Merge pull request #1323 from MODSetter/dev_mod

feat: updated agent harness

.gitignore

@@ -7,4 +7,5 @@ node_modules/
 .pnpm-store
 .DS_Store
 deepagents/
-debug.log
+debug.log
+opencode/

surfsense_backend/.env.example

@@ -247,3 +247,42 @@ LANGSMITH_TRACING=true
 LANGSMITH_ENDPOINT=https://api.smith.langchain.com
 LANGSMITH_API_KEY=lsv2_pt_.....
 LANGSMITH_PROJECT=surfsense
+
+
+# =============================================================================
+# OPTIONAL: New-chat agent feature flags
+# =============================================================================
+# Master kill-switch — when true, every flag below is forced OFF.
+# SURFSENSE_DISABLE_NEW_AGENT_STACK=false
+
+# Agent quality
+# SURFSENSE_ENABLE_CONTEXT_EDITING=false
+# SURFSENSE_ENABLE_COMPACTION_V2=false
+# SURFSENSE_ENABLE_RETRY_AFTER=false
+# SURFSENSE_ENABLE_MODEL_FALLBACK=false
+# SURFSENSE_ENABLE_MODEL_CALL_LIMIT=false
+# SURFSENSE_ENABLE_TOOL_CALL_LIMIT=false
+# SURFSENSE_ENABLE_TOOL_CALL_REPAIR=false
+# SURFSENSE_ENABLE_DOOM_LOOP=false   # leave OFF until UI handles permission='doom_loop'
+
+# Safety
+# SURFSENSE_ENABLE_PERMISSION=false
+# SURFSENSE_ENABLE_BUSY_MUTEX=false
+# SURFSENSE_ENABLE_LLM_TOOL_SELECTOR=false   # adds a per-turn LLM call
+
+# Observability — OTel (also requires OTEL_EXPORTER_OTLP_ENDPOINT)
+# SURFSENSE_ENABLE_OTEL=false
+
+# Skills + subagents
+# SURFSENSE_ENABLE_SKILLS=false
+# SURFSENSE_ENABLE_SPECIALIZED_SUBAGENTS=false
+# SURFSENSE_ENABLE_KB_PLANNER_RUNNABLE=false
+
+# Snapshot / revert
+# SURFSENSE_ENABLE_ACTION_LOG=false
+# SURFSENSE_ENABLE_REVERT_ROUTE=false        # Backend-only; flip when UI ships
+
+# Plugins
+# SURFSENSE_ENABLE_PLUGIN_LOADER=false
+# Comma-separated allowlist of plugin entry-point names
+# SURFSENSE_ALLOWED_PLUGINS=year_substituter

surfsense_backend/alembic/versions/130_add_agent_action_log.py

@@ -0,0 +1,94 @@
+"""130_add_agent_action_log
+
+Revision ID: 130
+Revises: 129
+Create Date: 2026-04-28
+
+Adds the append-only ``agent_action_log`` table that
+:class:`ActionLogMiddleware` writes to after every tool call. Each row
+optionally carries a ``reverse_descriptor`` payload used by
+``POST /api/threads/{thread_id}/revert/{action_id}`` to undo the action.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+from sqlalchemy.dialects import postgresql
+
+from alembic import op
+
+revision: str = "130"
+down_revision: str | None = "129"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        "agent_action_log",
+        sa.Column("id", sa.Integer(), primary_key=True, index=True),
+        sa.Column(
+            "thread_id",
+            sa.Integer(),
+            sa.ForeignKey("new_chat_threads.id", ondelete="CASCADE"),
+            nullable=False,
+            index=True,
+        ),
+        sa.Column(
+            "user_id",
+            postgresql.UUID(as_uuid=True),
+            sa.ForeignKey("user.id", ondelete="SET NULL"),
+            nullable=True,
+            index=True,
+        ),
+        sa.Column(
+            "search_space_id",
+            sa.Integer(),
+            sa.ForeignKey("searchspaces.id", ondelete="CASCADE"),
+            nullable=False,
+            index=True,
+        ),
+        sa.Column("turn_id", sa.String(length=64), nullable=True, index=True),
+        sa.Column("message_id", sa.String(length=128), nullable=True, index=True),
+        sa.Column("tool_name", sa.String(length=255), nullable=False, index=True),
+        sa.Column("args", postgresql.JSONB(astext_type=sa.Text()), nullable=True),
+        sa.Column("result_id", sa.String(length=255), nullable=True),
+        sa.Column(
+            "reversible",
+            sa.Boolean(),
+            nullable=False,
+            server_default=sa.text("false"),
+        ),
+        sa.Column(
+            "reverse_descriptor",
+            postgresql.JSONB(astext_type=sa.Text()),
+            nullable=True,
+        ),
+        sa.Column("error", postgresql.JSONB(astext_type=sa.Text()), nullable=True),
+        sa.Column(
+            "reverse_of",
+            sa.Integer(),
+            sa.ForeignKey("agent_action_log.id", ondelete="SET NULL"),
+            nullable=True,
+            index=True,
+        ),
+        sa.Column(
+            "created_at",
+            sa.TIMESTAMP(timezone=True),
+            nullable=False,
+            server_default=sa.text("(now() AT TIME ZONE 'utc')"),
+            index=True,
+        ),
+    )
+    op.create_index(
+        "ix_agent_action_log_thread_created",
+        "agent_action_log",
+        ["thread_id", "created_at"],
+    )
+
+
+def downgrade() -> None:
+    op.drop_index("ix_agent_action_log_thread_created", table_name="agent_action_log")
+    op.drop_table("agent_action_log")

surfsense_backend/alembic/versions/131_add_document_revisions.py

@@ -0,0 +1,119 @@
+"""131_add_document_revisions
+
+Revision ID: 131
+Revises: 130
+Create Date: 2026-04-28
+
+Adds two snapshot tables that back the per-action revert flow:
+
+* ``document_revisions``: pre-mutation snapshot of NOTE/FILE/EXTENSION docs.
+* ``folder_revisions``: pre-mutation snapshot of folder mkdir/move/delete.
+
+Both are written by :class:`KnowledgeBasePersistenceMiddleware` ahead of
+state-changing tool calls and consumed by ``revert_service.revert_action``.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+from sqlalchemy.dialects import postgresql
+
+from alembic import op
+
+revision: str = "131"
+down_revision: str | None = "130"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        "document_revisions",
+        sa.Column("id", sa.Integer(), primary_key=True, index=True),
+        sa.Column(
+            "document_id",
+            sa.Integer(),
+            sa.ForeignKey("documents.id", ondelete="CASCADE"),
+            nullable=False,
+            index=True,
+        ),
+        sa.Column(
+            "search_space_id",
+            sa.Integer(),
+            sa.ForeignKey("searchspaces.id", ondelete="CASCADE"),
+            nullable=False,
+            index=True,
+        ),
+        sa.Column("content_before", sa.Text(), nullable=True),
+        sa.Column("title_before", sa.String(), nullable=True),
+        sa.Column("folder_id_before", sa.Integer(), nullable=True),
+        sa.Column(
+            "chunks_before", postgresql.JSONB(astext_type=sa.Text()), nullable=True
+        ),
+        sa.Column(
+            "metadata_before", postgresql.JSONB(astext_type=sa.Text()), nullable=True
+        ),
+        sa.Column(
+            "created_by_turn_id", sa.String(length=64), nullable=True, index=True
+        ),
+        sa.Column(
+            "agent_action_id",
+            sa.Integer(),
+            sa.ForeignKey("agent_action_log.id", ondelete="SET NULL"),
+            nullable=True,
+            index=True,
+        ),
+        sa.Column(
+            "created_at",
+            sa.TIMESTAMP(timezone=True),
+            nullable=False,
+            server_default=sa.text("(now() AT TIME ZONE 'utc')"),
+            index=True,
+        ),
+    )
+
+    op.create_table(
+        "folder_revisions",
+        sa.Column("id", sa.Integer(), primary_key=True, index=True),
+        sa.Column(
+            "folder_id",
+            sa.Integer(),
+            sa.ForeignKey("folders.id", ondelete="CASCADE"),
+            nullable=False,
+            index=True,
+        ),
+        sa.Column(
+            "search_space_id",
+            sa.Integer(),
+            sa.ForeignKey("searchspaces.id", ondelete="CASCADE"),
+            nullable=False,
+            index=True,
+        ),
+        sa.Column("name_before", sa.String(length=255), nullable=True),
+        sa.Column("parent_id_before", sa.Integer(), nullable=True),
+        sa.Column("position_before", sa.String(length=50), nullable=True),
+        sa.Column(
+            "created_by_turn_id", sa.String(length=64), nullable=True, index=True
+        ),
+        sa.Column(
+            "agent_action_id",
+            sa.Integer(),
+            sa.ForeignKey("agent_action_log.id", ondelete="SET NULL"),
+            nullable=True,
+            index=True,
+        ),
+        sa.Column(
+            "created_at",
+            sa.TIMESTAMP(timezone=True),
+            nullable=False,
+            server_default=sa.text("(now() AT TIME ZONE 'utc')"),
+            index=True,
+        ),
+    )
+
+
+def downgrade() -> None:
+    op.drop_table("folder_revisions")
+    op.drop_table("document_revisions")

surfsense_backend/alembic/versions/132_add_agent_permission_rules.py

@@ -0,0 +1,81 @@
+"""132_add_agent_permission_rules
+
+Revision ID: 132
+Revises: 131
+Create Date: 2026-04-28
+
+Adds the persistent ``agent_permission_rules`` table consumed by
+:class:`PermissionMiddleware` at agent build time. Rules can be scoped
+at search-space (``user_id`` / ``thread_id`` NULL), user-wide
+(``user_id`` set, ``thread_id`` NULL), or per-thread (``thread_id`` set).
+"""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+from sqlalchemy.dialects import postgresql
+
+from alembic import op
+
+revision: str = "132"
+down_revision: str | None = "131"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        "agent_permission_rules",
+        sa.Column("id", sa.Integer(), primary_key=True, index=True),
+        sa.Column(
+            "search_space_id",
+            sa.Integer(),
+            sa.ForeignKey("searchspaces.id", ondelete="CASCADE"),
+            nullable=False,
+            index=True,
+        ),
+        sa.Column(
+            "user_id",
+            postgresql.UUID(as_uuid=True),
+            sa.ForeignKey("user.id", ondelete="CASCADE"),
+            nullable=True,
+            index=True,
+        ),
+        sa.Column(
+            "thread_id",
+            sa.Integer(),
+            sa.ForeignKey("new_chat_threads.id", ondelete="CASCADE"),
+            nullable=True,
+            index=True,
+        ),
+        sa.Column("permission", sa.String(length=255), nullable=False),
+        sa.Column(
+            "pattern",
+            sa.String(length=255),
+            nullable=False,
+            server_default="*",
+        ),
+        sa.Column("action", sa.String(length=16), nullable=False),
+        sa.Column(
+            "created_at",
+            sa.TIMESTAMP(timezone=True),
+            nullable=False,
+            server_default=sa.text("(now() AT TIME ZONE 'utc')"),
+            index=True,
+        ),
+        sa.UniqueConstraint(
+            "search_space_id",
+            "user_id",
+            "thread_id",
+            "permission",
+            "pattern",
+            "action",
+            name="uq_agent_permission_rules_scope",
+        ),
+    )
+
+
+def downgrade() -> None:
+    op.drop_table("agent_permission_rules")

surfsense_backend/alembic/versions/133_drop_documents_content_hash_unique.py

@@ -0,0 +1,105 @@
+"""133_drop_documents_content_hash_unique
+
+Revision ID: 133
+Revises: 132
+Create Date: 2026-04-29
+
+Drop the global UNIQUE constraint on ``documents.content_hash`` so the
+new-chat agent's ``write_file`` flow can persist legitimate file copies
+(two paths, identical content) without hitting a constraint that mirrors
+no real filesystem semantic.
+
+Path uniqueness still lives on ``documents.unique_identifier_hash`` (per
+search space), which is the right invariant — exactly like an inode at a
+given path on a POSIX filesystem.
+
+The non-unique INDEX on ``content_hash`` is preserved so connector
+indexers' "have we seen this content before?" lookup
+(:func:`app.tasks.document_processors.base.check_duplicate_document`,
+which already uses ``.scalars().first()`` and is therefore tolerant of
+duplicates) stays cheap.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+
+from sqlalchemy import inspect
+
+from alembic import op
+
+revision: str = "133"
+down_revision: str | None = "132"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _existing_constraint_names(bind, table: str) -> set[str]:
+    inspector = inspect(bind)
+    return {c["name"] for c in inspector.get_unique_constraints(table)}
+
+
+def _existing_index_names(bind, table: str) -> set[str]:
+    inspector = inspect(bind)
+    return {i["name"] for i in inspector.get_indexes(table)}
+
+
+def upgrade() -> None:
+    bind = op.get_bind()
+
+    # Both the named UniqueConstraint (added in revision 8) and the
+    # implicit-unique-index variant SQLAlchemy may emit need draining.
+    constraints = _existing_constraint_names(bind, "documents")
+    if "uq_documents_content_hash" in constraints:
+        op.drop_constraint("uq_documents_content_hash", "documents", type_="unique")
+
+    indexes = _existing_index_names(bind, "documents")
+    # Some Postgres versions surface the unique constraint via a unique
+    # index of the same name; check for that too.
+    for idx_name in ("uq_documents_content_hash",):
+        if idx_name in indexes:
+            op.drop_index(idx_name, table_name="documents")
+
+    # Ensure the non-unique index is present for fast lookups.
+    if "ix_documents_content_hash" not in indexes:
+        op.create_index(
+            "ix_documents_content_hash",
+            "documents",
+            ["content_hash"],
+            unique=False,
+        )
+
+
+def downgrade() -> None:
+    bind = op.get_bind()
+
+    # Re-applying UNIQUE is destructive: there may now be legitimate
+    # duplicates (e.g. two NOTE documents that share content because the
+    # user explicitly copied one to a new path). To avoid the migration
+    # silently deleting user data, we keep only the lowest-id row per
+    # content_hash — same strategy revision 8 used when first introducing
+    # the constraint.
+    op.execute(
+        """
+        DELETE FROM documents
+        WHERE id NOT IN (
+            SELECT MIN(id)
+            FROM documents
+            GROUP BY content_hash
+        )
+        """
+    )
+
+    indexes = _existing_index_names(bind, "documents")
+    if "ix_documents_content_hash" in indexes:
+        op.drop_index("ix_documents_content_hash", table_name="documents")
+
+    op.create_index(
+        "ix_documents_content_hash",
+        "documents",
+        ["content_hash"],
+        unique=False,
+    )
+    op.create_unique_constraint(
+        "uq_documents_content_hash", "documents", ["content_hash"]
+    )

surfsense_backend/app/agents/autocomplete/autocomplete_agent.py

@@ -0,0 +1,557 @@
+"""Vision autocomplete agent with scoped filesystem exploration.
+
+Converts the stateless single-shot vision autocomplete into an agent that
+seeds a virtual filesystem from KB search results and lets the vision LLM
+explore documents via ``ls``, ``read_file``, ``glob``, ``grep``, etc.
+before generating the final completion.
+
+Performance: KB search and agent graph compilation run in parallel so
+the only sequential latency is KB-search (or agent compile, whichever is
+slower) + the agent's LLM turns.  There is no separate "query extraction"
+LLM call — the window title is used directly as the KB search query.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+import re
+import uuid
+from collections.abc import AsyncGenerator
+from typing import Any
+
+from deepagents.graph import BASE_AGENT_PROMPT
+from deepagents.middleware.patch_tool_calls import PatchToolCallsMiddleware
+from langchain.agents import create_agent
+from langchain_anthropic.middleware import AnthropicPromptCachingMiddleware
+from langchain_core.language_models import BaseChatModel
+from langchain_core.messages import AIMessage, ToolMessage
+
+from app.agents.new_chat.document_xml import build_document_xml
+from app.agents.new_chat.middleware.filesystem import SurfSenseFilesystemMiddleware
+from app.agents.new_chat.middleware.knowledge_search import (
+    search_knowledge_base,
+)
+from app.agents.new_chat.path_resolver import (
+    DOCUMENTS_ROOT,
+    build_path_index,
+    doc_to_virtual_path,
+)
+from app.db import shielded_async_session
+from app.services.new_streaming_service import VercelStreamingService
+
+try:
+    from deepagents.backends.utils import create_file_data
+except Exception:  # pragma: no cover - defensive
+
+    def create_file_data(content: str) -> dict[str, Any]:
+        return {"content": content.split("\n")}
+
+
+async def _build_autocomplete_filesystem(
+    *,
+    documents: Any,
+    search_space_id: int,
+) -> tuple[dict[str, Any], dict[int, str]]:
+    """Build a ``state['files']``-shaped dict from KB search results.
+
+    This is the autocomplete-specific replacement for the previous
+    ``build_scoped_filesystem`` helper. It uses the canonical path resolver
+    so paths line up with the rest of the system, including collision
+    suffixes for duplicate titles.
+    """
+    files: dict[str, Any] = {}
+    doc_id_to_path: dict[int, str] = {}
+
+    if not documents:
+        return files, doc_id_to_path
+
+    async with shielded_async_session() as session:
+        index = await build_path_index(session, search_space_id)
+
+    for document in documents:
+        if not isinstance(document, dict):
+            continue
+        meta = document.get("document") or {}
+        doc_id = meta.get("id")
+        if not isinstance(doc_id, int):
+            continue
+        title = str(meta.get("title") or "untitled")
+        folder_id = meta.get("folder_id")
+        path = doc_to_virtual_path(
+            doc_id=doc_id, title=title, folder_id=folder_id, index=index
+        )
+        chunk_ids = document.get("matched_chunk_ids") or []
+        try:
+            matched_set = {int(c) for c in chunk_ids}
+        except (TypeError, ValueError):
+            matched_set = set()
+        xml = build_document_xml(document, matched_chunk_ids=matched_set)
+        files[path] = create_file_data(xml)
+        doc_id_to_path[doc_id] = path
+
+    if not files:
+        # Ensure the synthetic /documents folder is visible even when empty.
+        files.setdefault(f"{DOCUMENTS_ROOT}/.placeholder", create_file_data(""))
+
+    return files, doc_id_to_path
+
+
+logger = logging.getLogger(__name__)
+
+KB_TOP_K = 10
+
+# ---------------------------------------------------------------------------
+# System prompt
+# ---------------------------------------------------------------------------
+
+AUTOCOMPLETE_SYSTEM_PROMPT = """You are a smart writing assistant that analyzes the user's screen to draft or complete text.
+
+You will receive a screenshot of the user's screen. Your PRIMARY source of truth is the screenshot itself — the visual context determines what to write.
+
+Your job:
+1. Analyze the ENTIRE screenshot to understand what the user is working on (email thread, chat conversation, document, code editor, form, etc.).
+2. Identify the text area where the user will type.
+3. Generate the text the user most likely wants to write based on the visual context.
+
+You also have access to the user's knowledge base documents via filesystem tools. However:
+- ONLY consult the knowledge base if the screenshot clearly involves a topic where your KB documents are DIRECTLY relevant (e.g., the user is writing about a specific project/topic that matches a document title).
+- Do NOT explore documents just because they exist. Most autocomplete requests can be answered purely from the screenshot.
+- If you do read a document, only incorporate information that is 100% relevant to what the user is typing RIGHT NOW. Do not add extra details, background, or tangential information from the KB.
+- Keep your output SHORT — autocomplete should feel like a natural continuation, not an essay.
+
+Key behavior:
+- If the text area is EMPTY, draft a concise response or message based on what you see on screen (e.g., reply to an email, respond to a chat message, continue a document).
+- If the text area already has text, continue it naturally — typically just a sentence or two.
+
+Rules:
+- Be CONCISE. Prefer a single paragraph or a few sentences. Autocomplete is a quick assist, not a full draft.
+- Match the tone and formality of the surrounding context.
+- If the screen shows code, write code. If it shows a casual chat, be casual. If it shows a formal email, be formal.
+- Do NOT describe the screenshot or explain your reasoning.
+- Do NOT cite or reference documents explicitly — just let the knowledge inform your writing naturally.
+- If you cannot determine what to write, output an empty JSON array: []
+
+## Output Format
+
+You MUST provide exactly 3 different suggestion options. Each should be a distinct, plausible completion — vary the tone, detail level, or angle.
+
+Return your suggestions as a JSON array of exactly 3 strings. Output ONLY the JSON array, nothing else — no markdown fences, no explanation, no commentary.
+
+Example format:
+["First suggestion text here.", "Second suggestion — a different take.", "Third option with another approach."]
+
+## Filesystem Tools `ls`, `read_file`, `write_file`, `edit_file`, `glob`, `grep`
+
+All file paths must start with a `/`.
+- ls: list files and directories at a given path.
+- read_file: read a file from the filesystem.
+- write_file: create a temporary file in the session (not persisted).
+- edit_file: edit a file in the session (not persisted for /documents/ files).
+- glob: find files matching a pattern (e.g., "**/*.xml").
+- grep: search for text within files.
+
+## When to Use Filesystem Tools
+
+BEFORE reaching for any tool, ask yourself: "Can I write a good completion purely from the screenshot?" If yes, just write it — do NOT explore the KB.
+
+Only use tools when:
+- The user is clearly writing about a specific topic that likely has detailed information in their KB.
+- You need a specific fact, name, number, or reference that the screenshot doesn't provide.
+
+When you do use tools, be surgical:
+- Check the `ls` output first. If no document title looks relevant, stop — do not read files just to see what's there.
+- If a title looks relevant, read only the `<chunk_index>` (first ~20 lines) and jump to matched chunks. Do not read entire documents.
+- Extract only the specific information you need and move on to generating the completion.
+
+## Reading Documents Efficiently
+
+Documents are formatted as XML. Each document contains:
+- `<document_metadata>` — title, type, URL, etc.
+- `<chunk_index>` — a table of every chunk with its **line range** and a
+  `matched="true"` flag for chunks that matched the search query.
+- `<document_content>` — the actual chunks in original document order.
+
+**Workflow**: read the first ~20 lines to see the `<chunk_index>`, identify
+chunks marked `matched="true"`, then use `read_file(path, offset=<start_line>,
+limit=<lines>)` to jump directly to those sections."""
+
+APP_CONTEXT_BLOCK = """
+
+The user is currently working in "{app_name}" (window: "{window_title}"). Use this to understand the type of application and adapt your tone and format accordingly."""
+
+
+def _build_autocomplete_system_prompt(app_name: str, window_title: str) -> str:
+    prompt = AUTOCOMPLETE_SYSTEM_PROMPT
+    if app_name:
+        prompt += APP_CONTEXT_BLOCK.format(app_name=app_name, window_title=window_title)
+    return prompt
+
+
+# ---------------------------------------------------------------------------
+# Pre-compute KB filesystem (runs in parallel with agent compilation)
+# ---------------------------------------------------------------------------
+
+
+class _KBResult:
+    """Container for pre-computed KB filesystem results."""
+
+    __slots__ = ("files", "ls_ai_msg", "ls_tool_msg")
+
+    def __init__(
+        self,
+        files: dict[str, Any] | None = None,
+        ls_ai_msg: AIMessage | None = None,
+        ls_tool_msg: ToolMessage | None = None,
+    ) -> None:
+        self.files = files
+        self.ls_ai_msg = ls_ai_msg
+        self.ls_tool_msg = ls_tool_msg
+
+    @property
+    def has_documents(self) -> bool:
+        return bool(self.files)
+
+
+async def precompute_kb_filesystem(
+    search_space_id: int,
+    query: str,
+    top_k: int = KB_TOP_K,
+) -> _KBResult:
+    """Search the KB and build the scoped filesystem outside the agent.
+
+    This is designed to be called via ``asyncio.gather`` alongside agent
+    graph compilation so the two run concurrently.
+    """
+    if not query:
+        return _KBResult()
+
+    try:
+        search_results = await search_knowledge_base(
+            query=query,
+            search_space_id=search_space_id,
+            top_k=top_k,
+        )
+
+        if not search_results:
+            return _KBResult()
+
+        new_files, _ = await _build_autocomplete_filesystem(
+            documents=search_results,
+            search_space_id=search_space_id,
+        )
+
+        if not new_files:
+            return _KBResult()
+
+        doc_paths = [
+            p
+            for p, v in new_files.items()
+            if p.startswith("/documents/") and v is not None
+        ]
+        tool_call_id = f"auto_ls_{uuid.uuid4().hex[:12]}"
+        ai_msg = AIMessage(
+            content="",
+            tool_calls=[
+                {"name": "ls", "args": {"path": "/documents"}, "id": tool_call_id}
+            ],
+        )
+        tool_msg = ToolMessage(
+            content=str(doc_paths) if doc_paths else "No documents found.",
+            tool_call_id=tool_call_id,
+        )
+        return _KBResult(files=new_files, ls_ai_msg=ai_msg, ls_tool_msg=tool_msg)
+
+    except Exception:
+        logger.warning(
+            "KB pre-computation failed, proceeding without KB", exc_info=True
+        )
+        return _KBResult()
+
+
+# ---------------------------------------------------------------------------
+# Filesystem middleware — no save_document, no persistence
+# ---------------------------------------------------------------------------
+
+
+class AutocompleteFilesystemMiddleware(SurfSenseFilesystemMiddleware):
+    """Filesystem middleware for autocomplete — read-only exploration only.
+
+    Passes ``search_space_id=None`` so the new persistence pipeline is
+    bypassed; the autocomplete flow only reads, never commits to Postgres.
+    """
+
+    def __init__(self) -> None:
+        super().__init__(search_space_id=None, created_by_id=None)
+
+
+# ---------------------------------------------------------------------------
+# Agent factory
+# ---------------------------------------------------------------------------
+
+
+async def _compile_agent(
+    llm: BaseChatModel,
+    app_name: str,
+    window_title: str,
+) -> Any:
+    """Compile the agent graph (CPU-bound, runs in a thread)."""
+    system_prompt = _build_autocomplete_system_prompt(app_name, window_title)
+    final_system_prompt = system_prompt + "\n\n" + BASE_AGENT_PROMPT
+
+    middleware = [
+        AutocompleteFilesystemMiddleware(),
+        PatchToolCallsMiddleware(),
+        AnthropicPromptCachingMiddleware(unsupported_model_behavior="ignore"),
+    ]
+
+    agent = await asyncio.to_thread(
+        create_agent,
+        llm,
+        system_prompt=final_system_prompt,
+        tools=[],
+        middleware=middleware,
+    )
+    return agent.with_config({"recursion_limit": 200})
+
+
+async def create_autocomplete_agent(
+    llm: BaseChatModel,
+    *,
+    search_space_id: int,
+    kb_query: str,
+    app_name: str = "",
+    window_title: str = "",
+) -> tuple[Any, _KBResult]:
+    """Create the autocomplete agent and pre-compute KB in parallel.
+
+    Returns ``(agent, kb_result)`` so the caller can inject the pre-computed
+    filesystem into the agent's initial state without any middleware delay.
+    """
+    agent, kb = await asyncio.gather(
+        _compile_agent(llm, app_name, window_title),
+        precompute_kb_filesystem(search_space_id, kb_query),
+    )
+    return agent, kb
+
+
+# ---------------------------------------------------------------------------
+# JSON suggestion parsing (with fallback)
+# ---------------------------------------------------------------------------
+
+
+def _parse_suggestions(raw: str) -> list[str]:
+    """Extract a list of suggestion strings from the agent's output.
+
+    Tries, in order:
+      1. Direct ``json.loads``
+      2. Extract content between ```json ... ``` fences
+      3. Find the first ``[`` … ``]`` span
+    Falls back to wrapping the raw text as a single suggestion.
+    """
+    text = raw.strip()
+    if not text:
+        return []
+
+    for candidate in _json_candidates(text):
+        try:
+            parsed = json.loads(candidate)
+            if isinstance(parsed, list) and all(isinstance(s, str) for s in parsed):
+                return [s for s in parsed if s.strip()]
+        except (json.JSONDecodeError, ValueError):
+            continue
+
+    return [text]
+
+
+def _json_candidates(text: str) -> list[str]:
+    """Yield candidate JSON strings from raw text."""
+    candidates = [text]
+
+    fence = re.search(r"```(?:json)?\s*\n?(.*?)```", text, re.DOTALL)
+    if fence:
+        candidates.append(fence.group(1).strip())
+
+    bracket = re.search(r"\[.*]", text, re.DOTALL)
+    if bracket:
+        candidates.append(bracket.group(0))
+
+    return candidates
+
+
+# ---------------------------------------------------------------------------
+# Streaming helper
+# ---------------------------------------------------------------------------
+
+
+async def stream_autocomplete_agent(
+    agent: Any,
+    input_data: dict[str, Any],
+    streaming_service: VercelStreamingService,
+    *,
+    emit_message_start: bool = True,
+) -> AsyncGenerator[str, None]:
+    """Stream agent events as Vercel SSE, with thinking steps for tool calls.
+
+    When ``emit_message_start`` is False the caller has already sent the
+    ``message_start`` event (e.g. to show preparation steps before the agent
+    runs).
+    """
+    thread_id = uuid.uuid4().hex
+    config = {"configurable": {"thread_id": thread_id}}
+
+    text_buffer: list[str] = []
+    active_tool_depth = 0
+    thinking_step_counter = 0
+    tool_step_ids: dict[str, str] = {}
+    step_titles: dict[str, str] = {}
+    completed_step_ids: set[str] = set()
+    last_active_step_id: str | None = None
+
+    def next_thinking_step_id() -> str:
+        nonlocal thinking_step_counter
+        thinking_step_counter += 1
+        return f"autocomplete-step-{thinking_step_counter}"
+
+    def complete_current_step() -> str | None:
+        nonlocal last_active_step_id
+        if last_active_step_id and last_active_step_id not in completed_step_ids:
+            completed_step_ids.add(last_active_step_id)
+            title = step_titles.get(last_active_step_id, "Done")
+            event = streaming_service.format_thinking_step(
+                step_id=last_active_step_id,
+                title=title,
+                status="complete",
+            )
+            last_active_step_id = None
+            return event
+        return None
+
+    if emit_message_start:
+        yield streaming_service.format_message_start()
+
+    gen_step_id = next_thinking_step_id()
+    last_active_step_id = gen_step_id
+    step_titles[gen_step_id] = "Generating suggestions"
+    yield streaming_service.format_thinking_step(
+        step_id=gen_step_id,
+        title="Generating suggestions",
+        status="in_progress",
+    )
+
+    try:
+        async for event in agent.astream_events(
+            input_data, config=config, version="v2"
+        ):
+            event_type = event.get("event", "")
+            if event_type == "on_chat_model_stream":
+                if active_tool_depth > 0:
+                    continue
+                if "surfsense:internal" in event.get("tags", []):
+                    continue
+                chunk = event.get("data", {}).get("chunk")
+                if chunk and hasattr(chunk, "content"):
+                    content = chunk.content
+                    if content and isinstance(content, str):
+                        text_buffer.append(content)
+
+            elif event_type == "on_chat_model_end":
+                if active_tool_depth > 0:
+                    continue
+                if "surfsense:internal" in event.get("tags", []):
+                    continue
+                output = event.get("data", {}).get("output")
+                if output and hasattr(output, "content"):
+                    if getattr(output, "tool_calls", None):
+                        continue
+                    content = output.content
+                    if content and isinstance(content, str) and not text_buffer:
+                        text_buffer.append(content)
+
+            elif event_type == "on_tool_start":
+                active_tool_depth += 1
+                tool_name = event.get("name", "unknown_tool")
+                run_id = event.get("run_id", "")
+                tool_input = event.get("data", {}).get("input", {})
+
+                step_event = complete_current_step()
+                if step_event:
+                    yield step_event
+
+                tool_step_id = next_thinking_step_id()
+                tool_step_ids[run_id] = tool_step_id
+                last_active_step_id = tool_step_id
+
+                title, items = _describe_tool_call(tool_name, tool_input)
+                step_titles[tool_step_id] = title
+                yield streaming_service.format_thinking_step(
+                    step_id=tool_step_id,
+                    title=title,
+                    status="in_progress",
+                    items=items,
+                )
+
+            elif event_type == "on_tool_end":
+                active_tool_depth = max(0, active_tool_depth - 1)
+                run_id = event.get("run_id", "")
+                step_id = tool_step_ids.pop(run_id, None)
+                if step_id and step_id not in completed_step_ids:
+                    completed_step_ids.add(step_id)
+                    title = step_titles.get(step_id, "Done")
+                    yield streaming_service.format_thinking_step(
+                        step_id=step_id,
+                        title=title,
+                        status="complete",
+                    )
+                    if last_active_step_id == step_id:
+                        last_active_step_id = None
+
+        step_event = complete_current_step()
+        if step_event:
+            yield step_event
+
+        raw_text = "".join(text_buffer)
+        suggestions = _parse_suggestions(raw_text)
+
+        yield streaming_service.format_data("suggestions", {"options": suggestions})
+
+        yield streaming_service.format_finish()
+        yield streaming_service.format_done()
+
+    except Exception as e:
+        logger.error(f"Autocomplete agent streaming error: {e}", exc_info=True)
+        yield streaming_service.format_error("Autocomplete failed. Please try again.")
+        yield streaming_service.format_done()
+
+
+def _describe_tool_call(tool_name: str, tool_input: Any) -> tuple[str, list[str]]:
+    """Return a human-readable (title, items) for a tool call thinking step."""
+    inp = tool_input if isinstance(tool_input, dict) else {}
+    if tool_name == "ls":
+        path = inp.get("path", "/")
+        return "Listing files", [path]
+    if tool_name == "read_file":
+        fp = inp.get("file_path", "")
+        display = fp if len(fp) <= 80 else "…" + fp[-77:]
+        return "Reading file", [display]
+    if tool_name == "write_file":
+        fp = inp.get("file_path", "")
+        display = fp if len(fp) <= 80 else "…" + fp[-77:]
+        return "Writing file", [display]
+    if tool_name == "edit_file":
+        fp = inp.get("file_path", "")
+        display = fp if len(fp) <= 80 else "…" + fp[-77:]
+        return "Editing file", [display]
+    if tool_name == "glob":
+        pat = inp.get("pattern", "")
+        base = inp.get("path", "/")
+        return "Searching files", [f"{pat} in {base}"]
+    if tool_name == "grep":
+        pat = inp.get("pattern", "")
+        path = inp.get("path", "")
+        display_pat = pat[:60] + ("…" if len(pat) > 60 else "")
+        return "Searching content", [
+            f'"{display_pat}"' + (f" in {path}" if path else "")
+        ]
+    return f"Using {tool_name}", []

surfsense_backend/app/agents/new_chat/chat_deepagent.py

@@ -23,9 +23,16 @@ from deepagents import SubAgent, SubAgentMiddleware, __version__ as deepagents_v
 from deepagents.backends import StateBackend
 from deepagents.graph import BASE_AGENT_PROMPT
 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 import create_agent
-from langchain.agents.middleware import TodoListMiddleware
+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
@@ -33,24 +40,54 @@ from langgraph.types import Checkpointer
 from sqlalchemy.ext.asyncio import AsyncSession
 
 from app.agents.new_chat.context import SurfSenseContextSchema
+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 FilesystemSelection
+from app.agents.new_chat.filesystem_selection import FilesystemMode, FilesystemSelection
 from app.agents.new_chat.llm_config import AgentConfig
 from app.agents.new_chat.middleware import (
+    ActionLogMiddleware,
+    AnonymousDocumentMiddleware,
+    BusyMutexMiddleware,
+    ClearToolUsesEdit,
     DedupHITLToolCallsMiddleware,
+    DoomLoopMiddleware,
     FileIntentMiddleware,
-    KnowledgeBaseSearchMiddleware,
+    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.middleware.safe_summarization import (
+from app.agents.new_chat.permissions import Rule, Ruleset
-    create_safe_summarization_middleware,
+from app.agents.new_chat.plugin_loader import (
+    PluginContext,
+    load_allowed_plugin_names_from_env,
+    load_plugin_middlewares,
 )
+from app.agents.new_chat.subagents import build_specialized_subagents
 from app.agents.new_chat.system_prompt import (
     build_configurable_system_prompt,
     build_surfsense_system_prompt,
 )
-from app.agents.new_chat.tools.registry import build_tools_async, get_connector_gated_tools
+from app.agents.new_chat.tools.invalid_tool import (
+    INVALID_TOOL_NAME,
+    invalid_tool,
+)
+from app.agents.new_chat.tools.registry import (
+    BUILTIN_TOOLS,
+    build_tools_async,
+    get_connector_gated_tools,
+)
 from app.db import ChatVisibility
 from app.services.connector_service import ConnectorService
 from app.utils.perf import get_perf_logger
@@ -243,7 +280,12 @@ async def create_surfsense_deep_agent(
     """
     _t_agent_total = time.perf_counter()
     filesystem_selection = filesystem_selection or FilesystemSelection()
-    backend_resolver = build_backend_resolver(filesystem_selection)
+    backend_resolver = build_backend_resolver(
+        filesystem_selection,
+        search_space_id=search_space_id
+        if filesystem_selection.mode == FilesystemMode.CLOUD
+        else None,
+    )
 
     # Discover available connectors and document types for this search space
     available_connectors: list[str] | None = None
@@ -294,11 +336,11 @@ async def create_surfsense_deep_agent(
     }
 
     modified_disabled_tools = list(disabled_tools) if disabled_tools else []
-    modified_disabled_tools.extend(
+    modified_disabled_tools.extend(get_connector_gated_tools(available_connectors))
-        get_connector_gated_tools(available_connectors)
-    )
 
-    # Remove direct KB search tool; we now pre-seed a scoped filesystem via middleware.
+    # Remove direct KB search tool; KnowledgePriorityMiddleware now runs hybrid
+    # search per turn and surfaces hits as a <priority_documents> hint plus
+    # `<chunk_index matched="true">` markers inside lazy-loaded XML.
     if "search_knowledge_base" not in modified_disabled_tools:
         modified_disabled_tools.append("search_knowledge_base")
 
@@ -310,6 +352,18 @@ async def create_surfsense_deep_agent(
         disabled_tools=modified_disabled_tools,
         additional_tools=list(additional_tools) if additional_tools else None,
     )
+
+    # Register the ``invalid`` tool only when tool-call repair is on. It
+    # is dispatched only when :class:`ToolCallNameRepairMiddleware`
+    # rewrites a malformed call. We intentionally append it AFTER
+    # ``build_tools_async`` so it never appears in the system-prompt
+    # tool list (which is built from the registry, not the bound tool
+    # list).
+    _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,
@@ -328,7 +382,8 @@ async def create_surfsense_deep_agent(
         meta = getattr(t, "metadata", None) or {}
         if meta.get("mcp_is_generic") and meta.get("mcp_connector_name"):
             _mcp_connector_tools.setdefault(
-                meta["mcp_connector_name"], [],
+                meta["mcp_connector_name"],
+                [],
             ).append(t.name)
 
     if _mcp_connector_tools:
@@ -355,7 +410,139 @@ async def create_surfsense_deep_agent(
         "[create_agent] System prompt built in %.3fs", time.perf_counter() - _t0
     )
 
-    # -- Build the middleware stack (mirrors create_deep_agent internals) ------
+    # Combine system_prompt with BASE_AGENT_PROMPT (same as create_deep_agent)
+    final_system_prompt = system_prompt + "\n\n" + BASE_AGENT_PROMPT
+
+    # The middleware stack — and especially ``SubAgentMiddleware`` — is *not*
+    # cheap to build. ``SubAgentMiddleware.__init__`` calls ``create_agent``
+    # synchronously to compile the general-purpose subagent's full state graph
+    # (every tool + every middleware → pydantic schemas + langgraph compile).
+    # On gpt-5.x agents that's roughly 1.5-2s of pure CPU work. If we run it
+    # directly here it blocks the asyncio event loop for the whole streaming
+    # task (and any other coroutine sharing this loop), which is why
+    # "agent creation" wall-clock time used to stretch to ~3-4s. Move the
+    # entire middleware build + main-graph compile into a single
+    # ``asyncio.to_thread`` so the heavy CPU work runs off-loop and the
+    # event loop stays responsive.
+    _t0 = time.perf_counter()
+    agent = await asyncio.to_thread(
+        _build_compiled_agent_blocking,
+        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,
+    )
+    _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
+
+
+# Tools whose output is too costly / lossy to discard. Keep this
+# conservative — anything listed here is *never* pruned by
+# :class:`ContextEditingMiddleware`. The list is filtered against
+# actually-bound tool names so disabled connectors don't show up here.
+_PRUNE_PROTECTED_TOOL_NAMES: frozenset[str] = frozenset(
+    {
+        "generate_report",
+        "generate_resume",
+        "generate_podcast",
+        "generate_video_presentation",
+        "generate_image",
+        # Read-heavy connector reads — recomputing them is expensive
+        "read_email",
+        "search_emails",
+        # The fallback for malformed tool calls — keep its replies visible
+        "invalid",
+    }
+)
+
+
+def _safe_exclude_tools(tools: Sequence[BaseTool]) -> tuple[str, ...]:
+    """Return ``exclude_tools`` derived from the actually-bound tool list.
+
+    Filters :data:`_PRUNE_PROTECTED_TOOL_NAMES` against the bound tools
+    so we never list tools that don't exist (would be a silent no-op).
+    """
+    enabled = {t.name for t in tools}
+    return tuple(name for name in _PRUNE_PROTECTED_TOOL_NAMES if name in enabled)
+
+
+# Connector gating: any tool whose ``ToolDefinition.required_connector``
+# isn't actually wired up gets a synthesized permission deny rule so
+# execution attempts short-circuit with ``permission_denied`` instead of
+# bubbling up provider-specific 401/404 errors. Mirrors OpenCode's
+# ``Permission.disabled`` (declarative, per-tool gating) — replaces the
+# legacy binary ``_CONNECTOR_TYPE_TO_SEARCHABLE`` substring-heuristic.
+def _synthesize_connector_deny_rules(
+    *,
+    available_connectors: list[str] | None,
+    enabled_tool_names: set[str],
+) -> list[Rule]:
+    """Build deny rules for tools whose required connector is not enabled.
+
+    Source of truth is ``ToolDefinition.required_connector`` in
+    :data:`BUILTIN_TOOLS`. A tool only gets a deny rule when:
+
+    1. It is currently bound (``enabled_tool_names``).
+    2. It declares a ``required_connector``.
+    3. That connector is *not* in ``available_connectors``.
+    """
+    available = set(available_connectors or [])
+    deny: list[Rule] = []
+    for tool_def in BUILTIN_TOOLS:
+        if tool_def.name not in enabled_tool_names:
+            continue
+        rc = tool_def.required_connector
+        if rc and rc not in available:
+            deny.append(Rule(permission=tool_def.name, pattern="*", action="deny"))
+    return deny
+
+
+def _build_compiled_agent_blocking(
+    *,
+    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,
+):
+    """Build the middleware stack and compile the agent graph synchronously.
+
+    Runs in a worker thread (see ``asyncio.to_thread`` call site) so the heavy
+    CPU work — most notably ``SubAgentMiddleware.__init__`` eagerly calling
+    ``create_agent`` to compile the general-purpose subagent — does not block
+    the event loop.
+    """
     _memory_middleware = MemoryInjectionMiddleware(
         user_id=user_id,
         search_space_id=search_space_id,
@@ -363,18 +550,23 @@ async def create_surfsense_deep_agent(
     )
 
