."""
+    if not documents:
+        return "No web search results found."
+
+    parts: list[str] = []
+    total_chars = 0
+
+    for doc in documents:
+        doc_info = doc.get("document") or {}
+        metadata = doc_info.get("metadata") or {}
+        title = doc_info.get("title") or "Web Result"
+        url = metadata.get("url") or ""
+        content = (doc.get("content") or "").strip()
+        source = metadata.get("document_type") or doc.get("source") or "WEB_SEARCH"
+        if not content:
+            continue
+
+        metadata_json = json.dumps(metadata, ensure_ascii=False)
+        doc_xml = "\n".join(
+            [
+                "<document>",
+                "<document_metadata>",
+                f"  <document_type>{source}</document_type>",
+                f"  <title><![CDATA[{title}]]></title>",
+                f"  <url><![CDATA[{url}]]></url>",
+                f"  <metadata_json><![CDATA[{metadata_json}]]></metadata_json>",
+                "</document_metadata>",
+                "<document_content>",
+                f"  <chunk id='{url}'><![CDATA[{content}]]></chunk>",
+                "</document_content>",
+                "</document>",
+                "",
+            ]
+        )
+
+        if total_chars + len(doc_xml) > max_chars:
+            parts.append("<!-- Output truncated to fit context window -->")
+            break
+
+        parts.append(doc_xml)
+        total_chars += len(doc_xml)
+
+    return "\n".join(parts).strip() or "No web search results found."
+
+
+async def _search_live_connector(
+    connector: str,
+    query: str,
+    search_space_id: int,
+    top_k: int,
+    semaphore: asyncio.Semaphore,
+) -> list[dict[str, Any]]:
+    """Dispatch a single live-search connector (Tavily / Linkup / Baidu)."""
+    perf = get_perf_logger()
+    spec = _LIVE_CONNECTOR_SPECS.get(connector)
+    if spec is None:
+        return []
+
+    method_name, _includes_date_range, includes_top_k, extra_kwargs = spec
+    kwargs: dict[str, Any] = {
+        "user_query": query,
+        "search_space_id": search_space_id,
+        **extra_kwargs,
+    }
+    if includes_top_k:
+        kwargs["top_k"] = top_k
+
+    try:
+        t0 = time.perf_counter()
+        async with semaphore, shielded_async_session() as session:
+            svc = ConnectorService(session, search_space_id)
+            _, chunks = await getattr(svc, method_name)(**kwargs)
+            perf.info(
+                "[web_search] connector=%s results=%d in %.3fs",
+                connector,
+                len(chunks),
+                time.perf_counter() - t0,
+            )
+            return chunks
+    except Exception as e:
+        perf.warning("[web_search] connector=%s FAILED: %s", connector, e)
+        return []
+
+
+def create_web_search_tool(
+    search_space_id: int | None = None,
+    available_connectors: list[str] | None = None,
+) -> StructuredTool:
+    """Factory for the ``web_search`` tool.
+
+    Dispatches in parallel to the platform SearXNG instance and any
+    user-configured live-search connectors (Tavily, Linkup, Baidu).
+    """
+    active_live_connectors: list[str] = []
+    if available_connectors:
+        active_live_connectors = [
+            c for c in available_connectors if c in _LIVE_SEARCH_CONNECTORS
+        ]
+
+    engine_names = ["SearXNG (platform default)"]
+    engine_names.extend(_CONNECTOR_LABELS.get(c, c) for c in active_live_connectors)
+    engines_summary = ", ".join(engine_names)
+
+    description = (
+        "Search the web for real-time information. "
+        "Use this for current events, news, prices, weather, public facts, or any "
+        "question that requires up-to-date information from the internet.\n\n"
+        f"Active search engines: {engines_summary}.\n"
+        "All configured engines are queried in parallel and results are merged."
+    )
+
+    _search_space_id = search_space_id
+    _active_live = active_live_connectors
+
+    async def _web_search_impl(query: str, top_k: int = 10) -> str:
+        from app.services import web_search_service
+
+        perf = get_perf_logger()
+        t0 = time.perf_counter()
+        clamped_top_k = min(max(1, top_k), 50)
+
+        semaphore = asyncio.Semaphore(4)
+        tasks: list[asyncio.Task[list[dict[str, Any]]]] = []
+
+        if web_search_service.is_available():
+
+            async def _searxng() -> list[dict[str, Any]]:
+                async with semaphore:
+                    _result_obj, docs = await web_search_service.search(
+                        query=query,
+                        top_k=clamped_top_k,
+                    )
+                    return docs
+
+            tasks.append(asyncio.ensure_future(_searxng()))
+
+        if _search_space_id is not None:
+            for connector in _active_live:
+                tasks.append(
+                    asyncio.ensure_future(
+                        _search_live_connector(
+                            connector=connector,
+                            query=query,
+                            search_space_id=_search_space_id,
+                            top_k=clamped_top_k,
+                            semaphore=semaphore,
+                        )
+                    )
+                )
+
+        if not tasks:
+            return "Web search is not available — no search engines are configured."
+
+        results_lists = await asyncio.gather(*tasks, return_exceptions=True)
+
+        all_documents: list[dict[str, Any]] = []
+        for result in results_lists:
+            if isinstance(result, BaseException):
+                perf.warning("[web_search] a search engine failed: %s", result)
+                continue
+            all_documents.extend(result)
+
+        seen_urls: set[str] = set()
+        deduplicated: list[dict[str, Any]] = []
+        for doc in all_documents:
+            url = ((doc.get("document") or {}).get("metadata") or {}).get("url", "")
+            if url and url in seen_urls:
+                continue
+            if url:
+                seen_urls.add(url)
+            deduplicated.append(doc)
+
+        formatted = _format_web_results(deduplicated)
+
+        perf.info(
+            "[web_search] query=%r engines=%d results=%d deduped=%d chars=%d in %.3fs",
+            query[:60],
+            len(tasks),
+            len(all_documents),
+            len(deduplicated),
+            len(formatted),
+            time.perf_counter() - t0,
+        )
+        return formatted
+
+    return StructuredTool(
+        name="web_search",
+        description=description,
+        coroutine=_web_search_impl,
+        args_schema=WebSearchInput,
+    )