     # General-purpose subagent middleware
+    # Subagent omits AnonymousDocumentMiddleware, KnowledgeTreeMiddleware,
+    # KnowledgePriorityMiddleware, and KnowledgeBasePersistenceMiddleware - it
+    # inherits state and tools from the parent, but should not (a) re-load
+    # anon docs / re-render the tree / re-run hybrid search, or (b) commit at
+    # its own completion (only the top-level agent's aafter_agent commits).
     gp_middleware = [
         TodoListMiddleware(),
         _memory_middleware,
         FileIntentMiddleware(llm=llm),
         SurfSenseFilesystemMiddleware(
             backend=backend_resolver,
-            filesystem_mode=filesystem_selection.mode,
+            filesystem_mode=filesystem_mode,
             search_space_id=search_space_id,
             created_by_id=user_id,
             thread_id=thread_id,
         ),
-        create_safe_summarization_middleware(llm, StateBackend),
+        create_surfsense_compaction_middleware(llm, StateBackend),
         PatchToolCallsMiddleware(),
         AnthropicPromptCachingMiddleware(unsupported_model_behavior="ignore"),
     ]
@@ -386,48 +578,416 @@ async def create_surfsense_deep_agent(
         "middleware": gp_middleware,
     }
 
+    # Specialized user-facing subagents (explore, report_writer,
+    # connector_negotiator). Registered through SubAgentMiddleware alongside
+    # the general-purpose spec so the parent's `task` tool can address them
+    # by name. Off by default until the flag flips so existing deployments
+    # don't see new agent types in the task tool description.
+    specialized_subagents: list[SubAgent] = []
+    if flags.enable_specialized_subagents and not flags.disable_new_agent_stack:
+        try:
+            # Specialized subagents share the parent's filesystem +
+            # todo view so their system prompts (which promise
+            # ``read_file``, ``ls``, ``grep``, ``glob``, ``write_todos``)
+            # actually match runtime behavior. Build *fresh* instances
+            # rather than aliasing the parent's GP middleware to avoid
+            # subtle state coupling across compiled graphs.
+            subagent_extra_middleware: list = [
+                TodoListMiddleware(),
+                SurfSenseFilesystemMiddleware(
+                    backend=backend_resolver,
+                    filesystem_mode=filesystem_mode,
+                    search_space_id=search_space_id,
+                    created_by_id=user_id,
+                    thread_id=thread_id,
+                ),
+            ]
+            specialized_subagents = build_specialized_subagents(
+                tools=tools,
+                model=llm,
+                extra_middleware=subagent_extra_middleware,
+            )
+        except Exception as exc:  # pragma: no cover - defensive
+            logging.warning(
+                "Specialized subagent build failed; running without them: %s",
+                exc,
+            )
+            specialized_subagents = []
+
+    subagent_specs: list[SubAgent] = [general_purpose_spec, *specialized_subagents]
+
     # Main agent middleware
+    # Order: AnonDoc -> Tree -> Priority -> FileIntent -> Filesystem -> Persistence -> ...
+    # before_agent hooks run in declared order; later injections sit closer to
+    # the latest human turn. Tree (large + cacheable) is injected earliest so
+    # provider-side prefix caching has more material to hit; FileIntent (most
+    # actionable per-turn contract) is injected closest to the user message.
+    #
+    # ``wrap_model_call`` ordering: the FIRST middleware in the list is the
+    # OUTERMOST wrapper. To ensure prune executes before summarization,
+    # place ``SpillingContextEditingMiddleware`` before
+    # ``SurfSenseCompactionMiddleware``. Compaction is the canonical
+    # token-budget defense; the Bedrock buffer-empty defense is folded
+    # into ``SurfSenseCompactionMiddleware``.
+    summarization_mw = create_surfsense_compaction_middleware(llm, StateBackend)
+    _ = flags.enable_compaction_v2  # historical flag; retained for telemetry parity
+
+    # ContextEditing prune. Trigger at 55% of ``max_input_tokens``,
+    # earlier than summarization (~85%). When disabled, no edit runs.
+    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,
+        )
+
+    # Resilience knobs: header-aware retry, model fallback, and
+    # per-thread / per-run call-count limits. The fallback / limit
+    # middlewares are vanilla LangChain primitives; ``RetryAfter`` is
+    # SurfSense's header-aware variant (see its module docstring).
+    retry_mw = (
+        RetryAfterMiddleware(max_retries=3)
+        if flags.enable_retry_after and not flags.disable_new_agent_stack
+        else None
+    )
+    # Fallback chain — primary is the agent's own model; we add cheap
+    # alternatives. Off by default; only the first call site that
+    # configures the chain via env should enable it.
+    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
+    )
+
+    # Provider-compat ``_noop`` injection (mirrors OpenCode's
+    # ``llm.ts`` workaround for providers that reject empty assistant
+    # turns or alternating-role constraints).
+    noop_mw = (
+        NoopInjectionMiddleware()
+        if flags.enable_compaction_v2 and not flags.disable_new_agent_stack
+        else None
+    )
+
+    # Tool-call name repair (lowercase + ``invalid`` fallback).
+    #
+    # ``registered_tool_names`` MUST cover every tool the model can legitimately
+    # call. That includes the bound ``tools`` list AND every tool provided by
+    # middleware in the stack — ``FilesystemMiddleware`` (read_file, ls, grep,
+    # glob, edit_file, write_file, execute), ``TodoListMiddleware``
+    # (write_todos), ``SubAgentMiddleware`` (task), ``SkillsMiddleware`` (skill
+    # loaders), etc. If we only inspect ``tools`` here, every call to
+    # ``read_file`` / ``ls`` / ``grep`` from the model will be rewritten to
+    # ``invalid`` because the repair middleware doesn't recognize them. The
+    # built-in deepagents middleware aren't in scope yet at this point of the
+    # function but they're added unconditionally below, so we hard-code their
+    # canonical names alongside the dynamic ``tools`` set.
+    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}
+        # Tools owned by the standard deepagents middleware stack.
+        registered_names |= {
+            "write_todos",
+            "ls",
+            "read_file",
+            "write_file",
+            "edit_file",
+            "glob",
+            "grep",
+            "execute",
+            "task",
+        }
+        repair_mw = ToolCallNameRepairMiddleware(
+            registered_tool_names=registered_names,
+            # Disable fuzzy matching to avoid silent rewrites; the
+            # lowercase + ``invalid`` fallback alone covers >95% of
+            # observed model errors.
+            fuzzy_match_threshold=None,
+        )
+
+    # Doom-loop detector. Off by default until the frontend handles
+    # ``permission == "doom_loop"`` interrupts.
+    doom_loop_mw = (
+        DoomLoopMiddleware(threshold=3)
+        if flags.enable_doom_loop and not flags.disable_new_agent_stack
+        else None
+    )
+
+    # PermissionMiddleware. Layers, earliest -> latest (last match wins,
+    # same evaluation order as OpenCode's ``permission/index.ts``):
+    #
+    # 1. ``surfsense_defaults`` — single ``allow */*`` rule. SurfSense
+    #    already runs per-tool HITL (see ``tools/hitl.py``) for mutating
+    #    connector tools, so we only want PermissionMiddleware to *deny*
+    #    things the user has gated off; the default fallback in
+    #    ``permissions.evaluate`` is ``ask``, which would double-prompt
+    #    on every safe read-only call (``ls``, ``read_file``, ``grep``,
+    #    ``glob``, ``web_search`` …) and, on resume, replay the previous
+    #    reject decision into innocent calls.
+    # 2. ``connector_synthesized`` — deny rules for tools whose required
+    #    connector is not connected to this space. Overrides #1.
+    # 3. (future) user-defined rules from ``agent_permission_rules`` table
+    #    via the Agent Permissions UI. Loaded last so they override both.
+    permission_mw: PermissionMiddleware | None = None
+    if flags.enable_permission and not flags.disable_new_agent_stack:
+        synthesized = _synthesize_connector_deny_rules(
+            available_connectors=available_connectors,
+            enabled_tool_names={t.name for t in tools},
+        )
+        permission_mw = PermissionMiddleware(
+            rulesets=[
+                Ruleset(
+                    rules=[Rule(permission="*", pattern="*", action="allow")],
+                    origin="surfsense_defaults",
+                ),
+                Ruleset(rules=synthesized, origin="connector_synthesized"),
+            ],
+        )
+
+    # ActionLogMiddleware. Off by default until the ``agent_action_log``
+    # table is migrated. When enabled, persists one row per tool call
+    # with optional reverse_descriptor for
+    # ``POST /api/threads/{thread_id}/revert/{action_id}``. Sits inside
+    # ``permission`` so denied calls aren't logged as completions.
+    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
+
+    # Per-thread busy mutex (refuse a second concurrent turn on the same
+    # thread; see :class:`BusyMutexMiddleware` docstring).
+    busy_mutex_mw: BusyMutexMiddleware | None = (
+        BusyMutexMiddleware()
+        if flags.enable_busy_mutex and not flags.disable_new_agent_stack
+        else None
+    )
+
+    # OpenTelemetry spans (model.call + tool.call). Lives just inside
+    # BusyMutex so it spans every retry/fallback attempt of the current
+    # turn but never wraps a queued/blocked turn.
+    otel_mw: OtelSpanMiddleware | None = (
+        OtelSpanMiddleware()
+        if flags.enable_otel and not flags.disable_new_agent_stack
+        else None
+    )
+
+    # Plugin entry-point loader. Off by default; opt-in via the
+    # ``SURFSENSE_ENABLE_PLUGIN_LOADER`` flag. The allowlist is read from
+    # the ``SURFSENSE_ALLOWED_PLUGINS`` env var (comma-separated). A future
+    # PR can wire it through ``global_llm_config.yaml``.
+    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 = []
+
+    # SkillsMiddleware (deepagents) loads built-in + space-authored
+    # skills via a CompositeBackend. Sources are layered: built-in first,
+    # space last, so a search-space-authored skill of the same name
+    # overrides the bundled one.
+    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
+
+    # LangChain's LLM-driven tool selection — only enabled for stacks
+    # large enough to need narrowing (>30 tools).
+    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 = [
+        # BusyMutex is OUTERMOST: it must wrap the entire stream so no
+        # other turn can sneak in while this one is mid-flight.
+        busy_mutex_mw,
+        # OTel spans sit just inside BusyMutex so each retry attempt
+        # gets its own model.call / tool.call span.
+        otel_mw,
         TodoListMiddleware(),
         _memory_middleware,
-        FileIntentMiddleware(llm=llm),
+        AnonymousDocumentMiddleware(
-        KnowledgeBaseSearchMiddleware(
+            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_selection.mode,
+            filesystem_mode=filesystem_mode,
             available_connectors=available_connectors,
             available_document_types=available_document_types,
             mentioned_document_ids=mentioned_document_ids,
-            anon_session_id=anon_session_id,
         ),
+        FileIntentMiddleware(llm=llm),
         SurfSenseFilesystemMiddleware(
             backend=backend_resolver,
-            filesystem_mode=filesystem_selection.mode,
+            filesystem_mode=filesystem_mode,
             search_space_id=search_space_id,
             created_by_id=user_id,
             thread_id=thread_id,
         ),
-        SubAgentMiddleware(backend=StateBackend, subagents=[general_purpose_spec]),
+        KnowledgeBasePersistenceMiddleware(
-        create_safe_summarization_middleware(llm, StateBackend),
+            search_space_id=search_space_id,
+            created_by_id=user_id,
+            filesystem_mode=filesystem_mode,
+        )
+        if filesystem_mode == FilesystemMode.CLOUD
+        else None,
+        # Skill loader. Placed before SubAgentMiddleware so subagents
+        # inherit the same skill metadata (subagent specs reference the
+        # same source paths via ``default_skills_sources()``).
+        skills_mw,
+        SubAgentMiddleware(backend=StateBackend, subagents=subagent_specs),
+        # Tool selection (only when >30 tools and flag on).
+        selector_mw,
+        # Defensive caps, then prune, then summarize.
+        model_call_limit_mw,
+        tool_call_limit_mw,
+        context_edit_mw,
+        summarization_mw,
+        # Provider compatibility + retry chain — placed after prune/compact
+        # so retries happen on the already-trimmed payload.
+        noop_mw,
+        retry_mw,
+        fallback_mw,
+        # Tool-call repair must run after model emits but before
+        # permission / dedup / doom-loop interpret the calls.
+        repair_mw,
+        # Permission deny/ask BEFORE the calls are forwarded to tool nodes.
+        permission_mw,
+        doom_loop_mw,
+        # Action log sits inside permission so denied calls don't appear
+        # as completions, and outside dedup so each unique tool invocation
+        # gets its own row.
+        action_log_mw,
         PatchToolCallsMiddleware(),
-        DedupHITLToolCallsMiddleware(agent_tools=tools),
+        DedupHITLToolCallsMiddleware(agent_tools=list(tools)),
+        # Plugin slot — sits just before AnthropicCache so plugin-side
+        # transforms see the final tool result and run before any
+        # caching heuristics. Multiple plugins in declared order; loader
+        # filtered by the admin allowlist already.
+        *plugin_middlewares,
         AnthropicPromptCachingMiddleware(unsupported_model_behavior="ignore"),
     ]
+    deepagent_middleware = [m for m in deepagent_middleware if m is not None]
 
-    # Combine system_prompt with BASE_AGENT_PROMPT (same as create_deep_agent)
+    agent = create_agent(
-    final_system_prompt = system_prompt + "\n\n" + BASE_AGENT_PROMPT
-
-    _t0 = time.perf_counter()
-    agent = await asyncio.to_thread(
-        create_agent,
         llm,
         system_prompt=final_system_prompt,
-        tools=tools,
+        tools=list(tools),
         middleware=deepagent_middleware,
         context_schema=SurfSenseContextSchema,
         checkpointer=checkpointer,
     )
-    agent = agent.with_config(
+    return agent.with_config(
         {
             "recursion_limit": 10_000,
             "metadata": {
@@ -436,13 +996,3 @@ async def create_surfsense_deep_agent(
             },
         }
     )
-    _perf_log.info(
-        "[create_agent] Graph compiled (create_agent) 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/new_chat/document_xml.py

@@ -0,0 +1,103 @@
+"""Shared XML builder for KB documents.
+
+Produces the citation-friendly XML used by every read of a knowledge-base
+document (lazy-loaded by :class:`KBPostgresBackend` and synthetic anonymous
+files). The XML carries a ``<chunk_index>`` near the top so the LLM can jump
+directly to matched-chunk line ranges via ``read_file(offset=…, limit=…)``.
+
+Extracted from the original ``knowledge_search.py`` so the backend, the
+priority middleware, and any future renderer share a single implementation.
+"""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+
+
+def build_document_xml(
+    document: dict[str, Any],
+    matched_chunk_ids: set[int] | None = None,
+) -> str:
+    """Build citation-friendly XML with a ``<chunk_index>`` for smart seeking.
+
+    Args:
+        document: Dict shape produced by hybrid search / lazy-load helpers.
+            Expected keys: ``document`` (with ``id``, ``title``,
+            ``document_type``, ``metadata``) and ``chunks``
+            (list of ``{chunk_id, content}``).
+        matched_chunk_ids: Optional set of chunk IDs to flag as
+            ``matched="true"`` in the chunk index.
+    """
+    matched = matched_chunk_ids or set()
+
+    doc_meta = document.get("document") or {}
+    metadata = (doc_meta.get("metadata") or {}) if isinstance(doc_meta, dict) else {}
+    document_id = doc_meta.get("id", document.get("document_id", "unknown"))
+    document_type = doc_meta.get("document_type", document.get("source", "UNKNOWN"))
+    title = doc_meta.get("title") or metadata.get("title") or "Untitled Document"
+    url = (
+        metadata.get("url") or metadata.get("source") or metadata.get("page_url") or ""
+    )
+    metadata_json = json.dumps(metadata, ensure_ascii=False)
+
+    metadata_lines: list[str] = [
+        "<document>",
+        "<document_metadata>",
+        f"  <document_id>{document_id}</document_id>",
+        f"  <document_type>{document_type}</document_type>",
+        f"  <title><![CDATA[{title}]]></title>",
+        f"  <url><![CDATA[{url}]]></url>",
+        f"  <metadata_json><![CDATA[{metadata_json}]]></metadata_json>",
+        "</document_metadata>",
+        "",
+    ]
+
+    chunks = document.get("chunks") or []
+    chunk_entries: list[tuple[int | None, str]] = []
+    if isinstance(chunks, list):
+        for chunk in chunks:
+            if not isinstance(chunk, dict):
+                continue
+            chunk_id = chunk.get("chunk_id") or chunk.get("id")
+            chunk_content = str(chunk.get("content", "")).strip()
+            if not chunk_content:
+                continue
+            if chunk_id is None:
+                xml = f"  <chunk><![CDATA[{chunk_content}]]></chunk>"
+            else:
+                xml = f"  <chunk id='{chunk_id}'><![CDATA[{chunk_content}]]></chunk>"
+            chunk_entries.append((chunk_id, xml))
+
+    index_overhead = 1 + len(chunk_entries) + 1 + 1 + 1
+    first_chunk_line = len(metadata_lines) + index_overhead + 1
+
+    current_line = first_chunk_line
+    index_entry_lines: list[str] = []
+    for cid, xml_str in chunk_entries:
+        num_lines = xml_str.count("\n") + 1
+        end_line = current_line + num_lines - 1
+        matched_attr = ' matched="true"' if cid is not None and cid in matched else ""
+        if cid is not None:
+            index_entry_lines.append(
+                f'  <entry chunk_id="{cid}" lines="{current_line}-{end_line}"{matched_attr}/>'
+            )
+        else:
+            index_entry_lines.append(
+                f'  <entry lines="{current_line}-{end_line}"{matched_attr}/>'
+            )
+        current_line = end_line + 1
+
+    lines = metadata_lines.copy()
+    lines.append("<chunk_index>")
+    lines.extend(index_entry_lines)
+    lines.append("</chunk_index>")
+    lines.append("")
+    lines.append("<document_content>")
+    for _, xml_str in chunk_entries:
+        lines.append(xml_str)
+    lines.extend(["</document_content>", "</document>"])
+    return "\n".join(lines)
+
+
+__all__ = ["build_document_xml"]

surfsense_backend/app/agents/new_chat/errors.py

@@ -0,0 +1,95 @@
+"""
+Typed error taxonomy for the SurfSense agent stack.
+
+Used by:
+- :class:`RetryAfterMiddleware` — its ``retry_on`` callable consults
+  the error code to decide whether a retry is appropriate.
+- :class:`PermissionMiddleware` — emits ``code="permission_denied"``
+  errors when a deny rule trips.
+- All tools — return :class:`StreamingError` payloads in
+  ``ToolMessage.additional_kwargs["error"]`` so the model and the
+  retry/permission layers share a contract.
+"""
+
+from __future__ import annotations
+
+from typing import Literal
+
+from pydantic import BaseModel, Field
+
+ErrorCode = Literal[
+    "rate_limit",
+    "auth",
+    "tool_validation",
+    "tool_runtime",
+    "context_overflow",
+    "provider",
+    "permission_denied",
+    "doom_loop",
+    "busy",
+    "cancelled",
+]
+
+
+class StreamingError(BaseModel):
+    """Structured error payload attached to ``ToolMessage.additional_kwargs["error"]``.
+
+    Tools and middleware emit this so retry, permission, and routing
+    layers can decide what to do without parsing free-form strings.
+    """
+
+    code: ErrorCode
+    retryable: bool = False
+    suggestion: str | None = None
+    correlation_id: str | None = None
+    detail: str | None = Field(
+        default=None,
+        description="Free-form additional context. Not surfaced to the model.",
+    )
+
+    class Config:
+        frozen = True
+
+
+class RejectedError(Exception):
+    """Raised when the user rejects a permission ask without feedback.
+
+    Caught by :class:`PermissionMiddleware`; the agent stops the current
+    tool fan-out and surfaces a user-facing rejection.
+    """
+
+    def __init__(self, *, tool: str | None = None, pattern: str | None = None) -> None:
+        super().__init__(f"Permission rejected for tool {tool!r}, pattern {pattern!r}")
+        self.tool = tool
+        self.pattern = pattern
+
+
+class CorrectedError(Exception):
+    """Raised when the user rejects a permission ask *with* feedback.
+
+    The :class:`PermissionMiddleware` translates the feedback into a
+    synthetic ``ToolMessage`` so the model sees the user's correction
+    and can retry the request differently.
+    """
+
+    def __init__(self, feedback: str, *, tool: str | None = None) -> None:
+        super().__init__(feedback)
+        self.feedback = feedback
+        self.tool = tool
+
+
+class BusyError(Exception):
+    """Raised when a second prompt arrives while the same thread is mid-stream."""
+
+    def __init__(self, request_id: str | None = None) -> None:
+        super().__init__("Thread is busy with another request")
+        self.request_id = request_id
+
+
+__all__ = [
+    "BusyError",
+    "CorrectedError",
+    "ErrorCode",
+    "RejectedError",
+    "StreamingError",
+]

surfsense_backend/app/agents/new_chat/feature_flags.py

@@ -0,0 +1,199 @@
+"""
+Feature flags for the SurfSense new_chat agent stack.
+
+These flags gate the newer agent middleware (some ported from OpenCode,
+some sourced from ``langchain.agents.middleware`` / ``deepagents``, some
+SurfSense-native). They follow a "default-OFF for risky things,
+default-ON for safe upgrades, master kill-switch for everything new" model.
+
+All new middleware checks its flag at agent build time. If the master
+kill-switch ``SURFSENSE_DISABLE_NEW_AGENT_STACK`` is set, every new
+middleware is disabled regardless of its individual flag. This gives
+operators a single switch to revert to pre-port behavior.
+
+Examples
+--------
+
+Local development (recommended for trying everything except doom-loop / selector):
+
+    SURFSENSE_ENABLE_CONTEXT_EDITING=true
+    SURFSENSE_ENABLE_COMPACTION_V2=true
+    SURFSENSE_ENABLE_RETRY_AFTER=true
+    SURFSENSE_ENABLE_TOOL_CALL_REPAIR=true
+    SURFSENSE_ENABLE_PERMISSION=false   # default off, opt-in per deploy
+    SURFSENSE_ENABLE_DOOM_LOOP=false    # default off until UI ships
+    SURFSENSE_ENABLE_LLM_TOOL_SELECTOR=false
+
+Master kill-switch (overrides everything else):
+
+    SURFSENSE_DISABLE_NEW_AGENT_STACK=true
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+from dataclasses import dataclass
+
+logger = logging.getLogger(__name__)
+
+
+def _env_bool(name: str, default: bool) -> bool:
+    """Parse a boolean env var. Accepts ``1``/``true``/``yes``/``on`` (case-insensitive)."""
+    raw = os.environ.get(name)
+    if raw is None:
+        return default
+    return raw.strip().lower() in ("1", "true", "yes", "on")
+
+
+@dataclass(frozen=True)
+class AgentFeatureFlags:
+    """Resolved feature-flag state for one agent build.
+
+    Constructed via :meth:`from_env`. The dataclass is frozen so it can be
+    safely shared across coroutines.
+    """
+
+    # Master kill-switch — when true, every flag below resolves to False
+    # regardless of its env value. Used for rapid rollback.
+    disable_new_agent_stack: bool = False
+
+    # Agent quality — context budget, retry/limits, name-repair, doom-loop
+    enable_context_editing: bool = False
+    enable_compaction_v2: bool = False
+    enable_retry_after: bool = False
+    enable_model_fallback: bool = False
+    enable_model_call_limit: bool = False
+    enable_tool_call_limit: bool = False
+    enable_tool_call_repair: bool = False
+    enable_doom_loop: bool = (
+        False  # Default OFF until UI handles permission='doom_loop'
+    )
+
+    # Safety — permissions, concurrency, tool-set narrowing
+    enable_permission: bool = False  # Default OFF for first deploy
+    enable_busy_mutex: bool = False
+    enable_llm_tool_selector: bool = False  # Default OFF — adds per-turn LLM cost
+
+    # Skills + subagents
+    enable_skills: bool = False
+    enable_specialized_subagents: bool = False
+    enable_kb_planner_runnable: bool = False
+
+    # Snapshot / revert
+    enable_action_log: bool = False
+    enable_revert_route: bool = (
+        False  # Backend ships before UI; route returns 503 until this flips
+    )
+
+    # Plugins
+    enable_plugin_loader: bool = False
+
+    # Observability — OTel (orthogonal; also requires OTEL_EXPORTER_OTLP_ENDPOINT)
+    enable_otel: bool = False
+
+    @classmethod
+    def from_env(cls) -> AgentFeatureFlags:
+        """Read flags from environment.
+
+        Master kill-switch is evaluated first; when set, all other flags
+        force to False.
+        """
+        master_off = _env_bool("SURFSENSE_DISABLE_NEW_AGENT_STACK", False)
+        if master_off:
+            logger.info(
+                "SURFSENSE_DISABLE_NEW_AGENT_STACK is set: every new agent "
+                "middleware is forced OFF for this build."
+            )
+            return cls(disable_new_agent_stack=True)
+
+        return cls(
+            disable_new_agent_stack=False,
+            # Agent quality
+            enable_context_editing=_env_bool("SURFSENSE_ENABLE_CONTEXT_EDITING", False),
+            enable_compaction_v2=_env_bool("SURFSENSE_ENABLE_COMPACTION_V2", False),
+            enable_retry_after=_env_bool("SURFSENSE_ENABLE_RETRY_AFTER", False),
+            enable_model_fallback=_env_bool("SURFSENSE_ENABLE_MODEL_FALLBACK", False),
+            enable_model_call_limit=_env_bool(
+                "SURFSENSE_ENABLE_MODEL_CALL_LIMIT", False
+            ),
+            enable_tool_call_limit=_env_bool("SURFSENSE_ENABLE_TOOL_CALL_LIMIT", False),
+            enable_tool_call_repair=_env_bool(
+                "SURFSENSE_ENABLE_TOOL_CALL_REPAIR", False
+            ),
+            enable_doom_loop=_env_bool("SURFSENSE_ENABLE_DOOM_LOOP", False),
+            # Safety
+            enable_permission=_env_bool("SURFSENSE_ENABLE_PERMISSION", False),
+            enable_busy_mutex=_env_bool("SURFSENSE_ENABLE_BUSY_MUTEX", False),
+            enable_llm_tool_selector=_env_bool(
+                "SURFSENSE_ENABLE_LLM_TOOL_SELECTOR", False
+            ),
+            # Skills + subagents
+            enable_skills=_env_bool("SURFSENSE_ENABLE_SKILLS", False),
+            enable_specialized_subagents=_env_bool(
+                "SURFSENSE_ENABLE_SPECIALIZED_SUBAGENTS", False
+            ),
+            enable_kb_planner_runnable=_env_bool(
+                "SURFSENSE_ENABLE_KB_PLANNER_RUNNABLE", False
+            ),
+            # Snapshot / revert
+            enable_action_log=_env_bool("SURFSENSE_ENABLE_ACTION_LOG", False),
+            enable_revert_route=_env_bool("SURFSENSE_ENABLE_REVERT_ROUTE", False),
+            # Plugins
+            enable_plugin_loader=_env_bool("SURFSENSE_ENABLE_PLUGIN_LOADER", False),
+            # Observability
+            enable_otel=_env_bool("SURFSENSE_ENABLE_OTEL", False),
+        )
+
+    def any_new_middleware_enabled(self) -> bool:
+        """Return True if any new middleware flag is on."""
+        if self.disable_new_agent_stack:
+            return False
+        return any(
+            (
+                self.enable_context_editing,
+                self.enable_compaction_v2,
+                self.enable_retry_after,
+                self.enable_model_fallback,
+                self.enable_model_call_limit,
+                self.enable_tool_call_limit,
+                self.enable_tool_call_repair,
+                self.enable_doom_loop,
+                self.enable_permission,
+                self.enable_busy_mutex,
+                self.enable_llm_tool_selector,
+                self.enable_skills,
+                self.enable_specialized_subagents,
+                self.enable_kb_planner_runnable,
+                self.enable_action_log,
+                self.enable_revert_route,
+                self.enable_plugin_loader,
+            )
+        )
+
+
+# Module-level cache. Read once at import time so the values are consistent
+# across the process lifetime. Use ``reload_for_tests`` to reset in tests.
+_FLAGS: AgentFeatureFlags | None = None
+
+
+def get_flags() -> AgentFeatureFlags:
+    """Return the resolved feature-flag state, caching on first call."""
+    global _FLAGS
+    if _FLAGS is None:
+        _FLAGS = AgentFeatureFlags.from_env()
+    return _FLAGS
+
+
+def reload_for_tests() -> AgentFeatureFlags:
+    """Force a fresh read from env. Tests should call this after monkeypatching env."""
+    global _FLAGS
+    _FLAGS = AgentFeatureFlags.from_env()
+    return _FLAGS
+
+
+__all__ = [
+    "AgentFeatureFlags",
+    "get_flags",
+    "reload_for_tests",
+]

surfsense_backend/app/agents/new_chat/filesystem_backends.py

@@ -5,10 +5,12 @@ from __future__ import annotations
 from collections.abc import Callable
 from functools import lru_cache
 
+from deepagents.backends.protocol import BackendProtocol
 from deepagents.backends.state import StateBackend
 from langgraph.prebuilt.tool_node import ToolRuntime
 
 from app.agents.new_chat.filesystem_selection import FilesystemMode, FilesystemSelection
+from app.agents.new_chat.middleware.kb_postgres_backend import KBPostgresBackend
 from app.agents.new_chat.middleware.multi_root_local_folder_backend import (
     MultiRootLocalFolderBackend,
 )
@@ -23,8 +25,20 @@ def _cached_multi_root_backend(
 
 def build_backend_resolver(
     selection: FilesystemSelection,
-) -> Callable[[ToolRuntime], StateBackend | MultiRootLocalFolderBackend]:
+    *,
-    """Create deepagents backend resolver for the selected filesystem mode."""
+    search_space_id: int | None = None,
+) -> Callable[[ToolRuntime], BackendProtocol]:
+    """Create deepagents backend resolver for the selected filesystem mode.
+
+    In cloud mode the resolver returns a fresh :class:`KBPostgresBackend`
+    bound to the current ``runtime`` so the backend can read staging state
+    (``staged_dirs``, ``pending_moves``, ``files`` cache, ``kb_anon_doc``,
+    ``kb_matched_chunk_ids``) for each tool call. When no ``search_space_id``
+    is provided, the resolver falls back to :class:`StateBackend` (used by
+    sub-agents and tests that don't need DB-backed reads).
+
+    Desktop-local mode unchanged.
+    """
 
     if selection.mode == FilesystemMode.DESKTOP_LOCAL_FOLDER and selection.local_mounts:
 
@@ -36,7 +50,14 @@ def build_backend_resolver(
 
         return _resolve_local
 
-    def _resolve_cloud(runtime: ToolRuntime) -> StateBackend:
+    if search_space_id is not None:
+
+        def _resolve_kb(runtime: ToolRuntime) -> BackendProtocol:
+            return KBPostgresBackend(search_space_id, runtime)
+
+        return _resolve_kb
+
+    def _resolve_state(runtime: ToolRuntime) -> StateBackend:
         return StateBackend(runtime)
 
-    return _resolve_cloud
+    return _resolve_state

surfsense_backend/app/agents/new_chat/filesystem_state.py

@@ -0,0 +1,113 @@
+"""LangGraph state schema additions used by the SurfSense filesystem agent.
+
+This schema extends deepagents' upstream :class:`FilesystemState` with the
+extra fields needed to implement Postgres-backed virtual filesystem semantics:
+
+* ``cwd`` — current working directory (per-thread checkpointed).
+* ``staged_dirs`` — pending mkdir requests (cloud only).
+* ``pending_moves`` — pending move_file requests (cloud only).
+* ``doc_id_by_path`` — virtual_path -> Document.id, populated by lazy reads.
+* ``dirty_paths`` — paths whose state file content differs from DB.
+* ``kb_priority`` — top-K priority hints rendered into a system message.
+* ``kb_matched_chunk_ids`` — internal hand-off for matched-chunk highlighting.
+* ``kb_anon_doc`` — Redis-loaded anonymous document (if any).
+* ``tree_version`` — bumped by persistence; invalidates the tree render cache.
+
+Tools mutate these fields ONLY via ``Command(update=...)`` returns; the
+reducers in :mod:`app.agents.new_chat.state_reducers` handle merging.
+"""
+
+from __future__ import annotations
+
+from typing import Annotated, Any, NotRequired
+
+from deepagents.middleware.filesystem import FilesystemState
+from typing_extensions import TypedDict
+
+from app.agents.new_chat.state_reducers import (
+    _add_unique_reducer,
+    _dict_merge_with_tombstones_reducer,
+    _list_append_reducer,
+    _replace_reducer,
+)
+
+
+class PendingMove(TypedDict):
+    """A staged move_file operation pending end-of-turn commit."""
+
+    source: str
+    dest: str
+    overwrite: bool
+
+
+class KbPriorityEntry(TypedDict, total=False):
+    path: str
+    score: float
+    document_id: int | None
+    title: str
+    mentioned: bool
+
+
+class KbAnonDoc(TypedDict, total=False):
+    """In-memory anonymous-session document loaded from Redis."""
+
+    path: str
+    title: str
+    content: str
+    chunks: list[dict[str, Any]]
+
+
+class SurfSenseFilesystemState(FilesystemState):
+    """Filesystem state used by the SurfSense agent (cloud + desktop).
+
+    Extends deepagents' :class:`FilesystemState` (which provides ``files``)
+    with cloud-mode staging fields and search-priority hints. All extra fields
+    are reducer-backed so that ``Command(update=...)`` payloads merge cleanly
+    across agent steps and across checkpoints.
+    """
+
+    cwd: NotRequired[Annotated[str, _replace_reducer]]
+    """Current working directory.
+
+    Defaults to ``"/documents"`` in cloud mode and ``"/"`` (or first mount) in
+    desktop mode. Initialized once per thread by ``KnowledgeTreeMiddleware``.
+    """
+
+    staged_dirs: NotRequired[Annotated[list[str], _add_unique_reducer]]
+    """mkdir paths staged for end-of-turn folder creation (cloud only)."""
+
+    pending_moves: NotRequired[Annotated[list[PendingMove], _list_append_reducer]]
+    """move_file ops staged for end-of-turn commit (cloud only)."""
+
+    doc_id_by_path: NotRequired[
+        Annotated[dict[str, int], _dict_merge_with_tombstones_reducer]
+    ]
+    """virtual_path -> ``Document.id`` for lazily loaded files.
+
+    Populated on first read of a KB document. Used by edit_file/move_file/
+    aafter_agent to map paths back to a real DB row. ``None`` values delete
+    the key (tombstones).
+    """
+
+    dirty_paths: NotRequired[Annotated[list[str], _add_unique_reducer]]
+    """Paths whose ``state["files"]`` content has been modified this turn."""
+
+    kb_priority: NotRequired[Annotated[list[KbPriorityEntry], _replace_reducer]]
+    """Top-K priority hints rendered as a system message before the user turn."""
+
+    kb_matched_chunk_ids: NotRequired[Annotated[dict[int, list[int]], _replace_reducer]]
+    """Internal: ``Document.id`` -> list of matched chunk IDs from hybrid search."""
+
+    kb_anon_doc: NotRequired[Annotated[KbAnonDoc | None, _replace_reducer]]
+    """Anonymous-session document loaded from Redis (read-only, no DB row)."""
+
+    tree_version: NotRequired[Annotated[int, _replace_reducer]]
+    """Monotonically increasing counter; bumped when commits change the KB tree."""
+
+
+__all__ = [
+    "KbAnonDoc",
+    "KbPriorityEntry",
+    "PendingMove",
+    "SurfSenseFilesystemState",
+]

surfsense_backend/app/agents/new_chat/middleware/__init__.py

@@ -1,25 +1,83 @@
 """Middleware components for the SurfSense new chat agent."""
 
+from app.agents.new_chat.middleware.action_log import ActionLogMiddleware
+from app.agents.new_chat.middleware.anonymous_document import (
+    AnonymousDocumentMiddleware,
+)
+from app.agents.new_chat.middleware.busy_mutex import BusyMutexMiddleware
+from app.agents.new_chat.middleware.compaction import (
+    SurfSenseCompactionMiddleware,
+    create_surfsense_compaction_middleware,
+)
+from app.agents.new_chat.middleware.context_editing import (
+    ClearToolUsesEdit,
+    SpillingContextEditingMiddleware,
+    SpillToBackendEdit,
+)
 from app.agents.new_chat.middleware.dedup_tool_calls import (
     DedupHITLToolCallsMiddleware,
 )
+from app.agents.new_chat.middleware.doom_loop import DoomLoopMiddleware
+from app.agents.new_chat.middleware.file_intent import (
+    FileIntentMiddleware,
+)
 from app.agents.new_chat.middleware.filesystem import (
     SurfSenseFilesystemMiddleware,
 )
-from app.agents.new_chat.middleware.file_intent import (
+from app.agents.new_chat.middleware.kb_persistence import (
-    FileIntentMiddleware,
+    KnowledgeBasePersistenceMiddleware,
+    commit_staged_filesystem_state,
 )
 from app.agents.new_chat.middleware.knowledge_search import (
     KnowledgeBaseSearchMiddleware,
+    KnowledgePriorityMiddleware,
+)
+from app.agents.new_chat.middleware.knowledge_tree import (
+    KnowledgeTreeMiddleware,
 )
 from app.agents.new_chat.middleware.memory_injection import (
     MemoryInjectionMiddleware,
 )
+from app.agents.new_chat.middleware.noop_injection import NoopInjectionMiddleware
+from app.agents.new_chat.middleware.otel_span import OtelSpanMiddleware
+from app.agents.new_chat.middleware.permission import PermissionMiddleware
+from app.agents.new_chat.middleware.retry_after import RetryAfterMiddleware
+from app.agents.new_chat.middleware.skills_backends import (
+    BuiltinSkillsBackend,
+    SearchSpaceSkillsBackend,
+    build_skills_backend_factory,
+    default_skills_sources,
+)
+from app.agents.new_chat.middleware.tool_call_repair import (
+    ToolCallNameRepairMiddleware,
+)
 
 __all__ = [
+    "ActionLogMiddleware",
+    "AnonymousDocumentMiddleware",
+    "BuiltinSkillsBackend",
+    "BusyMutexMiddleware",
+    "ClearToolUsesEdit",
     "DedupHITLToolCallsMiddleware",
+    "DoomLoopMiddleware",
     "FileIntentMiddleware",
+    "KnowledgeBasePersistenceMiddleware",
     "KnowledgeBaseSearchMiddleware",
+    "KnowledgePriorityMiddleware",
+    "KnowledgeTreeMiddleware",
     "MemoryInjectionMiddleware",
+    "NoopInjectionMiddleware",
+    "OtelSpanMiddleware",
+    "PermissionMiddleware",
+    "RetryAfterMiddleware",
+    "SearchSpaceSkillsBackend",
+    "SpillToBackendEdit",
+    "SpillingContextEditingMiddleware",
+    "SurfSenseCompactionMiddleware",
     "SurfSenseFilesystemMiddleware",
+    "ToolCallNameRepairMiddleware",
+    "build_skills_backend_factory",
+    "commit_staged_filesystem_state",
+    "create_surfsense_compaction_middleware",
+    "default_skills_sources",
 ]

surfsense_backend/app/agents/new_chat/middleware/action_log.py

@@ -0,0 +1,292 @@
+"""Append-only action-log middleware for the SurfSense agent.
+
+Wraps every tool call via :meth:`AgentMiddleware.awrap_tool_call` and writes
+a row to :class:`~app.db.AgentActionLog` after the tool returns. Tools opt
+into reversibility by declaring a ``reverse`` callable on their
+:class:`~app.agents.new_chat.tools.registry.ToolDefinition`; the rendered
+descriptor is persisted in ``reverse_descriptor`` for use by
+``/api/threads/{thread_id}/revert/{action_id}``.
+
+Design points:
+
+* **Defensive.** Logging never blocks the agent. We catch every exception
+  on the DB write path and emit a warning; the tool's ``ToolMessage``
+  result is always returned untouched.
+* **Lightweight payload.** Only the tool ``name`` + ``args`` (capped) +
+  ``result_id`` + ``reverse_descriptor`` are stored. Tool output text
+  remains in the LangGraph checkpoint / spilled tool-output files.
+* **Best-effort reversibility.** We invoke ``reverse(args, result_obj)``
+  with the parsed JSON result when the tool's content is a JSON object;
+  otherwise the raw text is passed. Exceptions in the reverse callable
+  are swallowed and logged — a failed descriptor render simply means the
+  action is NOT marked reversible.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+from collections.abc import Awaitable, Callable
+from typing import TYPE_CHECKING, Any
+
+from langchain.agents.middleware import AgentMiddleware
+from langchain_core.messages import ToolMessage
+
+from app.agents.new_chat.feature_flags import get_flags
+from app.agents.new_chat.tools.registry import ToolDefinition
+
+if TYPE_CHECKING:  # pragma: no cover - type-only
+    from langchain.agents.middleware.types import ToolCallRequest
+    from langgraph.types import Command
+
+
+logger = logging.getLogger(__name__)
+
+
+# Cap for the persisted ``args`` JSON to avoid bloating the action log with
+# accidentally-huge inputs. Values are truncated and a flag is set in the
+# stored payload so consumers can detect truncation.
+_MAX_ARGS_PERSIST_BYTES = 32 * 1024  # 32KB
+
+
+class ActionLogMiddleware(AgentMiddleware):
+    """Persist a row in :class:`AgentActionLog` after every tool call.
+
+    Should be placed near the OUTERMOST end of the tool-call wrapping stack
+    so that it sees the *final* :class:`ToolMessage` after all retries,
+    permission checks, and dedup logic have run. In practice that means
+    placing it just inside :class:`PermissionMiddleware` and outside
+    :class:`DedupHITLToolCallsMiddleware`.
+
+    The middleware is fully a no-op when:
+
+    * the master kill-switch ``SURFSENSE_DISABLE_NEW_AGENT_STACK`` is set
+      (checked via :func:`get_flags`),
+    * the per-feature flag ``enable_action_log`` is off, or
+    * persistence raises (defensive: tool-call dispatch always succeeds).
+
+    Args:
+        thread_id: The current chat thread's primary-key id. Required to
+            persist a row; if ``None`` the middleware silently no-ops.
+        search_space_id: Search-space id for cascade-on-delete safety.
+        user_id: UUID string of the user driving this turn (nullable in
+            anonymous mode).
+        tool_definitions: Optional mapping of tool name -> :class:`ToolDefinition`
+            so the middleware can look up the tool's ``reverse`` callable.
+            When omitted, no actions are marked reversible.
+    """
+
+    tools = ()
+
+    def __init__(
+        self,
+        *,
+        thread_id: int | None,
+        search_space_id: int,
+        user_id: str | None,
+        tool_definitions: dict[str, ToolDefinition] | None = None,
+    ) -> None:
+        super().__init__()
+        self._thread_id = thread_id
+        self._search_space_id = search_space_id
+        self._user_id = user_id
+        self._tool_definitions = dict(tool_definitions or {})
+
+    def _enabled(self) -> bool:
+        flags = get_flags()
+        if flags.disable_new_agent_stack:
+            return False
+        return bool(flags.enable_action_log) and self._thread_id is not None
+
+    async def awrap_tool_call(
+        self,
+        request: ToolCallRequest,
+        handler: Callable[[ToolCallRequest], Awaitable[ToolMessage | Command[Any]]],
+    ) -> ToolMessage | Command[Any]:
+        if not self._enabled():
+            return await handler(request)
+
+        result: ToolMessage | Command[Any]
+        error_payload: dict[str, Any] | None = None
+        try:
+            result = await handler(request)
+        except Exception as exc:
+            # Persist the failure too so revert/audit can see it, then
+            # re-raise so downstream middleware (RetryAfter, etc.) handles it.
+            error_payload = {"type": type(exc).__name__, "message": str(exc)}
+            await self._record(
+                request=request,
+                result=None,
+                error_payload=error_payload,
+            )
+            raise
+
+        await self._record(request=request, result=result, error_payload=None)
+        return result
+
+    async def _record(
+        self,
+        *,
+        request: ToolCallRequest,
+        result: ToolMessage | Command[Any] | None,
+        error_payload: dict[str, Any] | None,
+    ) -> None:
+        """Persist one ``agent_action_log`` row. Defensive: never raises."""
+        try:
+            from app.db import AgentActionLog, shielded_async_session
+
+            tool_name = _resolve_tool_name(request)
+            args_payload = _resolve_args_payload(request)
+            result_id = _resolve_result_id(result)
+            reverse_descriptor, reversible = self._render_reverse(
+                tool_name=tool_name,
+                args=_resolve_args_dict(request),
+                result=result,
+            )
+
+            row = AgentActionLog(
+                thread_id=self._thread_id,
+                user_id=self._user_id,
+                search_space_id=self._search_space_id,
+                turn_id=_resolve_turn_id(request),
+                message_id=_resolve_message_id(request),
+                tool_name=tool_name,
+                args=args_payload,
+                result_id=result_id,
+                reversible=reversible,
+                reverse_descriptor=reverse_descriptor,
+                error=error_payload,
+            )
+            async with shielded_async_session() as session:
+                session.add(row)
+                await session.commit()
+        except Exception:
+            logger.warning(
+                "ActionLogMiddleware failed to persist action log row",
+                exc_info=True,
+            )
+
+    def _render_reverse(
+        self,
+        *,
+        tool_name: str,
+        args: dict[str, Any] | None,
+        result: ToolMessage | Command[Any] | None,
+    ) -> tuple[dict[str, Any] | None, bool]:
+        """Run the tool's ``reverse`` callable and return its descriptor.
+
+        Returns a tuple of ``(descriptor_or_None, reversible_bool)``. When
+        the tool has no ``reverse`` callable, or when the callable raises,
+        the action is marked non-reversible.
+        """
+        if not result or not isinstance(result, ToolMessage):
+            return None, False
+        if args is None:
+            return None, False
+        tool_def = self._tool_definitions.get(tool_name)
+        if tool_def is None or tool_def.reverse is None:
+            return None, False
+        try:
+            parsed_result = _parse_tool_result_content(result)
+            descriptor = tool_def.reverse(args, parsed_result)
+        except Exception:
+            logger.warning(
+                "Reverse descriptor render failed for tool %s",
+                tool_name,
+                exc_info=True,
+            )
+            return None, False
+        if not isinstance(descriptor, dict):
+            return None, False
+        return descriptor, True
+
+
+# ---------------------------------------------------------------------------
+# Resolution helpers — defensive against tool_call request shape variation.
+# ---------------------------------------------------------------------------
+
+
+def _resolve_tool_name(request: Any) -> str:
+    try:
+        tool = getattr(request, "tool", None)
+        if tool is not None:
+            name = getattr(tool, "name", None)
+            if isinstance(name, str) and name:
+                return name
+        call = getattr(request, "tool_call", None) or {}
+        if isinstance(call, dict):
+            name = call.get("name")
+            if isinstance(name, str) and name:
+                return name
+    except Exception:  # pragma: no cover - defensive
+        pass
+    return "unknown"
+
+
+def _resolve_args_dict(request: Any) -> dict[str, Any] | None:
+    try:
+        call = getattr(request, "tool_call", None)
+        if not isinstance(call, dict):
+            return None
+        args = call.get("args")
+        if isinstance(args, dict):
+            return args
+        return None
+    except Exception:  # pragma: no cover - defensive
+        return None
+
+
+def _resolve_args_payload(request: Any) -> dict[str, Any] | None:
+    """Return a JSON-serializable args dict, truncated if too big."""
+    args = _resolve_args_dict(request)
+    if args is None:
+        return None
+    try:
+        encoded = json.dumps(args, default=str)
+    except Exception:
+        return {"_repr": repr(args)[:_MAX_ARGS_PERSIST_BYTES]}
+    if len(encoded) <= _MAX_ARGS_PERSIST_BYTES:
+        return args
+    return {
+        "_truncated": True,
+        "_size": len(encoded),
+        "_preview": encoded[:_MAX_ARGS_PERSIST_BYTES],
+    }
+
+
+def _resolve_turn_id(request: Any) -> str | None:
+    try:
+        call = getattr(request, "tool_call", None) or {}
+        if isinstance(call, dict):
+            tid = call.get("id")
+            if isinstance(tid, str):
+                return tid
+    except Exception:  # pragma: no cover
+        pass
+    return None
+
+
+def _resolve_message_id(request: Any) -> str | None:
+    """Tool-call IDs serve as best-available message correlator at this layer."""
+    return _resolve_turn_id(request)
+
+
+def _resolve_result_id(result: Any) -> str | None:
+    if isinstance(result, ToolMessage):
+        msg_id = getattr(result, "id", None)
+        if isinstance(msg_id, str):
+            return msg_id
+    return None
+
+
+def _parse_tool_result_content(result: ToolMessage) -> Any:
+    content = result.content
+    if isinstance(content, str):
+        try:
+            return json.loads(content)
+        except (json.JSONDecodeError, ValueError):
+            return content
+    return content
+
+
+__all__ = ["ActionLogMiddleware"]

surfsense_backend/app/agents/new_chat/middleware/anonymous_document.py

@@ -0,0 +1,91 @@
+"""Lightweight middleware that loads the anonymous-session document into state.
+
+Anonymous chats receive a single uploaded document via Redis (no DB row,
+read-only). This middleware loads it once on the first turn into
+``state['kb_anon_doc']`` so:
+
+* :class:`KnowledgeTreeMiddleware` can render the synthetic ``/documents``
+  view without touching the DB.
+* :class:`KnowledgePriorityMiddleware` skips hybrid search and emits a
+  degenerate priority list.
+* :class:`KBPostgresBackend` (``als_info`` / ``aread`` / ``_load_file_data``)
+  recognises the synthetic path.
+
+The middleware is a no-op when ``anon_session_id`` is not provided or when
+the document is already cached in state.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+from typing import Any
+
+from langchain.agents.middleware import AgentMiddleware, AgentState
+from langgraph.runtime import Runtime
+
+from app.agents.new_chat.filesystem_state import SurfSenseFilesystemState
+from app.agents.new_chat.path_resolver import DOCUMENTS_ROOT, safe_filename
+
+logger = logging.getLogger(__name__)
+
+
+class AnonymousDocumentMiddleware(AgentMiddleware):  # type: ignore[type-arg]
+    """Load the anonymous user's uploaded document from Redis into state."""
+
+    tools = ()
+    state_schema = SurfSenseFilesystemState
+
+    def __init__(self, *, anon_session_id: str | None) -> None:
+        self.anon_session_id = anon_session_id
+
+    async def abefore_agent(  # type: ignore[override]
+        self,
+        state: AgentState,
+        runtime: Runtime[Any],
+    ) -> dict[str, Any] | None:
+        del runtime
+        if not self.anon_session_id:
+            return None
+        if state.get("kb_anon_doc"):
+            return None
+
+        anon_doc = await self._load_anon_document()
+        if anon_doc is None:
+            return None
+        return {"kb_anon_doc": anon_doc}
+
+    async def _load_anon_document(self) -> dict[str, Any] | None:
+        """Read ``anon:doc:<session_id>`` from Redis."""
+        try:
+            import redis.asyncio as aioredis  # local import to keep cold paths cheap
+
+            from app.config import config
+
+            redis_client = aioredis.from_url(
+                config.REDIS_APP_URL, decode_responses=True
+            )
+            try:
+                redis_key = f"anon:doc:{self.anon_session_id}"
+                data = await redis_client.get(redis_key)
+                if not data:
+                    return None
+                payload = json.loads(data)
+            finally:
+                await redis_client.aclose()
+        except Exception as exc:
+            logger.warning("Failed to load anonymous document from Redis: %s", exc)
+            return None
+
+        title = str(payload.get("filename") or "uploaded_document")
+        content = str(payload.get("content") or "")
+        path = f"{DOCUMENTS_ROOT}/{safe_filename(title)}"
+        return {
+            "path": path,
+            "title": title,
+            "content": content,
+            "chunks": [{"chunk_id": -1, "content": content}] if content else [],
+        }
+
+
+__all__ = ["AnonymousDocumentMiddleware"]

surfsense_backend/app/agents/new_chat/middleware/busy_mutex.py

@@ -0,0 +1,236 @@
+"""
+BusyMutexMiddleware — per-thread asyncio lock + cancel token.
+
+LangChain has no built-in concept of "this thread is already running a
+turn — refuse the second concurrent request". Without it, a user
+double-clicking "send" or refreshing the page mid-stream can spawn two
+turns racing on the same checkpoint, producing duplicated tool calls
+and mangled state.
+
+Ported from OpenCode's ``Stream.scoped(AbortController)`` pattern: a
+single-process, in-memory lock + cooperative cancellation token keyed by
+``thread_id``. For multi-worker deployments a distributed lock backend
+(Redis or PostgreSQL advisory locks) is a phase-2 follow-up.
+
+What this provides:
+- A ``WeakValueDictionary[str, asyncio.Lock]`` keyed by ``thread_id``;
+  acquiring the lock during ``before_agent`` blocks any concurrent
+  prompt on the same thread until release.
+- A per-thread ``asyncio.Event`` (``cancel_event``) that long-running
+  tools can poll to abort cooperatively. The event is reset between
+  turns. Tools should check ``runtime.context.cancel_event.is_set()``
+  in tight inner loops.
+- A typed :class:`~app.agents.new_chat.errors.BusyError` raised when a
+  second turn arrives while the lock is held.
+
+Note: SurfSense's ``stream_new_chat`` is the call site that should
+acquire/release. Wiring this as middleware means the contract is
+explicit and the lock manager is shared with subagents that compile
+their own ``create_agent`` runnables.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import weakref
+from typing import Any
+
+from langchain.agents.middleware.types import (
+    AgentMiddleware,
+    AgentState,
+    ContextT,
+    ResponseT,
+)
+from langgraph.config import get_config
+from langgraph.runtime import Runtime
+
+from app.agents.new_chat.errors import BusyError
+
+logger = logging.getLogger(__name__)
+
+
+class _ThreadLockManager:
+    """Process-local registry of per-thread asyncio locks + cancel events."""
+
+    def __init__(self) -> None:
+        self._locks: weakref.WeakValueDictionary[str, asyncio.Lock] = (
+            weakref.WeakValueDictionary()
+        )
+        self._cancel_events: dict[str, asyncio.Event] = {}
+
+    def lock_for(self, thread_id: str) -> asyncio.Lock:
+        lock = self._locks.get(thread_id)
+        if lock is None:
+            lock = asyncio.Lock()
+            self._locks[thread_id] = lock
+        return lock
+
+    def cancel_event(self, thread_id: str) -> asyncio.Event:
+        event = self._cancel_events.get(thread_id)
+        if event is None:
+            event = asyncio.Event()
+            self._cancel_events[thread_id] = event
+        return event
+
+    def request_cancel(self, thread_id: str) -> bool:
+        event = self._cancel_events.get(thread_id)
+        if event is None:
+            return False
+        event.set()
+        return True
+
+    def reset(self, thread_id: str) -> None:
+        event = self._cancel_events.get(thread_id)
+        if event is not None:
+            event.clear()
+
+
+# Module-level singleton — process-local but reused across all agent
+# instances built in this process. Subagents created in nested
+# ``create_agent`` calls also get this so locks are coherent.
+manager = _ThreadLockManager()
+
+
+def get_cancel_event(thread_id: str) -> asyncio.Event:
+    """Public accessor used by long-running tools to poll cancellation."""
+    return manager.cancel_event(thread_id)
+
+
+def request_cancel(thread_id: str) -> bool:
+    """Trip the cancel event for ``thread_id``. Returns True if found."""
+    return manager.request_cancel(thread_id)
+
+
+def reset_cancel(thread_id: str) -> None:
+    """Reset the cancel event for ``thread_id`` (called between turns)."""
+    manager.reset(thread_id)
+
+
+class BusyMutexMiddleware(AgentMiddleware[AgentState[ResponseT], ContextT, ResponseT]):
+    """Block concurrent prompts on the same thread.
+
+    Acquires the thread's lock in ``abefore_agent`` and releases in
+    ``aafter_agent``. If the lock is held, raises :class:`BusyError`
+    so the caller can emit a ``surfsense.busy`` SSE event with the
+    in-flight request id.
+
+    Args:
+        require_thread_id: When True, raise :class:`BusyError` if no
+            ``thread_id`` can be resolved from the active
+            ``RunnableConfig``. Default is False — we treat a missing
+            thread_id as "this turn has nothing to lock against" and
+            no-op the mutex. Set True only when you trust the call
+            site to always provide ``configurable.thread_id`` (e.g.
+            in production where ``stream_new_chat`` always does).
+    """
+
+    def __init__(self, *, require_thread_id: bool = False) -> None:
+        super().__init__()
+        self._require_thread_id = require_thread_id
+        self.tools = []
+        # Per-call locks owned by this middleware. We track them as
+        # an instance attribute so ``aafter_agent`` knows which lock
+        # to release.
+        self._held_locks: dict[str, asyncio.Lock] = {}
+
+    @staticmethod
+    def _thread_id(runtime: Runtime[ContextT]) -> str | None:
+        """Extract ``thread_id`` from the active LangGraph ``RunnableConfig``.
+
+        ``langgraph.runtime.Runtime`` deliberately does NOT expose ``config``.
+        The runnable config (where ``configurable.thread_id`` lives) must be
+        fetched via :func:`langgraph.config.get_config` from inside a node /
+        middleware. We fall back to ``getattr(runtime, "config", None)`` for
+        unit tests / legacy runtimes that synthesize a config-bearing stub.
+        """
+
+        def _from_dict(cfg: Any) -> str | None:
+            if not isinstance(cfg, dict):
+                return None
+            tid = (cfg.get("configurable") or {}).get("thread_id")
+            return str(tid) if tid is not None else None
+
+        # Preferred path: real LangGraph runtime context.
+        try:
+            tid = _from_dict(get_config())
+        except Exception:
+            tid = None
+        if tid is not None:
+            return tid
+
+        # Fallback for tests and any runtime that surfaces a config dict
+        # directly on the runtime instance.
+        return _from_dict(getattr(runtime, "config", None))
+
+    async def abefore_agent(  # type: ignore[override]
+        self,
+        state: AgentState[Any],
+        runtime: Runtime[ContextT],
+    ) -> dict[str, Any] | None:
+        del state
+        thread_id = self._thread_id(runtime)
+        if thread_id is None:
+            if self._require_thread_id:
+                raise BusyError("no thread_id configured")
+            logger.debug(
+                "BusyMutexMiddleware: no thread_id resolved from RunnableConfig; "
+                "skipping per-thread lock for this turn."
+            )
+            return None
+
+        lock = manager.lock_for(thread_id)
+        if lock.locked():
+            raise BusyError(request_id=thread_id)
+        await lock.acquire()
+        self._held_locks[thread_id] = lock
+        # Reset the cancel event so this turn starts fresh
+        reset_cancel(thread_id)
+        return None
+
+    async def aafter_agent(  # type: ignore[override]
+        self,
+        state: AgentState[Any],
+        runtime: Runtime[ContextT],
+    ) -> dict[str, Any] | None:
+        del state
+        thread_id = self._thread_id(runtime)
+        if thread_id is None:
+            return None
+        lock = self._held_locks.pop(thread_id, None)
+        if lock is not None and lock.locked():
+            lock.release()
+        # Always clear cancel event between turns so a stale signal
+        # doesn't leak into the next request.
+        reset_cancel(thread_id)
+        return None
+
+    # Provide sync no-ops because the middleware base class allows them
+    def before_agent(  # type: ignore[override]
+        self, state: AgentState[Any], runtime: Runtime[ContextT]
+    ) -> dict[str, Any] | None:
+        # Sync path: no asyncio.Lock to acquire. Best we can do is reject
+        # if anyone else is in flight.
+        thread_id = self._thread_id(runtime)
+        if thread_id is None:
+            if self._require_thread_id:
+                raise BusyError("no thread_id configured")
+            return None
+        lock = manager.lock_for(thread_id)
+        if lock.locked():
+            raise BusyError(request_id=thread_id)
+        return None
+
+    def after_agent(  # type: ignore[override]
+        self, state: AgentState[Any], runtime: Runtime[ContextT]
+    ) -> dict[str, Any] | None:
+        return None
+
+
+__all__ = [
+    "BusyMutexMiddleware",
+    "get_cancel_event",
+    "manager",
+    "request_cancel",
+    "reset_cancel",
+]

surfsense_backend/app/agents/new_chat/middleware/compaction.py

@@ -0,0 +1,254 @@
+"""
+SurfSense compaction middleware.
+
+Subclasses :class:`deepagents.middleware.summarization.SummarizationMiddleware`
+to add SurfSense-specific behavior:
+
+1. **Structured summary template** (OpenCode-style ``## Goal / Constraints /
+   Progress / Key Decisions / Next Steps / Critical Context / Relevant Files``)
+   — see :data:`SURFSENSE_SUMMARY_PROMPT` below. The base
+   ``SummarizationMiddleware`` only ships a freeform "summarize this"
+   prompt; the structured template is ported from OpenCode's
+   ``compaction.ts``.
+2. **Protect SurfSense-specific SystemMessages** so injected hints
+   (``<priority_documents>``, ``<workspace_tree>``, ``<file_operation_contract>``,
+   ``<user_memory>``, ``<team_memory>``, ``<user_name>``, ``<memory_warning>``)
+   are *not* summarized away and are kept verbatim in the post-summary
+   message list. Mirrors OpenCode's ``PRUNE_PROTECTED_TOOLS`` philosophy
+   (some message types are part of the agent's contract and must survive
+   compaction unchanged).
+3. **Sanitize ``content=None``** when feeding messages into ``get_buffer_string``
+   (Azure OpenAI / LiteLLM defense — when a provider streams an AIMessage
+   containing only tool_calls and no text, ``content`` can be ``None`` and
+   ``get_buffer_string`` crashes iterating over ``None``). SurfSense-specific.
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import TYPE_CHECKING, Any
+
+from deepagents.middleware.summarization import (
+    SummarizationMiddleware,
+    compute_summarization_defaults,
+)
+from langchain_core.messages import SystemMessage
+
+from app.observability import otel as ot
+
+if TYPE_CHECKING:
+    from deepagents.backends.protocol import BACKEND_TYPES
+    from langchain_core.language_models import BaseChatModel
+    from langchain_core.messages import AnyMessage
+
+logger = logging.getLogger(__name__)
+
+# Structured summary template ported from OpenCode's
+# ``opencode/packages/opencode/src/session/compaction.ts:40-75``. Kept as a
+# module-level constant so unit tests can assert on its sections.
+SURFSENSE_SUMMARY_PROMPT = """<role>
+SurfSense Conversation Compaction Assistant
+</role>
+
+<primary_objective>
+Extract the most important context from the conversation history below into a structured summary that will replace the older messages.
+</primary_objective>
+
+<instructions>
+You are running because the conversation has grown beyond the model's input window. The conversation history below will be summarized and replaced with your output. Use the structured template that follows; keep each section concise but comprehensive enough that the agent can resume work without losing context. Each section is a checklist — populate it with relevant content or write "None" if there is nothing to report.
+
+## Goal
+What is the user's primary goal or request? State it in one or two sentences.
+
+## Constraints
+What boundaries must the agent respect (citations rules, visibility scope, allowed tools, user-imposed style, deadlines, deny-listed topics)?
+
+## Progress
+What has the agent already accomplished? List each completed step succinctly. Do not reproduce tool output; just record the conclusion.
+
+## Key Decisions
+What choices were made and why? Include rejected alternatives and the reasoning behind selecting the current path.
+
+## Next Steps
+What specific tasks remain to achieve the goal? Order them by dependency.
+
+## Critical Context
+What facts, IDs, document titles, query keywords, error messages, or partial answers must persist into the next turn? Include verbatim quotes only when the exact wording matters (e.g. a precise filter clause or a literal name).
+
+## Relevant Files
+What documents or paths in the SurfSense knowledge base are in play? Use ``/documents/...`` paths exactly as they appeared in the workspace tree.
+</instructions>
+
+<messages>
+Messages to summarize:
+{messages}
+</messages>
+
+Respond ONLY with the structured summary. Do not include any text before or after.
+"""
+
+# SystemMessage prefixes that must NOT be summarized away. They are
+# re-injected on every turn by the corresponding middleware, but the
+# compaction step happens *before* re-injection in some paths, so we
+# must preserve them verbatim across the cutoff.
+PROTECTED_SYSTEM_PREFIXES: tuple[str, ...] = (
+    "<priority_documents>",  # KnowledgePriorityMiddleware
+    "<workspace_tree>",  # KnowledgeTreeMiddleware
+    "<file_operation_contract>",  # FileIntentMiddleware
+    "<user_memory>",  # MemoryInjectionMiddleware
+    "<team_memory>",  # MemoryInjectionMiddleware
+    "<user_name>",  # MemoryInjectionMiddleware
+    "<memory_warning>",  # MemoryInjectionMiddleware
+)
+
+
+def _is_protected_system_message(msg: AnyMessage) -> bool:
+    """Return True if ``msg`` is a SystemMessage we must not summarize."""
+    if not isinstance(msg, SystemMessage):
+        return False
+    content = msg.content
+    if not isinstance(content, str):
+        return False
+    stripped = content.lstrip()
+    return any(stripped.startswith(prefix) for prefix in PROTECTED_SYSTEM_PREFIXES)
+
+
+def _sanitize_message_content(msg: AnyMessage) -> AnyMessage:
+    """Return ``msg`` with ``content=None`` coerced to ``""``.
+
+    Folds in the historical defense from ``safe_summarization.py`` —
+    ``get_buffer_string`` reads ``m.text`` which iterates ``self.content``,
+    so a ``None`` content (Azure OpenAI / LiteLLM streaming a tool-only
+    AIMessage) explodes. We return a copy with empty string content so
+    downstream consumers see an empty body without mutating the original.
+    """
+    if getattr(msg, "content", "not-missing") is not None:
+        return msg
+    try:
+        return msg.model_copy(update={"content": ""})
+    except AttributeError:
+        import copy
+
+        new_msg = copy.copy(msg)
+        try:
+            new_msg.content = ""
+        except Exception:
+            logger.debug(
+                "Could not sanitize content=None on message of type %s",
+                type(msg).__name__,
+            )
+            return msg
+        return new_msg
+
+
+class SurfSenseCompactionMiddleware(SummarizationMiddleware):
+    """SummarizationMiddleware tuned for SurfSense.
+
+    Notes
+    -----
+    - Overrides :meth:`_partition_messages` so protected SystemMessages
+      survive into the ``preserved_messages`` half regardless of cutoff.
+    - Overrides :meth:`_filter_summary_messages` so the buffer-string path
+      never iterates ``None`` content.
+    - Inherits everything else (auto-trigger, backend offload,
+      ``_summarization_event`` plumbing, ``ContextOverflowError`` fallback).
+    """
+
+    def _partition_messages(  # type: ignore[override]
+        self,
+        conversation_messages: list[AnyMessage],
+        cutoff_index: int,
+    ) -> tuple[list[AnyMessage], list[AnyMessage]]:
+        """Split messages but always preserve SurfSense protected SystemMessages.
+
+        Mirrors OpenCode's ``PRUNE_PROTECTED_TOOLS`` philosophy
+        (``opencode/packages/opencode/src/session/compaction.ts``): some
+        message types are always kept verbatim because they are part of the
+        agent's working contract, not transient output.
+
+        Also opens a ``compaction.run`` OTel span (no-op when OTel is off)
+        so dashboards can count compaction events and message-volume
+        without having to instrument upstream callers.
+        """
+        # Opening a span here is appropriate because partitioning is the
+        # first call SummarizationMiddleware makes when it has decided to
+        # summarize; we record the volume and then close as a normal span.
+        with ot.compaction_span(
+            reason="auto",
+            messages_in=len(conversation_messages),
+            extra={"compaction.cutoff_index": int(cutoff_index)},
+        ):
+            messages_to_summarize, preserved_messages = super()._partition_messages(
+                conversation_messages, cutoff_index
+            )
+
+            protected: list[AnyMessage] = []
+            kept_for_summary: list[AnyMessage] = []
+            for msg in messages_to_summarize:
+                if _is_protected_system_message(msg):
+                    protected.append(msg)
+                else:
+                    kept_for_summary.append(msg)
+
+            # Place protected blocks at the *front* of preserved_messages so
+            # they keep their original ordering relative to the summary
+            # HumanMessage that precedes the rest of the preserved tail.
+            return kept_for_summary, [*protected, *preserved_messages]
+
+    def _filter_summary_messages(  # type: ignore[override]
+        self, messages: list[AnyMessage]
+    ) -> list[AnyMessage]:
+        """Filter previous summaries AND sanitize ``content=None``.
+
+        Folds the ``safe_summarization.py`` defense in: when the buffer
+        builder iterates ``m.text`` over ``None`` it explodes; sanitizing
+        here covers both the sync and async offload paths.
+        """
+        filtered = super()._filter_summary_messages(messages)
+        return [_sanitize_message_content(m) for m in filtered]
+
+
+def create_surfsense_compaction_middleware(
+    model: BaseChatModel,
+    backend: BACKEND_TYPES,
+    *,
+    summary_prompt: str | None = None,
+    history_path_prefix: str = "/conversation_history",
+    **overrides: Any,
+) -> SurfSenseCompactionMiddleware:
+    """Build a :class:`SurfSenseCompactionMiddleware` with sensible defaults.
+
+    Pulls profile-aware ``trigger`` / ``keep`` / ``truncate_args_settings``
+    via :func:`deepagents.middleware.summarization.compute_summarization_defaults`
+    so callers get the same behavior as ``create_summarization_middleware``
+    plus our overrides.
+
+    Args:
+        model: Chat model to call for summary generation.
+        backend: Backend instance or factory for offloading conversation history.
+        summary_prompt: Optional override; defaults to :data:`SURFSENSE_SUMMARY_PROMPT`.
+        history_path_prefix: Path prefix for offloaded conversation history.
+        **overrides: Forwarded to :class:`SurfSenseCompactionMiddleware`.
+    """
+    defaults = compute_summarization_defaults(model)
+    return SurfSenseCompactionMiddleware(
+        model=model,
+        backend=backend,
+        trigger=overrides.pop("trigger", defaults["trigger"]),
+        keep=overrides.pop("keep", defaults["keep"]),
+        trim_tokens_to_summarize=overrides.pop("trim_tokens_to_summarize", None),
+        truncate_args_settings=overrides.pop(
+            "truncate_args_settings", defaults["truncate_args_settings"]
+        ),
+        summary_prompt=summary_prompt or SURFSENSE_SUMMARY_PROMPT,
+        history_path_prefix=history_path_prefix,
+        **overrides,
+    )
+
+
+__all__ = [
+    "PROTECTED_SYSTEM_PREFIXES",
+    "SURFSENSE_SUMMARY_PROMPT",
+    "SurfSenseCompactionMiddleware",
+    "create_surfsense_compaction_middleware",
+]

surfsense_backend/app/agents/new_chat/middleware/context_editing.py

@@ -0,0 +1,350 @@
+"""
+SpillToBackendEdit + SpillingContextEditingMiddleware.
+
+LangChain's :class:`ClearToolUsesEdit` discards old ``ToolMessage.content``
+when the context-editing budget triggers, replacing the body with a fixed
+placeholder. That's lossy: anything the agent might want to revisit is
+gone. The spill-to-disk pattern (originally from OpenCode's
+``opencode/packages/opencode/src/tool/truncate.ts``) keeps the prune
+behavior but writes the full original payload to the runtime backend
+under ``/tool_outputs/{thread_id}/{message_id}.txt`` first. The
+placeholder is then upgraded to point at the spill path so the agent
+(or a subagent) can read it back on demand.
+
+Why this is a middleware subclass instead of a plain ``ContextEdit``:
+``ContextEdit.apply`` is sync, but writing to the backend is async. We
+capture the spill payloads inside ``apply`` and flush them via
+``await backend.aupload_files(...)`` from ``awrap_model_call`` *before*
+delegating to the handler, so the explore subagent can always read what
+the placeholder advertises.
+"""
+
+from __future__ import annotations
+
+import logging
+import threading
+from collections.abc import Awaitable, Callable, Sequence
+from copy import deepcopy
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING, Any
+
+from langchain.agents.middleware.context_editing import (
+    ClearToolUsesEdit,
+    ContextEdit,
+    ContextEditingMiddleware,
+    TokenCounter,
+)
+from langchain_core.messages import (
+    AIMessage,
+    AnyMessage,
+    BaseMessage,
+    ToolMessage,
+)
+from langchain_core.messages.utils import count_tokens_approximately
+from langgraph.config import get_config
+
+if TYPE_CHECKING:
+    from deepagents.backends.protocol import BackendProtocol
+    from langchain.agents.middleware.types import (
+        ModelRequest,
+        ModelResponse,
+    )
+
+logger = logging.getLogger(__name__)
+
+DEFAULT_SPILL_PREFIX = "/tool_outputs"
+
+
+def _build_spill_placeholder(spill_path: str) -> str:
+    """Build the user-facing placeholder text shown to the model."""
+    return (
+        f"[cleared — full output at {spill_path}; ask the explore subagent to read it]"
+    )
+
+
+def _get_thread_id_or_session() -> str:
+    """Best-effort thread_id discovery for the spill path.
+
+    Falls back to a process-stable string if no LangGraph config is
+    available (e.g. unit tests). The exact value doesn't matter as long
+    as it's stable within one stream so the placeholder paths line up
+    with the actual upload path.
+    """
+    try:
+        config = get_config()
+        thread_id = config.get("configurable", {}).get("thread_id")
+        if thread_id is not None:
+            return str(thread_id)
+    except RuntimeError:
+        pass
+    return "no_thread"
+
+
+@dataclass(slots=True)
+class SpillToBackendEdit(ContextEdit):
+    """Capture-and-replace context edit that spills full tool output to the backend.
+
+    Behaves like :class:`ClearToolUsesEdit` (same trigger / keep / exclude
+    semantics) **and** records the original ``ToolMessage.content`` in
+    :attr:`pending_spills` so the wrapping middleware can flush them
+    before the model call.
+
+    Args:
+        trigger: Token threshold above which the edit fires.
+        clear_at_least: Minimum number of tokens to reclaim (best effort).
+        keep: Number of most-recent ``ToolMessage`` instances to leave
+            untouched.
+        exclude_tools: Names of tools whose output is NOT spilled.
+        clear_tool_inputs: Also clear the originating ``AIMessage.tool_calls``
+            args when their pair is cleared.
+        path_prefix: Path under the backend where spills are written.
+            Default ``"/tool_outputs"``.
+    """
+
+    trigger: int = 100_000
+    clear_at_least: int = 0
+    keep: int = 3
+    clear_tool_inputs: bool = False
+    exclude_tools: Sequence[str] = ()
+    path_prefix: str = DEFAULT_SPILL_PREFIX
+
+    pending_spills: list[tuple[str, bytes]] = field(default_factory=list)
+    _lock: threading.Lock = field(default_factory=threading.Lock)
+
+    def drain_pending(self) -> list[tuple[str, bytes]]:
+        """Return and clear the pending-spill list atomically."""
+        with self._lock:
+            out = list(self.pending_spills)
+            self.pending_spills.clear()
+        return out
+
+    def apply(
+        self,
+        messages: list[AnyMessage],
+        *,
+        count_tokens: TokenCounter,
+    ) -> None:
+        """Mirror ``ClearToolUsesEdit.apply`` but capture originals first."""
+        tokens = count_tokens(messages)
+        if tokens <= self.trigger:
+            return
+
+        candidates = [
+            (idx, msg)
+            for idx, msg in enumerate(messages)
+            if isinstance(msg, ToolMessage)
+        ]
+        if self.keep >= len(candidates):
+            return
+        if self.keep:
+            candidates = candidates[: -self.keep]
+
+        thread_id = _get_thread_id_or_session()
+        excluded_tools = set(self.exclude_tools)
+
+        for idx, tool_message in candidates:
+            if tool_message.response_metadata.get("context_editing", {}).get("cleared"):
+                continue
+
+            ai_message = next(
+                (m for m in reversed(messages[:idx]) if isinstance(m, AIMessage)),
+                None,
+            )
+            if ai_message is None:
+                continue
+
+            tool_call = next(
+                (
+                    call
+                    for call in ai_message.tool_calls
+                    if call.get("id") == tool_message.tool_call_id
+                ),
+                None,
+            )
+            if tool_call is None:
+                continue
+
+            tool_name = tool_message.name or tool_call["name"]
+            if tool_name in excluded_tools:
+                continue
+
+            message_id = tool_message.id or tool_message.tool_call_id or "unknown"
+            spill_path = f"{self.path_prefix}/{thread_id}/{message_id}.txt"
+
+            original = tool_message.content
+            payload = self._encode_payload(original)
+            with self._lock:
+                self.pending_spills.append((spill_path, payload))
+
+            messages[idx] = tool_message.model_copy(
+                update={
+                    "artifact": None,
+                    "content": _build_spill_placeholder(spill_path),
+                    "response_metadata": {
+                        **tool_message.response_metadata,
+                        "context_editing": {
+                            "cleared": True,
+                            "strategy": "spill_to_backend",
+                            "spill_path": spill_path,
+                        },
+                    },
+                }
+            )
+
+            if self.clear_tool_inputs:
+                ai_idx = messages.index(ai_message)
+                messages[ai_idx] = self._clear_input_args(
+                    ai_message, tool_message.tool_call_id or ""
+                )
+
+            if self.clear_at_least > 0:
+                new_token_count = count_tokens(messages)
+                cleared_tokens = max(0, tokens - new_token_count)
+                if cleared_tokens >= self.clear_at_least:
+                    break
+
+    @staticmethod
+    def _encode_payload(content: Any) -> bytes:
+        """Serialize ``ToolMessage.content`` to bytes for upload."""
+        if isinstance(content, bytes):
+            return content
+        if isinstance(content, str):
+            return content.encode("utf-8")
+        try:
+            import json
+
+            return json.dumps(content, default=str).encode("utf-8")
+        except Exception:
+            return str(content).encode("utf-8")
+
+    @staticmethod
+    def _clear_input_args(message: AIMessage, tool_call_id: str) -> AIMessage:
+        updated_tool_calls: list[dict[str, Any]] = []
+        cleared_any = False
+        for tool_call in message.tool_calls:
+            updated = dict(tool_call)
+            if updated.get("id") == tool_call_id:
+                updated["args"] = {}
+                cleared_any = True
+            updated_tool_calls.append(updated)
+
+        metadata = dict(getattr(message, "response_metadata", {}))
+        if cleared_any:
+            ctx = dict(metadata.get("context_editing", {}))
+            ids = set(ctx.get("cleared_tool_inputs", []))
+            ids.add(tool_call_id)
+            ctx["cleared_tool_inputs"] = sorted(ids)
+            metadata["context_editing"] = ctx
+        return message.model_copy(
+            update={
+                "tool_calls": updated_tool_calls,
+                "response_metadata": metadata,
+            }
+        )
+
+
+BackendResolver = "Callable[[Any], BackendProtocol] | BackendProtocol"
+
+
+class SpillingContextEditingMiddleware(ContextEditingMiddleware):
+    """:class:`ContextEditingMiddleware` that flushes :class:`SpillToBackendEdit` writes.
+
+    Runs the configured edits as the parent does, then flushes any
+    pending spills via the supplied backend resolver before delegating
+    to the model handler. Spill failures are logged but never abort the
+    model call — the placeholder text is already in the message, so the
+    worst case is the agent gets a placeholder it cannot follow up on.
+    """
+
+    def __init__(
+        self,
+        *,
+        edits: Sequence[ContextEdit],
+        backend_resolver: BackendResolver | None = None,
+        token_count_method: str = "approximate",
+    ) -> None:
+        super().__init__(edits=list(edits), token_count_method=token_count_method)  # type: ignore[arg-type]
+        self._backend_resolver = backend_resolver
+
+    def _resolve_backend(self, request: ModelRequest) -> BackendProtocol | None:
+        if self._backend_resolver is None:
+            return None
+        if callable(self._backend_resolver):
+            try:
+                from langchain.tools import ToolRuntime
+
+                tool_runtime = ToolRuntime(
+                    state=getattr(request, "state", {}),
+                    context=getattr(request.runtime, "context", None),
+                    stream_writer=getattr(request.runtime, "stream_writer", None),
+                    store=getattr(request.runtime, "store", None),
+                    config=getattr(request.runtime, "config", None) or {},
+                    tool_call_id=None,
+                )
+                return self._backend_resolver(tool_runtime)
+            except Exception:
+                logger.exception("Failed to resolve spill backend")
+                return None
+        return self._backend_resolver  # type: ignore[return-value]
+
+    def _collect_pending(self) -> list[tuple[str, bytes]]:
+        out: list[tuple[str, bytes]] = []
+        for edit in self.edits:
+            if isinstance(edit, SpillToBackendEdit):
+                out.extend(edit.drain_pending())
+        return out
+
+    async def awrap_model_call(  # type: ignore[override]
+        self,
+        request: ModelRequest,
+        handler: Callable[[ModelRequest], Awaitable[ModelResponse]],
+    ) -> Any:
+        if not request.messages:
+            return await handler(request)
+
+        if self.token_count_method == "approximate":
+
+            def count_tokens(messages: Sequence[BaseMessage]) -> int:
+                return count_tokens_approximately(messages)
+
+        else:
+            system_msg = [request.system_message] if request.system_message else []
+
+            def count_tokens(messages: Sequence[BaseMessage]) -> int:
+                return request.model.get_num_tokens_from_messages(
+                    system_msg + list(messages), request.tools
+                )
+
+        edited_messages = deepcopy(list(request.messages))
+        for edit in self.edits:
+            edit.apply(edited_messages, count_tokens=count_tokens)
+
+        pending = self._collect_pending()
+        if pending:
+            backend = self._resolve_backend(request)
+            if backend is not None:
+                try:
+                    await backend.aupload_files(pending)
+                except Exception:
+                    logger.exception(
+                        "Spill-to-backend upload failed (%d files); placeholders "
+                        "remain in messages but content is unrecoverable",
+                        len(pending),
+                    )
+            else:
+                logger.warning(
+                    "SpillToBackendEdit produced %d pending spills but no backend "
+                    "resolver was configured; content is unrecoverable",
+                    len(pending),
+                )
+
+        return await handler(request.override(messages=edited_messages))
+
+
+__all__ = [
+    "DEFAULT_SPILL_PREFIX",
+    "ClearToolUsesEdit",
+    "SpillToBackendEdit",
+    "SpillingContextEditingMiddleware",
+    "_build_spill_placeholder",
+]

surfsense_backend/app/agents/new_chat/middleware/dedup_tool_calls.py

@@ -2,17 +2,27 @@
 
 When the LLM emits multiple calls to the same HITL tool with the same
 primary argument (e.g. two ``delete_calendar_event("Doctor Appointment")``),
-only the first call is kept.  Non-HITL tools are never touched.
+only the first call is kept. Non-HITL tools are never touched.
 
 This runs in the ``after_model`` hook — **before** any tool executes — so
 the duplicate call is stripped from the AIMessage that gets checkpointed.
 That means it is also safe across LangGraph ``interrupt()`` boundaries:
 the removed call will never appear on graph resume.
+
+Dedup-key resolution order:
+
+1. :class:`ToolDefinition.dedup_key` — callable provided by the registry
+   entry. This is the canonical mechanism.
+2. ``tool.metadata["hitl_dedup_key"]`` — string with a primary arg name;
+   used by MCP / Composio tools whose schemas the registry doesn't see.
+
+A tool with no resolver from either path simply opts out of dedup.
 """
 
 from __future__ import annotations
 
 import logging
+from collections.abc import Callable
 from typing import Any
 
 from langchain.agents.middleware import AgentMiddleware, AgentState
@@ -20,81 +30,83 @@ from langgraph.runtime import Runtime
 
 logger = logging.getLogger(__name__)
 
-_NATIVE_HITL_TOOL_DEDUP_KEYS: dict[str, str] = {
+# Resolver type — given the tool ``args`` dict returns a stable
-    # Gmail
+# string used to dedupe consecutive calls. ``None`` means no dedup.
-    "send_gmail_email": "subject",
+DedupResolver = Callable[[dict[str, Any]], str]
-    "create_gmail_draft": "subject",
+
-    "update_gmail_draft": "draft_subject_or_id",
+
-    "trash_gmail_email": "email_subject_or_id",
+def wrap_dedup_key_by_arg_name(arg_name: str) -> DedupResolver:
-    # Google Calendar
+    """Adapt a string-arg name into a :data:`DedupResolver`.
-    "create_calendar_event": "title",
+
-    "update_calendar_event": "event_title_or_id",
+    Convenience helper used by registry entries that just want to dedupe
-    "delete_calendar_event": "event_title_or_id",
+    on a single arg's lowercased value (the most common case for native
-    # Google Drive
+    HITL tools like ``send_gmail_email`` keyed on ``subject``).
-    "create_google_drive_file": "file_name",
+
-    "delete_google_drive_file": "file_name",
+    Example::
-    # OneDrive
+
-    "create_onedrive_file": "file_name",
+        ToolDefinition(
-    "delete_onedrive_file": "file_name",
+            name="send_gmail_email",
-    # Dropbox
+            ...,
-    "create_dropbox_file": "file_name",
+            dedup_key=wrap_dedup_key_by_arg_name("subject"),
-    "delete_dropbox_file": "file_name",
+        )
-    # Notion
+    """
-    "create_notion_page": "title",
+
-    "update_notion_page": "page_title",
+    def _resolver(args: dict[str, Any]) -> str:
-    "delete_notion_page": "page_title",
+        return str(args.get(arg_name, "")).lower()
-    # Linear
+
-    "create_linear_issue": "title",
+    return _resolver
-    "update_linear_issue": "issue_ref",
+
-    "delete_linear_issue": "issue_ref",
+
-    # Jira
+# Backwards-compatible alias for code that imported the original
-    "create_jira_issue": "summary",
+# private name. New callers should use :func:`wrap_dedup_key_by_arg_name`.
-    "update_jira_issue": "issue_title_or_key",
+_wrap_string_key = wrap_dedup_key_by_arg_name
-    "delete_jira_issue": "issue_title_or_key",
-    # Confluence
-    "create_confluence_page": "title",
-    "update_confluence_page": "page_title_or_id",
-    "delete_confluence_page": "page_title_or_id",
-}
 
 
 class DedupHITLToolCallsMiddleware(AgentMiddleware):  # type: ignore[type-arg]
     """Remove duplicate HITL tool calls from a single LLM response.
 
-    Only the **first** occurrence of each (tool-name, primary-arg-value)
+    Only the **first** occurrence of each ``(tool-name, dedup_key)``
     pair is kept; subsequent duplicates are silently dropped.
 