surfsense_backend/app/agents/multi_agent_chat/subagents/connectors/airtable/agent.py

@@ -0,0 +1,55 @@
+"""`airtable` route: ``SubAgent`` spec for deepagents."""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+from typing import Any
+
+from deepagents import SubAgent
+from langchain_core.language_models import BaseChatModel
+
+from app.agents.multi_agent_chat.subagents.shared.md_file_reader import (
+    read_md_file,
+)
+from app.agents.multi_agent_chat.subagents.shared.permissions import (
+    ToolsPermissions,
+    merge_tools_permissions,
+    middleware_gated_interrupt_on,
+)
+from app.agents.multi_agent_chat.subagents.shared.subagent_builder import (
+    pack_subagent,
+)
+
+from .tools.index import load_tools
+
+NAME = "airtable"
+
+
+def build_subagent(
+    *,
+    dependencies: dict[str, Any],
+    model: BaseChatModel | None = None,
+    extra_middleware: Sequence[Any] | None = None,
+    extra_tools_bucket: ToolsPermissions | None = None,
+) -> SubAgent:
+    buckets = load_tools(dependencies=dependencies)
+    merged_tools_bucket = merge_tools_permissions(buckets, extra_tools_bucket)
+    tools = [
+        row["tool"]
+        for row in (*merged_tools_bucket["allow"], *merged_tools_bucket["ask"])
+        if row.get("tool") is not None
+    ]
+    interrupt_on = middleware_gated_interrupt_on(merged_tools_bucket)
+    description = read_md_file(__package__, "description").strip()
+    if not description:
+        description = "Handles airtable tasks for this workspace."
+    system_prompt = read_md_file(__package__, "system_prompt").strip()
+    return pack_subagent(
+        name=NAME,
+        description=description,
+        system_prompt=system_prompt,
+        tools=tools,
+        interrupt_on=interrupt_on,
+        model=model,
+        extra_middleware=extra_middleware,
+    )

surfsense_backend/app/agents/multi_agent_chat/subagents/connectors/airtable/description.md

@@ -0,0 +1 @@
+Use for Airtable structured data operations: locate bases/tables and create/read/update records.