-    The dedup map is built from two sources:
+    The dedup-resolver map is built from two sources, in priority order:
 
-    1. A comprehensive list of native HITL tools (hardcoded above).
+    1. ``tool.metadata["dedup_key"]`` — callable provided by the registry's
-    2. Any ``StructuredTool`` instances passed via *agent_tools* whose
+       ``ToolDefinition.dedup_key``. Receives the args dict and returns
-       ``metadata`` contains ``{"hitl": True, "hitl_dedup_key": "..."}``.
+       a string signature. This is the canonical mechanism.
-       This is how MCP tools automatically get dedup support.
+    2. ``tool.metadata["hitl_dedup_key"]`` — string with a primary arg
+       name; primarily used by MCP / Composio tools.
     """
 
     tools = ()
 
     def __init__(self, *, agent_tools: list[Any] | None = None) -> None:
-        self._dedup_keys: dict[str, str] = dict(_NATIVE_HITL_TOOL_DEDUP_KEYS)
+        self._resolvers: dict[str, DedupResolver] = {}
+
         for t in agent_tools or []:
             meta = getattr(t, "metadata", None) or {}
+            callable_key = meta.get("dedup_key")
+            if callable(callable_key):
+                self._resolvers[t.name] = callable_key
+                continue
             if meta.get("hitl") and meta.get("hitl_dedup_key"):
-                self._dedup_keys[t.name] = meta["hitl_dedup_key"]
+                self._resolvers[t.name] = wrap_dedup_key_by_arg_name(
+                    meta["hitl_dedup_key"]
+                )
 
     def after_model(
         self, state: AgentState, runtime: Runtime[Any]
     ) -> dict[str, Any] | None:
-        return self._dedup(state, self._dedup_keys)
+        return self._dedup(state, self._resolvers)
 
     async def aafter_model(
         self, state: AgentState, runtime: Runtime[Any]
     ) -> dict[str, Any] | None:
-        return self._dedup(state, self._dedup_keys)
+        return self._dedup(state, self._resolvers)
 
     @staticmethod
     def _dedup(
         state: AgentState,
-        dedup_keys: dict[str, str],  # type: ignore[type-arg]
+        resolvers: dict[str, DedupResolver],
     ) -> dict[str, Any] | None:
         messages = state.get("messages")
         if not messages:
@@ -110,9 +122,16 @@ class DedupHITLToolCallsMiddleware(AgentMiddleware):  # type: ignore[type-arg]
 
         for tc in tool_calls:
             name = tc.get("name", "")
-            dedup_key_arg = dedup_keys.get(name)
+            resolver = resolvers.get(name)
-            if dedup_key_arg is not None:
+            if resolver is not None:
-                arg_val = str(tc.get("args", {}).get(dedup_key_arg, "")).lower()
+                try:
+                    arg_val = resolver(tc.get("args", {}) or {})
+                except Exception:
+                    logger.exception(
+                        "Dedup resolver for tool %s raised; keeping call", name
+                    )
+                    deduped.append(tc)
+                    continue
                 key = (name, arg_val)
                 if key in seen:
                     logger.info(

surfsense_backend/app/agents/new_chat/middleware/doom_loop.py

@@ -0,0 +1,237 @@
+"""
+DoomLoopMiddleware — pattern-based detector for repeated identical tool calls.
+
+LangChain has :class:`ToolCallLimitMiddleware` which caps the *total* number
+of tool calls per turn — but it can't tell apart "10 distinct, useful
+calls" from "the same call 10 times in a row". This middleware fills that
+gap with a sliding-window check on tool-call signatures, ported from
+OpenCode's ``packages/opencode/src/session/processor.ts``.
+
+When the same tool with the same arguments is called N times in a row,
+the agent has likely entered an infinite loop. We surface this to the
+user as an interrupt with ``permission="doom_loop"`` so the UI can
+render an "Are you stuck? Continue / cancel?" affordance.
+
+This ships **OFF by default** until the frontend explicitly handles
+``context.permission == "doom_loop"`` interrupts.
+
+Wire format: uses SurfSense's existing ``interrupt()`` payload shape
+(see ``app/agents/new_chat/tools/hitl.py``):
+
+    {
+        "type": "permission_ask",
+        "action": {"tool": <name>, "params": <args>},
+        "context": {"permission": "doom_loop", "recent_signatures": [...]},
+    }
+
+so the frontend that already handles HITL prompts can render this with
+no changes beyond a string check.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import logging
+from collections import deque
+from typing import Any
+
+from langchain.agents.middleware.types import (
+    AgentMiddleware,
+    AgentState,
+    ContextT,
+    ResponseT,
+)
+from langchain_core.messages import AIMessage
+from langgraph.config import get_config
+from langgraph.runtime import Runtime
+from langgraph.types import interrupt
+
+from app.observability import otel as ot
+
+logger = logging.getLogger(__name__)
+
+
+def _signature(name: str, args: Any) -> str:
+    """Hash a tool call ``(name, args)`` to a short signature."""
+    try:
+        canonical = json.dumps(args, sort_keys=True, default=str)
+    except (TypeError, ValueError):
+        canonical = repr(args)
+    digest = hashlib.sha1(f"{name}::{canonical}".encode()).hexdigest()
+    return digest[:16]
+
+
+class DoomLoopMiddleware(AgentMiddleware[AgentState[ResponseT], ContextT, ResponseT]):
+    """Detect repeated identical tool calls and prompt the user.
+
+    Tracks a sliding window of the most-recent ``threshold`` tool-call
+    signatures across the live request. When all entries match, raise
+    a SurfSense-style HITL interrupt with ``permission="doom_loop"``.
+
+    Args:
+        threshold: How many consecutive identical signatures count as a
+            doom loop. Default 3 (matches OpenCode's processor.ts).
+    """
+
+    def __init__(self, *, threshold: int = 3) -> None:
+        super().__init__()
+        if threshold < 2:
+            raise ValueError("DoomLoopMiddleware threshold must be >= 2")
+        self._threshold = threshold
+        self.tools = []
+        # Per-thread sliding windows. We can't put this in graph state
+        # without state-schema gymnastics; for one process-lifetime it's
+        # fine to keep an in-memory map keyed by thread_id.
+        self._windows: dict[str, deque[str]] = {}
+
+    @staticmethod
+    def _thread_id_from_runtime(runtime: Runtime[ContextT]) -> str:
+        """Resolve the thread id for sliding-window keying.
+
+        Prefer LangGraph's ``get_config()`` (the only way to read
+        ``RunnableConfig`` inside a node — :class:`Runtime` does NOT carry
+        a ``config`` attribute). Fall back to ``runtime.config`` for unit
+        tests that synthesize a config-bearing stub. Default
+        ``"no_thread"`` is intentionally only used when both lookups fail
+        — it would collapse all threads into one window so we keep the
+        debug log loud.
+        """
+
+        def _from_dict(cfg: Any) -> str | None:
+            if not isinstance(cfg, dict):
+                return None
+            tid = (cfg.get("configurable") or {}).get("thread_id")
+            return str(tid) if tid is not None else None
+
+        try:
+            tid = _from_dict(get_config())
+        except Exception:
+            tid = None
+        if tid is not None:
+            return tid
+
+        tid = _from_dict(getattr(runtime, "config", None))
+        if tid is not None:
+            return tid
+
+        logger.debug(
+            "DoomLoopMiddleware: no thread_id resolved from RunnableConfig; "
+            "falling back to shared 'no_thread' window."
+        )
+        return "no_thread"
+
+    def _window(self, thread_id: str) -> deque[str]:
+        win = self._windows.get(thread_id)
+        if win is None:
+            win = deque(maxlen=self._threshold)
+            self._windows[thread_id] = win
+        return win
+
+    def _detect(
+        self, message: AIMessage, runtime: Runtime[ContextT]
+    ) -> tuple[bool, list[str], dict[str, Any] | None]:
+        if not message.tool_calls:
+            return False, [], None
+
+        thread_id = self._thread_id_from_runtime(runtime)
+        window = self._window(thread_id)
+
+        triggered_call: dict[str, Any] | None = None
+        for call in message.tool_calls:
+            name = (
+                call.get("name")
+                if isinstance(call, dict)
+                else getattr(call, "name", None)
+            )
+            args = (
+                call.get("args")
+                if isinstance(call, dict)
+                else getattr(call, "args", {})
+            )
+            if not isinstance(name, str):
+                continue
+            sig = _signature(name, args)
+            window.append(sig)
+            if len(window) >= self._threshold and len(set(window)) == 1:
+                triggered_call = {"name": name, "params": args or {}}
+                break
+
+        if triggered_call is None:
+            return False, list(window), None
+        return True, list(window), triggered_call
+
+    def after_model(  # type: ignore[override]
+        self,
+        state: AgentState[ResponseT],
+        runtime: Runtime[ContextT],
+    ) -> dict[str, Any] | None:
+        messages = state.get("messages") or []
+        if not messages:
+            return None
+        last = messages[-1]
+        if not isinstance(last, AIMessage):
+            return None
+
+        triggered, signatures, action = self._detect(last, runtime)
+        if not triggered:
+            return None
+
+        logger.warning(
+            "Doom loop detected: tool %s called %d times in a row (sig=%s)",
+            action["name"] if action else "<unknown>",
+            self._threshold,
+            signatures[-1] if signatures else "<empty>",
+        )
+
+        # Open an interrupt.raised span with permission=doom_loop attribute
+        # so dashboards can break out doom-loop interrupts from regular
+        # permission asks via the ``interrupt.permission`` attribute.
+        with ot.interrupt_span(
+            interrupt_type="permission_ask",
+            extra={
+                "interrupt.permission": "doom_loop",
+                "interrupt.threshold": self._threshold,
+                "interrupt.tool": (action or {}).get("tool", "<unknown>"),
+            },
+        ):
+            decision = interrupt(
+                {
+                    "type": "permission_ask",
+                    "action": action or {"tool": "<unknown>", "params": {}},
+                    "context": {
+                        "permission": "doom_loop",
+                        "recent_signatures": signatures,
+                        "threshold": self._threshold,
+                    },
+                }
+            )
+
+        # Reset window so the next decision (continue/cancel) starts fresh.
+        thread_id = self._thread_id_from_runtime(runtime)
+        self._windows.pop(thread_id, None)
+
+        # Decision shape mirrors ``tools/hitl.py``: {"decision_type": "..."}
+        # If the user cancelled, jump to end. Otherwise return ``None`` so the
+        # tool call proceeds. The frontend's exact reply names may differ —
+        # we tolerate any shape that contains a string with "reject"/"cancel".
+        if isinstance(decision, dict):
+            kind = str(
+                decision.get("decision_type") or decision.get("type") or ""
+            ).lower()
+            if "reject" in kind or "cancel" in kind:
+                return {"jump_to": "end"}
+        return None
+
+    async def aafter_model(  # type: ignore[override]
+        self,
+        state: AgentState[ResponseT],
+        runtime: Runtime[ContextT],
+    ) -> dict[str, Any] | None:
+        return self.after_model(state, runtime)
+
+
+__all__ = [
+    "DoomLoopMiddleware",
+    "_signature",
+]

surfsense_backend/app/agents/new_chat/middleware/file_intent.py

@@ -21,7 +21,7 @@ from typing import Any
 
 from langchain.agents.middleware import AgentMiddleware, AgentState
 from langchain_core.language_models import BaseChatModel
-from langchain_core.messages import BaseMessage, HumanMessage, SystemMessage
+from langchain_core.messages import AIMessage, BaseMessage, HumanMessage, SystemMessage
 from langgraph.runtime import Runtime
 from pydantic import BaseModel, Field, ValidationError
 
@@ -213,10 +213,23 @@ def _build_classifier_prompt(*, recent_conversation: str, user_text: str) -> str
     )
 
 
-def _build_recent_conversation(messages: list[BaseMessage], *, max_messages: int = 6) -> str:
+def _build_recent_conversation(
+    messages: list[BaseMessage], *, max_messages: int = 6
+) -> str:
     rows: list[str] = []
-    for msg in messages[-max_messages:]:
+    filtered: list[tuple[str, BaseMessage]] = []
-        role = "user" if isinstance(msg, HumanMessage) else "assistant"
+    for msg in messages:
+        role: str | None = None
+        if isinstance(msg, HumanMessage):
+            role = "user"
+        elif isinstance(msg, AIMessage):
+            if getattr(msg, "tool_calls", None):
+                continue
+            role = "assistant"
+        else:
+            continue
+        filtered.append((role, msg))
+    for role, msg in filtered[-max_messages:]:
         text = re.sub(r"\s+", " ", _extract_text_from_message(msg)).strip()
         if text:
             rows.append(f"{role}: {text[:280]}")
@@ -246,7 +259,9 @@ class FileIntentMiddleware(AgentMiddleware):  # type: ignore[type-arg]
                 [HumanMessage(content=prompt)],
                 config={"tags": ["surfsense:internal"]},
             )
-            payload = json.loads(_extract_json_payload(_extract_text_from_message(response)))
+            payload = json.loads(
+                _extract_json_payload(_extract_text_from_message(response))
+            )
             plan = FileIntentPlan.model_validate(payload)
             return plan
         except (json.JSONDecodeError, ValidationError, ValueError) as exc:
@@ -317,4 +332,3 @@ class FileIntentMiddleware(AgentMiddleware):  # type: ignore[type-arg]
         insert_at = max(len(new_messages) - 1, 0)
         new_messages.insert(insert_at, contract_msg)
         return {"messages": new_messages, "file_operation_contract": contract}
-

surfsense_backend/app/agents/new_chat/middleware/kb_persistence.py

@@ -0,0 +1,645 @@
+"""End-of-turn persistence for the cloud-mode SurfSense filesystem.
+
+This middleware runs ``aafter_agent`` once per turn (cloud only). It commits
+all staged folder creations, file moves, and content writes/edits to
+Postgres in a single ordered pass:
+
+1. Materialize ``staged_dirs`` into ``Folder`` rows.
+2. Apply ``pending_moves`` in order (chained moves resolved via
+   ``doc_id_by_path``).
+3. Normalize ``dirty_paths`` through ``pending_moves`` so write-then-move
+   sequences commit at the final path.
+4. Commit content writes / edits for ``/documents/*`` paths, skipping
+   ``temp_*`` basenames.
+
+The commit body is exposed as a free function ``commit_staged_filesystem_state``
+so the optional stream-task fallback (``stream_new_chat.py``) can call the
+exact same routine when ``aafter_agent`` was skipped (e.g. client disconnect).
+"""
+
+from __future__ import annotations
+
+import logging
+from datetime import UTC, datetime
+from typing import Any
+
+from fractional_indexing import generate_key_between
+from langchain.agents.middleware import AgentMiddleware, AgentState
+from langchain_core.callbacks import dispatch_custom_event
+from langgraph.runtime import Runtime
+from sqlalchemy import delete, select
+from sqlalchemy.exc import IntegrityError
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.agents.new_chat.filesystem_selection import FilesystemMode
+from app.agents.new_chat.filesystem_state import SurfSenseFilesystemState
+from app.agents.new_chat.path_resolver import (
+    DOCUMENTS_ROOT,
+    parse_documents_path,
+    safe_folder_segment,
+    virtual_path_to_doc,
+)
+from app.agents.new_chat.state_reducers import _CLEAR
+from app.db import (
+    Chunk,
+    Document,
+    DocumentType,
+    Folder,
+    shielded_async_session,
+)
+from app.indexing_pipeline.document_chunker import chunk_text
+from app.utils.document_converters import (
+    embed_texts,
+    generate_content_hash,
+    generate_unique_identifier_hash,
+)
+
+logger = logging.getLogger(__name__)
+
+
+_TEMP_PREFIX = "temp_"
+
+
+def _basename(path: str) -> str:
+    return path.rsplit("/", 1)[-1]
+
+
+# ---------------------------------------------------------------------------
+# Folder helpers
+# ---------------------------------------------------------------------------
+
+
+async def _ensure_folder_hierarchy(
+    session: AsyncSession,
+    *,
+    search_space_id: int,
+    created_by_id: str | None,
+    folder_parts: list[str],
+) -> int | None:
+    """Ensure a chain of folder names exists under the search space.
+
+    Returns the leaf folder id, or ``None`` if ``folder_parts`` is empty
+    (i.e. a document directly under ``/documents/``).
+    """
+    if not folder_parts:
+        return None
+    parent_id: int | None = None
+    for raw in folder_parts:
+        name = safe_folder_segment(str(raw))
+        query = select(Folder).where(
+            Folder.search_space_id == search_space_id,
+            Folder.name == name,
+        )
+        if parent_id is None:
+            query = query.where(Folder.parent_id.is_(None))
+        else:
+            query = query.where(Folder.parent_id == parent_id)
+        result = await session.execute(query)
+        folder = result.scalar_one_or_none()
+        if folder is None:
+            sibling_query = (
+                select(Folder.position).order_by(Folder.position.desc()).limit(1)
+            )
+            sibling_query = sibling_query.where(
+                Folder.search_space_id == search_space_id
+            )
+            if parent_id is None:
+                sibling_query = sibling_query.where(Folder.parent_id.is_(None))
+            else:
+                sibling_query = sibling_query.where(Folder.parent_id == parent_id)
+            sibling_result = await session.execute(sibling_query)
+            last_position = sibling_result.scalar_one_or_none()
+            folder = Folder(
+                name=name,
+                position=generate_key_between(last_position, None),
+                parent_id=parent_id,
+                search_space_id=search_space_id,
+                created_by_id=created_by_id,
+                updated_at=datetime.now(UTC),
+            )
+            session.add(folder)
+            await session.flush()
+        parent_id = folder.id
+    return parent_id
+
+
+# ---------------------------------------------------------------------------
+# Document helpers
+# ---------------------------------------------------------------------------
+
+
+async def _create_document(
+    session: AsyncSession,
+    *,
+    virtual_path: str,
+    content: str,
+    search_space_id: int,
+    created_by_id: str | None,
+) -> Document:
+    """Create a NOTE Document + Chunks for ``virtual_path``."""
+    folder_parts, title = parse_documents_path(virtual_path)
+    if not title:
+        raise ValueError(f"invalid /documents path '{virtual_path}'")
+    folder_id = await _ensure_folder_hierarchy(
+        session,
+        search_space_id=search_space_id,
+        created_by_id=created_by_id,
+        folder_parts=folder_parts,
+    )
+    unique_identifier_hash = generate_unique_identifier_hash(
+        DocumentType.NOTE,
+        virtual_path,
+        search_space_id,
+    )
+    # Filesystem-parity invariant: the only thing that *must* be unique is
+    # the path. Two notes can legitimately share content (e.g. ``cp a b``).
+    # Guard against the path-derived ``unique_identifier_hash`` constraint
+    # so we surface a clean ValueError instead of letting the INSERT poison
+    # the session with an IntegrityError.
+    path_collision = await session.execute(
+        select(Document.id).where(
+            Document.search_space_id == search_space_id,
+            Document.unique_identifier_hash == unique_identifier_hash,
+        )
+    )
+    if path_collision.scalar_one_or_none() is not None:
+        raise ValueError(
+            f"a document already exists at path '{virtual_path}' "
+            "(unique_identifier_hash collision)"
+        )
+    # ``content_hash`` is intentionally NOT checked for uniqueness here.
+    # In a real filesystem two files at different paths can hold identical
+    # bytes, and the agent's ``write_file`` path needs that semantic to
+    # support copy/duplicate operations. The hash remains useful as a
+    # change-detection hint for connector indexers, which still consult it
+    # via :func:`check_duplicate_document` but do so with a non-unique
+    # lookup (``.first()``).
+    content_hash = generate_content_hash(content, search_space_id)
+    doc = Document(
+        title=title,
+        document_type=DocumentType.NOTE,
+        document_metadata={"virtual_path": virtual_path},
+        content=content,
+        content_hash=content_hash,
+        unique_identifier_hash=unique_identifier_hash,
+        source_markdown=content,
+        search_space_id=search_space_id,
+        folder_id=folder_id,
+        created_by_id=created_by_id,
+        updated_at=datetime.now(UTC),
+    )
+    session.add(doc)
+    await session.flush()
+
+    summary_embedding = embed_texts([content])[0]
+    doc.embedding = summary_embedding
+    chunks = chunk_text(content)
+    if chunks:
+        chunk_embeddings = embed_texts(chunks)
+        session.add_all(
+            [
+                Chunk(document_id=doc.id, content=text, embedding=embedding)
+                for text, embedding in zip(chunks, chunk_embeddings, strict=True)
+            ]
+        )
+    return doc
+
+
+async def _update_document(
+    session: AsyncSession,
+    *,
+    doc_id: int,
+    content: str,
+    virtual_path: str,
+    search_space_id: int,
+) -> Document | None:
+    """Update an existing Document's content + chunks."""
+    result = await session.execute(
+        select(Document).where(
+            Document.id == doc_id,
+            Document.search_space_id == search_space_id,
+        )
+    )
+    document = result.scalar_one_or_none()
+    if document is None:
+        return None
+
+    document.content = content
+    document.source_markdown = content
+    document.content_hash = generate_content_hash(content, search_space_id)
+    document.updated_at = datetime.now(UTC)
+    metadata = dict(document.document_metadata or {})
+    metadata["virtual_path"] = virtual_path
+    document.document_metadata = metadata
+    document.unique_identifier_hash = generate_unique_identifier_hash(
+        DocumentType.NOTE,
+        virtual_path,
+        search_space_id,
+    )
+
+    summary_embedding = embed_texts([content])[0]
+    document.embedding = summary_embedding
+
+    await session.execute(delete(Chunk).where(Chunk.document_id == document.id))
+    chunks = chunk_text(content)
+    if chunks:
+        chunk_embeddings = embed_texts(chunks)
+        session.add_all(
+            [
+                Chunk(document_id=document.id, content=text, embedding=embedding)
+                for text, embedding in zip(chunks, chunk_embeddings, strict=True)
+            ]
+        )
+    return document
+
+
+# ---------------------------------------------------------------------------
+# Move helpers
+# ---------------------------------------------------------------------------
+
+
+async def _apply_move(
+    session: AsyncSession,
+    *,
+    search_space_id: int,
+    created_by_id: str | None,
+    move: dict[str, Any],
+    doc_id_by_path: dict[str, int],
+    doc_id_path_tombstones: dict[str, int | None],
+) -> dict[str, Any] | None:
+    """Apply a single staged move; updates the in-memory mapping for chain resolution."""
+    source = str(move.get("source") or "")
+    dest = str(move.get("dest") or "")
+    if not source or not dest or source == dest:
+        return None
+
+    if not source.startswith(DOCUMENTS_ROOT + "/") or not dest.startswith(
+        DOCUMENTS_ROOT + "/"
+    ):
+        return None
+
+    doc_id: int | None = doc_id_by_path.get(source)
+    document: Document | None = None
+    if doc_id is not None:
+        result = await session.execute(
+            select(Document).where(
+                Document.id == doc_id,
+                Document.search_space_id == search_space_id,
+            )
+        )
+        document = result.scalar_one_or_none()
+    if document is None:
+        document = await virtual_path_to_doc(
+            session,
+            search_space_id=search_space_id,
+            virtual_path=source,
+        )
+    if document is None:
+        logger.info(
+            "kb_persistence: skipping move %s -> %s (source not found)",
+            source,
+            dest,
+        )
+        return None
+
+    folder_parts, new_title = parse_documents_path(dest)
+    if not new_title:
+        return None
+    folder_id = await _ensure_folder_hierarchy(
+        session,
+        search_space_id=search_space_id,
+        created_by_id=created_by_id,
+        folder_parts=folder_parts,
+    )
+
+    document.title = new_title
+    document.folder_id = folder_id
+    metadata = dict(document.document_metadata or {})
+    metadata["virtual_path"] = dest
+    document.document_metadata = metadata
+    document.unique_identifier_hash = generate_unique_identifier_hash(
+        DocumentType.NOTE,
+        dest,
+        search_space_id,
+    )
+    document.updated_at = datetime.now(UTC)
+
+    doc_id_by_path.pop(source, None)
+    doc_id_by_path[dest] = document.id
+    doc_id_path_tombstones[source] = None
+    doc_id_path_tombstones[dest] = document.id
+    return {"id": document.id, "source": source, "dest": dest, "title": new_title}
+
+
+# ---------------------------------------------------------------------------
+# Commit body
+# ---------------------------------------------------------------------------
+
+
+async def commit_staged_filesystem_state(
+    state: dict[str, Any] | AgentState,
+    *,
+    search_space_id: int,
+    created_by_id: str | None,
+    filesystem_mode: FilesystemMode = FilesystemMode.CLOUD,
+    dispatch_events: bool = True,
+) -> dict[str, Any] | None:
+    """Commit all staged filesystem changes; return the state delta for reducers.
+
+    Shared between :class:`KnowledgeBasePersistenceMiddleware.aafter_agent`
+    and the optional stream-task fallback.
+    """
+    if filesystem_mode != FilesystemMode.CLOUD:
+        return None
+
+    state_dict: dict[str, Any] = (
+        dict(state)
+        if isinstance(state, dict)
+        else dict(getattr(state, "values", {}) or {})
+    )
+
+    files: dict[str, Any] = state_dict.get("files") or {}
+    staged_dirs: list[str] = list(state_dict.get("staged_dirs") or [])
+    pending_moves: list[dict[str, Any]] = list(state_dict.get("pending_moves") or [])
+    dirty_paths: list[str] = list(state_dict.get("dirty_paths") or [])
+    doc_id_by_path: dict[str, int] = dict(state_dict.get("doc_id_by_path") or {})
+    kb_anon_doc = state_dict.get("kb_anon_doc")
+
+    if kb_anon_doc:
+        temp_paths = [
+            p
+            for p in files
+            if isinstance(p, str) and _basename(p).startswith(_TEMP_PREFIX)
+        ]
+        return {
+            "dirty_paths": [_CLEAR],
+            "staged_dirs": [_CLEAR],
+            "pending_moves": [_CLEAR],
+            "files": dict.fromkeys(temp_paths),
+        }
+
+    if not (staged_dirs or pending_moves or dirty_paths):
+        return None
+
+    committed_creates: list[dict[str, Any]] = []
+    committed_updates: list[dict[str, Any]] = []
+    discarded: list[str] = []
+    applied_moves: list[dict[str, Any]] = []
+    doc_id_path_tombstones: dict[str, int | None] = {}
+    tree_changed = False
+
+    try:
+        async with shielded_async_session() as session:
+            for folder_path in staged_dirs:
+                if not isinstance(folder_path, str):
+                    continue
+                if not folder_path.startswith(DOCUMENTS_ROOT):
+                    continue
+                rel = folder_path[len(DOCUMENTS_ROOT) :].strip("/")
+                folder_parts_full = [p for p in rel.split("/") if p]
+                if not folder_parts_full:
+                    continue
+                await _ensure_folder_hierarchy(
+                    session,
+                    search_space_id=search_space_id,
+                    created_by_id=created_by_id,
+                    folder_parts=folder_parts_full,
+                )
+                tree_changed = True
+
+            for move in pending_moves:
+                applied = await _apply_move(
+                    session,
+                    search_space_id=search_space_id,
+                    created_by_id=created_by_id,
+                    move=move,
+                    doc_id_by_path=doc_id_by_path,
+                    doc_id_path_tombstones=doc_id_path_tombstones,
+                )
+                if applied:
+                    applied_moves.append(applied)
+                    tree_changed = True
+
+            move_alias = {
+                m["source"]: m["dest"] for m in pending_moves if m.get("source")
+            }
+
+            def _final_path(path: str) -> str:
+                seen: set[str] = set()
+                while path in move_alias and path not in seen:
+                    seen.add(path)
+                    path = move_alias[path]
+                return path
+
+            kb_dirty_seen: set[str] = set()
+            kb_dirty: list[str] = []
+            for raw in dirty_paths:
+                if not isinstance(raw, str):
+                    continue
+                final = _final_path(raw)
+                if not final.startswith(DOCUMENTS_ROOT + "/"):
+                    continue
+                if final in kb_dirty_seen:
+                    continue
+                kb_dirty_seen.add(final)
+                kb_dirty.append(final)
+
+            for path in kb_dirty:
+                basename = _basename(path)
+                if basename.startswith(_TEMP_PREFIX):
+                    discarded.append(path)
+                    continue
+                file_data = files.get(path)
+                if not isinstance(file_data, dict):
+                    continue
+                content = "\n".join(file_data.get("content") or [])
+                doc_id = doc_id_by_path.get(path)
+                if doc_id is None:
+                    # The in-memory ``doc_id_by_path`` is per-thread and starts
+                    # empty in every new chat. If the agent writes to a path
+                    # that already exists in the DB (e.g. a previous chat's
+                    # ``notes.md``), we must NOT try to INSERT — it would hit
+                    # ``unique_identifier_hash`` (path-derived). Look up the
+                    # existing doc and update it in place instead.
+                    existing = await virtual_path_to_doc(
+                        session,
+                        search_space_id=search_space_id,
+                        virtual_path=path,
+                    )
+                    if existing is not None:
+                        doc_id = existing.id
+                        doc_id_by_path[path] = existing.id
+                if doc_id is not None:
+                    updated = await _update_document(
+                        session,
+                        doc_id=doc_id,
+                        content=content,
+                        virtual_path=path,
+                        search_space_id=search_space_id,
+                    )
+                    if updated is not None:
+                        committed_updates.append(
+                            {
+                                "id": updated.id,
+                                "title": updated.title,
+                                "documentType": DocumentType.NOTE.value,
+                                "searchSpaceId": search_space_id,
+                                "folderId": updated.folder_id,
+                                "createdById": str(created_by_id)
+                                if created_by_id
+                                else None,
+                                "virtualPath": path,
+                            }
+                        )
+                else:
+                    # Wrap each create in a SAVEPOINT so a residual
+                    # ``IntegrityError`` (e.g. a deployment that hasn't run
+                    # migration 133 yet, where ``documents.content_hash``
+                    # still carries its legacy global UNIQUE constraint)
+                    # rolls back only this one create instead of poisoning
+                    # the whole turn's transaction.
+                    try:
+                        async with session.begin_nested():
+                            new_doc = await _create_document(
+                                session,
+                                virtual_path=path,
+                                content=content,
+                                search_space_id=search_space_id,
+                                created_by_id=created_by_id,
+                            )
+                    except ValueError as exc:
+                        logger.warning(
+                            "kb_persistence: skipping %s create: %s", path, exc
+                        )
+                        continue
+                    except IntegrityError as exc:
+                        # The path-uniqueness check above already protected
+                        # against ``unique_identifier_hash`` collisions, so
+                        # the most likely culprit is the legacy
+                        # ``ix_documents_content_hash`` UNIQUE constraint
+                        # that migration 133 drops. Log loudly so operators
+                        # know to run the migration; do NOT silently swallow.
+                        msg = str(exc.orig) if exc.orig is not None else str(exc)
+                        logger.error(
+                            "kb_persistence: IntegrityError creating %s: %s. "
+                            "If this mentions content_hash, run alembic "
+                            "upgrade to apply migration 133 which drops the "
+                            "global UNIQUE constraint on documents.content_hash.",
+                            path,
+                            msg,
+                        )
+                        continue
+                    doc_id_by_path[path] = new_doc.id
+                    committed_creates.append(
+                        {
+                            "id": new_doc.id,
+                            "title": new_doc.title,
+                            "documentType": DocumentType.NOTE.value,
+                            "searchSpaceId": search_space_id,
+                            "folderId": new_doc.folder_id,
+                            "createdById": str(created_by_id)
+                            if created_by_id
+                            else None,
+                            "virtualPath": path,
+                        }
+                    )
+                    tree_changed = True
+
+            await session.commit()
+    except Exception:  # pragma: no cover - rollback safety net
+        logger.exception(
+            "kb_persistence: commit failed (search_space=%s)", search_space_id
+        )
+        return None
+
+    if dispatch_events:
+        for payload in committed_creates:
+            try:
+                dispatch_custom_event("document_created", payload)
+            except Exception:
+                logger.exception(
+                    "kb_persistence: failed to dispatch document_created event"
+                )
+        for payload in committed_updates:
+            try:
+                dispatch_custom_event("document_updated", payload)
+            except Exception:
+                logger.exception(
+                    "kb_persistence: failed to dispatch document_updated event"
+                )
+
+    temp_paths = [
+        p for p in files if isinstance(p, str) and _basename(p).startswith(_TEMP_PREFIX)
+    ]
+
+    doc_id_update: dict[str, int | None] = {**doc_id_path_tombstones}
+    for payload in committed_creates:
+        doc_id_update[str(payload.get("virtualPath") or "")] = int(payload["id"])
+
+    delta: dict[str, Any] = {
+        "dirty_paths": [_CLEAR],
+        "staged_dirs": [_CLEAR],
+        "pending_moves": [_CLEAR],
+    }
+    if temp_paths:
+        delta["files"] = dict.fromkeys(temp_paths)
+    if doc_id_update:
+        delta["doc_id_by_path"] = doc_id_update
+    if tree_changed:
+        delta["tree_version"] = int(state_dict.get("tree_version") or 0) + 1
+
+    logger.info(
+        "kb_persistence: commit (search_space=%s) creates=%d updates=%d "
+        "moves=%d staged_dirs=%d discarded=%d",
+        search_space_id,
+        len(committed_creates),
+        len(committed_updates),
+        len(applied_moves),
+        len(staged_dirs),
+        len(discarded),
+    )
+    return delta
+
+
+# ---------------------------------------------------------------------------
+# Middleware
+# ---------------------------------------------------------------------------
+
+
+class KnowledgeBasePersistenceMiddleware(AgentMiddleware):  # type: ignore[type-arg]
+    """End-of-turn cloud persistence for the SurfSense filesystem agent."""
+
+    tools = ()
+    state_schema = SurfSenseFilesystemState
+
+    def __init__(
+        self,
+        *,
+        search_space_id: int,
+        created_by_id: str | None,
+        filesystem_mode: FilesystemMode,
+    ) -> None:
+        self.search_space_id = search_space_id
+        self.created_by_id = created_by_id
+        self.filesystem_mode = filesystem_mode
+
+    async def aafter_agent(  # type: ignore[override]
+        self,
+        state: AgentState,
+        runtime: Runtime[Any],
+    ) -> dict[str, Any] | None:
+        del runtime
+        if self.filesystem_mode != FilesystemMode.CLOUD:
+            return None
+        return await commit_staged_filesystem_state(
+            state,
+            search_space_id=self.search_space_id,
+            created_by_id=self.created_by_id,
+            filesystem_mode=self.filesystem_mode,
+        )
+
+
+__all__ = [
+    "KnowledgeBasePersistenceMiddleware",
+    "commit_staged_filesystem_state",
+]

surfsense_backend/app/agents/new_chat/middleware/kb_postgres_backend.py

@@ -0,0 +1,963 @@
+"""Postgres-backed virtual filesystem for the SurfSense agent (cloud mode).
+
+The backend is **strictly conforming** to deepagents'
+:class:`BackendProtocol`. It returns ``WriteResult`` / ``EditResult`` / list
+shapes exactly as upstream expects (no extra fields). All side-state
+plumbing — ``dirty_paths``, ``doc_id_by_path``, ``staged_dirs``,
+``pending_moves``, ``files`` cache — is appended by the overridden tool
+wrappers in :class:`SurfSenseFilesystemMiddleware` via ``Command.update``.
+
+The backend never writes to Postgres. End-of-turn persistence is handled by
+:class:`KnowledgeBasePersistenceMiddleware`. This module is purely a
+read-side and a state-merging helper.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import contextlib
+import fnmatch
+import logging
+import re
+from datetime import UTC
+from typing import Any
+
+from deepagents.backends.protocol import (
+    BackendProtocol,
+    EditResult,
+    FileDownloadResponse,
+    FileInfo,
+    FileUploadResponse,
+    GrepMatch,
+    WriteResult,
+)
+from deepagents.backends.utils import (
+    create_file_data,
+    file_data_to_string,
+    format_read_response,
+    perform_string_replacement,
+    update_file_data,
+)
+from langchain.tools import ToolRuntime
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.agents.new_chat.document_xml import build_document_xml
+from app.agents.new_chat.path_resolver import (
+    DOCUMENTS_ROOT,
+    build_path_index,
+    doc_to_virtual_path,
+    virtual_path_to_doc,
+)
+from app.db import Chunk, Document, shielded_async_session
+
+logger = logging.getLogger(__name__)
+
+_TEMP_PREFIX = "temp_"
+_GREP_MAX_TOTAL_MATCHES = 50
+_GREP_MAX_PER_DOC = 5
+
+
+def _basename(path: str) -> str:
+    return path.rsplit("/", 1)[-1]
+
+
+def _is_under(child: str, parent: str) -> bool:
+    """Return True iff ``child`` is at-or-under ``parent`` (directory semantics)."""
+    if parent == "/":
+        return child.startswith("/")
+    return child == parent or child.startswith(parent.rstrip("/") + "/")
+
+
+def paginate_listing(
+    infos: list[FileInfo],
+    *,
+    offset: int = 0,
+    limit: int | None = None,
+) -> list[FileInfo]:
+    """Paginate a listing produced by :meth:`KBPostgresBackend.als_info`."""
+    if offset < 0:
+        offset = 0
+    end: int | None
+    end = None if limit is None or limit < 0 else offset + limit
+    return list(infos[offset:end])
+
+
+class KBPostgresBackend(BackendProtocol):
+    """Lazy, read-only Postgres view for ``/documents/*`` virtual paths.
+
+    The backend exposes a virtual ``/documents/`` namespace mirroring the
+    ``Folder``/``Document`` graph. Reads materialize XML on first access and
+    cache it via the overriding tool wrappers (NOT here). Writes never touch
+    the DB — they return ``files_update`` deltas that the wrappers turn into
+    Command updates, and the persistence middleware commits them at end of
+    turn.
+    """
+
+    _IMAGE_EXTENSIONS = frozenset({".png", ".jpg", ".jpeg", ".gif", ".webp"})
+
+    def __init__(self, search_space_id: int, runtime: ToolRuntime) -> None:
+        self.search_space_id = search_space_id
+        self.runtime = runtime
+
+    @property
+    def state(self) -> dict[str, Any]:
+        return getattr(self.runtime, "state", {}) or {}
+
+    # ------------------------------------------------------------------ helpers
+
+    def _state_files(self) -> dict[str, Any]:
+        return dict(self.state.get("files") or {})
+
+    def _staged_dirs(self) -> list[str]:
+        return list(self.state.get("staged_dirs") or [])
+
+    def _pending_moves(self) -> list[dict[str, Any]]:
+        return list(self.state.get("pending_moves") or [])
+
+    def _kb_anon_doc(self) -> dict[str, Any] | None:
+        anon = self.state.get("kb_anon_doc")
+        return anon if isinstance(anon, dict) else None
+
+    def _matched_chunk_ids(self, doc_id: int) -> set[int]:
+        mapping = self.state.get("kb_matched_chunk_ids") or {}
+        try:
+            return set(mapping.get(doc_id, []) or [])
+        except TypeError:
+            return set()
+
+    @staticmethod
+    def _file_data_size(file_data: dict[str, Any]) -> int:
+        try:
+            return len("\n".join(file_data.get("content") or []))
+        except Exception:
+            return 0
+
+    def _normalize_listing_path(self, path: str) -> str:
+        if not path:
+            return DOCUMENTS_ROOT
+        if path == "/":
+            return path
+        return path.rstrip("/") if path != "/" else path
+
+    def _moved_view_paths(
+        self,
+        existing: dict[str, dict[str, Any]],
+    ) -> tuple[set[str], dict[str, str]]:
+        """Apply ``pending_moves`` to a path set and return ``(removed, alias)``.
+
+        Removed paths should disappear from listings; ``alias[source] = dest``
+        means a virtual entry should appear at ``dest`` even if no DB row is
+        yet there.
+        """
+        removed: set[str] = set()
+        alias: dict[str, str] = {}
+        for move in self._pending_moves():
+            src = move.get("source")
+            dst = move.get("dest")
+            if not src or not dst:
+                continue
+            removed.add(src)
+            alias[src] = dst
+            existing.pop(src, None)
+        return removed, alias
+
+    # ------------------------------------------------------------------ ls/read
+
+    async def als_info(self, path: str) -> list[FileInfo]:  # type: ignore[override]
+        normalized = self._normalize_listing_path(path)
+        infos: list[FileInfo] = []
+        seen: set[str] = set()
+
+        anon = self._kb_anon_doc()
+        if anon:
+            anon_path = str(anon.get("path") or "")
+            if (
+                anon_path
+                and _is_under(anon_path, normalized)
+                and anon_path != normalized
+                and anon_path not in seen
+            ):
+                infos.append(
+                    FileInfo(
+                        path=anon_path,
+                        is_dir=False,
+                        size=len(str(anon.get("content") or "")),
+                        modified_at="",
+                    )
+                )
+                seen.add(anon_path)
+
+        files = self._state_files()
+        moved_removed, moved_alias = self._moved_view_paths(files)
+
+        if normalized.startswith(DOCUMENTS_ROOT) or normalized == "/":
+            try:
+                async with shielded_async_session() as session:
+                    db_infos, subdir_paths = await self._list_db_directory(
+                        session, normalized
+                    )
+            except Exception as exc:  # pragma: no cover - defensive
+                logger.warning("KBPostgresBackend.als_info DB error: %s", exc)
+                db_infos, subdir_paths = [], set()
+
+            for info in db_infos:
+                p = info.get("path", "")
+                if not p or p in seen or p in moved_removed:
+                    continue
+                infos.append(info)
+                seen.add(p)
+
+            for src, dst in moved_alias.items():
+                if src not in seen:
+                    if not _is_under(dst, normalized):
+                        continue
+                    rel = (
+                        dst[len(normalized) :].lstrip("/")
+                        if normalized != "/"
+                        else dst.lstrip("/")
+                    )
+                    if "/" in rel:
+                        subdir_paths.add(
+                            (normalized.rstrip("/") + "/" + rel.split("/", 1)[0])
+                            if normalized != "/"
+                            else "/" + rel.split("/", 1)[0]
+                        )
+                        continue
+                    if dst in seen:
+                        continue
+                    fd = files.get(dst)
+                    size = self._file_data_size(fd) if isinstance(fd, dict) else 0
+                    infos.append(
+                        FileInfo(
+                            path=dst,
+                            is_dir=False,
+                            size=int(size),
+                            modified_at=fd.get("modified_at", "")
+                            if isinstance(fd, dict)
+                            else "",
+                        )
+                    )
+                    seen.add(dst)
+
+            for staged in self._staged_dirs():
+                if not staged or not staged.startswith(DOCUMENTS_ROOT):
+                    continue
+                if staged == normalized:
+                    continue
+                if not _is_under(staged, normalized):
+                    continue
+                rel = (
+                    staged[len(normalized) :].lstrip("/")
+                    if normalized != "/"
+                    else staged.lstrip("/")
+                )
+                if not rel:
+                    continue
+                first = rel.split("/", 1)[0]
+                immediate = (
+                    normalized.rstrip("/") + "/" + first
+                    if normalized != "/"
+                    else "/" + first
+                )
+                subdir_paths.add(immediate)
+
+            for sub in sorted(subdir_paths):
+                if sub in seen:
+                    continue
+                infos.append(FileInfo(path=sub, is_dir=True, size=0, modified_at=""))
+                seen.add(sub)
+
+        for path_key, fd in files.items():
+            if not isinstance(path_key, str) or path_key in seen:
+                continue
+            if not _is_under(path_key, normalized) or path_key == normalized:
+                continue
+            if normalized == "/":
+                rel = path_key.lstrip("/")
+            else:
+                rel = path_key[len(normalized) :].lstrip("/")
+            if not rel:
+                continue
+            if "/" in rel:
+                first = rel.split("/", 1)[0]
+                immediate = (
+                    normalized.rstrip("/") + "/" + first
+                    if normalized != "/"
+                    else "/" + first
+                )
+                if immediate not in seen:
+                    infos.append(
+                        FileInfo(path=immediate, is_dir=True, size=0, modified_at="")
+                    )
+                    seen.add(immediate)
+                continue
+            include = path_key.startswith(DOCUMENTS_ROOT) or _basename(
+                path_key
+            ).startswith(_TEMP_PREFIX)
+            if not include:
+                continue
+            size = self._file_data_size(fd) if isinstance(fd, dict) else 0
+            infos.append(
+                FileInfo(
+                    path=path_key,
+                    is_dir=False,
+                    size=int(size),
+                    modified_at=fd.get("modified_at", "")
+                    if isinstance(fd, dict)
+                    else "",
+                )
+            )
+            seen.add(path_key)
+
+        infos.sort(key=lambda fi: (not fi.get("is_dir", False), fi.get("path", "")))
+        return infos
+
+    def ls_info(self, path: str) -> list[FileInfo]:  # type: ignore[override]
+        return asyncio.run(self.als_info(path))
+
+    async def _list_db_directory(
+        self,
+        session: AsyncSession,
+        normalized_path: str,
+    ) -> tuple[list[FileInfo], set[str]]:
+        """List immediate Folders + Documents at ``normalized_path``.
+
+        Returns ``(file_infos, subdirectory_paths)``. ``normalized_path`` may
+        be ``/`` (synthesizes ``/documents``) or a path under ``/documents``.
+        """
+        if normalized_path == "/":
+            return (
+                [],
+                {DOCUMENTS_ROOT},
+            )
+
+        if not normalized_path.startswith(DOCUMENTS_ROOT):
+            return [], set()
+
+        index = await build_path_index(session, self.search_space_id)
+        target_folder_id: int | None = None
+        if normalized_path != DOCUMENTS_ROOT:
+            target_path = normalized_path
+            matches = [
+                fid for fid, fpath in index.folder_paths.items() if fpath == target_path
+            ]
+            if not matches:
+                return [], set()
+            target_folder_id = matches[0]
+
+        result = await session.execute(
+            select(Document.id, Document.title, Document.folder_id, Document.updated_at)
+            .where(Document.search_space_id == self.search_space_id)
+            .where(
+                Document.folder_id == target_folder_id
+                if target_folder_id is not None
+                else Document.folder_id.is_(None)
+            )
+        )
+        rows = result.all()
+
+        file_infos: list[FileInfo] = []
+        for row in rows:
+            path = doc_to_virtual_path(
+                doc_id=row.id,
+                title=str(row.title or "untitled"),
+                folder_id=row.folder_id,
+                index=index,
+            )
+            modified = ""
+            if row.updated_at is not None:
+                with contextlib.suppress(Exception):
+                    modified = row.updated_at.astimezone(UTC).isoformat()
+            file_infos.append(
+                FileInfo(
+                    path=path,
+                    is_dir=False,
+                    size=0,
+                    modified_at=modified,
+                )
+            )
+
+        subdirs: set[str] = set()
+        for _fid, fpath in index.folder_paths.items():
+            if fpath == normalized_path:
+                continue
+            base = normalized_path.rstrip("/")
+            if not fpath.startswith(base + "/"):
+                continue
+            rel = fpath[len(base) + 1 :]
+            if "/" in rel:
+                continue
+            subdirs.add(base + "/" + rel)
+        return file_infos, subdirs
+
+    async def aread(  # type: ignore[override]
+        self,
+        file_path: str,
+        offset: int = 0,
+        limit: int = 2000,
+    ) -> str:
+        files = self._state_files()
+        file_data = files.get(file_path)
+        if file_data is not None:
+            return format_read_response(file_data, offset, limit)
+
+        loaded = await self._load_file_data(file_path)
+        if loaded is None:
+            return f"Error: File '{file_path}' not found"
+        file_data, _ = loaded
+        return format_read_response(file_data, offset, limit)
+
+    def read(self, file_path: str, offset: int = 0, limit: int = 2000) -> str:  # type: ignore[override]
+        return asyncio.run(self.aread(file_path, offset, limit))
+
+    async def _load_file_data(
+        self,
+        path: str,
+    ) -> tuple[dict[str, Any], int | None] | None:
+        """Lazy-load a virtual KB document into a deepagents ``FileData``.
+
+        Returns ``(file_data, doc_id)`` or ``None`` if the path doesn't map
+        to any known document. ``doc_id`` is ``None`` for the synthetic
+        anonymous document so the caller doesn't track it as a DB-backed file.
+        """
+        anon = self._kb_anon_doc()
+        if anon and str(anon.get("path") or "") == path:
+            doc_payload = {
+                "document_id": -1,
+                "chunks": list(anon.get("chunks") or []),
+                "matched_chunk_ids": [],
+                "document": {
+                    "id": -1,
+                    "title": anon.get("title") or "uploaded_document",
+                    "document_type": "FILE",
+                    "metadata": {"source": "anonymous_upload"},
+                },
+                "source": "FILE",
+            }
+            xml = build_document_xml(doc_payload, matched_chunk_ids=set())
+            file_data = create_file_data(xml)
+            return file_data, None
+
+        if not path.startswith(DOCUMENTS_ROOT):
+            return None
+
+        async with shielded_async_session() as session:
+            document = await virtual_path_to_doc(
+                session,
+                search_space_id=self.search_space_id,
+                virtual_path=path,
+            )
+            if document is None:
+                return None
+            chunk_rows = await session.execute(
+                select(Chunk.id, Chunk.content)
+                .where(Chunk.document_id == document.id)
+                .order_by(Chunk.id)
+            )
+            chunks = [
+                {"chunk_id": row.id, "content": row.content} for row in chunk_rows.all()
+            ]
+
+        doc_payload = {
+            "document_id": document.id,
+            "chunks": chunks,
+            "matched_chunk_ids": list(self._matched_chunk_ids(document.id)),
+            "document": {
+                "id": document.id,
+                "title": document.title,
+                "document_type": (
+                    document.document_type.value
+                    if getattr(document, "document_type", None) is not None
+                    else "UNKNOWN"
+                ),
+                "metadata": dict(document.document_metadata or {}),
+            },
+            "source": (
+                document.document_type.value
+                if getattr(document, "document_type", None) is not None
+                else "UNKNOWN"
+            ),
+        }
+        xml = build_document_xml(
+            doc_payload,
+            matched_chunk_ids=self._matched_chunk_ids(document.id),
+        )
+        file_data = create_file_data(xml)
+        return file_data, document.id
+
+    # ------------------------------------------------------------------ writes
+
+    async def awrite(self, file_path: str, content: str) -> WriteResult:  # type: ignore[override]
+        files = self._state_files()
+        if file_path in files:
+            return WriteResult(
+                error=(
+                    f"Cannot write to {file_path} because it already exists. "
+                    "Read and then make an edit, or write to a new path."
+                )
+            )
+        new_file_data = create_file_data(content)
+        return WriteResult(path=file_path, files_update={file_path: new_file_data})
+
+    def write(self, file_path: str, content: str) -> WriteResult:  # type: ignore[override]
+        return asyncio.run(self.awrite(file_path, content))
+
+    async def aedit(  # type: ignore[override]
+        self,
+        file_path: str,
+        old_string: str,
+        new_string: str,
+        replace_all: bool = False,
+    ) -> EditResult:
+        files = self._state_files()
+        file_data = files.get(file_path)
+        if file_data is None:
+            loaded = await self._load_file_data(file_path)
+            if loaded is None:
+                return EditResult(error=f"Error: File '{file_path}' not found")
+            file_data, _ = loaded
+
+        content = file_data_to_string(file_data)
+        result = perform_string_replacement(
+            content, old_string, new_string, replace_all
+        )
+        if isinstance(result, str):
+            return EditResult(error=result)
+
+        new_content, occurrences = result
+        new_file_data = update_file_data(file_data, new_content)
+        return EditResult(
+            path=file_path,
+            files_update={file_path: new_file_data},
+            occurrences=int(occurrences),
+        )
+
+    def edit(  # type: ignore[override]
+        self,
+        file_path: str,
+        old_string: str,
+        new_string: str,
+        replace_all: bool = False,
+    ) -> EditResult:
+        return asyncio.run(self.aedit(file_path, old_string, new_string, replace_all))
+
+    # ------------------------------------------------------------------ glob/grep
+
+    async def aglob_info(self, pattern: str, path: str = "/") -> list[FileInfo]:  # type: ignore[override]
+        normalized = self._normalize_listing_path(path)
+        results: list[FileInfo] = []
+        seen: set[str] = set()
+
+        files = self._state_files()
+        moved_removed, _ = self._moved_view_paths(files)
+        regex = re.compile(fnmatch.translate(pattern))
+        for path_key, fd in files.items():
+            if path_key in moved_removed:
+                continue
+            if not _is_under(path_key, normalized):
+                continue
+            rel = (
+                path_key[len(normalized) :].lstrip("/")
+                if normalized != "/"
+                else path_key.lstrip("/")
+            )
+            if not regex.match(rel) and not regex.match(path_key):
+                continue
+            if path_key in seen:
+                continue
+            size = self._file_data_size(fd) if isinstance(fd, dict) else 0
+            results.append(
+                FileInfo(
+                    path=path_key,
+                    is_dir=False,
+                    size=int(size),
+                    modified_at=fd.get("modified_at", "")
+                    if isinstance(fd, dict)
+                    else "",
+                )
+            )
+            seen.add(path_key)
+
+        if normalized.startswith(DOCUMENTS_ROOT) or normalized == "/":
+            try:
+                async with shielded_async_session() as session:
+                    index = await build_path_index(session, self.search_space_id)
+                    rows = await session.execute(
+                        select(Document.id, Document.title, Document.folder_id).where(
+                            Document.search_space_id == self.search_space_id
+                        )
+                    )
+                    for row in rows.all():
+                        candidate = doc_to_virtual_path(
+                            doc_id=row.id,
+                            title=str(row.title or "untitled"),
+                            folder_id=row.folder_id,
+                            index=index,
+                        )
+                        if candidate in seen or candidate in moved_removed:
+                            continue
+                        if not _is_under(candidate, normalized):
+                            continue
+                        rel = (
+                            candidate[len(normalized) :].lstrip("/")
+                            if normalized != "/"
+                            else candidate.lstrip("/")
+                        )
+                        if not regex.match(rel) and not regex.match(candidate):
+                            continue
+                        results.append(
+                            FileInfo(
+                                path=candidate, is_dir=False, size=0, modified_at=""
+                            )
+                        )
+                        seen.add(candidate)
+            except Exception as exc:  # pragma: no cover - defensive
+                logger.warning("KBPostgresBackend.aglob_info DB error: %s", exc)
+
+        results.sort(key=lambda fi: fi.get("path", ""))
+        return results
+
+    def glob_info(self, pattern: str, path: str = "/") -> list[FileInfo]:  # type: ignore[override]
+        return asyncio.run(self.aglob_info(pattern, path))
+
+    async def agrep_raw(  # type: ignore[override]
+        self,
+        pattern: str,
+        path: str | None = None,
+        glob: str | None = None,
+    ) -> list[GrepMatch] | str:
+        if not pattern:
+            return "Error: pattern cannot be empty"
+
+        normalized = self._normalize_listing_path(path or "/")
+        matches: list[GrepMatch] = []
+
+        files = self._state_files()
+        moved_removed, _ = self._moved_view_paths(files)
+        glob_re = re.compile(fnmatch.translate(glob)) if glob else None
+        for path_key, fd in files.items():
+            if path_key in moved_removed:
+                continue
+            if not _is_under(path_key, normalized):
+                continue
+            if glob_re is not None and not glob_re.match(_basename(path_key)):
+                continue
+            if not isinstance(fd, dict):
+                continue
+            for line_no, line in enumerate(fd.get("content") or [], 1):
+                if pattern in line:
+                    matches.append(
+                        GrepMatch(path=path_key, line=int(line_no), text=str(line))
+                    )
+                    if len(matches) >= _GREP_MAX_TOTAL_MATCHES:
+                        return matches
+
+        if normalized.startswith(DOCUMENTS_ROOT) or normalized == "/":
+            try:
+                async with shielded_async_session() as session:
+                    index = await build_path_index(session, self.search_space_id)
+                    sub = (
+                        select(Chunk.document_id, Chunk.id, Chunk.content)
+                        .join(Document, Document.id == Chunk.document_id)
+                        .where(Document.search_space_id == self.search_space_id)
+                        .where(Chunk.content.ilike(f"%{pattern}%"))
+                        .order_by(Chunk.document_id, Chunk.id)
+                    )
+                    chunk_rows = await session.execute(sub)
+                    per_doc: dict[int, int] = {}
+                    doc_id_to_path: dict[int, str] = {}
+                    needed_doc_ids: set[int] = set()
+                    chunk_buffer: list[tuple[int, int, str]] = []
+                    for row in chunk_rows.all():
+                        per_doc.setdefault(row.document_id, 0)
+                        if per_doc[row.document_id] >= _GREP_MAX_PER_DOC:
+                            continue
+                        per_doc[row.document_id] += 1
+                        chunk_buffer.append((row.document_id, row.id, row.content))
+                        needed_doc_ids.add(row.document_id)
+                        if sum(per_doc.values()) >= _GREP_MAX_TOTAL_MATCHES - len(
+                            matches
+                        ):
+                            break
+                    if needed_doc_ids:
+                        doc_rows = await session.execute(
+                            select(
+                                Document.id, Document.title, Document.folder_id
+                            ).where(Document.id.in_(list(needed_doc_ids)))
+                        )
+                        for row in doc_rows.all():
+                            doc_id_to_path[row.id] = doc_to_virtual_path(
+                                doc_id=row.id,
+                                title=str(row.title or "untitled"),
+                                folder_id=row.folder_id,
+                                index=index,
+                            )
+                    for doc_id, chunk_id, content in chunk_buffer:
+                        candidate = doc_id_to_path.get(doc_id)
+                        if not candidate or candidate in moved_removed:
+                            continue
+                        if not _is_under(candidate, normalized):
+                            continue
+                        if glob_re is not None and not glob_re.match(
+                            _basename(candidate)
+                        ):
+                            continue
+                        snippet = " ".join(str(content).split())[:240]
+                        matches.append(
+                            GrepMatch(
+                                path=candidate,
+                                line=0,
+                                text=(
+                                    f"<chunk-match in {candidate} chunk_id={chunk_id}>: "
+                                    f"{snippet}"
+                                ),
+                            )
+                        )
+                        if len(matches) >= _GREP_MAX_TOTAL_MATCHES:
+                            break
+            except Exception as exc:  # pragma: no cover - defensive
+                logger.warning("KBPostgresBackend.agrep_raw DB error: %s", exc)
+
+        return matches
+
+    def grep_raw(  # type: ignore[override]
+        self,
+        pattern: str,
+        path: str | None = None,
+        glob: str | None = None,
+    ) -> list[GrepMatch] | str:
+        return asyncio.run(self.agrep_raw(pattern, path, glob))
+
+    # ------------------------------------------------------------------ list_tree (helper)
+
+    async def alist_tree_listing(
+        self,
+        path: str = DOCUMENTS_ROOT,
+        *,
+        max_depth: int | None = 8,
+        page_size: int = 500,
+        include_files: bool = True,
+        include_dirs: bool = True,
+    ) -> dict[str, Any]:
+        """Recursive tree listing for cloud mode.
+
+        Mirrors the shape returned by :class:`MultiRootLocalFolderBackend.list_tree`:
+        ``{"entries": [{path, is_dir, size, modified_at, depth}, ...], "truncated": bool}``.
+        """
+        normalized = self._normalize_listing_path(path or DOCUMENTS_ROOT)
+        if not normalized.startswith(DOCUMENTS_ROOT) and normalized != "/":
+            return {"error": "Error: path must be under /documents/"}
+
+        entries: list[dict[str, Any]] = []
+        truncated = False
+
+        try:
+            async with shielded_async_session() as session:
+                index = await build_path_index(session, self.search_space_id)
+                doc_rows_raw = await session.execute(
+                    select(
+                        Document.id,
+                        Document.title,
+                        Document.folder_id,
+                        Document.updated_at,
+                    ).where(Document.search_space_id == self.search_space_id)
+                )
+                doc_rows = list(doc_rows_raw.all())
+        except Exception as exc:  # pragma: no cover
+            logger.warning("KBPostgresBackend.alist_tree_listing DB error: %s", exc)
+            return {"entries": [], "truncated": False}
+
+        files = self._state_files()
+        moved_removed, _ = self._moved_view_paths(files)
+        anon = self._kb_anon_doc()
+        anon_path = str(anon.get("path") or "") if anon else ""
+
+        def _depth_of(p: str) -> int:
+            if p == DOCUMENTS_ROOT:
+                return 0
+            rel_root = (
+                p[len(DOCUMENTS_ROOT) :].lstrip("/")
+                if normalized.startswith(DOCUMENTS_ROOT)
+                else p.lstrip("/")
+            )
+            return len([part for part in rel_root.split("/") if part])
+
+        def _add_entry(entry: dict[str, Any]) -> bool:
+            nonlocal truncated
+            if len(entries) >= page_size:
+                truncated = True
+                return False
+            entries.append(entry)
+            return True
+
+        if include_dirs:
+            for _fid, fpath in sorted(index.folder_paths.items(), key=lambda kv: kv[1]):
+                if not _is_under(fpath, normalized):
+                    continue
+                depth = _depth_of(fpath)
+                if max_depth is not None and depth > max_depth:
+                    continue
+                if not _add_entry(
+                    {
+                        "path": fpath,
+                        "is_dir": True,
+                        "size": 0,
+                        "modified_at": "",
+                        "depth": depth,
+                    }
+                ):
+                    return {"entries": entries, "truncated": True}
+            for staged in self._staged_dirs():
+                if not _is_under(staged, normalized):
+                    continue
+                depth = _depth_of(staged)
+                if max_depth is not None and depth > max_depth:
+                    continue
+                if any(e["path"] == staged for e in entries):
+                    continue
+                if not _add_entry(
+                    {
+                        "path": staged,
+                        "is_dir": True,
+                        "size": 0,
+                        "modified_at": "",
+                        "depth": depth,
+                    }
+                ):
+                    return {"entries": entries, "truncated": True}
+
+        if include_files:
+            for row in sorted(doc_rows, key=lambda r: str(r.title or "")):
+                candidate = doc_to_virtual_path(
+                    doc_id=row.id,
+                    title=str(row.title or "untitled"),
+                    folder_id=row.folder_id,
+                    index=index,
+                )
+                if candidate in moved_removed:
+                    continue
+                if not _is_under(candidate, normalized):
+                    continue
+                depth = _depth_of(candidate)
+                if max_depth is not None and depth > max_depth:
+                    continue
+                modified = ""
+                if row.updated_at is not None:
+                    with contextlib.suppress(Exception):
+                        modified = row.updated_at.astimezone(UTC).isoformat()
+                if not _add_entry(
+                    {
+                        "path": candidate,
+                        "is_dir": False,
+                        "size": 0,
+                        "modified_at": modified,
+                        "depth": depth,
+                    }
+                ):
+                    return {"entries": entries, "truncated": True}
+
+            if anon_path and _is_under(anon_path, normalized):
+                depth = _depth_of(anon_path)
+                if (max_depth is None or depth <= max_depth) and not _add_entry(
+                    {
+                        "path": anon_path,
+                        "is_dir": False,
+                        "size": len(str(anon.get("content") or "")),
+                        "modified_at": "",
+                        "depth": depth,
+                    }
+                ):
+                    return {"entries": entries, "truncated": True}
+
+            for path_key, fd in files.items():
+                if not isinstance(path_key, str):
+                    continue
+                if not _is_under(path_key, normalized):
+                    continue
+                if any(e["path"] == path_key for e in entries):
+                    continue
+                if not (
+                    path_key.startswith(DOCUMENTS_ROOT)
+                    or _basename(path_key).startswith(_TEMP_PREFIX)
+                ):
+                    continue
+                depth = _depth_of(path_key)
+                if max_depth is not None and depth > max_depth:
+                    continue
+                size = self._file_data_size(fd) if isinstance(fd, dict) else 0
+                if not _add_entry(
+                    {
+                        "path": path_key,
+                        "is_dir": False,
+                        "size": int(size),
+                        "modified_at": fd.get("modified_at", "")
+                        if isinstance(fd, dict)
+                        else "",
+                        "depth": depth,
+                    }
+                ):
+                    return {"entries": entries, "truncated": True}
+
+        return {"entries": entries, "truncated": truncated}
+
+    # ------------------------------------------------------------------ uploads (unsupported)
+
+    def upload_files(  # type: ignore[override]
+        self, files: list[tuple[str, bytes]]
+    ) -> list[FileUploadResponse]:
+        msg = "KBPostgresBackend does not support upload_files."
+        raise NotImplementedError(msg)
+
+    def download_files(  # type: ignore[override]
+        self, paths: list[str]
+    ) -> list[FileDownloadResponse]:
+        responses: list[FileDownloadResponse] = []
+        files = self._state_files()
+        for path in paths:
+            fd = files.get(path)
+            if fd is None:
+                responses.append(
+                    FileDownloadResponse(
+                        path=path, content=None, error="file_not_found"
+                    )
+                )
+                continue
+            content_str = file_data_to_string(fd)
+            responses.append(
+                FileDownloadResponse(
+                    path=path,
+                    content=content_str.encode("utf-8"),
+                    error=None,
+                )
+            )
+        return responses
+
+
+# --- module-level small helpers ---------------------------------------------
+
+
+async def list_tree_listing(
+    backend: KBPostgresBackend,
+    path: str,
+    *,
+    max_depth: int | None = 8,
+    page_size: int = 500,
+    include_files: bool = True,
+    include_dirs: bool = True,
+) -> dict[str, Any]:
+    """Async helper used by the overridden ``list_tree`` tool wrapper."""
+    return await backend.alist_tree_listing(
+        path,
+        max_depth=max_depth,
+        page_size=page_size,
+        include_files=include_files,
+        include_dirs=include_dirs,
+    )
+
+
+__all__ = [
+    "KBPostgresBackend",
+    "list_tree_listing",
+    "paginate_listing",
+]

surfsense_backend/app/agents/new_chat/middleware/knowledge_search.py

@@ -1,10 +1,24 @@
-"""Knowledge-base pre-search middleware for the SurfSense new chat agent.
+"""Hybrid-search priority middleware for the SurfSense new chat agent.
 
-This middleware runs before the main agent loop and seeds a virtual filesystem
+This middleware runs ``before_agent`` on every turn and writes:
-(`files` state) with relevant documents retrieved via hybrid search.  On each
+
-turn the filesystem is *expanded* — new results merge with documents loaded
+* ``state["kb_priority"]`` — the top-K most relevant documents for the
-during prior turns — and a synthetic ``ls`` result is injected into the message
+  current user message, used to render a ``<priority_documents>`` system
-history so the LLM is immediately aware of the current filesystem structure.
+  message immediately before the user turn.
+* ``state["kb_matched_chunk_ids"]`` — internal hand-off mapping
+  (``Document.id`` → matched chunk IDs) consumed by
+  :class:`KBPostgresBackend._load_file_data` when the agent first reads each
+  document, so the XML wrapper can flag matched sections in
+  ``<chunk_index>``.
+
+The previous "scoped filesystem" behaviour (synthetic ``ls`` + state
+``files`` seeding) is intentionally removed: documents are now lazy-loaded
+from Postgres on demand, with the full workspace tree rendered separately
+by :class:`KnowledgeTreeMiddleware`.
+
+In anonymous mode the middleware skips hybrid search entirely and emits a
+single-entry priority list pointing at the Redis-loaded document
+(``state["kb_anon_doc"]``).
 """
 
 from __future__ import annotations
@@ -13,27 +27,33 @@ import asyncio
 import json
 import logging
 import re
-import uuid
 from collections.abc import Sequence
 from datetime import UTC, datetime
 from typing import Any
 
+from langchain.agents import create_agent
 from langchain.agents.middleware import AgentMiddleware, AgentState
 from langchain_core.language_models import BaseChatModel
-from langchain_core.messages import AIMessage, BaseMessage, HumanMessage, ToolMessage
+from langchain_core.messages import AIMessage, BaseMessage, HumanMessage, SystemMessage
+from langchain_core.runnables import Runnable
 from langgraph.runtime import Runtime
 from litellm import token_counter
 from pydantic import BaseModel, Field, ValidationError
 from sqlalchemy import select
-from sqlalchemy.ext.asyncio import AsyncSession
 
-from app.agents.new_chat.utils import parse_date_or_datetime, resolve_date_range
+from app.agents.new_chat.feature_flags import get_flags
 from app.agents.new_chat.filesystem_selection import FilesystemMode
+from app.agents.new_chat.filesystem_state import SurfSenseFilesystemState
+from app.agents.new_chat.path_resolver import (
+    PathIndex,
+    build_path_index,
+    doc_to_virtual_path,
+)
+from app.agents.new_chat.utils import parse_date_or_datetime, resolve_date_range
 from app.db import (
     NATIVE_TO_LEGACY_DOCTYPE,
     Chunk,
     Document,
-    Folder,
     shielded_async_session,
 )
 from app.retriever.chunks_hybrid_search import ChucksHybridSearchRetriever
@@ -70,7 +90,6 @@ class KBSearchPlan(BaseModel):
 
 
 def _extract_text_from_message(message: BaseMessage) -> str:
-    """Extract plain text from a message content."""
     content = getattr(message, "content", "")
     if isinstance(content, str):
         return content
@@ -85,19 +104,6 @@ def _extract_text_from_message(message: BaseMessage) -> str:
     return str(content)
 
 
-def _safe_filename(value: str, *, fallback: str = "untitled.xml") -> str:
-    """Convert arbitrary text into a filesystem-safe filename."""
-    name = re.sub(r"[\\/:*?\"<>|]+", "_", value).strip()
-    name = re.sub(r"\s+", " ", name)
-    if not name:
-        name = fallback
-    if len(name) > 180:
-        name = name[:180].rstrip()
-    if not name.lower().endswith(".xml"):
-        name = f"{name}.xml"
-    return name
-
-
 def _render_recent_conversation(
     messages: Sequence[BaseMessage],
     *,
@@ -107,10 +113,9 @@ def _render_recent_conversation(
 ) -> str:
     """Render recent dialogue for internal planning under a token budget.
 
-    Prefers the latest messages and uses the project's existing model-aware
+    Filters to ``HumanMessage`` and ``AIMessage`` (without tool_calls) so that
-    token budgeting hooks when available on the LLM (`_count_tokens`,
+    injected ``SystemMessage`` artefacts (priority list, workspace tree,
-    `_get_max_input_tokens`). Falls back to the prior fixed-message heuristic
+    file-write contract) don't pollute the planner prompt.
-    if token counting is unavailable.
     """
     rendered: list[tuple[str, str]] = []
     for message in messages:
@@ -133,8 +138,6 @@ def _render_recent_conversation(
     if not rendered:
         return ""
 
-    # Exclude the latest user message from "recent conversation" because it is
-    # already passed separately as "Latest user message" in the planner prompt.
     if rendered and rendered[-1][0] == "user" and rendered[-1][1] == user_text.strip():
         rendered = rendered[:-1]
 
@@ -216,8 +219,6 @@ def _render_recent_conversation(
             selected_lines = candidate_lines
             continue
 
-        # If the full message does not fit, keep as much of this most-recent
-        # older message as possible via binary search.
         lo, hi = 1, len(text)
         best_line: str | None = None
         while lo <= hi:
@@ -249,7 +250,6 @@ def _build_kb_planner_prompt(
     recent_conversation: str,
     user_text: str,
 ) -> str:
-    """Build a compact internal prompt for KB query rewriting and date scoping."""
     today = datetime.now(UTC).date().isoformat()
     return (
         "You optimize internal knowledge-base search inputs for document retrieval.\n"
@@ -275,12 +275,10 @@ def _build_kb_planner_prompt(
 
 
 def _extract_json_payload(text: str) -> str:
-    """Extract a JSON object from a raw LLM response."""
     stripped = text.strip()
     fenced = re.search(r"```(?:json)?\s*(\{.*?\})\s*```", stripped, re.DOTALL)
     if fenced:
         return fenced.group(1)
-
     start = stripped.find("{")
     end = stripped.rfind("}")
     if start != -1 and end != -1 and end > start:
@@ -289,7 +287,6 @@ def _extract_json_payload(text: str) -> str:
 
 
 def _parse_kb_search_plan_response(response_text: str) -> KBSearchPlan:
-    """Parse and validate the planner's JSON response."""
     payload = json.loads(_extract_json_payload(response_text))
     return KBSearchPlan.model_validate(payload)
 
@@ -298,212 +295,19 @@ def _normalize_optional_date_range(
     start_date: str | None,
     end_date: str | None,
 ) -> tuple[datetime | None, datetime | None]:
-    """Normalize optional planner dates into a UTC datetime range."""
     parsed_start = parse_date_or_datetime(start_date) if start_date else None
     parsed_end = parse_date_or_datetime(end_date) if end_date else None
 
     if parsed_start is None and parsed_end is None:
         return None, None
 
-    resolved_start, resolved_end = resolve_date_range(parsed_start, parsed_end)
+    return resolve_date_range(parsed_start, parsed_end)
-    return resolved_start, resolved_end
-
-
-def _build_document_xml(
-    document: dict[str, Any],
-    matched_chunk_ids: set[int] | None = None,
-) -> str:
-    """Build citation-friendly XML with a ``<chunk_index>`` for smart seeking.
-
-    The ``<chunk_index>`` at the top of each document lists every chunk with its
-    line range inside ``<document_content>`` and flags chunks that directly
-    matched the search query (``matched="true"``).  This lets the LLM jump
-    straight to the most relevant section via ``read_file(offset=…, limit=…)``
-    instead of reading sequentially from the start.
-    """
-    matched = matched_chunk_ids or set()
-
-    doc_meta = document.get("document") or {}
-    metadata = (doc_meta.get("metadata") or {}) if isinstance(doc_meta, dict) else {}
-    document_id = doc_meta.get("id", document.get("document_id", "unknown"))
-    document_type = doc_meta.get("document_type", document.get("source", "UNKNOWN"))
-    title = doc_meta.get("title") or metadata.get("title") or "Untitled Document"
-    url = (
-        metadata.get("url") or metadata.get("source") or metadata.get("page_url") or ""
-    )
-    metadata_json = json.dumps(metadata, ensure_ascii=False)
-
-    # --- 1. Metadata header (fixed structure) ---
-    metadata_lines: list[str] = [
-        "<document>",
-        "<document_metadata>",
-        f"  <document_id>{document_id}</document_id>",
-        f"  <document_type>{document_type}</document_type>",
-        f"  <title><![CDATA[{title}]]></title>",
-        f"  <url><![CDATA[{url}]]></url>",
-        f"  <metadata_json><![CDATA[{metadata_json}]]></metadata_json>",
-        "</document_metadata>",
-        "",
-    ]
-
-    # --- 2. Pre-build chunk XML strings to compute line counts ---
-    chunks = document.get("chunks") or []
-    chunk_entries: list[tuple[int | None, str]] = []  # (chunk_id, xml_string)
-    if isinstance(chunks, list):
-        for chunk in chunks:
-            if not isinstance(chunk, dict):
-                continue
-            chunk_id = chunk.get("chunk_id") or chunk.get("id")
-            chunk_content = str(chunk.get("content", "")).strip()
-            if not chunk_content:
-                continue
-            if chunk_id is None:
-                xml = f"  <chunk><![CDATA[{chunk_content}]]></chunk>"
-            else:
-                xml = f"  <chunk id='{chunk_id}'><![CDATA[{chunk_content}]]></chunk>"
-            chunk_entries.append((chunk_id, xml))
-
-    # --- 3. Compute line numbers for every chunk ---
-    # Layout (1-indexed lines for read_file):
-    #   metadata_lines          -> len(metadata_lines) lines
-    #   <chunk_index>           -> 1 line
-    #   index entries           -> len(chunk_entries) lines
-    #   </chunk_index>          -> 1 line
-    #   (empty line)            -> 1 line
-    #   <document_content>      -> 1 line
-    #   chunk xml lines…
-    #   </document_content>     -> 1 line
-    #   </document>             -> 1 line
-    index_overhead = (
-        1 + len(chunk_entries) + 1 + 1 + 1
-    )  # tags + empty + <document_content>
-    first_chunk_line = len(metadata_lines) + index_overhead + 1  # 1-indexed
-
-    current_line = first_chunk_line
-    index_entry_lines: list[str] = []
-    for cid, xml_str in chunk_entries:
-        num_lines = xml_str.count("\n") + 1
-        end_line = current_line + num_lines - 1
-        matched_attr = ' matched="true"' if cid is not None and cid in matched else ""
-        if cid is not None:
-            index_entry_lines.append(
-                f'  <entry chunk_id="{cid}" lines="{current_line}-{end_line}"{matched_attr}/>'
-            )
-        else:
-            index_entry_lines.append(
-                f'  <entry lines="{current_line}-{end_line}"{matched_attr}/>'
-            )
-        current_line = end_line + 1
-
-    # --- 4. Assemble final XML ---
-    lines = metadata_lines.copy()
-    lines.append("<chunk_index>")
-    lines.extend(index_entry_lines)
-    lines.append("</chunk_index>")
-    lines.append("")
-    lines.append("<document_content>")
-    for _, xml_str in chunk_entries:
-        lines.append(xml_str)
-    lines.extend(["</document_content>", "</document>"])
-    return "\n".join(lines)
-
-
-async def _get_folder_paths(
-    session: AsyncSession, search_space_id: int
-) -> dict[int, str]:
-    """Return a map of folder_id -> virtual folder path under /documents."""
-    result = await session.execute(
-        select(Folder.id, Folder.name, Folder.parent_id).where(
-            Folder.search_space_id == search_space_id
-        )
-    )
-    rows = result.all()
-    by_id = {row.id: {"name": row.name, "parent_id": row.parent_id} for row in rows}
-
-    cache: dict[int, str] = {}
-
-    def resolve_path(folder_id: int) -> str:
-        if folder_id in cache:
-            return cache[folder_id]
-        parts: list[str] = []
-        cursor: int | None = folder_id
-        visited: set[int] = set()
-        while cursor is not None and cursor in by_id and cursor not in visited:
-            visited.add(cursor)
-            entry = by_id[cursor]
-            parts.append(
-                _safe_filename(str(entry["name"]), fallback="folder").removesuffix(
-                    ".xml"
-                )
-            )
-            cursor = entry["parent_id"]
-        parts.reverse()
-        path = "/documents/" + "/".join(parts) if parts else "/documents"
-        cache[folder_id] = path
-        return path
-
-    for folder_id in by_id:
-        resolve_path(folder_id)
-    return cache
-
-
-def _build_synthetic_ls(
-    existing_files: dict[str, Any] | None,
-    new_files: dict[str, Any],
-    *,
-    mentioned_paths: set[str] | None = None,
-) -> tuple[AIMessage, ToolMessage]:
-    """Build a synthetic ls("/documents") tool-call + result for the LLM context.
-
-    Mentioned files are listed first.  A separate header tells the LLM which
-    files the user explicitly selected; the path list itself stays clean so
-    paths can be passed directly to ``read_file`` without stripping tags.
-    """
-    _mentioned = mentioned_paths or set()
-    merged: dict[str, Any] = {**(existing_files or {}), **new_files}
-    doc_paths = [
-        p for p, v in merged.items() if p.startswith("/documents/") and v is not None
-    ]
-
-    new_set = set(new_files)
-    mentioned_list = [p for p in doc_paths if p in _mentioned]
-    new_non_mentioned = [p for p in doc_paths if p in new_set and p not in _mentioned]
-    old_paths = [p for p in doc_paths if p not in new_set]
-    ordered = mentioned_list + new_non_mentioned + old_paths
-
-    parts: list[str] = []
-    if mentioned_list:
-        parts.append(
-            "USER-MENTIONED documents (read these thoroughly before answering):"
-        )
-        for p in mentioned_list:
-            parts.append(f"  {p}")
-        parts.append("")
-    parts.append(str(ordered) if ordered else "No documents found.")
-
-    tool_call_id = f"auto_ls_{uuid.uuid4().hex[:12]}"
-    ai_msg = AIMessage(
-        content="",
-        tool_calls=[{"name": "ls", "args": {"path": "/documents"}, "id": tool_call_id}],
-    )
-    tool_msg = ToolMessage(
-        content="\n".join(parts),
-        tool_call_id=tool_call_id,
-    )
-    return ai_msg, tool_msg
 
 
 def _resolve_search_types(
     available_connectors: list[str] | None,
     available_document_types: list[str] | None,
 ) -> list[str] | None:
-    """Build a flat list of document-type strings for the chunk retriever.
-
-    Includes legacy equivalents from ``NATIVE_TO_LEGACY_DOCTYPE`` so that
-    old documents indexed under Composio names are still found.
-
-    Returns ``None`` when no filtering is desired (search all types).
-    """
     types: set[str] = set()
     if available_document_types:
         types.update(available_document_types)
@@ -531,13 +335,8 @@ async def browse_recent_documents(
     start_date: datetime | None = None,
     end_date: datetime | None = None,
 ) -> list[dict[str, Any]]:
-    """Return documents ordered by recency (newest first), no relevance ranking.
+    """Return documents ordered by recency (newest first), no relevance ranking."""
-
+    from sqlalchemy import func
-    Used when the user's intent is temporal ("latest file", "most recent upload")
-    and hybrid search would produce poor results because the query has no
-    meaningful topical signal.
-    """
-    from sqlalchemy import func, select
 
     from app.db import DocumentType
 
@@ -581,7 +380,6 @@ async def browse_recent_documents(
             return []
 
         doc_ids = [d.id for d in documents]
-
         numbered = (
             select(
                 Chunk.id.label("chunk_id"),
@@ -632,6 +430,7 @@ async def browse_recent_documents(
                         else None
                     ),
                     "metadata": metadata,
+                    "folder_id": getattr(doc, "folder_id", None),
                 },
                 "source": (
                     doc.document_type.value
@@ -640,12 +439,6 @@ async def browse_recent_documents(
                 ),
             }
         )
-
-    logger.info(
-        "browse_recent_documents: %d docs returned for space=%d",
-        len(results),
-        search_space_id,
-    )
     return results
 
 
@@ -659,17 +452,11 @@ async def search_knowledge_base(
     start_date: datetime | None = None,
     end_date: datetime | None = None,
 ) -> list[dict[str, Any]]:
-    """Run a single unified hybrid search against the knowledge base.
+    """Run a single unified hybrid search against the knowledge base."""
-
-    Uses one ``ChucksHybridSearchRetriever`` call across all document types
-    instead of fanning out per-connector.  This reduces the number of DB
-    queries from ~10 to 2 (one RRF query + one chunk fetch).
-    """
     if not query:
         return []
 
     [embedding] = embed_texts([query])
-
     doc_types = _resolve_search_types(available_connectors, available_document_types)
     retriever_top_k = min(top_k * 3, 30)
 
@@ -693,14 +480,7 @@ async def fetch_mentioned_documents(
     document_ids: list[int],
     search_space_id: int,
 ) -> list[dict[str, Any]]:
-    """Fetch explicitly mentioned documents with *all* their chunks.
+    """Fetch explicitly mentioned documents."""
-
-    Returns the same dict structure as ``search_knowledge_base`` so results
-    can be merged directly into ``build_scoped_filesystem``.  Unlike search
-    results, every chunk is included (no top-K limiting) and none are marked
-    as ``matched`` since the entire document is relevant by virtue of the
-    user's explicit mention.
-    """
     if not document_ids:
         return []
 
@@ -750,6 +530,7 @@ async def fetch_mentioned_documents(
                         else None
                     ),
                     "metadata": metadata,
+                    "folder_id": getattr(doc, "folder_id", None),
                 },
                 "source": (
                     doc.document_type.value
@@ -762,96 +543,36 @@ async def fetch_mentioned_documents(
     return results
 
 
-async def build_scoped_filesystem(
+def _render_priority_message(priority: list[dict[str, Any]]) -> SystemMessage:
-    *,
+    """Render the priority list as a single ``<priority_documents>`` system message."""
-    documents: Sequence[dict[str, Any]],
+    if not priority:
-    search_space_id: int,
+        body = "(no priority documents for this turn)"
-) -> tuple[dict[str, dict[str, str]], dict[int, str]]:
+    else:
-    """Build a StateBackend-compatible files dict from search results.
+        lines: list[str] = []
-
+        for entry in priority:
-    Returns ``(files, doc_id_to_path)`` so callers can reliably map a
+            score = entry.get("score")
-    document id back to its filesystem path without guessing by title.
+            mentioned = entry.get("mentioned")
-    Paths are collision-proof: when two documents resolve to the same
+            score_str = f"{score:.3f}" if isinstance(score, int | float) else "n/a"
-    path the doc-id is appended to disambiguate.
+            mark = " [USER-MENTIONED]" if mentioned else ""
-    """
+            lines.append(f"- {entry.get('path', '')} (score={score_str}){mark}")
-    async with shielded_async_session() as session:
+        body = "\n".join(lines)
-        folder_paths = await _get_folder_paths(session, search_space_id)
+    return SystemMessage(
-        doc_ids = [
+        content=(
-            (doc.get("document") or {}).get("id")
+            "<priority_documents>\n"
-            for doc in documents
+            "These documents are most relevant to the latest user message; "
-            if isinstance(doc, dict)
+            "read them first. Matched sections are flagged inside each "
-        ]
+            "document's <chunk_index>.\n"
-        doc_ids = [doc_id for doc_id in doc_ids if isinstance(doc_id, int)]
+            f"{body}\n"
-        folder_by_doc_id: dict[int, int | None] = {}
+            "</priority_documents>"
-        if doc_ids:
+        )
-            doc_rows = await session.execute(
+    )
-                select(Document.id, Document.folder_id).where(
-                    Document.search_space_id == search_space_id,
-                    Document.id.in_(doc_ids),
-                )
-            )
-            folder_by_doc_id = {
-                row.id: row.folder_id for row in doc_rows.all() if row.id is not None
-            }
-
-    files: dict[str, dict[str, str]] = {}
-    doc_id_to_path: dict[int, str] = {}
-    for document in documents:
-        doc_meta = document.get("document") or {}
-        title = str(doc_meta.get("title") or "untitled")
-        doc_id = doc_meta.get("id")
-        folder_id = folder_by_doc_id.get(doc_id) if isinstance(doc_id, int) else None
-        base_folder = folder_paths.get(folder_id, "/documents")
-        file_name = _safe_filename(title)
-        path = f"{base_folder}/{file_name}"
-        if path in files:
-            stem = file_name.removesuffix(".xml")
-            path = f"{base_folder}/{stem} ({doc_id}).xml"
-        matched_ids = set(document.get("matched_chunk_ids") or [])
-        xml_content = _build_document_xml(document, matched_chunk_ids=matched_ids)
-        files[path] = {
-            "content": xml_content.split("\n"),
-            "encoding": "utf-8",
-            "created_at": "",
-            "modified_at": "",
-        }
-        if isinstance(doc_id, int):
-            doc_id_to_path[doc_id] = path
-    return files, doc_id_to_path
 
 
-def _build_anon_scoped_filesystem(
+class KnowledgePriorityMiddleware(AgentMiddleware):  # type: ignore[type-arg]
-    documents: Sequence[dict[str, Any]],
+    """Compute hybrid-search priority hints for the current turn."""
-) -> dict[str, dict[str, str]]:
-    """Build a scoped filesystem for anonymous documents without DB queries.
-
-    Anonymous uploads have no folders, so all files go under /documents.
-    """
-    files: dict[str, dict[str, str]] = {}
-    for document in documents:
-        doc_meta = document.get("document") or {}
-        title = str(doc_meta.get("title") or "untitled")
-        file_name = _safe_filename(title)
-        path = f"/documents/{file_name}"
-        if path in files:
-            doc_id = doc_meta.get("id", "dup")
-            stem = file_name.removesuffix(".xml")
-            path = f"/documents/{stem} ({doc_id}).xml"
-        matched_ids = set(document.get("matched_chunk_ids") or [])
-        xml_content = _build_document_xml(document, matched_chunk_ids=matched_ids)
-        files[path] = {
-            "content": xml_content.split("\n"),
-            "encoding": "utf-8",
-            "created_at": "",
-            "modified_at": "",
-        }
-    return files
-
-
-class KnowledgeBaseSearchMiddleware(AgentMiddleware):  # type: ignore[type-arg]
-    """Pre-agent middleware that always searches the KB and seeds a scoped filesystem."""
 
     tools = ()
+    state_schema = SurfSenseFilesystemState
 
     def __init__(
         self,
@@ -863,7 +584,6 @@ class KnowledgeBaseSearchMiddleware(AgentMiddleware):  # type: ignore[type-arg]
         available_document_types: list[str] | None = None,
         top_k: int = 10,
         mentioned_document_ids: list[int] | None = None,
-        anon_session_id: str | None = None,
     ) -> None:
         self.llm = llm
         self.search_space_id = search_space_id
@@ -872,7 +592,51 @@ class KnowledgeBaseSearchMiddleware(AgentMiddleware):  # type: ignore[type-arg]
         self.available_document_types = available_document_types
         self.top_k = top_k
         self.mentioned_document_ids = mentioned_document_ids or []
-        self.anon_session_id = anon_session_id
+        # Build the kb-planner private Runnable ONCE here so we don't pay
+        # the ``create_agent`` compile cost (50-200ms) on every turn.
+        # Disabled by default behind ``enable_kb_planner_runnable``; when
+        # off the planner falls back to the legacy ``self.llm.ainvoke``
+        # path.
+        self._planner: Runnable | None = None
+        self._planner_compile_failed = False
+
+    def _build_kb_planner_runnable(self) -> Runnable | None:
+        """Compile the kb-planner private :class:`Runnable` once.
+
+        Returns ``None`` when the feature flag is disabled, when the LLM is
+        unavailable, or when ``create_agent`` raises (we fall back to the
+        legacy ``self.llm.ainvoke`` path in that case). Compilation happens
+        lazily on first call, then memoized via ``self._planner``.
+
+        The compiled agent is constructed without tools — the planner's
+        contract is "answer with structured JSON" — but it inherits the
+        :class:`RetryAfterMiddleware` so transient rate-limit errors
+        from the planner LLM call don't fail the whole turn.
+        """
+        if self._planner is not None or self._planner_compile_failed:
+            return self._planner
+        if self.llm is None:
+            return None
+        flags = get_flags()
+        if not flags.enable_kb_planner_runnable or flags.disable_new_agent_stack:
+            return None
+
+        from app.agents.new_chat.middleware.retry_after import RetryAfterMiddleware
+
+        try:
+            self._planner = create_agent(
+                self.llm,
+                tools=[],
+                middleware=[RetryAfterMiddleware(max_retries=2)],
+            )
+        except Exception as exc:  # pragma: no cover - defensive
+            logger.warning(
+                "kb-planner Runnable compile failed; falling back to llm.ainvoke: %s",
+                exc,
+            )
+            self._planner_compile_failed = True
+            self._planner = None
+        return self._planner
 
     async def _plan_search_inputs(
         self,
@@ -880,10 +644,6 @@ class KnowledgeBaseSearchMiddleware(AgentMiddleware):  # type: ignore[type-arg]
         messages: Sequence[BaseMessage],
         user_text: str,
     ) -> tuple[str, datetime | None, datetime | None, bool]:
-        """Rewrite the KB query and infer optional date filters with the LLM.
-
-        Returns (optimized_query, start_date, end_date, is_recency_query).
-        """
         if self.llm is None:
             return user_text, None, None, False
 
@@ -899,11 +659,32 @@ class KnowledgeBaseSearchMiddleware(AgentMiddleware):  # type: ignore[type-arg]
         loop = asyncio.get_running_loop()
         t0 = loop.time()
 
+        # Prefer the compiled-once planner Runnable when enabled; otherwise
+        # fall back to ``self.llm.ainvoke``. The ``surfsense:internal`` tag
+        # is preserved on both paths so ``_stream_agent_events`` still
+        # suppresses the planner's intermediate events from the UI.
+        planner = self._build_kb_planner_runnable()
         try:
-            response = await self.llm.ainvoke(
+            if planner is not None:
-                [HumanMessage(content=prompt)],
+                planner_state = await planner.ainvoke(
-                config={"tags": ["surfsense:internal"]},
+                    {"messages": [HumanMessage(content=prompt)]},
-            )
+                    config={"tags": ["surfsense:internal"]},
+                )
+                response_messages = (
+                    planner_state.get("messages", [])
+                    if isinstance(planner_state, dict)
+                    else []
+                )
+                response = (
+                    response_messages[-1]
+                    if response_messages
+                    else AIMessage(content="")
+                )
+            else:
+                response = await self.llm.ainvoke(
+                    [HumanMessage(content=prompt)],
+                    config={"tags": ["surfsense:internal"]},
+                )
             plan = _parse_kb_search_plan_response(_extract_text_from_message(response))
             optimized_query = (
                 re.sub(r"\s+", " ", plan.optimized_query).strip() or user_text
@@ -914,7 +695,7 @@ class KnowledgeBaseSearchMiddleware(AgentMiddleware):  # type: ignore[type-arg]
             )
             is_recency = plan.is_recency_query
             _perf_log.info(
-                "[kb_fs_middleware] planner in %.3fs query=%r optimized=%r "
+                "[kb_priority] planner in %.3fs query=%r optimized=%r "
                 "start=%s end=%s recency=%s",
                 loop.time() - t0,
                 user_text[:80],
@@ -946,106 +727,68 @@ class KnowledgeBaseSearchMiddleware(AgentMiddleware):  # type: ignore[type-arg]
             pass
         return asyncio.run(self.abefore_agent(state, runtime))
 
-    async def _load_anon_document(self) -> dict[str, Any] | None:
-        """Load the anonymous user's uploaded document from Redis."""
-        if not self.anon_session_id:
-            return None
-        try:
-            import redis.asyncio as aioredis
-
-            from app.config import config
-
-            redis_client = aioredis.from_url(
-                config.REDIS_APP_URL, decode_responses=True
-            )
-            try:
-                redis_key = f"anon:doc:{self.anon_session_id}"
-                data = await redis_client.get(redis_key)
-                if not data:
-                    return None
-                doc = json.loads(data)
-                return {
-                    "document_id": -1,
-                    "content": doc.get("content", ""),
-                    "score": 1.0,
-                    "chunks": [
-                        {
-                            "chunk_id": -1,
-                            "content": doc.get("content", ""),
-                        }
-                    ],
-                    "matched_chunk_ids": [-1],
-                    "document": {
-                        "id": -1,
-                        "title": doc.get("filename", "uploaded_document"),
-                        "document_type": "FILE",
-                        "metadata": {"source": "anonymous_upload"},
-                    },
-                    "source": "FILE",
-                    "_user_mentioned": True,
-                }
-            finally:
-                await redis_client.aclose()
-        except Exception as exc:
-            logger.warning("Failed to load anonymous document from Redis: %s", exc)
-            return None
-
     async def abefore_agent(  # type: ignore[override]
         self,
         state: AgentState,
         runtime: Runtime[Any],
     ) -> dict[str, Any] | None:
         del runtime
+        if self.filesystem_mode != FilesystemMode.CLOUD:
+            return None
+
         messages = state.get("messages") or []
         if not messages:
             return None
-        if self.filesystem_mode != FilesystemMode.CLOUD:
-            # Local-folder mode should not seed cloud KB documents into filesystem.
-            return None
 
-        last_human = None
+        last_human: HumanMessage | None = None
         for msg in reversed(messages):
             if isinstance(msg, HumanMessage):
                 last_human = msg
                 break
         if last_human is None:
             return None
-
         user_text = _extract_text_from_message(last_human).strip()
         if not user_text:
             return None
 
-        t0 = _perf_log and asyncio.get_event_loop().time()
+        anon_doc = state.get("kb_anon_doc")
-        existing_files = state.get("files")
+        if anon_doc:
+            return self._anon_priority(state, anon_doc)
 
-        # --- Anonymous session: load Redis doc and skip DB queries ---
+        return await self._authenticated_priority(state, messages, user_text)
-        if self.anon_session_id:
-            merged: list[dict[str, Any]] = []
-            anon_doc = await self._load_anon_document()
-            if anon_doc:
-                merged.append(anon_doc)
 
-            if merged:
+    def _anon_priority(
-                new_files = _build_anon_scoped_filesystem(merged)
+        self,
-                mentioned_paths = set(new_files.keys())
+        state: AgentState,
-            else:
+        anon_doc: dict[str, Any],
-                new_files = {}
+    ) -> dict[str, Any]:
-                mentioned_paths = set()
+        path = str(anon_doc.get("path") or "")
+        title = str(anon_doc.get("title") or "uploaded_document")
+        priority = [
+            {
+                "path": path,
+                "score": 1.0,
+                "document_id": None,
+                "title": title,
+                "mentioned": True,
+            }
+        ]
+        new_messages = list(state.get("messages") or [])
+        insert_at = max(len(new_messages) - 1, 0)
+        new_messages.insert(insert_at, _render_priority_message(priority))
+        return {
+            "kb_priority": priority,
+            "kb_matched_chunk_ids": {},
+            "messages": new_messages,
+        }
 
-            ai_msg, tool_msg = _build_synthetic_ls(
+    async def _authenticated_priority(
-                existing_files,
+        self,
-                new_files,
+        state: AgentState,
-                mentioned_paths=mentioned_paths,
+        messages: Sequence[BaseMessage],
-            )
+        user_text: str,
-            if t0 is not None:
+    ) -> dict[str, Any]:
-                _perf_log.info(
+        t0 = asyncio.get_event_loop().time()
-                    "[kb_fs_middleware] anon completed in %.3fs new_files=%d",
-                    asyncio.get_event_loop().time() - t0,
-                    len(new_files),
-                )
-            return {"files": new_files, "messages": [ai_msg, tool_msg]}
-
-        # --- Authenticated session: full KB search ---
         (
             planned_query,
             start_date,
@@ -1056,7 +799,6 @@ class KnowledgeBaseSearchMiddleware(AgentMiddleware):  # type: ignore[type-arg]
             user_text=user_text,
         )
 
-        # --- 1. Fetch mentioned documents (user-selected, all chunks) ---
         mentioned_results: list[dict[str, Any]] = []
         if self.mentioned_document_ids:
             mentioned_results = await fetch_mentioned_documents(
@@ -1065,7 +807,6 @@ class KnowledgeBaseSearchMiddleware(AgentMiddleware):  # type: ignore[type-arg]
             )
             self.mentioned_document_ids = []
 
-        # --- 2. Run KB search (recency browse or hybrid) ---
         if is_recency:
             doc_types = _resolve_search_types(
                 self.available_connectors, self.available_document_types
@@ -1088,48 +829,108 @@ class KnowledgeBaseSearchMiddleware(AgentMiddleware):  # type: ignore[type-arg]
                 end_date=end_date,
             )
 
-        # --- 3. Merge: mentioned first, then search (dedup by doc id) ---
         seen_doc_ids: set[int] = set()
-        merged_auth: list[dict[str, Any]] = []
+        merged: list[dict[str, Any]] = []
         for doc in mentioned_results:
             doc_id = (doc.get("document") or {}).get("id")
-            if doc_id is not None:
+            if isinstance(doc_id, int):
                 seen_doc_ids.add(doc_id)
-            merged_auth.append(doc)
+            merged.append(doc)
         for doc in search_results:
             doc_id = (doc.get("document") or {}).get("id")
-            if doc_id is not None and doc_id in seen_doc_ids:
+            if isinstance(doc_id, int) and doc_id in seen_doc_ids:
                 continue
-            merged_auth.append(doc)
+            merged.append(doc)
 
-        # --- 4. Build scoped filesystem ---
+        priority, matched_chunk_ids = await self._materialize_priority(merged)
-        new_files, doc_id_to_path = await build_scoped_filesystem(
+
-            documents=merged_auth,
+        new_messages = list(messages)
-            search_space_id=self.search_space_id,
+        insert_at = max(len(new_messages) - 1, 0)
+        new_messages.insert(insert_at, _render_priority_message(priority))
+
+        _perf_log.info(
+            "[kb_priority] completed in %.3fs query=%r priority=%d mentioned=%d",
+            asyncio.get_event_loop().time() - t0,
+            user_text[:80],
+            len(priority),
+            len(mentioned_results),
         )
 
-        mentioned_doc_ids = {
+        return {
-            (d.get("document") or {}).get("id") for d in mentioned_results
+            "kb_priority": priority,
-        }
+            "kb_matched_chunk_ids": matched_chunk_ids,
-        mentioned_paths = {
+            "messages": new_messages,
-            doc_id_to_path[did] for did in mentioned_doc_ids if did in doc_id_to_path
         }
 
-        ai_msg, tool_msg = _build_synthetic_ls(
+    async def _materialize_priority(
-            existing_files,
+        self, merged: list[dict[str, Any]]
-            new_files,
+    ) -> tuple[list[dict[str, Any]], dict[int, list[int]]]:
-            mentioned_paths=mentioned_paths,
+        """Resolve canonical paths and matched chunk ids for the priority list."""
-        )
+        priority: list[dict[str, Any]] = []
+        matched_chunk_ids: dict[int, list[int]] = {}
 
-        if t0 is not None:
+        if not merged:
-            _perf_log.info(
+            return priority, matched_chunk_ids
-                "[kb_fs_middleware] completed in %.3fs query=%r optimized=%r "
+
-                "mentioned=%d new_files=%d total=%d",
+        async with shielded_async_session() as session:
-                asyncio.get_event_loop().time() - t0,
+            index: PathIndex = await build_path_index(session, self.search_space_id)
-                user_text[:80],
+            doc_ids = [
-                planned_query[:120],
+                (doc.get("document") or {}).get("id")
-                len(mentioned_results),
+                for doc in merged
-                len(new_files),
+                if isinstance(doc, dict)
-                len(new_files) + len(existing_files or {}),
+            ]
+            doc_ids = [doc_id for doc_id in doc_ids if isinstance(doc_id, int)]
+            folder_by_doc_id: dict[int, int | None] = {}
+            if doc_ids:
+                folder_rows = await session.execute(
+                    select(Document.id, Document.folder_id).where(
+                        Document.search_space_id == self.search_space_id,
+                        Document.id.in_(doc_ids),
+                    )
+                )
+                folder_by_doc_id = {row.id: row.folder_id for row in folder_rows.all()}
+
+        for doc in merged:
+            doc_meta = doc.get("document") or {}
+            doc_id = doc_meta.get("id")
+            title = doc_meta.get("title") or "untitled"
+            folder_id = (
+                folder_by_doc_id.get(doc_id)
+                if isinstance(doc_id, int)
+                else doc_meta.get("folder_id")
             )
-        return {"files": new_files, "messages": [ai_msg, tool_msg]}
+            path = doc_to_virtual_path(
+                doc_id=doc_id if isinstance(doc_id, int) else None,
+                title=str(title),
+                folder_id=folder_id if isinstance(folder_id, int) else None,
+                index=index,
+            )
+            priority.append(
+                {
+                    "path": path,
+                    "score": float(doc.get("score") or 0.0),
+                    "document_id": doc_id if isinstance(doc_id, int) else None,
+                    "title": str(title),
+                    "mentioned": bool(doc.get("_user_mentioned")),
+                }
+            )
+            if isinstance(doc_id, int):
+                chunk_ids = doc.get("matched_chunk_ids") or []
+                if chunk_ids:
+                    matched_chunk_ids[doc_id] = [
+                        int(cid) for cid in chunk_ids if isinstance(cid, int | str)
+                    ]
+        return priority, matched_chunk_ids
+
+
+# Backwards-compatible alias for any external imports.
+KnowledgeBaseSearchMiddleware = KnowledgePriorityMiddleware
+
+
+__all__ = [
+    "KnowledgeBaseSearchMiddleware",
+    "KnowledgePriorityMiddleware",
+    "browse_recent_documents",
+    "fetch_mentioned_documents",
+    "search_knowledge_base",
+]

surfsense_backend/app/agents/new_chat/middleware/knowledge_tree.py

@@ -0,0 +1,272 @@
+"""Workspace-tree middleware for the SurfSense agent.
+
+Renders the full ``Folder``+``Document`` tree under ``/documents/`` once per
+turn (cloud only), caches it by ``(search_space_id, tree_version)``, and
+injects the result as a ``<workspace_tree>`` system message immediately
+before the latest human turn.
+
+The render is bounded by two truncation layers:
+
+1. **Entry cap** — at most ``MAX_TREE_ENTRIES`` lines. The remainder is
+   replaced with a "use ls" hint.
+2. **Token cap** — at most ``MAX_TREE_TOKENS`` tokens (using the LLM's
+   token-count profile when available). If the entry-truncated tree still
+   exceeds the token cap we fall back to a root-only summary.
+
+Anonymous mode renders only ``state['kb_anon_doc']`` (no DB calls).
+
+This middleware also performs a one-time initialization of ``state['cwd']``
+to ``"/documents"`` so subsequent middlewares and tools always see a valid
+cwd in cloud mode.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+from typing import Any
+
+from langchain.agents.middleware import AgentMiddleware, AgentState
+from langchain_core.language_models import BaseChatModel
+from langchain_core.messages import SystemMessage
+from langgraph.runtime import Runtime
+from sqlalchemy import select
+
+from app.agents.new_chat.filesystem_selection import FilesystemMode
+from app.agents.new_chat.filesystem_state import SurfSenseFilesystemState
+from app.agents.new_chat.path_resolver import (
+    DOCUMENTS_ROOT,
+    PathIndex,
+    build_path_index,
+    doc_to_virtual_path,
+)
+from app.db import Document, shielded_async_session
+
+try:
+    from litellm import token_counter
+except Exception:  # pragma: no cover - optional dep
+    token_counter = None  # type: ignore[assignment]
+
+logger = logging.getLogger(__name__)
+
+
+MAX_TREE_ENTRIES = 500
+MAX_TREE_TOKENS = 4000
+
+
+def _approx_tokens(text: str) -> int:
+    """Cheap fallback token estimate (1 token ~= 4 chars)."""
+    return max(1, (len(text) + 3) // 4)
+
+
+def _count_tokens(text: str, *, llm: BaseChatModel | None) -> int:
+    if llm is None:
+        return _approx_tokens(text)
+    count_fn = getattr(llm, "_count_tokens", None)
+    if callable(count_fn):
+        try:
+            return int(count_fn([{"role": "user", "content": text}]))
+        except Exception:
+            pass
+    profile = getattr(llm, "profile", None)
+    model_names: list[str] = []
+    if isinstance(profile, dict):
+        tcms = profile.get("token_count_models")
+        if isinstance(tcms, list):
+            model_names.extend(name for name in tcms if isinstance(name, str) and name)
+        tcm = profile.get("token_count_model")
+        if isinstance(tcm, str) and tcm and tcm not in model_names:
+            model_names.append(tcm)
+    model_name = model_names[0] if model_names else getattr(llm, "model", None)
+    if not isinstance(model_name, str) or not model_name or token_counter is None:
+        return _approx_tokens(text)
+    try:
+        return int(
+            token_counter(
+                messages=[{"role": "user", "content": text}],
+                model=model_name,
+            )
+        )
+    except Exception:
+        return _approx_tokens(text)
+
+
+class KnowledgeTreeMiddleware(AgentMiddleware):  # type: ignore[type-arg]
+    """Inject the workspace folder/document tree into the agent's context."""
+
+    tools = ()
+    state_schema = SurfSenseFilesystemState
+
+    def __init__(
+        self,
+        *,
+        search_space_id: int,
+        filesystem_mode: FilesystemMode,
+        llm: BaseChatModel | None = None,
+        max_entries: int = MAX_TREE_ENTRIES,
+        max_tokens: int = MAX_TREE_TOKENS,
+    ) -> None:
+        self.search_space_id = search_space_id
+        self.filesystem_mode = filesystem_mode
+        self.llm = llm
+        self.max_entries = max_entries
+        self.max_tokens = max_tokens
+        self._cache: dict[tuple[int, int, bool], str] = {}
+
+    async def abefore_agent(  # type: ignore[override]
+        self,
+        state: AgentState,
+        runtime: Runtime[Any],
+    ) -> dict[str, Any] | None:
+        del runtime
+        if self.filesystem_mode != FilesystemMode.CLOUD:
+            return None
+
+        update: dict[str, Any] = {}
+        if not state.get("cwd"):
+            update["cwd"] = DOCUMENTS_ROOT
+
+        anon_doc = state.get("kb_anon_doc")
+        if anon_doc:
+            tree_msg = self._render_anon_tree(anon_doc)
+        else:
+            tree_msg = await self._render_kb_tree(state)
+
+        messages = list(state.get("messages") or [])
+        insert_at = max(len(messages) - 1, 0)
+        messages.insert(insert_at, SystemMessage(content=tree_msg))
+        update["messages"] = messages
+        return update
+
+    def before_agent(  # type: ignore[override]
+        self,
+        state: AgentState,
+        runtime: Runtime[Any],
+    ) -> dict[str, Any] | None:
+        try:
+            loop = asyncio.get_running_loop()
+            if loop.is_running():
+                return None
+        except RuntimeError:
+            pass
+        return asyncio.run(self.abefore_agent(state, runtime))
+
+    # ------------------------------------------------------------------ render
+
+    def _render_anon_tree(self, anon_doc: dict[str, Any]) -> str:
+        path = str(anon_doc.get("path") or "")
+        title = str(anon_doc.get("title") or "uploaded_document")
+        return (
+            "<workspace_tree>\n"
+            "Anonymous session — only one read-only document is available.\n"
+            f"{DOCUMENTS_ROOT}/\n"
+            f"  {path} — {title}\n"
+            "</workspace_tree>"
+        )
+
+    async def _render_kb_tree(self, state: AgentState) -> str:
+        version = int(state.get("tree_version") or 0)
+        cache_key = (self.search_space_id, version, False)
+        cached = self._cache.get(cache_key)
+        if cached is not None:
+            return cached
+
+        try:
+            async with shielded_async_session() as session:
+                index = await build_path_index(session, self.search_space_id)
+                doc_rows = await session.execute(
+                    select(Document.id, Document.title, Document.folder_id).where(
+                        Document.search_space_id == self.search_space_id
+                    )
+                )
+                docs = list(doc_rows.all())
+        except Exception as exc:  # pragma: no cover - defensive
+            logger.warning("knowledge_tree: DB error %s", exc)
+            return "<workspace_tree>\n(unavailable)\n</workspace_tree>"
+
+        rendered = self._format_tree(index, docs)
+        self._cache[cache_key] = rendered
+        return rendered
+
+    def _format_tree(self, index: PathIndex, docs: list[Any]) -> str:
+        folder_paths = sorted(set(index.folder_paths.values()))
+        doc_paths = sorted(
+            doc_to_virtual_path(
+                doc_id=row.id,
+                title=str(row.title or "untitled"),
+                folder_id=row.folder_id,
+                index=index,
+            )
+            for row in docs
+        )
+        all_paths = sorted(set(folder_paths + doc_paths + [DOCUMENTS_ROOT]))
+
+        lines: list[str] = []
+        for path in all_paths:
+            depth = (
+                0
+                if path == DOCUMENTS_ROOT
+                else len([p for p in path[len(DOCUMENTS_ROOT) :].split("/") if p])
+            )
+            indent = "  " * depth
+            is_dir = path == DOCUMENTS_ROOT or path in folder_paths
+            display = (
+                path.rsplit("/", 1)[-1] if path != DOCUMENTS_ROOT else "/documents"
+            )
+            if is_dir:
+                lines.append(f"{indent}{display}/")
+            else:
+                lines.append(f"{indent}{display}")
+            if len(lines) >= self.max_entries:
+                remaining = len(all_paths) - len(lines)
+                if remaining > 0:
+                    lines.append(
+                        f"... {remaining} more entries — use "
+                        "ls('/documents/<folder>', offset, limit) to expand"
+                    )
+                break
+
+        body = "\n".join(lines)
+        rendered = f"<workspace_tree>\n{body}\n</workspace_tree>"
+
+        token_count = _count_tokens(rendered, llm=self.llm)
+        if token_count <= self.max_tokens:
+            return rendered
+
+        return self._format_root_summary(folder_paths, doc_paths)
+
+    def _format_root_summary(
+        self, folder_paths: list[str], doc_paths: list[str]
+    ) -> str:
+        top_level: dict[str, int] = {}
+        loose_docs = 0
+        for path in doc_paths:
+            rel = path[len(DOCUMENTS_ROOT) :].lstrip("/")
+            if "/" in rel:
+                top = rel.split("/", 1)[0]
+                top_level[top] = top_level.get(top, 0) + 1
+            else:
+                loose_docs += 1
+        for path in folder_paths:
+            rel = path[len(DOCUMENTS_ROOT) :].lstrip("/")
+            if not rel:
+                continue
+            top = rel.split("/", 1)[0]
+            top_level.setdefault(top, 0)
+
+        lines = [DOCUMENTS_ROOT + "/"]
+        for name in sorted(top_level):
+            count = top_level[name]
+            lines.append(f"  {name}/ ({count} document{'s' if count != 1 else ''})")
+        if loose_docs:
+            lines.append(
+                f"  ({loose_docs} loose document{'s' if loose_docs != 1 else ''})"
+            )
+        lines.append(
+            "Tree is large; use list_tree('/documents/<folder>') to drill in "
+            "or ls('/documents/<folder>', offset, limit) for paginated listings."
+        )
+        return "<workspace_tree>\n" + "\n".join(lines) + "\n</workspace_tree>"
+
+
+__all__ = ["KnowledgeTreeMiddleware"]

surfsense_backend/app/agents/new_chat/middleware/local_folder_backend.py

@@ -120,7 +120,9 @@ class LocalFolderBackend:
         if not target.exists() or not target.is_dir():
             return []
         infos: list[FileInfo] = []
-        for child in sorted(target.iterdir(), key=lambda p: (not p.is_dir(), p.name.lower())):
+        for child in sorted(
+            target.iterdir(), key=lambda p: (not p.is_dir(), p.name.lower())
+        ):
             infos.append(
                 FileInfo(
                     path=self._to_virtual(child, self._root),
@@ -317,7 +319,9 @@ class LocalFolderBackend:
             return WriteResult(error="Error: source and destination paths are the same")
         with self._acquire_path_locks(source_path, destination_path):
             if not source.exists():
-                return WriteResult(error=f"Error: source path '{source_path}' not found")
+                return WriteResult(
+                    error=f"Error: source path '{source_path}' not found"
+                )
             if destination.exists():
                 if not overwrite:
                     return WriteResult(
@@ -339,8 +343,12 @@ class LocalFolderBackend:
                 else:
                     source.rename(destination)
             except OSError as exc:
-                return WriteResult(error=f"Error: failed to move '{source_path}': {exc}")
+                return WriteResult(
-        return WriteResult(path=self._to_virtual(destination, self._root), files_update=None)
+                    error=f"Error: failed to move '{source_path}': {exc}"
+                )
+        return WriteResult(
+            path=self._to_virtual(destination, self._root), files_update=None
+        )
 
     async def amove(
         self,
@@ -368,12 +376,16 @@ class LocalFolderBackend:
             if not path.exists() or not path.is_file():
                 return EditResult(error=f"Error: File '{file_path}' not found")
             content = path.read_text(encoding="utf-8", errors="replace")
-            result = perform_string_replacement(content, old_string, new_string, replace_all)
+            result = perform_string_replacement(
+                content, old_string, new_string, replace_all
+            )
             if isinstance(result, str):
                 return EditResult(error=result)
             updated_content, occurrences = result
             self._write_text_atomic(path, updated_content)
-        return EditResult(path=file_path, files_update=None, occurrences=int(occurrences))
+        return EditResult(
+            path=file_path, files_update=None, occurrences=int(occurrences)
+        )
 
     async def aedit(
         self,
@@ -447,7 +459,9 @@ class LocalFolderBackend:
         matches: list[GrepMatch] = []
         for file_path in self._iter_candidate_files(path, glob):
             try:
-                lines = file_path.read_text(encoding="utf-8", errors="replace").splitlines()
+                lines = file_path.read_text(
+                    encoding="utf-8", errors="replace"
+                ).splitlines()
             except Exception:
                 continue
             for idx, line in enumerate(lines, start=1):
@@ -481,12 +495,18 @@ class LocalFolderBackend:
                     FileUploadResponse(path=virtual_path, error=_FILE_NOT_FOUND)
                 )
             except IsADirectoryError:
-                responses.append(FileUploadResponse(path=virtual_path, error=_IS_DIRECTORY))
+                responses.append(
+                    FileUploadResponse(path=virtual_path, error=_IS_DIRECTORY)
+                )
             except Exception:
-                responses.append(FileUploadResponse(path=virtual_path, error=_INVALID_PATH))
+                responses.append(
+                    FileUploadResponse(path=virtual_path, error=_INVALID_PATH)
+                )
         return responses
 
-    async def aupload_files(self, files: list[tuple[str, bytes]]) -> list[FileUploadResponse]:
+    async def aupload_files(
+        self, files: list[tuple[str, bytes]]
+    ) -> list[FileUploadResponse]:
         return await asyncio.to_thread(self.upload_files, files)
 
     def download_files(self, paths: list[str]) -> list[FileDownloadResponse]:
@@ -515,7 +535,9 @@ class LocalFolderBackend:
                 )
             except Exception:
                 responses.append(
-                    FileDownloadResponse(path=virtual_path, content=None, error=_INVALID_PATH)
+                    FileDownloadResponse(
+                        path=virtual_path, content=None, error=_INVALID_PATH
+                    )
                 )
         return responses
 

surfsense_backend/app/agents/new_chat/middleware/multi_root_local_folder_backend.py

@@ -127,7 +127,9 @@ class MultiRootLocalFolderBackend:
             mount, local_path = self._split_mount_path(path)
         except ValueError:
             return []
-        return self._transform_infos(mount, self._mount_to_backend[mount].ls_info(local_path))
+        return self._transform_infos(
+            mount, self._mount_to_backend[mount].ls_info(local_path)
+        )
 
     async def als_info(self, path: str) -> list[FileInfo]:
         return await asyncio.to_thread(self.ls_info, path)
@@ -355,7 +357,9 @@ class MultiRootLocalFolderBackend:
                 all_matches.extend(
                     [
                         GrepMatch(
-                            path=self._prefix_mount_path(mount, self._get_str(match, "path")),
+                            path=self._prefix_mount_path(
+                                mount, self._get_str(match, "path")
+                            ),
                             line=self._get_int(match, "line"),
                             text=self._get_str(match, "text"),
                         )
@@ -394,7 +398,9 @@ class MultiRootLocalFolderBackend:
             try:
                 mount, local_path = self._split_mount_path(virtual_path)
             except ValueError:
-                invalid.append(FileUploadResponse(path=virtual_path, error=_INVALID_PATH))
+                invalid.append(
+                    FileUploadResponse(path=virtual_path, error=_INVALID_PATH)
+                )
                 continue
             grouped.setdefault(mount, []).append((local_path, content))
 
@@ -404,7 +410,9 @@ class MultiRootLocalFolderBackend:
             responses.extend(
                 [
                     FileUploadResponse(
-                        path=self._prefix_mount_path(mount, self._get_str(item, "path")),
+                        path=self._prefix_mount_path(
+                            mount, self._get_str(item, "path")
+                        ),
                         error=self._get_str(item, "error") or None,
                     )
                     for item in result
@@ -412,7 +420,9 @@ class MultiRootLocalFolderBackend:
             )
         return responses
 
-    async def aupload_files(self, files: list[tuple[str, bytes]]) -> list[FileUploadResponse]:
+    async def aupload_files(
+        self, files: list[tuple[str, bytes]]
+    ) -> list[FileUploadResponse]:
         return await asyncio.to_thread(self.upload_files, files)
 
     def download_files(self, paths: list[str]) -> list[FileDownloadResponse]:
@@ -423,7 +433,9 @@ class MultiRootLocalFolderBackend:
                 mount, local_path = self._split_mount_path(virtual_path)
             except ValueError:
                 invalid.append(
-                    FileDownloadResponse(path=virtual_path, content=None, error=_INVALID_PATH)
+                    FileDownloadResponse(
+                        path=virtual_path, content=None, error=_INVALID_PATH
+                    )
                 )
                 continue
             grouped.setdefault(mount, []).append(local_path)
@@ -434,7 +446,9 @@ class MultiRootLocalFolderBackend:
             responses.extend(
                 [
                     FileDownloadResponse(
-                        path=self._prefix_mount_path(mount, self._get_str(item, "path")),
+                        path=self._prefix_mount_path(
+                            mount, self._get_str(item, "path")
+                        ),
                         content=self._get_value(item, "content"),
                         error=self._get_str(item, "error") or None,
                     )

surfsense_backend/app/agents/new_chat/middleware/noop_injection.py

@@ -0,0 +1,141 @@
+"""
+``_noop`` provider-compatibility tool + injection middleware.
+
+Some providers (LiteLLM, Bedrock, Copilot) 400 when a model call has
+empty ``tools`` but the message history includes prior ``tool_calls`` —
+they treat that shape as malformed even though it's perfectly valid
+LangChain. SurfSense hits this on the compaction summarize call (no
+tools, history full of tool calls).
+
+Ported from OpenCode's ``packages/opencode/src/session/llm.ts:209-228``,
+which discovered and codified the workaround: inject a no-op tool *only*
+on those provider shapes so the request validates without ever being
+called.
+
+Operation: a :class:`NoopInjectionMiddleware` ``wrap_model_call`` checks
+if the request has zero tools but the last AI message in history includes
+``tool_calls``. If yes, it injects the ``_noop`` tool only — never
+globally — mirroring OpenCode's gating exactly. The :func:`noop_tool`
+returns empty content when called (which it should never be in
+practice).
+"""
+
+from __future__ import annotations
+
+import logging
+from collections.abc import Awaitable, Callable
+from typing import Any
+
+from langchain.agents.middleware.types import (
+    AgentMiddleware,
+    AgentState,
+    ContextT,
+    ModelRequest,
+    ModelResponse,
+    ResponseT,
+)
+from langchain_core.messages import AIMessage
+from langchain_core.tools import tool
+
+logger = logging.getLogger(__name__)
+
+NOOP_TOOL_NAME = "_noop"
+NOOP_TOOL_DESCRIPTION = "Do not call this tool. It exists only for API compatibility."
+
+
+@tool(name_or_callable=NOOP_TOOL_NAME, description=NOOP_TOOL_DESCRIPTION)
+def noop_tool() -> str:
+    """Return empty content. Never expected to be called."""
+    return ""
+
+
+# Provider markers that benefit from ``_noop`` injection. These match
+# OpenCode's gating list (``llm.ts:209-228``). We also accept any string
+# containing one of these substrings so e.g. ``litellm`` matches
+# ``ChatLiteLLM``.
+_NOOP_NEEDED_PROVIDERS: tuple[str, ...] = (
+    "litellm",
+    "bedrock",
+    "copilot",
+)
+
+
+def _provider_needs_noop(model: Any) -> bool:
+    """Heuristic: does this model's provider need the _noop injection?"""
+    try:
+        ls_params = model._get_ls_params()
+        provider = str(ls_params.get("ls_provider", "")).lower()
+    except Exception:
+        provider = ""
+
+    if not provider:
+        cls_name = type(model).__name__.lower()
+        provider = cls_name
+
+    return any(needle in provider for needle in _NOOP_NEEDED_PROVIDERS)
+
+
+def _last_ai_has_tool_calls(messages: list[Any]) -> bool:
+    for msg in reversed(messages):
+        if isinstance(msg, AIMessage):
+            return bool(msg.tool_calls)
+    return False
+
+
+class NoopInjectionMiddleware(
+    AgentMiddleware[AgentState[ResponseT], ContextT, ResponseT]
+):
+    """Inject the ``_noop`` tool only when the provider would otherwise 400.
+
+    The check fires per model call, not at agent build time, because the
+    summarization path generates a no-tool subcall at runtime. The
+    extra tool is appended to ``request.tools`` as an instance — the
+    actual ``langchain_core.tools.BaseTool`` is bound on every call site
+    that creates the agent.
+    """
+
+    def __init__(self, *, noop_tool_instance: Any | None = None) -> None:
+        super().__init__()
+        self._noop_tool = noop_tool_instance or noop_tool
+        self.tools = []
+
+    def _should_inject(self, request: ModelRequest[ContextT]) -> bool:
+        if request.tools:
+            return False
+        if not _last_ai_has_tool_calls(request.messages):
+            return False
+        return _provider_needs_noop(request.model)
+
+    def _augmented(self, request: ModelRequest[ContextT]) -> ModelRequest[ContextT]:
+        return request.override(tools=[self._noop_tool])
+
+    def wrap_model_call(  # type: ignore[override]
+        self,
+        request: ModelRequest[ContextT],
+        handler: Callable[[ModelRequest[ContextT]], ModelResponse[ResponseT]],
+    ) -> Any:
+        if self._should_inject(request):
+            logger.debug("Injecting _noop tool for provider compatibility")
+            return handler(self._augmented(request))
+        return handler(request)
+
+    async def awrap_model_call(  # type: ignore[override]
+        self,
+        request: ModelRequest[ContextT],
+        handler: Callable[
+            [ModelRequest[ContextT]], Awaitable[ModelResponse[ResponseT]]
+        ],
+    ) -> Any:
+        if self._should_inject(request):
+            logger.debug("Injecting _noop tool for provider compatibility")
+            return await handler(self._augmented(request))
+        return await handler(request)
+
+
+__all__ = [
+    "NOOP_TOOL_DESCRIPTION",
+    "NOOP_TOOL_NAME",
+    "NoopInjectionMiddleware",
+    "_provider_needs_noop",
+    "noop_tool",
+]

surfsense_backend/app/agents/new_chat/middleware/otel_span.py

@@ -0,0 +1,202 @@
+"""
+OpenTelemetry span middleware for the SurfSense ``new_chat`` agent.
+
+Wraps both ``model.call`` (LLM invocations) and ``tool.call`` (tool
+executions) with OTel spans, attaching low-cardinality span names and
+high-cardinality identifiers as attributes.
+
+This middleware is intentionally a thin adapter over
+:mod:`app.observability.otel`; when OTel is not configured all spans
+collapse to no-ops and the wrapper adds <1µs overhead per call. When
+OTel **is** configured (``OTEL_EXPORTER_OTLP_ENDPOINT`` set), every
+model and tool call gets a span with the standard attributes our
+dashboards expect.
+"""
+
+from __future__ import annotations
+
+import logging
+from collections.abc import Awaitable, Callable
+from typing import TYPE_CHECKING, Any
+
+from langchain.agents.middleware import AgentMiddleware
+from langchain_core.messages import AIMessage, ToolMessage
+
+from app.observability import otel as ot
+
+if TYPE_CHECKING:  # pragma: no cover — type-only
+    from langchain.agents.middleware.types import (
+        ModelRequest,
+        ModelResponse,
+        ToolCallRequest,
+    )
+    from langgraph.types import Command
+
+logger = logging.getLogger(__name__)
+
+
+class OtelSpanMiddleware(AgentMiddleware):
+    """Emit ``model.call`` and ``tool.call`` OTel spans for every invocation.
+
+    Should be placed near the **outer** end of the middleware list so
+    that the spans encompass retry/fallback wrapper effects (i.e. ``N``
+    model.call spans for ``N`` retry attempts) but inside any concurrency/
+    auth gate. Empirically this means **between** ``BusyMutex`` and
+    ``RetryAfter``.
+    """
+
+    def __init__(self, *, instrumentation_name: str = "surfsense.new_chat") -> None:
+        super().__init__()
+        self._instrumentation_name = instrumentation_name
+
+    # ------------------------------------------------------------------
+    # Model call spans
+    # ------------------------------------------------------------------
+
+    async def awrap_model_call(
+        self,
+        request: ModelRequest,
+        handler: Callable[[ModelRequest], Awaitable[ModelResponse | AIMessage | Any]],
+    ) -> ModelResponse | AIMessage | Any:
+        if not ot.is_enabled():
+            return await handler(request)
+
+        model_id, provider = _resolve_model_attrs(request)
+        with ot.model_call_span(model_id=model_id, provider=provider) as sp:
+            try:
+                result = await handler(request)
+            except Exception:
+                # span context manager records + re-raises
+                raise
+            else:
+                _annotate_model_response(sp, result)
+                return result
+
+    # ------------------------------------------------------------------
+    # Tool call spans
+    # ------------------------------------------------------------------
+
+    async def awrap_tool_call(
+        self,
+        request: ToolCallRequest,
+        handler: Callable[[ToolCallRequest], Awaitable[ToolMessage | Command[Any]]],
+    ) -> ToolMessage | Command[Any]:
+        if not ot.is_enabled():
+            return await handler(request)
+
+        tool_name = _resolve_tool_name(request)
+        input_size = _resolve_input_size(request)
+
+        with ot.tool_call_span(tool_name, input_size=input_size) as sp:
+            result = await handler(request)
+            _annotate_tool_result(sp, result)
+            return result
+
+
+# ---------------------------------------------------------------------------
+# Attribute helpers (kept defensive; we never want OTel bookkeeping to break
+# a real model/tool call).
+# ---------------------------------------------------------------------------
+
+
+def _resolve_model_attrs(request: Any) -> tuple[str | None, str | None]:
+    """Extract ``model.id`` and ``model.provider`` from a ``ModelRequest``."""
+    model_id: str | None = None
+    provider: str | None = None
+    try:
+        model = getattr(request, "model", None)
+        if model is None:
+            return None, None
+        # langchain BaseChatModel exposes a few different identifiers
+        for attr in ("model_name", "model", "model_id"):
+            value = getattr(model, attr, None)
+            if value:
+                model_id = str(value)
+                break
+        # provider sometimes lives on ``_llm_type`` (legacy) or ``provider``
+        for attr in ("provider", "_llm_type"):
+            value = getattr(model, attr, None)
+            if value:
+                provider = str(value)
+                break
+    except Exception:  # pragma: no cover — defensive
+        pass
+    return model_id, provider
+
+
+def _resolve_tool_name(request: Any) -> str:
+    try:
+        tool = getattr(request, "tool", None)
+        if tool is not None:
+            name = getattr(tool, "name", None)
+            if isinstance(name, str) and name:
+                return name
+        # Fall back to the tool_call dict
+        call = getattr(request, "tool_call", None) or {}
+        name = call.get("name") if isinstance(call, dict) else None
+        if isinstance(name, str) and name:
+            return name
+    except Exception:  # pragma: no cover — defensive
+        pass
+    return "unknown"
+
+
+def _resolve_input_size(request: Any) -> int | None:
+    try:
+        call = getattr(request, "tool_call", None)
+        if not isinstance(call, dict) or not call:
+            return None
+        args = call.get("args")
+        if args is None:
+            return None
+        return len(repr(args))
+    except Exception:  # pragma: no cover — defensive
+        return None
+
+
+def _annotate_model_response(span: Any, result: Any) -> None:
+    """Best-effort: attach prompt/completion token counts when available."""
+    try:
+        # ModelResponse may be a dataclass with .result containing AIMessage
+        msg: Any
+        if isinstance(result, AIMessage):
+            msg = result
+        else:
+            inner = getattr(result, "result", None)
+            msg = inner[-1] if isinstance(inner, list) and inner else inner
+        if msg is None:
+            return
+        usage = getattr(msg, "usage_metadata", None) or {}
+        if isinstance(usage, dict):
+            if (n := usage.get("input_tokens")) is not None:
+                span.set_attribute("tokens.prompt", int(n))
+            if (n := usage.get("output_tokens")) is not None:
+                span.set_attribute("tokens.completion", int(n))
+            if (n := usage.get("total_tokens")) is not None:
+                span.set_attribute("tokens.total", int(n))
+        tool_calls = getattr(msg, "tool_calls", None) or []
+        span.set_attribute("model.tool_calls", len(tool_calls))
+    except Exception:  # pragma: no cover — defensive
+        pass
+
+
+def _annotate_tool_result(span: Any, result: Any) -> None:
+    try:
+        if isinstance(result, ToolMessage):
+            content = (
+                result.content
+                if isinstance(result.content, str)
+                else repr(result.content)
+            )
+            span.set_attribute("tool.output.size", len(content))
+            status = getattr(result, "status", None)
+            if isinstance(status, str):
+                span.set_attribute("tool.status", status)
+            kwargs = getattr(result, "additional_kwargs", None) or {}
+            if isinstance(kwargs, dict) and kwargs.get("error"):
+                span.set_attribute("tool.error", True)
+    except Exception:  # pragma: no cover — defensive
+        pass
+
+
+__all__ = ["OtelSpanMiddleware"]

surfsense_backend/app/agents/new_chat/middleware/permission.py

@@ -0,0 +1,358 @@
+"""
+PermissionMiddleware — pattern-based allow/deny/ask with HITL fallback.
+
+LangChain's :class:`HumanInTheLoopMiddleware` only supports a static
+"this tool always asks" decision per tool. There's no rule-based
+allow/deny/ask layered ruleset, no glob patterns, no per-search-space or
+per-thread overrides, and no auto-deny synthesis.
+
+This middleware ports OpenCode's ``packages/opencode/src/permission/index.ts``
+ruleset model on top of SurfSense's existing ``interrupt({type, action,
+context})`` payload shape (see ``app/agents/new_chat/tools/hitl.py``) so
+the frontend keeps working unchanged.
+
+Operation:
+1. ``aafter_model`` inspects the latest ``AIMessage.tool_calls``.
+2. For each call, the middleware builds a list of ``patterns`` (the
+   tool name plus any tool-specific patterns from the resolver). It
+   evaluates each pattern against the layered rulesets and aggregates
+   the results: ``deny`` > ``ask`` > ``allow``.
+3. On ``deny``: replaces the call with a synthetic ``ToolMessage``
+   containing a :class:`StreamingError`.
+4. On ``ask``: raises a SurfSense-style ``interrupt(...)``. The reply
+   shape is ``{"decision_type": "once|always|reject", "feedback"?: str}``.
+   - ``once``: proceed.
+   - ``always``: also persist allow rules for ``request.always`` patterns.
+   - ``reject`` w/o feedback: raise :class:`RejectedError`.
+   - ``reject`` w/  feedback: raise :class:`CorrectedError`.
+5. On ``allow``: proceed unchanged.
+
+The middleware also performs a *pre-model* tool-filter step (the
+``before_model`` hook) so globally denied tools are stripped from the
+exposed tool list before the model gets to see them. This mirrors
+OpenCode's ``Permission.disabled`` and dramatically reduces the chance
+the model emits a deny-only call.
+"""
+
+from __future__ import annotations
+
+import logging
+from collections.abc import Callable
+from typing import Any
+
+from langchain.agents.middleware.types import (
+    AgentMiddleware,
+    AgentState,
+    ContextT,
+)
+from langchain_core.messages import AIMessage, ToolMessage
+from langgraph.runtime import Runtime
+from langgraph.types import interrupt
+
+from app.agents.new_chat.errors import (
+    CorrectedError,
+    RejectedError,
+    StreamingError,
+)
+from app.agents.new_chat.permissions import (
+    Rule,
+    Ruleset,
+    aggregate_action,
+    evaluate_many,
+)
+from app.observability import otel as ot
+
+logger = logging.getLogger(__name__)
+
+
+# Mapping ``tool_name -> resolver`` that converts ``args`` to a list of
+# patterns to evaluate. The first pattern is conventionally the bare
+# tool name; later entries narrow down to specific resources.
+PatternResolver = Callable[[dict[str, Any]], list[str]]
+
+
+def _default_pattern_resolver(name: str) -> PatternResolver:
+    def _resolve(args: dict[str, Any]) -> list[str]:
+        # Bare name covers the default catch-all; primary-arg fallbacks
+        # are best added per-tool by callers.
+        del args
+        return [name]
+
+    return _resolve
+
+
+class PermissionMiddleware(AgentMiddleware):  # type: ignore[type-arg]
+    """Allow/deny/ask layer over the agent's tool calls.
+
+    Args:
+        rulesets: Layered rulesets to evaluate. Earlier entries are
+            overridden by later ones (last-match-wins). Typical layering:
+            ``defaults < global < space < thread < runtime_approved``.
+        pattern_resolvers: Optional per-tool callables that return a list
+            of patterns to evaluate. When a tool isn't listed, the bare
+            tool name is used as the only pattern.
+        runtime_ruleset: Mutable :class:`Ruleset` that the middleware
+            extends in-place when the user replies ``"always"`` to an
+            ask interrupt. Reused across all calls in the same agent
+            instance so newly-allowed rules apply to subsequent calls.
+        always_emit_interrupt_payload: If True, every ask uses the
+            SurfSense interrupt wire format (default). Set False to
+            disable interrupts and treat ``ask`` as ``deny`` for
+            non-interactive deployments.
+    """
+
+    tools = ()
+
+    def __init__(
+        self,
+        *,
+        rulesets: list[Ruleset] | None = None,
+        pattern_resolvers: dict[str, PatternResolver] | None = None,
+        runtime_ruleset: Ruleset | None = None,
+        always_emit_interrupt_payload: bool = True,
+    ) -> None:
+        super().__init__()
+        self._static_rulesets: list[Ruleset] = list(rulesets or [])
+        self._pattern_resolvers: dict[str, PatternResolver] = dict(
+            pattern_resolvers or {}
+        )
+        self._runtime_ruleset: Ruleset = runtime_ruleset or Ruleset(
+            origin="runtime_approved"
+        )
+        self._emit_interrupt = always_emit_interrupt_payload
+
+    # ------------------------------------------------------------------
+    # Tool-filter step (mirrors OpenCode's ``Permission.disabled``)
+    # ------------------------------------------------------------------
+
+    def _globally_denied(self, tool_name: str) -> bool:
+        """Return True if a deny rule with no narrowing pattern matches."""
+        rules = evaluate_many(tool_name, ["*"], *self._all_rulesets())
+        return aggregate_action(rules) == "deny"
+
+    def _all_rulesets(self) -> list[Ruleset]:
+        return [*self._static_rulesets, self._runtime_ruleset]
+
+    # NOTE: ``before_model`` filtering of the tools list is left to the
+    # agent factory. This middleware only blocks at execution time — and
+    # only via the rule-evaluator path, not by mutating ``request.tools``.
+    # Mutating ``request.tools`` per-call would invalidate provider
+    # prompt-cache prefixes (see Operational risks: prompt-cache regression).
+
+    # ------------------------------------------------------------------
+    # Tool-call evaluation
+    # ------------------------------------------------------------------
+
+    def _resolve_patterns(self, tool_name: str, args: dict[str, Any]) -> list[str]:
+        resolver = self._pattern_resolvers.get(
+            tool_name, _default_pattern_resolver(tool_name)
+        )
+        try:
+            patterns = resolver(args or {})
+        except Exception:
+            logger.exception(
+                "Pattern resolver for %s raised; using bare name", tool_name
+            )
+            patterns = [tool_name]
+        if not patterns:
+            patterns = [tool_name]
+        return patterns
+
+    def _evaluate(
+        self, tool_name: str, args: dict[str, Any]
+    ) -> tuple[str, list[str], list[Rule]]:
+        patterns = self._resolve_patterns(tool_name, args)
+        rules = evaluate_many(tool_name, patterns, *self._all_rulesets())
+        action = aggregate_action(rules)
+        return action, patterns, rules
+
+    # ------------------------------------------------------------------
+    # HITL ask flow — SurfSense wire format
+    # ------------------------------------------------------------------
+
+    def _raise_interrupt(
+        self,
+        *,
+        tool_name: str,
+        args: dict[str, Any],
+        patterns: list[str],
+        rules: list[Rule],
+    ) -> dict[str, Any]:
+        """Block on user approval via SurfSense's ``interrupt`` shape."""
+        if not self._emit_interrupt:
+            return {"decision_type": "reject"}
+
+        # ``params`` (NOT ``args``) is what SurfSense's streaming
+        # normalizer forwards. Other fields move into ``context``.
+        payload = {
+            "type": "permission_ask",
+            "action": {"tool": tool_name, "params": args or {}},
+            "context": {
+                "patterns": patterns,
+                "rules": [
+                    {
+                        "permission": r.permission,
+                        "pattern": r.pattern,
+                        "action": r.action,
+                    }
+                    for r in rules
+                ],
+                # Rules of thumb for the frontend: surface the patterns
+                # the user can promote to "always" with a single reply.
+                "always": patterns,
+            },
+        }
+        # Open ``permission.asked`` + ``interrupt.raised`` OTel spans
+        # (no-op when OTel is disabled) so dashboards can correlate
+        # "we asked X" with "interrupt was actually delivered".
+        with (
+            ot.permission_asked_span(
+                permission=tool_name,
+                pattern=patterns[0] if patterns else None,
+                extra={"permission.patterns": list(patterns)},
+            ),
+            ot.interrupt_span(interrupt_type="permission_ask"),
+        ):
+            decision = interrupt(payload)
+        if isinstance(decision, dict):
+            return decision
+        # Tolerate a plain string reply ("once", "always", "reject")
+        if isinstance(decision, str):
+            return {"decision_type": decision}
+        return {"decision_type": "reject"}
+
+    def _persist_always(self, tool_name: str, patterns: list[str]) -> None:
+        """Promote ``always`` reply into runtime allow rules.
+
+        Persistence to ``agent_permission_rules`` is done by the
+        streaming layer (``stream_new_chat``) once it observes the
+        ``always`` reply — the middleware just keeps an in-memory
+        copy so subsequent calls in the same stream see the rule.
+        """
+        for pattern in patterns:
+            self._runtime_ruleset.rules.append(
+                Rule(permission=tool_name, pattern=pattern, action="allow")
+            )
+
+    # ------------------------------------------------------------------
+    # Synthesizing deny -> ToolMessage
+    # ------------------------------------------------------------------
+
+    @staticmethod
+    def _deny_message(
+        tool_call: dict[str, Any],
+        rule: Rule,
+    ) -> ToolMessage:
+        err = StreamingError(
+            code="permission_denied",
+            retryable=False,
+            suggestion=(
+                f"rule permission={rule.permission!r} pattern={rule.pattern!r} "
+                f"blocked this call"
+            ),
+        )
+        return ToolMessage(
+            content=(
+                f"Permission denied: rule {rule.permission}/{rule.pattern} "
+                f"blocked tool {tool_call.get('name')!r}."
+            ),
+            tool_call_id=tool_call.get("id") or "",
+            name=tool_call.get("name"),
+            status="error",
+            additional_kwargs={"error": err.model_dump()},
+        )
+
+    # ------------------------------------------------------------------
+    # The hook: aafter_model
+    # ------------------------------------------------------------------
+
+    def _process(
+        self,
+        state: AgentState,
+        runtime: Runtime[Any],
+    ) -> dict[str, Any] | None:
+        del runtime  # unused
+        messages = state.get("messages") or []
+        if not messages:
+            return None
+        last = messages[-1]
+        if not isinstance(last, AIMessage) or not last.tool_calls:
+            return None
+
+        deny_messages: list[ToolMessage] = []
+        kept_calls: list[dict[str, Any]] = []
+        any_change = False
+
+        for raw in last.tool_calls:
+            call = (
+                dict(raw)
+                if isinstance(raw, dict)
+                else {
+                    "name": getattr(raw, "name", None),
+                    "args": getattr(raw, "args", {}),
+                    "id": getattr(raw, "id", None),
+                    "type": "tool_call",
+                }
+            )
+            name = call.get("name") or ""
+            args = call.get("args") or {}
+            action, patterns, rules = self._evaluate(name, args)
+
+            if action == "deny":
+                # Find the deny rule for the suggestion text
+                deny_rule = next((r for r in rules if r.action == "deny"), rules[0])
+                deny_messages.append(self._deny_message(call, deny_rule))
+                any_change = True
+                continue
+
+            if action == "ask":
+                decision = self._raise_interrupt(
+                    tool_name=name, args=args, patterns=patterns, rules=rules
+                )
+                kind = str(decision.get("decision_type") or "reject").lower()
+                if kind == "once":
+                    kept_calls.append(call)
+                elif kind == "always":
+                    self._persist_always(name, patterns)
+                    kept_calls.append(call)
+                elif kind == "reject":
+                    feedback = decision.get("feedback")
+                    if isinstance(feedback, str) and feedback.strip():
+                        raise CorrectedError(feedback, tool=name)
+                    raise RejectedError(
+                        tool=name, pattern=patterns[0] if patterns else None
+                    )
+                else:
+                    logger.warning(
+                        "Unknown permission decision %r; treating as reject", kind
+                    )
+                    raise RejectedError(tool=name)
+                continue
+
+            # allow
+            kept_calls.append(call)
+
+        if not any_change and len(kept_calls) == len(last.tool_calls):
+            return None
+
+        updated = last.model_copy(update={"tool_calls": kept_calls})
+        result_messages: list[Any] = [updated]
+        if deny_messages:
+            result_messages.extend(deny_messages)
+        return {"messages": result_messages}
+
+    def after_model(  # type: ignore[override]
+        self, state: AgentState, runtime: Runtime[ContextT]
+    ) -> dict[str, Any] | None:
+        return self._process(state, runtime)
+
+    async def aafter_model(  # type: ignore[override]
+        self, state: AgentState, runtime: Runtime[ContextT]
+    ) -> dict[str, Any] | None:
+        return self._process(state, runtime)
+
+
+__all__ = [
+    "PatternResolver",
+    "PermissionMiddleware",
+]

surfsense_backend/app/agents/new_chat/middleware/retry_after.py

@@ -0,0 +1,257 @@
+"""
+RetryAfterMiddleware — Header-aware retry with custom backoff and SSE eventing.
+
+LangChain's :class:`ModelRetryMiddleware` retries on exceptions but ignores
+the ``Retry-After`` HTTP header — it just runs its own exponential backoff.
+That wastes time when a provider has explicitly told us how long to wait.
+This middleware honors the header (mirroring OpenCode's
+``packages/opencode/src/session/llm.ts`` retry pathway) and emits an SSE
+event so the UI can show "rate-limited, retrying in Ns".
+
+We can't subclass ``ModelRetryMiddleware`` cleanly because its loop calls a
+module-level ``calculate_delay`` inline (no overridable
+``_calculate_delay`` hook), so this is a standalone implementation.
+
+Behaviour:
+- Extracts ``Retry-After`` / ``retry-after-ms`` from
+  ``litellm.exceptions.RateLimitError.response.headers`` (or any exception
+  exposing a similar shape).
+- Sleeps ``max(exponential_backoff, header_delay)`` between retries.
+- Returns ``False`` from ``retry_on`` for ``ContextWindowExceededError`` /
+  ``ContextOverflowError`` so :class:`SurfSenseCompactionMiddleware` (or
+  the LangChain summarization fallback path) handles those instead.
+- Emits ``surfsense.retrying`` via ``adispatch_custom_event`` on each retry
+  so ``stream_new_chat`` can forward it to clients as an SSE event.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import random
+import re
+import time
+from collections.abc import Awaitable, Callable
+from typing import Any
+
+from langchain.agents.middleware.types import (
+    AgentMiddleware,
+    AgentState,
+    ContextT,
+    ModelRequest,
+    ModelResponse,
+    ResponseT,
+)
+from langchain_core.callbacks import adispatch_custom_event, dispatch_custom_event
+from langchain_core.messages import AIMessage
+
+logger = logging.getLogger(__name__)
+
+# Names of exception classes for which a retry would not help — context
+# overflow needs compaction, auth needs human intervention, etc. Detected
+# by class-name substring so we don't have to import LiteLLM/Anthropic
+# here (which would tie this module to optional deps).
+_NON_RETRYABLE_NAME_HINTS: tuple[str, ...] = (
+    "ContextWindowExceeded",
+    "ContextOverflow",
+    "AuthenticationError",
+    "InvalidRequestError",
+    "PermissionDenied",
+    "InvalidApiKey",
+    "ContextLimit",
+)
+
+
+def _is_non_retryable(exc: BaseException) -> bool:
+    name = type(exc).__name__
+    return any(hint in name for hint in _NON_RETRYABLE_NAME_HINTS)
+
+
+def _extract_retry_after_seconds(exc: BaseException) -> float | None:
+    """Return seconds-to-wait suggested by the provider, if any.
+
+    Looks at ``exc.response.headers`` or ``exc.headers`` for the standard
+    HTTP ``Retry-After`` header (in seconds) or its millisecond cousin
+    ``retry-after-ms`` (sometimes used by Anthropic / OpenAI). Falls back
+    to a regex on the exception message for shapes like
+    ``"Please retry after 30s"``.
+    """
+    headers: dict[str, Any] | None = None
+    response = getattr(exc, "response", None)
+    if response is not None:
+        headers = getattr(response, "headers", None)
+    if headers is None:
+        headers = getattr(exc, "headers", None)
+
+    if isinstance(headers, dict):
+        # Normalize keys to lowercase for case-insensitive matching
+        norm = {str(k).lower(): v for k, v in headers.items()}
+        ms = norm.get("retry-after-ms")
+        if ms is not None:
+            try:
+                return float(ms) / 1000.0
+            except (TypeError, ValueError):
+                pass
+        seconds = norm.get("retry-after")
+        if seconds is not None:
+            try:
+                return float(seconds)
+            except (TypeError, ValueError):
+                pass
+
+    # Last resort: scan the message for "retry after Xs" or "X seconds"
+    msg = str(exc)
+    match = re.search(r"retry\s+after\s+([0-9]+(?:\.[0-9]+)?)", msg, re.IGNORECASE)
+    if match:
+        try:
+            return float(match.group(1))
+        except ValueError:
+            return None
+    return None
+
+
+def _exponential_delay(
+    attempt: int,
+    *,
+    initial_delay: float,
+    backoff_factor: float,
+    max_delay: float,
+    jitter: bool,
+) -> float:
+    """Compute an exponential-backoff delay with optional ±25% jitter."""
+    delay = (
+        initial_delay * (backoff_factor**attempt) if backoff_factor else initial_delay
+    )
+    delay = min(delay, max_delay)
+    if jitter and delay > 0:
+        delay *= 1 + random.uniform(-0.25, 0.25)
+    return max(delay, 0.0)
+
+
+class RetryAfterMiddleware(AgentMiddleware[AgentState[ResponseT], ContextT, ResponseT]):
+    """Retry middleware that honors provider-issued Retry-After hints.
+
+    Drop-in replacement for :class:`langchain.agents.middleware.ModelRetryMiddleware`
+    when working with LiteLLM/Anthropic/OpenAI providers that surface
+    rate-limit hints in headers. Always emits ``surfsense.retrying`` SSE
+    events so the UI can show a friendly "rate limited, retrying in Xs"
+    indicator.
+
+    Args:
+        max_retries: Maximum retries after the initial attempt (default 3).
+        initial_delay: Initial backoff delay in seconds.
+        backoff_factor: Exponential growth factor for backoff.
+        max_delay: Cap on per-attempt delay in seconds.
+        jitter: Whether to add ±25% jitter.
+        retry_on: Optional callable that returns True for retryable
+            exceptions. The default retries everything except known
+            non-retryable classes (context overflow, auth, etc.).
+    """
+
+    def __init__(
+        self,
+        *,
+        max_retries: int = 3,
+        initial_delay: float = 1.0,
+        backoff_factor: float = 2.0,
+        max_delay: float = 60.0,
+        jitter: bool = True,
+        retry_on: Callable[[BaseException], bool] | None = None,
+    ) -> None:
+        super().__init__()
+        self.max_retries = max_retries
+        self.initial_delay = initial_delay
+        self.backoff_factor = backoff_factor
+        self.max_delay = max_delay
+        self.jitter = jitter
+        self._retry_on: Callable[[BaseException], bool] = retry_on or (
+            lambda exc: not _is_non_retryable(exc)
+        )
+
+    def _should_retry(self, exc: BaseException) -> bool:
+        try:
+            return bool(self._retry_on(exc))
+        except Exception:
+            logger.exception("retry_on callable raised; defaulting to False")
+            return False
+
+    def _delay_for_attempt(self, attempt: int, exc: BaseException) -> float:
+        backoff = _exponential_delay(
+            attempt,
+            initial_delay=self.initial_delay,
+            backoff_factor=self.backoff_factor,
+            max_delay=self.max_delay,
+            jitter=self.jitter,
+        )
+        header = _extract_retry_after_seconds(exc) or 0.0
+        return max(backoff, header)
+
+    def wrap_model_call(  # type: ignore[override]
+        self,
+        request: ModelRequest[ContextT],
+        handler: Callable[[ModelRequest[ContextT]], ModelResponse[ResponseT]],
+    ) -> ModelResponse[ResponseT] | AIMessage:
+        for attempt in range(self.max_retries + 1):
+            try:
+                return handler(request)
+            except Exception as exc:
+                if not self._should_retry(exc) or attempt >= self.max_retries:
+                    raise
+                delay = self._delay_for_attempt(attempt, exc)
+                try:
+                    dispatch_custom_event(
+                        "surfsense.retrying",
+                        {
+                            "attempt": attempt + 1,
+                            "max_retries": self.max_retries,
+                            "delay_ms": int(delay * 1000),
+                            "reason": type(exc).__name__,
+                        },
+                    )
+                except Exception:
+                    logger.debug(
+                        "dispatch_custom_event failed; suppressed", exc_info=True
+                    )
+                if delay > 0:
+                    time.sleep(delay)
+        # Unreachable
+        raise RuntimeError("RetryAfterMiddleware: retry loop exited without resolution")
+
+    async def awrap_model_call(  # type: ignore[override]
+        self,
+        request: ModelRequest[ContextT],
+        handler: Callable[
+            [ModelRequest[ContextT]], Awaitable[ModelResponse[ResponseT]]
+        ],
+    ) -> ModelResponse[ResponseT] | AIMessage:
+        for attempt in range(self.max_retries + 1):
+            try:
+                return await handler(request)
+            except Exception as exc:
+                if not self._should_retry(exc) or attempt >= self.max_retries:
+                    raise
+                delay = self._delay_for_attempt(attempt, exc)
+                try:
+                    await adispatch_custom_event(
+                        "surfsense.retrying",
+                        {
+                            "attempt": attempt + 1,
+                            "max_retries": self.max_retries,
+                            "delay_ms": int(delay * 1000),
+                            "reason": type(exc).__name__,
+                        },
+                    )
+                except Exception:
+                    logger.debug(
+                        "adispatch_custom_event failed; suppressed", exc_info=True
+                    )
+                if delay > 0:
+                    await asyncio.sleep(delay)
+        raise RuntimeError("RetryAfterMiddleware: retry loop exited without resolution")
+
+
+__all__ = [
+    "RetryAfterMiddleware",
+    "_extract_retry_after_seconds",
+    "_is_non_retryable",
+]

surfsense_backend/app/agents/new_chat/middleware/safe_summarization.py

@@ -1,123 +0,0 @@
-"""Safe wrapper around deepagents' SummarizationMiddleware.
-
-Upstream issue
---------------
-`deepagents.middleware.summarization.SummarizationMiddleware._aoffload_to_backend`
-(and its sync counterpart) call
-``get_buffer_string(filtered_messages)`` before writing the evicted history
-to the backend file. In recent ``langchain-core`` versions, ``get_buffer_string``
-accesses ``m.text`` which iterates ``self.content`` — this raises
-``TypeError: 'NoneType' object is not iterable`` whenever an ``AIMessage``
-has ``content=None`` (common when a model returns *only* tool_calls, seen
-frequently with Azure OpenAI ``gpt-5.x`` responses streamed through
-LiteLLM).
-
-The exception aborts the whole agent turn, so the user just sees "Error during
-chat" with no assistant response.
-
-Fix
----
-We subclass ``SummarizationMiddleware`` and override
-``_filter_summary_messages`` — the only call site that feeds messages into
-``get_buffer_string`` — to return *copies* of messages whose ``content`` is
-``None`` with ``content=""``. The originals flowing through the rest of the
-agent state are untouched.
-
-We also expose a drop-in ``create_safe_summarization_middleware`` factory
-that mirrors ``deepagents.middleware.summarization.create_summarization_middleware``
-but instantiates our safe subclass.
-"""
-
-from __future__ import annotations
-
-import logging
-from typing import TYPE_CHECKING
-
-from deepagents.middleware.summarization import (
-    SummarizationMiddleware,
-    compute_summarization_defaults,
-)
-
-if TYPE_CHECKING:
-    from deepagents.backends.protocol import BACKEND_TYPES
-    from langchain_core.language_models import BaseChatModel
-    from langchain_core.messages import AnyMessage
-
-logger = logging.getLogger(__name__)
-
-
-def _sanitize_message_content(msg: AnyMessage) -> AnyMessage:
-    """Return ``msg`` with ``content`` coerced to a non-``None`` value.
-
-    ``get_buffer_string`` reads ``m.text`` which iterates ``self.content``;
-    when a provider streams back an ``AIMessage`` with only tool_calls and
-    no text, ``content`` can be ``None`` and the iteration explodes. We
-    replace ``None`` with an empty string so downstream consumers that only
-    care about text see an empty body.
-
-    The original message is left untouched — we return a copy via
-    pydantic's ``model_copy`` when available, otherwise we fall back to
-    re-setting the attribute on a shallow copy.
-    """
-
-    if getattr(msg, "content", "not-missing") is not None:
-        return msg
-
-    try:
-        return msg.model_copy(update={"content": ""})
-    except AttributeError:
-        import copy
-
-        new_msg = copy.copy(msg)
-        try:
-            new_msg.content = ""
-        except Exception:  # pragma: no cover - defensive
-            logger.debug(
-                "Could not sanitize content=None on message of type %s",
-                type(msg).__name__,
-            )
-            return msg
-        return new_msg
-
-
-class SafeSummarizationMiddleware(SummarizationMiddleware):
-    """`SummarizationMiddleware` that tolerates messages with ``content=None``.
-
-    Only ``_filter_summary_messages`` is overridden — this is the single
-    helper invoked by both the sync and async offload paths immediately
-    before ``get_buffer_string``. Normalising here means we get coverage
-    for both without having to copy the (long, rapidly-changing) offload
-    implementations from upstream.
-    """
-
-    def _filter_summary_messages(self, messages: list[AnyMessage]) -> list[AnyMessage]:
-        filtered = super()._filter_summary_messages(messages)
-        return [_sanitize_message_content(m) for m in filtered]
-
-
-def create_safe_summarization_middleware(
-    model: BaseChatModel,
-    backend: BACKEND_TYPES,
-) -> SafeSummarizationMiddleware:
-    """Drop-in replacement for ``create_summarization_middleware``.
-
-    Mirrors the defaults computed by ``deepagents`` but returns our
-    ``SafeSummarizationMiddleware`` subclass so the
-    ``content=None`` crash in ``get_buffer_string`` is avoided.
-    """
-
-    defaults = compute_summarization_defaults(model)
-    return SafeSummarizationMiddleware(
-        model=model,
-        backend=backend,
-        trigger=defaults["trigger"],
-        keep=defaults["keep"],
-        trim_tokens_to_summarize=None,
-        truncate_args_settings=defaults["truncate_args_settings"],
-    )
-
-
-__all__ = [
-    "SafeSummarizationMiddleware",
-    "create_safe_summarization_middleware",
-]

surfsense_backend/app/agents/new_chat/middleware/skills_backends.py

@@ -0,0 +1,337 @@
+"""Skills backends for SurfSense.
+
+Implements two minimal :class:`deepagents.backends.protocol.BackendProtocol`
+subclasses tailored for use with :class:`deepagents.middleware.skills.SkillsMiddleware`.
+
+The middleware only needs four methods to load skills from a backend:
+
+* ``ls_info`` / ``als_info`` — list directories under a source path.
+* ``download_files`` / ``adownload_files`` — fetch ``SKILL.md`` bytes.
+
+Other ``BackendProtocol`` methods (``read``/``write``/``edit``/``grep_raw`` …)
+default to ``NotImplementedError`` from the base class. They are never reached
+by the skills middleware because skill content is rendered into the system
+prompt at agent build time, not edited at runtime.
+
+Two backends are provided:
+
+* :class:`BuiltinSkillsBackend` — disk-backed read of bundled skills from
+  ``app/agents/new_chat/skills/builtin/``.
+* :class:`SearchSpaceSkillsBackend` — a thin read-only wrapper over
+  :class:`KBPostgresBackend` that filters notes under the privileged folder
+  ``/documents/_skills/``.
+
+Both backends are intentionally read-only: skill authoring happens out of band
+(via filesystem or a search-space-admin route), so we never expose
+``write`` / ``edit`` / ``upload_files``. The base class' ``NotImplementedError``
+gives a clean failure mode if anything tries.
+"""
+
+from __future__ import annotations
+
+import contextlib
+import logging
+from collections.abc import Callable
+from dataclasses import replace
+from pathlib import Path
+from typing import TYPE_CHECKING
+
+from deepagents.backends.composite import CompositeBackend
+from deepagents.backends.protocol import (
+    BackendProtocol,
+    FileDownloadResponse,
+    FileInfo,
+)
+from deepagents.backends.state import StateBackend
+
+if TYPE_CHECKING:
+    from langchain.tools import ToolRuntime
+
+    from app.agents.new_chat.middleware.kb_postgres_backend import KBPostgresBackend
+
+logger = logging.getLogger(__name__)
+
+
+# Limit per Agent Skills spec; matches deepagents.middleware.skills.MAX_SKILL_FILE_SIZE.
+_MAX_SKILL_FILE_SIZE = 10 * 1024 * 1024
+
+
+def _default_builtin_root() -> Path:
+    """Return the absolute path to the bundled builtin skills directory.
+
+    Located at ``app/agents/new_chat/skills/builtin/`` relative to this module.
+    """
+    return (Path(__file__).resolve().parent.parent / "skills" / "builtin").resolve()
+
+
+class BuiltinSkillsBackend(BackendProtocol):
+    """Read-only disk-backed skills source.
+
+    Maps a virtual ``/skills/builtin/`` namespace onto a directory on local disk,
+    where each skill is its own subdirectory containing a ``SKILL.md`` file::
+
+        <root>/<skill-name>/SKILL.md
+
+    The middleware calls :meth:`als_info` with the source path and expects a
+    ``list[FileInfo]`` whose ``is_dir=True`` entries are descended into. Then it
+    calls :meth:`adownload_files` with the synthesized ``SKILL.md`` paths and
+    parses YAML frontmatter from the returned ``content`` bytes.
+
+    Mounting under :class:`~deepagents.backends.composite.CompositeBackend` at
+    prefix ``/skills/builtin/`` means the middleware can issue paths like
+    ``/skills/builtin/kb-research/SKILL.md`` which the composite strips down to
+    ``/kb-research/SKILL.md`` before forwarding here. We treat any leading
+    slash as anchoring at :attr:`root`.
+    """
+
+    def __init__(self, root: Path | str | None = None) -> None:
+        self.root: Path = Path(root).resolve() if root else _default_builtin_root()
+        if not self.root.exists():
+            logger.info(
+                "BuiltinSkillsBackend root %s does not exist; skills will be empty.",
+                self.root,
+            )
+
+    def _resolve(self, path: str) -> Path:
+        """Resolve a virtual posix path under :attr:`root`, refusing escapes."""
+        bare = path.lstrip("/")
+        candidate = (self.root / bare).resolve() if bare else self.root
+        # Refuse symlink/.. traversal that escapes the root.
+        try:
+            candidate.relative_to(self.root)
+        except ValueError as exc:
+            raise ValueError(f"path {path!r} escapes builtin skills root") from exc
+        return candidate
+
+    def ls_info(self, path: str) -> list[FileInfo]:
+        try:
+            target = self._resolve(path)
+        except ValueError as exc:
+            logger.warning("BuiltinSkillsBackend.ls_info refused: %s", exc)
+            return []
+        if not target.exists() or not target.is_dir():
+            return []
+
+        infos: list[FileInfo] = []
+        # Build virtual paths anchored at "/" because CompositeBackend already
+        # stripped the route prefix before calling us.
+        target_virtual = (
+            "/"
+            if target == self.root
+            else ("/" + str(target.relative_to(self.root)).replace("\\", "/"))
+        )
+        for child in sorted(target.iterdir()):
+            child_virtual = (
+                target_virtual.rstrip("/") + "/" + child.name
+                if target_virtual != "/"
+                else "/" + child.name
+            )
+            info: FileInfo = {
+                "path": child_virtual,
+                "is_dir": child.is_dir(),
+            }
+            if child.is_file():
+                with contextlib.suppress(OSError):  # pragma: no cover - defensive
+                    info["size"] = child.stat().st_size
+            infos.append(info)
+        return infos
+
+    def download_files(self, paths: list[str]) -> list[FileDownloadResponse]:
+        responses: list[FileDownloadResponse] = []
+        for p in paths:
+            try:
+                target = self._resolve(p)
+            except ValueError:
+                responses.append(FileDownloadResponse(path=p, error="invalid_path"))
+                continue
+            if not target.exists():
+                responses.append(FileDownloadResponse(path=p, error="file_not_found"))
+                continue
+            if target.is_dir():
+                responses.append(FileDownloadResponse(path=p, error="is_directory"))
+                continue
+            try:
+                # Hard cap to avoid loading rogue mega-files into memory.
+                size = target.stat().st_size
+                if size > _MAX_SKILL_FILE_SIZE:
+                    logger.warning(
+                        "Builtin skill file %s exceeds %d bytes; truncating.",
+                        target,
+                        _MAX_SKILL_FILE_SIZE,
+                    )
+                    with target.open("rb") as fh:
+                        content = fh.read(_MAX_SKILL_FILE_SIZE)
+                else:
+                    content = target.read_bytes()
+            except PermissionError:
+                responses.append(
+                    FileDownloadResponse(path=p, error="permission_denied")
+                )
+                continue
+            except OSError as exc:  # pragma: no cover - defensive
+                logger.warning("Builtin skill read failed %s: %s", target, exc)
+                responses.append(FileDownloadResponse(path=p, error="file_not_found"))
+                continue
+            responses.append(FileDownloadResponse(path=p, content=content, error=None))
+        return responses
+
+
+class SearchSpaceSkillsBackend(BackendProtocol):
+    """Read-only view of search-space-authored skills.
+
+    Wraps a :class:`KBPostgresBackend` and only ever reads under the privileged
+    folder ``/documents/_skills/`` (configurable). The folder is intended to be
+    writable only by search-space admins; this backend never writes.
+
+    The skills middleware expects a layout like::
+
+        /<source_root>/<skill-name>/SKILL.md
+
+    But the KB stores documents like ``/documents/_skills/<name>/SKILL.md``.
+    We expose the inner namespace by remapping each path. When mounted under
+    :class:`CompositeBackend` at prefix ``/skills/space/`` the paths the
+    middleware sees become ``/skills/space/<name>/SKILL.md``; the composite
+    strips ``/skills/space/`` and hands us ``/<name>/SKILL.md``, which we
+    rewrite to ``/documents/_skills/<name>/SKILL.md`` before forwarding to the
+    KB.
+
+    No new database table is needed: the privileged folder convention is
+    enforced server-side outside of this class. We intentionally swallow any
+    write/edit attempts (the base class raises ``NotImplementedError``).
+    """
+
+    DEFAULT_KB_ROOT: str = "/documents/_skills"
+
+    def __init__(
+        self,
+        kb_backend: KBPostgresBackend,
+        *,
+        kb_root: str = DEFAULT_KB_ROOT,
+    ) -> None:
+        self._kb = kb_backend
+        # Normalize trailing slash off so we can join cleanly.
+        self._kb_root = kb_root.rstrip("/") or "/"
+
+    def _to_kb(self, path: str) -> str:
+        """Rewrite a virtual path into the underlying KB namespace."""
+        bare = path.lstrip("/")
+        if not bare:
+            return self._kb_root
+        return f"{self._kb_root}/{bare}"
+
+    def _from_kb(self, kb_path: str) -> str:
+        """Rewrite a KB path back into our virtual namespace."""
+        if not kb_path.startswith(self._kb_root):
+            return kb_path  # pragma: no cover - defensive
+        rel = kb_path[len(self._kb_root) :]
+        return rel if rel.startswith("/") else "/" + rel
+
+    def ls_info(self, path: str) -> list[FileInfo]:
+        # KBPostgresBackend exposes only the async API meaningfully; the sync
+        # path falls back to ``asyncio.to_thread(...)`` in the base class. We
+        # keep this stub to satisfy abstract resolution; the middleware calls
+        # ``als_info``.
+        raise NotImplementedError("SearchSpaceSkillsBackend is async-only")
+
+    async def als_info(self, path: str) -> list[FileInfo]:
+        kb_path = self._to_kb(path)
+        try:
+            infos = await self._kb.als_info(kb_path)
+        except Exception as exc:  # pragma: no cover - defensive
+            logger.warning("SearchSpaceSkillsBackend.als_info failed: %s", exc)
+            return []
+        remapped: list[FileInfo] = []
+        for info in infos:
+            kb_p = info.get("path", "")
+            if not kb_p.startswith(self._kb_root):
+                continue
+            remapped.append({**info, "path": self._from_kb(kb_p)})
+        return remapped
+
+    def download_files(self, paths: list[str]) -> list[FileDownloadResponse]:
+        raise NotImplementedError("SearchSpaceSkillsBackend is async-only")
+
+    async def adownload_files(self, paths: list[str]) -> list[FileDownloadResponse]:
+        kb_paths = [self._to_kb(p) for p in paths]
+        responses = await self._kb.adownload_files(kb_paths)
+        # Re-map response paths back to the virtual namespace so the middleware
+        # correlates them to the input list correctly.
+        remapped: list[FileDownloadResponse] = []
+        for original, resp in zip(paths, responses, strict=True):
+            remapped.append(replace(resp, path=original))
+        return remapped
+
+
+SKILLS_BUILTIN_PREFIX = "/skills/builtin/"
+SKILLS_SPACE_PREFIX = "/skills/space/"
+
+
+def build_skills_backend_factory(
+    *,
+    builtin_root: Path | str | None = None,
+    search_space_id: int | None = None,
+) -> Callable[[ToolRuntime], BackendProtocol]:
+    """Return a runtime-aware factory for the skills :class:`CompositeBackend`.
+
+    When ``search_space_id`` is provided the composite includes a
+    :class:`SearchSpaceSkillsBackend` route at ``/skills/space/`` over a fresh
+    per-runtime :class:`KBPostgresBackend`, mirroring how
+    :func:`build_backend_resolver` constructs the main filesystem backend.
+
+    When ``search_space_id`` is ``None`` (e.g., desktop-local mode or unit
+    tests) only the bundled :class:`BuiltinSkillsBackend` is exposed.
+
+    Returning a factory rather than a fixed instance is intentional: the
+    underlying KB backend depends on per-call ``ToolRuntime`` state
+    (``staged_dirs``, ``files`` cache, runtime config), so a single shared
+    instance cannot serve multiple concurrent agent runs.
+    """
+    builtin = BuiltinSkillsBackend(builtin_root)
+
+    if search_space_id is None:
+
+        def _factory_builtin_only(runtime: ToolRuntime) -> BackendProtocol:
+            # Default StateBackend is intentionally inert: any path outside the
+            # ``/skills/builtin/`` route resolves to an empty per-runtime state
+            # so the SkillsMiddleware can iterate sources without raising.
+            return CompositeBackend(
+                default=StateBackend(runtime),
+                routes={SKILLS_BUILTIN_PREFIX: builtin},
+            )
+
+        return _factory_builtin_only
+
+    def _factory_with_space(runtime: ToolRuntime) -> BackendProtocol:
+        # Imported lazily to avoid a hard dependency at module import time:
+        # ``KBPostgresBackend`` pulls in DB models, which are unnecessary for
+        # the unit-tested builtin path.
+        from app.agents.new_chat.middleware.kb_postgres_backend import (
+            KBPostgresBackend,
+        )
+
+        kb = KBPostgresBackend(search_space_id, runtime)
+        space = SearchSpaceSkillsBackend(kb)
+        return CompositeBackend(
+            default=StateBackend(runtime),
+            routes={
+                SKILLS_BUILTIN_PREFIX: builtin,
+                SKILLS_SPACE_PREFIX: space,
+            },
+        )
+
+    return _factory_with_space
+
+
+def default_skills_sources() -> list[str]:
+    """Return the canonical source list for SkillsMiddleware (built-in then space)."""
+    return [SKILLS_BUILTIN_PREFIX, SKILLS_SPACE_PREFIX]
+
+
+__all__ = [
+    "SKILLS_BUILTIN_PREFIX",
+    "SKILLS_SPACE_PREFIX",
+    "BuiltinSkillsBackend",
+    "SearchSpaceSkillsBackend",
+    "build_skills_backend_factory",
+    "default_skills_sources",
+]

surfsense_backend/app/agents/new_chat/middleware/tool_call_repair.py

@@ -0,0 +1,193 @@
+"""
+ToolCallNameRepairMiddleware — two-stage tool-name repair.
+
+Operation:
+1. **Stage 1 — lowercase repair:** if a tool call's ``name`` is not in
+   the registry but ``name.lower()`` is, rewrite in place. Catches
+   models that emit ``Search`` instead of ``search``.
+2. **Stage 2 — invalid fallback:** if still unmatched, rewrite the call
+   to ``invalid`` with ``args={"tool": original_name, "error": <error>}``
+   so the registered :func:`invalid_tool` returns the error to the model
+   for self-correction.
+
+Ported from OpenCode's ``packages/opencode/src/session/llm.ts:339-358``
++ ``packages/opencode/src/tool/invalid.ts``. LangChain has no equivalent:
+:class:`deepagents.middleware.PatchToolCallsMiddleware` patches
+*dangling* tool calls (no matching ToolMessage) but does nothing about
+wrong names, and the model framework's default behavior on an unknown
+name is to crash the turn rather than route to a self-correction
+fallback.
+"""
+
+from __future__ import annotations
+
+import difflib
+import logging
+from typing import Any
+
+from langchain.agents.middleware.types import (
+    AgentMiddleware,
+    AgentState,
+    ContextT,
+    ResponseT,
+)
+from langchain_core.messages import AIMessage
+from langgraph.runtime import Runtime
+
+from app.agents.new_chat.tools.invalid_tool import INVALID_TOOL_NAME
+
+logger = logging.getLogger(__name__)
+
+
+def _coerce_existing_tool_call(call: Any) -> dict[str, Any]:
+    """Normalize a tool call entry to a mutable dict."""
+    if isinstance(call, dict):
+        return dict(call)
+    return {
+        "name": getattr(call, "name", None),
+        "args": getattr(call, "args", {}),
+        "id": getattr(call, "id", None),
+        "type": "tool_call",
+    }
+
+
+class ToolCallNameRepairMiddleware(
+    AgentMiddleware[AgentState[ResponseT], ContextT, ResponseT]
+):
+    """Two-stage tool-name repair on the most recent ``AIMessage``.
+
+    Args:
+        registered_tool_names: Set of canonically-registered tool names.
+            ``invalid`` should be in this set so the fallback dispatches.
+        fuzzy_match_threshold: Optional ``difflib`` ratio (0-1) for the
+            fuzzy-match step that runs *between* lowercase and invalid.
+            Set to ``None`` to disable fuzzy matching (default in
+            OpenCode; we mirror that to avoid silent rewrites).
+    """
+
+    def __init__(
+        self,
+        *,
+        registered_tool_names: set[str],
+        fuzzy_match_threshold: float | None = 0.85,
+    ) -> None:
+        super().__init__()
+        self._registered = set(registered_tool_names)
+        self._registered_lower = {name.lower(): name for name in self._registered}
+        self._fuzzy_threshold = fuzzy_match_threshold
+        self.tools = []
+
+    def _registered_for_runtime(self, runtime: Runtime[ContextT]) -> set[str]:
+        """Allow runtime overrides to expand the set (e.g. dynamic MCP tools)."""
+        ctx_tools = getattr(runtime.context, "registered_tool_names", None)
+        if isinstance(ctx_tools, set | frozenset):
+            return self._registered | set(ctx_tools)
+        if isinstance(ctx_tools, list | tuple):
+            return self._registered | set(ctx_tools)
+        return self._registered
+
+    def _repair_one(
+        self,
+        call: dict[str, Any],
+        registered: set[str],
+    ) -> dict[str, Any]:
+        name = call.get("name")
+        if not isinstance(name, str):
+            return call
+
+        if name in registered:
+            return call
+
+        # Stage 1 — lowercase
+        lowered = name.lower()
+        if lowered in registered:
+            call["name"] = lowered
+            metadata = dict(call.get("response_metadata") or {})
+            metadata.setdefault("repair", "lowercase")
+            call["response_metadata"] = metadata
+            return call
+
+        # Optional fuzzy step (off by default — see class docstring)
+        if self._fuzzy_threshold is not None:
+            close = difflib.get_close_matches(
+                name, registered, n=1, cutoff=self._fuzzy_threshold
+            )
+            if close:
+                call["name"] = close[0]
+                metadata = dict(call.get("response_metadata") or {})
+                metadata.setdefault("repair", f"fuzzy:{name}->{close[0]}")
+                call["response_metadata"] = metadata
+                return call
+
+        # Stage 2 — invalid fallback
+        if INVALID_TOOL_NAME in registered:
+            original_args = call.get("args") or {}
+            error_msg = (
+                f"Tool name '{name}' is not registered. "
+                f"Original arguments were: {original_args!r}."
+            )
+            call["name"] = INVALID_TOOL_NAME
+            call["args"] = {"tool": name, "error": error_msg}
+            metadata = dict(call.get("response_metadata") or {})
+            metadata.setdefault("repair", f"invalid_fallback:{name}")
+            call["response_metadata"] = metadata
+        else:
+            logger.warning(
+                "Could not repair unknown tool call %r; 'invalid' tool not registered",
+                name,
+            )
+        return call
+
+    def _maybe_repair(
+        self,
+        message: AIMessage,
+        registered: set[str],
+    ) -> AIMessage | None:
+        if not message.tool_calls:
+            return None
+
+        new_calls: list[dict[str, Any]] = []
+        any_changed = False
+        for raw in message.tool_calls:
+            call = _coerce_existing_tool_call(raw)
+            before = (call.get("name"), call.get("args"))
+            repaired = self._repair_one(call, registered)
+            after = (repaired.get("name"), repaired.get("args"))
+            if before != after:
+                any_changed = True
+            new_calls.append(repaired)
+
+        if not any_changed:
+            return None
+
+        return message.model_copy(update={"tool_calls": new_calls})
+
+    def after_model(  # type: ignore[override]
+        self,
+        state: AgentState[ResponseT],
+        runtime: Runtime[ContextT],
+    ) -> dict[str, Any] | None:
+        messages = state.get("messages") or []
+        if not messages:
+            return None
+        last = messages[-1]
+        if not isinstance(last, AIMessage):
+            return None
+
+        registered = self._registered_for_runtime(runtime)
+        repaired = self._maybe_repair(last, registered)
+        if repaired is None:
+            return None
+        return {"messages": [repaired]}
+
+    async def aafter_model(  # type: ignore[override]
+        self,
+        state: AgentState[ResponseT],
+        runtime: Runtime[ContextT],
+    ) -> dict[str, Any] | None:
+        return self.after_model(state, runtime)
+
+
+__all__ = [
+    "ToolCallNameRepairMiddleware",
+]

surfsense_backend/app/agents/new_chat/path_resolver.py

@@ -0,0 +1,351 @@
+"""Canonical virtual-path resolver for SurfSense knowledge-base documents.
+
+This module is the single source of truth for mapping ``Document`` rows to
+virtual paths under ``/documents/`` and back. It is used by:
+
+* :class:`KnowledgeTreeMiddleware` (rendering the workspace tree)
+* :class:`KnowledgePriorityMiddleware` (computing priority paths)
+* :class:`KBPostgresBackend` (``als_info`` / ``aread`` / move operations)
+* :class:`KnowledgeBasePersistenceMiddleware` (resolving moves and creates)
+
+Centralising the logic ensures that title-collision suffixes, folder paths,
+and ``unique_identifier_hash`` lookups never drift between renders and
+commits.
+"""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass, field
+
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.db import Document, DocumentType, Folder
+from app.utils.document_converters import generate_unique_identifier_hash
+
+DOCUMENTS_ROOT = "/documents"
+"""Root virtual folder for all KB documents."""
+
+_INVALID_FILENAME_CHARS = re.compile(r"[\\/:*?\"<>|]+")
+_WHITESPACE_RUN = re.compile(r"\s+")
+
+
+def safe_filename(value: str, *, fallback: str = "untitled.xml") -> str:
+    """Convert arbitrary text into a filesystem-safe ``.xml`` filename."""
+    name = _INVALID_FILENAME_CHARS.sub("_", value).strip()
+    name = _WHITESPACE_RUN.sub(" ", name)
+    if not name:
+        name = fallback
+    if len(name) > 180:
+        name = name[:180].rstrip()
+    if not name.lower().endswith(".xml"):
+        name = f"{name}.xml"
+    return name
+
+
+def safe_folder_segment(value: str, *, fallback: str = "folder") -> str:
+    """Sanitize a single folder name into a path-safe segment."""
+    name = _INVALID_FILENAME_CHARS.sub("_", value).strip()
+    name = _WHITESPACE_RUN.sub(" ", name)
+    if not name:
+        return fallback
+    if len(name) > 180:
+        name = name[:180].rstrip()
+    return name
+
+
+def _suffix_with_doc_id(filename: str, doc_id: int | None) -> str:
+    if doc_id is None:
+        return filename
+    if not filename.lower().endswith(".xml"):
+        return f"{filename} ({doc_id}).xml"
+    stem = filename[:-4]
+    return f"{stem} ({doc_id}).xml"
+
+
+_SUFFIX_PATTERN = re.compile(r"\s\((\d+)\)\.xml$", re.IGNORECASE)
+
+
+def parse_doc_id_suffix(filename: str) -> tuple[str, int | None]:
+    """Strip a trailing ``" (<doc_id>).xml"`` suffix; return ``(stem, doc_id)``.
+
+    If no suffix is present, returns ``(stem_without_xml_extension, None)``.
+    """
+    match = _SUFFIX_PATTERN.search(filename)
+    if match:
+        doc_id = int(match.group(1))
+        stem = filename[: match.start()]
+        return stem, doc_id
+    if filename.lower().endswith(".xml"):
+        return filename[:-4], None
+    return filename, None
+
+
+@dataclass
+class PathIndex:
+    """In-memory occupancy snapshot used by :func:`doc_to_virtual_path`.
+
+    Built once per call site so collision handling is deterministic and so
+    we don't perform N folder lookups per render.
+    """
+
+    folder_paths: dict[int, str] = field(default_factory=dict)
+    """``Folder.id`` -> absolute virtual folder path under ``/documents``."""
+
+    occupants: dict[str, int] = field(default_factory=dict)
+    """virtual path -> ``Document.id`` already occupying that path (this render)."""
+
+
+async def _build_folder_paths(
+    session: AsyncSession,
+    search_space_id: int,
+) -> dict[int, str]:
+    """Compute ``Folder.id`` -> absolute virtual path under ``/documents``."""
+    result = await session.execute(
+        select(Folder.id, Folder.name, Folder.parent_id).where(
+            Folder.search_space_id == search_space_id
+        )
+    )
+    rows = result.all()
+    by_id = {row.id: {"name": row.name, "parent_id": row.parent_id} for row in rows}
+    cache: dict[int, str] = {}
+
+    def resolve(folder_id: int) -> str:
+        if folder_id in cache:
+            return cache[folder_id]
+        parts: list[str] = []
+        cursor: int | None = folder_id
+        visited: set[int] = set()
+        while cursor is not None and cursor in by_id and cursor not in visited:
+            visited.add(cursor)
+            entry = by_id[cursor]
+            parts.append(safe_folder_segment(str(entry["name"])))
+            cursor = entry["parent_id"]
+        parts.reverse()
+        path = f"{DOCUMENTS_ROOT}/" + "/".join(parts) if parts else DOCUMENTS_ROOT
+        cache[folder_id] = path
+        return path
+
+    for folder_id in by_id:
+        resolve(folder_id)
+    return cache
+
+
+async def build_path_index(
+    session: AsyncSession,
+    search_space_id: int,
+    *,
+    populate_occupants: bool = True,
+) -> PathIndex:
+    """Build a :class:`PathIndex` for a search space.
+
+    ``populate_occupants`` controls whether the occupancy map is pre-seeded
+    from existing ``Document`` rows. Most callers want this so that
+    :func:`doc_to_virtual_path` can detect collisions across the whole space;
+    the persistence middleware sets this to ``False`` when it is iterating to
+    decide where to place fresh documents.
+    """
+    folder_paths = await _build_folder_paths(session, search_space_id)
+    occupants: dict[str, int] = {}
+    if populate_occupants:
+        rows = await session.execute(
+            select(Document.id, Document.title, Document.folder_id).where(
+                Document.search_space_id == search_space_id,
+            )
+        )
+        for row in rows.all():
+            base = folder_paths.get(row.folder_id, DOCUMENTS_ROOT)
+            filename = safe_filename(str(row.title or "untitled"))
+            path = f"{base}/{filename}"
+            if path in occupants and occupants[path] != row.id:
+                path = f"{base}/{_suffix_with_doc_id(filename, row.id)}"
+            occupants[path] = row.id
+    return PathIndex(folder_paths=folder_paths, occupants=occupants)
+
+
+def doc_to_virtual_path(
+    *,
+    doc_id: int | None,
+    title: str,
+    folder_id: int | None,
+    index: PathIndex,
+) -> str:
+    """Return the canonical virtual path for a document.
+
+    Mutates ``index.occupants`` so subsequent calls see this assignment and
+    deterministically pick a different suffix for the next colliding doc.
+    """
+    base = index.folder_paths.get(folder_id, DOCUMENTS_ROOT)
+    filename = safe_filename(str(title or "untitled"))
+    path = f"{base}/{filename}"
+    occupant = index.occupants.get(path)
+    if occupant is not None and occupant != doc_id:
+        path = f"{base}/{_suffix_with_doc_id(filename, doc_id)}"
+    if doc_id is not None:
+        index.occupants[path] = doc_id
+    return path
+
+
+async def virtual_path_to_doc(
+    session: AsyncSession,
+    *,
+    search_space_id: int,
+    virtual_path: str,
+) -> Document | None:
+    """Resolve a virtual path back to a ``Document`` row.
+
+    Resolution order:
+    1. ``Document.unique_identifier_hash`` lookup (fast path for paths created
+       by SurfSense itself — every NOTE write goes through this hash).
+    2. If the basename carries a ``" (<doc_id>).xml"`` disambiguation suffix,
+       try a direct id lookup constrained to the search space.
+    3. Title-from-basename + folder-resolution lookup as a last resort.
+    """
+    if not virtual_path or not virtual_path.startswith(DOCUMENTS_ROOT):
+        return None
+
+    unique_hash = generate_unique_identifier_hash(
+        DocumentType.NOTE,
+        virtual_path,
+        search_space_id,
+    )
+    result = await session.execute(
+        select(Document).where(
+            Document.search_space_id == search_space_id,
+            Document.unique_identifier_hash == unique_hash,
+        )
+    )
+    document = result.scalar_one_or_none()
+    if document is not None:
+        return document
+
+    rel = virtual_path[len(DOCUMENTS_ROOT) :].lstrip("/")
+    if not rel:
+        return None
+    parts = [p for p in rel.split("/") if p]
+    if not parts:
+        return None
+    basename = parts[-1]
+    folder_parts = parts[:-1]
+
+    stem, suffix_doc_id = parse_doc_id_suffix(basename)
+    if suffix_doc_id is not None:
+        result = await session.execute(
+            select(Document).where(
+                Document.search_space_id == search_space_id,
+                Document.id == suffix_doc_id,
+            )
+        )
+        document = result.scalar_one_or_none()
+        if document is not None:
+            return document
+
+    folder_id = await _resolve_folder_id(
+        session, search_space_id=search_space_id, folder_parts=folder_parts
+    )
+    title_candidates: list[str] = []
+    raw_title = stem
+    title_candidates.append(raw_title)
+    if raw_title.endswith(".xml"):
+        title_candidates.append(raw_title[:-4])
+
+    for candidate in dict.fromkeys(title_candidates):
+        if not candidate:
+            continue
+        query = select(Document).where(
+            Document.search_space_id == search_space_id,
+            Document.title == candidate,
+        )
+        if folder_id is None:
+            query = query.where(Document.folder_id.is_(None))
+        else:
+            query = query.where(Document.folder_id == folder_id)
+        result = await session.execute(query)
+        document = result.scalars().first()
+        if document is not None:
+            return document
+
+    # Fallback: title-as-string lookup misses when the real DB title contains
+    # characters that ``safe_filename`` lossily replaces (``:``, ``/``, ``*``,
+    # etc.) — common for connector-imported docs (Google Calendar/Drive etc.).
+    # The workspace tree shows the lossy filename, so the agent passes that
+    # filename back here. Scan all documents in the resolved folder and match
+    # by ``safe_filename(title)`` to recover the original document.
+    folder_scan = select(Document).where(
+        Document.search_space_id == search_space_id,
+    )
+    if folder_id is None:
+        folder_scan = folder_scan.where(Document.folder_id.is_(None))
+    else:
+        folder_scan = folder_scan.where(Document.folder_id == folder_id)
+    result = await session.execute(folder_scan)
+    for candidate_doc in result.scalars().all():
+        encoded = safe_filename(str(candidate_doc.title or "untitled"))
+        if encoded == basename:
+            return candidate_doc
+    return None
+
+
+async def _resolve_folder_id(
+    session: AsyncSession,
+    *,
+    search_space_id: int,
+    folder_parts: list[str],
+) -> int | None:
+    """Look up the leaf folder id for a chain of folder names; return ``None`` if missing."""
+    if not folder_parts:
+        return None
+    parent_id: int | None = None
+    for raw in folder_parts:
+        name = safe_folder_segment(raw)
+        query = select(Folder.id).where(
+            Folder.search_space_id == search_space_id,
+            Folder.name == name,
+        )
+        if parent_id is None:
+            query = query.where(Folder.parent_id.is_(None))
+        else:
+            query = query.where(Folder.parent_id == parent_id)
+        result = await session.execute(query)
+        row = result.first()
+        if row is None:
+            return None
+        parent_id = row[0]
+    return parent_id
+
+
+def parse_documents_path(virtual_path: str) -> tuple[list[str], str]:
+    """Parse a ``/documents/...`` path into ``(folder_parts, document_title)``.
+
+    The title has any ``.xml`` extension and trailing ``" (<doc_id>)"``
+    disambiguation suffix stripped.
+    """
+    if not virtual_path or not virtual_path.startswith(DOCUMENTS_ROOT):
+        return [], ""
+    rel = virtual_path[len(DOCUMENTS_ROOT) :].strip("/")
+    if not rel:
+        return [], ""
+    parts = [p for p in rel.split("/") if p]
+    if not parts:
+        return [], ""
+    folder_parts = parts[:-1]
+    basename = parts[-1]
+    stem, _ = parse_doc_id_suffix(basename)
+    title = stem
+    if title.endswith(".xml"):
+        title = title[:-4]
+    return folder_parts, title
+
+
+__all__ = [
+    "DOCUMENTS_ROOT",
+    "PathIndex",
+    "build_path_index",
+    "doc_to_virtual_path",
+    "parse_doc_id_suffix",
+    "parse_documents_path",
+    "safe_filename",
+    "safe_folder_segment",
+    "virtual_path_to_doc",
+]

surfsense_backend/app/agents/new_chat/permissions.py

@@ -0,0 +1,203 @@
+"""
+Wildcard pattern matching + rule evaluation for the SurfSense permission system.
+
+Ported from OpenCode's ``packages/opencode/src/permission/evaluate.ts`` and
+``packages/opencode/src/util/wildcard.ts``. LangChain has no rule-based
+permission evaluator, so we keep OpenCode's semantics intact:
+
+- ``Wildcard.match`` matches both the ``permission`` and the ``pattern``
+  fields of a rule against the requested ``(permission, pattern)`` pair.
+  ``*`` matches any segment, ``**`` matches across separators.
+- The evaluator runs ``findLast`` over the **flattened** list of rules
+  from all rulesets — last matching rule wins.
+- The default fallback is ``ask`` (NOT deny), matching OpenCode.
+- Multi-pattern requests AND together: if ANY pattern resolves to
+  ``deny``, the whole request is denied; if ANY needs ``ask``, an
+  interrupt is raised; only when all patterns ``allow`` does the
+  request proceed.
+"""
+
+from __future__ import annotations
+
+import re
+from collections.abc import Iterable
+from dataclasses import dataclass, field
+from typing import Literal
+
+RuleAction = Literal["allow", "deny", "ask"]
+
+
+@dataclass(frozen=True)
+class Rule:
+    """A single permission rule.
+
+    Attributes:
+        permission: A wildcard-matched permission identifier
+            (e.g. ``"edit"``, ``"linear_*"``, ``"mcp:*"``,
+            ``"doom_loop"``). Anchored at start AND end of the input.
+        pattern: A wildcard-matched pattern over the request payload
+            (e.g. ``"/documents/secrets/**"``, ``"page_id=123"``,
+            ``"*"``). Anchored at start AND end.
+        action: One of ``"allow"`` / ``"deny"`` / ``"ask"``.
+    """
+
+    permission: str
+    pattern: str
+    action: RuleAction
+
+
+@dataclass
+class Ruleset:
+    """A list of rules with an associated origin used for debugging."""
+
+    rules: list[Rule] = field(default_factory=list)
+    origin: str = "unknown"  # e.g. "defaults", "global", "space", "thread", "runtime"
+
+
+# -----------------------------------------------------------------------------
+# Wildcard matcher
+# -----------------------------------------------------------------------------
+
+
+_GLOB_TOKEN = re.compile(r"\*\*|\*|[^*]+")
+
+
+def _wildcard_to_regex(pattern: str) -> re.Pattern[str]:
+    """Translate an opencode-style wildcard pattern to a compiled regex.
+
+    Rules:
+    - ``**`` matches any sequence of any characters (including separators).
+    - ``*`` matches any sequence of characters that does **not** include
+      the path separator ``/`` — same as glob.
+    - All other characters match literally.
+    - The pattern is anchored at both ends (``^...$``).
+    """
+    parts: list[str] = ["^"]
+    for token in _GLOB_TOKEN.findall(pattern):
+        if token == "**":
+            parts.append(r".*")
+        elif token == "*":
+            parts.append(r"[^/]*")
+        else:
+            parts.append(re.escape(token))
+    parts.append("$")
+    return re.compile("".join(parts))
+
+
+_REGEX_CACHE: dict[str, re.Pattern[str]] = {}
+
+
+def wildcard_match(value: str, pattern: str) -> bool:
+    """Return True if ``value`` matches the wildcard ``pattern``.
+
+    Special case: a bare ``"*"`` pattern matches any value, including
+    those containing ``/`` separators. This mirrors opencode's
+    ``Wildcard.match`` short-circuit and matches the convention that
+    ``pattern="*"`` means "any pattern" in permission rules.
+    """
+    if pattern == "*":
+        return True
+    compiled = _REGEX_CACHE.get(pattern)
+    if compiled is None:
+        compiled = _wildcard_to_regex(pattern)
+        _REGEX_CACHE[pattern] = compiled
+    return compiled.match(value) is not None
+
+
+# -----------------------------------------------------------------------------
+# Evaluator
+# -----------------------------------------------------------------------------
+
+
+def evaluate(
+    permission: str,
+    pattern: str,
+    *rulesets: Ruleset | Iterable[Rule],
+) -> Rule:
+    """Find the last rule matching ``(permission, pattern)`` from ``rulesets``.
+
+    Mirrors opencode ``permission/evaluate.ts:9-15`` precisely:
+    - Flatten rulesets in argument order.
+    - Walk the flat list **in reverse**.
+    - First reverse-match wins (i.e. the last specified rule wins).
+    - When no rule matches, default to ``Rule(permission, "*", "ask")``.
+
+    Args:
+        permission: The permission identifier being requested
+            (e.g. tool name, ``"edit"``, ``"doom_loop"``).
+        pattern: The request-specific pattern (e.g. file path,
+            primary arg value). Use ``"*"`` when no specific pattern
+            applies.
+        *rulesets: Layered rulesets, applied earliest to latest. Later
+            rulesets override earlier ones.
+
+    Returns:
+        The matched :class:`Rule`, or the default ask fallback.
+    """
+    flat: list[Rule] = []
+    for rs in rulesets:
+        if isinstance(rs, Ruleset):
+            flat.extend(rs.rules)
+        else:
+            flat.extend(rs)
+
+    for rule in reversed(flat):
+        if wildcard_match(permission, rule.permission) and wildcard_match(
+            pattern, rule.pattern
+        ):
+            return rule
+
+    return Rule(permission=permission, pattern="*", action="ask")
+
+
+def evaluate_many(
+    permission: str,
+    patterns: Iterable[str],
+    *rulesets: Ruleset | Iterable[Rule],
+) -> list[Rule]:
+    """Evaluate ``permission`` against each of ``patterns`` (multi-pattern AND).
+
+    Returns the list of resolved rules in the same order as ``patterns``.
+    The caller is responsible for combining the results — opencode-style
+    multi-pattern AND collapses ``deny`` first, then ``ask``, then
+    ``allow``.
+    """
+    return [evaluate(permission, p, *rulesets) for p in patterns]
+
+
+def aggregate_action(rules: Iterable[Rule]) -> RuleAction:
+    """Collapse a list of per-pattern rules into one action.
+
+    Order:
+    1. If any rule is ``deny`` -> ``deny``.
+    2. Else if any rule is ``ask`` -> ``ask``.
+    3. Else if at least one rule is ``allow`` -> ``allow``.
+    4. Else (empty input) -> ``ask`` (safe default mirroring ``evaluate``).
+
+    Mirrors opencode's behavior in ``permission/index.ts:180-272``.
+    """
+    saw_ask = False
+    saw_allow = False
+    for rule in rules:
+        if rule.action == "deny":
+            return "deny"
+        if rule.action == "ask":
+            saw_ask = True
+        elif rule.action == "allow":
+            saw_allow = True
+    if saw_ask:
+        return "ask"
+    if saw_allow:
+        return "allow"
+    return "ask"
+
+
+__all__ = [
+    "Rule",
+    "RuleAction",
+    "Ruleset",
+    "aggregate_action",
+    "evaluate",
+    "evaluate_many",
+    "wildcard_match",
+]

surfsense_backend/app/agents/new_chat/plugin_loader.py

@@ -0,0 +1,158 @@
+"""Entry-point based plugin loader for SurfSense agent middleware.
+
+LangChain's :class:`AgentMiddleware` ABC already covers the practical
+surface most plugins need (``before_agent`` / ``before_model`` /
+``wrap_tool_call`` / their async counterparts), so a SurfSense-specific
+plugin protocol would be redundant. We just need a way to discover and
+admit third-party middleware safely.
+
+A plugin is therefore just an installable Python package that registers a
+factory callable under the ``surfsense.plugins`` entry-point group:
+
+.. code-block:: toml
+
+    # in a plugin package's pyproject.toml
+    [project.entry-points."surfsense.plugins"]
+    year_substituter = "my_plugin:make_middleware"
+
+The factory has the signature ``Callable[[PluginContext], AgentMiddleware]``.
+It receives a small, sanitized :class:`PluginContext` with the IDs and the
+LLM the plugin is allowed to talk to — and **never** raw secrets, DB
+sessions, or other connectors.
+
+## Trust model
+
+Plugins are loaded **only if** their entry-point ``name`` appears in
+``allowed_plugins`` (admin-controlled, sourced from
+``global_llm_config.yaml`` or :func:`load_allowed_plugin_names_from_env`).
+There is **no env-driven auto-load**. A plugin failure is logged and
+isolated; it does not break agent construction.
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+from collections.abc import Iterable
+from importlib.metadata import entry_points
+from typing import TYPE_CHECKING
+
+from langchain.agents.middleware import AgentMiddleware
+
+if TYPE_CHECKING:  # pragma: no cover - type-only
+    from langchain_core.language_models import BaseChatModel
+
+    from app.db import ChatVisibility
+
+
+logger = logging.getLogger(__name__)
+
+
+PLUGIN_ENTRY_POINT_GROUP = "surfsense.plugins"
+
+
+class PluginContext(dict):
+    """Sanitized DI bag handed to each plugin factory.
+
+    Backed by ``dict`` so plugins can inspect the keys they care about
+    without coupling to a concrete dataclass shape. Required keys:
+
+    * ``search_space_id`` (int)
+    * ``user_id`` (str | None)
+    * ``thread_visibility`` (:class:`app.db.ChatVisibility`)
+    * ``llm`` (:class:`langchain_core.language_models.BaseChatModel`)
+
+    The context **never** carries DB sessions, raw secrets, or other
+    connectors. If a future plugin genuinely needs DB access, that
+    integration goes through a rate-limited service interface, not
+    through this bag.
+    """
+
+    @classmethod
+    def build(
+        cls,
+        *,
+        search_space_id: int,
+        user_id: str | None,
+        thread_visibility: ChatVisibility,
+        llm: BaseChatModel,
+    ) -> PluginContext:
+        return cls(
+            search_space_id=search_space_id,
+            user_id=user_id,
+            thread_visibility=thread_visibility,
+            llm=llm,
+        )
+
+
+def load_plugin_middlewares(
+    ctx: PluginContext,
+    allowed_plugin_names: Iterable[str],
+) -> list[AgentMiddleware]:
+    """Discover, allowlist-filter, and instantiate plugin middleware.
+
+    For each entry-point in :data:`PLUGIN_ENTRY_POINT_GROUP` whose name is
+    in ``allowed_plugin_names``, load the factory and call it with ``ctx``.
+    The factory's return value must be an :class:`AgentMiddleware` instance;
+    anything else is logged and skipped.
+
+    Errors are isolated — a plugin that raises during ``ep.load()`` or
+    factory invocation is logged at ``ERROR`` and ignored. Agent
+    construction continues with whatever plugins did succeed.
+    """
+    allowed = {name for name in allowed_plugin_names if name}
+    if not allowed:
+        return []
+
+    out: list[AgentMiddleware] = []
+    try:
+        eps = entry_points(group=PLUGIN_ENTRY_POINT_GROUP)
+    except Exception:  # pragma: no cover - defensive (entry_points is robust)
+        logger.exception("Failed to enumerate plugin entry points")
+        return []
+
+    for ep in eps:
+        if ep.name not in allowed:
+            logger.info("Skipping non-allowlisted plugin %s", ep.name)
+            continue
+        try:
+            factory = ep.load()
+        except Exception:
+            logger.exception("Failed to load plugin %s", ep.name)
+            continue
+        try:
+            mw = factory(ctx)
+        except Exception:
+            logger.exception("Plugin %s factory raised", ep.name)
+            continue
+        if not isinstance(mw, AgentMiddleware):
+            logger.warning(
+                "Plugin %s returned %s, expected AgentMiddleware; skipping",
+                ep.name,
+                type(mw).__name__,
+            )
+            continue
+        out.append(mw)
+        logger.info("Loaded plugin %s as %s", ep.name, type(mw).__name__)
+    return out
+
+
+def load_allowed_plugin_names_from_env() -> set[str]:
+    """Read ``SURFSENSE_ALLOWED_PLUGINS`` (comma-separated) into a set.
+
+    Provided as a thin convenience for deployments that don't surface plugins
+    through ``global_llm_config.yaml`` yet. Whitespace is stripped and empty
+    entries are dropped.
+    """
+    raw = os.environ.get("SURFSENSE_ALLOWED_PLUGINS", "").strip()
+    if not raw:
+        return set()
+    return {token.strip() for token in raw.split(",") if token.strip()}
+
+
+__all__ = [
+    "PLUGIN_ENTRY_POINT_GROUP",
+    "PluginContext",
+    "load_allowed_plugin_names_from_env",
+    "load_plugin_middlewares",
+]

surfsense_backend/app/agents/new_chat/plugins/__init__.py

@@ -0,0 +1,6 @@
+"""Reference plugins bundled with SurfSense.
+
+These plugins are intentionally small and demonstrative. They are NOT
+auto-loaded — they ship as examples that a deployment can opt into via
+``global_llm_config.yaml`` or ``SURFSENSE_ALLOWED_PLUGINS``.
+"""

surfsense_backend/app/agents/new_chat/plugins/year_substituter.py

@@ -0,0 +1,88 @@
+"""Reference plugin: substitute ``{{year}}`` in tool descriptions.
+
+Demonstrates the :meth:`AgentMiddleware.awrap_tool_call` hook -- the
+plugin sees every tool invocation and can rewrite the request *or* the
+result. This particular plugin is read-only and only transforms the
+*description* the user might see in error messages (no request
+mutation).
+
+The plugin is built as a factory function so the entry-point loader can
+inject :class:`PluginContext` (containing the agent's LLM, search-space
+ID, etc.). The factory signature
+``Callable[[PluginContext], AgentMiddleware]`` is the only contract --
+SurfSense doesn't define a custom plugin protocol on top of LangChain's
+:class:`AgentMiddleware`.
+
+Wire-up in ``pyproject.toml`` (illustrative; the in-repo plugin doesn't
+need this -- it's already on the import path)::
+
+    [project.entry-points."surfsense.plugins"]
+    year_substituter = "app.agents.new_chat.plugins.year_substituter:make_middleware"
+"""
+
+from __future__ import annotations
+
+import logging
+from collections.abc import Awaitable, Callable
+from datetime import UTC, datetime
+from typing import TYPE_CHECKING, Any
+
+from langchain.agents.middleware import AgentMiddleware
+
+if TYPE_CHECKING:  # pragma: no cover - type-only
+    from langchain.agents.middleware.types import ToolCallRequest
+    from langchain_core.messages import ToolMessage
+    from langgraph.types import Command
+
+    from app.agents.new_chat.plugin_loader import PluginContext
+
+
+logger = logging.getLogger(__name__)
+
+
+class _YearSubstituterMiddleware(AgentMiddleware):
+    """Replace ``{{year}}`` in the result text with the current UTC year."""
+
+    tools = ()
+
+    def __init__(self, year: int | None = None) -> None:
+        super().__init__()
+        self._year = str(year if year is not None else datetime.now(UTC).year)
+
+    async def awrap_tool_call(
+        self,
+        request: ToolCallRequest,
+        handler: Callable[[ToolCallRequest], Awaitable[ToolMessage | Command[Any]]],
+    ) -> ToolMessage | Command[Any]:
+        result = await handler(request)
+        try:
+            from langchain_core.messages import ToolMessage
+
+            if (
+                isinstance(result, ToolMessage)
+                and isinstance(result.content, str)
+                and "{{year}}" in result.content
+            ):
+                new_text = result.content.replace("{{year}}", self._year)
+                result = ToolMessage(
+                    content=new_text,
+                    tool_call_id=result.tool_call_id,
+                    id=result.id,
+                    name=result.name,
+                    status=result.status,
+                    artifact=result.artifact,
+                )
+        except Exception:  # pragma: no cover - defensive
+            logger.exception("year_substituter plugin failed; passing original result")
+        return result
+
+
+def make_middleware(ctx: PluginContext) -> AgentMiddleware:
+    """Plugin factory used by :func:`load_plugin_middlewares`."""
+    # Plugin is intentionally small so it has no state to threading-protect
+    # and ignores ``ctx`` beyond demonstrating that the loader passes it in.
+    _ = ctx
+    return _YearSubstituterMiddleware()
+
+
+__all__ = ["make_middleware"]

surfsense_backend/app/agents/new_chat/prompts/__init__.py

@@ -0,0 +1,7 @@
+"""SurfSense agent prompt fragments.
+
+The prompt is composed at runtime by :mod:`composer` from the markdown
+fragments under ``base/``, ``providers/``, ``tools/``, ``examples/``, and
+``routing/``. ``system_prompt.py`` is now a thin wrapper that delegates
+to :func:`composer.compose_system_prompt`.
+"""

surfsense_backend/app/agents/new_chat/prompts/base/__init__.py

@@ -0,0 +1 @@
+

surfsense_backend/app/agents/new_chat/prompts/base/agent_private.md

@@ -0,0 +1,7 @@
+You are SurfSense, a reasoning and acting AI agent designed to answer user questions using the user's personal knowledge base.
+
+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/new_chat/prompts/base/agent_team.md

@@ -0,0 +1,9 @@
+You are SurfSense, a reasoning and acting AI agent designed to answer questions in this team space using the team's shared knowledge base.
+
+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/new_chat/prompts/base/citations_off.md

@@ -0,0 +1,16 @@
+<citation_instructions>
+IMPORTANT: Citations are DISABLED for this configuration.
+
+DO NOT include any citations in your responses. Specifically:
+1. Do NOT use the [citation:chunk_id] format anywhere in your response.
+2. Do NOT reference document IDs, chunk IDs, or source IDs.
+3. Simply provide the information naturally without any citation markers.
+4. Write your response as if you're having a normal conversation, incorporating the information from your knowledge seamlessly.
+
+When answering questions based on documents from the knowledge base:
+- Present the information directly and confidently
+- Do not mention that information comes from specific documents or chunks
+- Integrate facts naturally into your response without attribution markers
+
+Your goal is to provide helpful, informative answers in a clean, readable format without any citation notation.
+</citation_instructions>

surfsense_backend/app/agents/new_chat/prompts/base/citations_on.md

@@ -0,0 +1,90 @@
+<citation_instructions>
+CRITICAL CITATION REQUIREMENTS:
+
+1. For EVERY piece of information you include from the documents, add a citation in the format [citation:chunk_id] where chunk_id is the exact value from the `<chunk id='...'>` tag inside `<document_content>`.
+2. Make sure ALL factual statements from the documents have proper citations.
+3. If multiple chunks support the same point, include all relevant citations [citation:chunk_id1], [citation:chunk_id2].
+4. You MUST use the exact chunk_id values from the `<chunk id='...'>` attributes. Do not create your own citation numbers.
+5. Every citation MUST be in the format [citation:chunk_id] where chunk_id is the exact chunk id value.
+6. Never modify or change the chunk_id - always use the original values exactly as provided in the chunk tags.
+7. Do not return citations as clickable links.
+8. Never format citations as markdown links like "([citation:5](https://example.com))". Always use plain square brackets only.
+9. Citations must ONLY appear as [citation:chunk_id] or [citation:chunk_id1], [citation:chunk_id2] format - never with parentheses, hyperlinks, or other formatting.
+10. Never make up chunk IDs. Only use chunk_id values that are explicitly provided in the `<chunk id='...'>` tags.
+11. If you are unsure about a chunk_id, do not include a citation rather than guessing or making one up.
+
+<document_structure_example>
+The documents you receive are structured like this:
+
+**Knowledge base documents (numeric chunk IDs):**
+<document>
+<document_metadata>
+  <document_id>42</document_id>
+  <document_type>GITHUB_CONNECTOR</document_type>
+  <title><![CDATA[Some repo / file / issue title]]></title>
+  <url><![CDATA[https://example.com]]></url>
+  <metadata_json><![CDATA[{{"any":"other metadata"}}]]></metadata_json>
+</document_metadata>
+
+<document_content>
+  <chunk id='123'><![CDATA[First chunk text...]]></chunk>
+  <chunk id='124'><![CDATA[Second chunk text...]]></chunk>
+</document_content>
+</document>
+
+**Web search results (URL chunk IDs):**
+<document>
+<document_metadata>
+  <document_type>WEB_SEARCH</document_type>
+  <title><![CDATA[Some web search result]]></title>
+  <url><![CDATA[https://example.com/article]]></url>
+</document_metadata>
+
+<document_content>
+  <chunk id='https://example.com/article'><![CDATA[Content from web search...]]></chunk>
+</document_content>
+</document>
+
+IMPORTANT: You MUST cite using the EXACT chunk ids from the `<chunk id='...'>` tags.
+- For knowledge base documents, chunk ids are numeric (e.g. 123, 124) or prefixed (e.g. doc-45).
+- For live web search results, chunk ids are URLs (e.g. https://example.com/article).
+Do NOT cite document_id. Always use the chunk id.
+</document_structure_example>
+
+<citation_format>
+- Every fact from the documents must have a citation in the format [citation:chunk_id] where chunk_id is the EXACT id value from a `<chunk id='...'>` tag
+- Citations should appear at the end of the sentence containing the information they support
+- Multiple citations should be separated by commas: [citation:chunk_id1], [citation:chunk_id2], [citation:chunk_id3]
+- No need to return references section. Just citations in answer.
+- NEVER create your own citation format - use the exact chunk_id values from the documents in the [citation:chunk_id] format
+- NEVER format citations as clickable links or as markdown links like "([citation:5](https://example.com))". Always use plain square brackets only
+- NEVER make up chunk IDs if you are unsure about the chunk_id. It is better to omit the citation than to guess
+- Copy the EXACT chunk id from the XML - if it says `<chunk id='doc-123'>`, use [citation:doc-123]
+- If the chunk id is a URL like `<chunk id='https://example.com/page'>`, use [citation:https://example.com/page]
+</citation_format>
+
+<citation_examples>
+CORRECT citation formats:
+- [citation:5] (numeric chunk ID from knowledge base)
+- [citation:doc-123] (for Surfsense documentation chunks)
+- [citation:https://example.com/article] (URL chunk ID from web search results)
+- [citation:chunk_id1], [citation:chunk_id2], [citation:chunk_id3] (multiple citations)
+
+INCORRECT citation formats (DO NOT use):
+- Using parentheses and markdown links: ([citation:5](https://github.com/MODSetter/SurfSense))
+- Using parentheses around brackets: ([citation:5])
+- Using hyperlinked text: [link to source 5](https://example.com)
+- Using footnote style: ... library¹
+- Making up source IDs when source_id is unknown
+- Using old IEEE format: [1], [2], [3]
+- Using source types instead of IDs: [citation:GITHUB_CONNECTOR] instead of [citation:5]
+</citation_examples>
+
+<citation_output_example>
+Based on your GitHub repositories and video content, Python's asyncio library provides tools for writing concurrent code using the async/await syntax [citation:5]. It's particularly useful for I/O-bound and high-level structured network code [citation:5].
+
+According to web search results, the key advantage of asyncio is that it can improve performance by allowing other code to run while waiting for I/O operations to complete [citation:https://docs.python.org/3/library/asyncio.html]. This makes it excellent for scenarios like web scraping, API calls, database operations, or any situation where your program spends time waiting for external resources.
+
+However, from your video learning, it's important to note that asyncio is not suitable for CPU-bound tasks as it runs on a single thread [citation:12]. For computationally intensive work, you'd want to use multiprocessing instead.
+</citation_output_example>
+</citation_instructions>

surfsense_backend/app/agents/new_chat/prompts/base/kb_only_policy_private.md

@@ -0,0 +1,15 @@
+<knowledge_base_only_policy>
+CRITICAL RULE — KNOWLEDGE BASE FIRST, NEVER DEFAULT TO GENERAL KNOWLEDGE:
+- You MUST answer questions ONLY using information retrieved from the user's knowledge base, web search results, scraped webpages, or other tool outputs.
+- You MUST NOT answer factual or informational questions from your own training data or general knowledge unless the user explicitly grants permission.
+- If the knowledge base search returns no relevant results AND no other tool provides the answer, you MUST:
+  1. Inform the user that you could not find relevant information in their knowledge base.
+  2. Ask the user: "Would you like me to answer from my general knowledge instead?"
+  3. ONLY provide a general-knowledge answer AFTER the user explicitly says yes.
+- This policy does NOT apply to:
+  * Casual conversation, greetings, or meta-questions about SurfSense itself (e.g., "what can you do?")
+  * Formatting, summarization, or analysis of content already present in the conversation
+  * Following user instructions that are clearly task-oriented (e.g., "rewrite this in bullet points")
+  * Tool-usage actions like generating reports, podcasts, images, or scraping webpages
+  * Queries about services that have direct tools (Linear, ClickUp, Jira, Slack, Airtable) — see <tool_routing> below
+</knowledge_base_only_policy>

surfsense_backend/app/agents/new_chat/prompts/base/kb_only_policy_team.md

@@ -0,0 +1,15 @@
+<knowledge_base_only_policy>
+CRITICAL RULE — KNOWLEDGE BASE FIRST, NEVER DEFAULT TO GENERAL KNOWLEDGE:
+- You MUST answer questions ONLY using information retrieved from the team's shared knowledge base, web search results, scraped webpages, or other tool outputs.
+- You MUST NOT answer factual or informational questions from your own training data or general knowledge unless a team member explicitly grants permission.
+- If the knowledge base search returns no relevant results AND no other tool provides the answer, you MUST:
+  1. Inform the team that you could not find relevant information in the shared knowledge base.
+  2. Ask: "Would you like me to answer from my general knowledge instead?"
+  3. ONLY provide a general-knowledge answer AFTER a team member explicitly says yes.
+- This policy does NOT apply to:
+  * Casual conversation, greetings, or meta-questions about SurfSense itself (e.g., "what can you do?")
+  * Formatting, summarization, or analysis of content already present in the conversation
+  * Following user instructions that are clearly task-oriented (e.g., "rewrite this in bullet points")
+  * Tool-usage actions like generating reports, podcasts, images, or scraping webpages
+  * Queries about services that have direct tools (Linear, ClickUp, Jira, Slack, Airtable) — see <tool_routing> below
+</knowledge_base_only_policy>

surfsense_backend/app/agents/new_chat/prompts/base/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/new_chat/prompts/base/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/new_chat/prompts/base/parameter_resolution.md

@@ -0,0 +1,39 @@
+<parameter_resolution>
+Some service tools require identifiers or context you do not have (account IDs,
+workspace names, channel IDs, project keys, etc.). NEVER ask the user for raw
+IDs or technical identifiers — they cannot memorise them.
+
+Instead, follow this discovery pattern:
+1. Call a listing/discovery tool to find available options.
+2. ONE result → use it silently, no question to the user.
+3. MULTIPLE results → present the options by their display names and let the
+   user choose. Never show raw UUIDs — always use friendly names.
+
+Discovery tools by level:
+- Which account/workspace? → get_connected_accounts("<service>")
+- Which Jira site (cloudId)? → getAccessibleAtlassianResources
+- Which Jira project?  → getVisibleJiraProjects (after resolving cloudId)
+- Which Jira issue type? → getJiraProjectIssueTypesMetadata (after resolving project)
+- Which channel?  → slack_search_channels
+- Which base?     → list_bases
+- Which table?    → list_tables_for_base (after resolving baseId)
+- Which task?     → clickup_search
+- Which issue?    → list_issues (Linear) or searchJiraIssuesUsingJql (Jira)
+
+For Jira specifically: ALWAYS call getAccessibleAtlassianResources first to
+obtain the cloudId, then pass it to other Jira tools. When creating an issue,
+chain: getAccessibleAtlassianResources → getVisibleJiraProjects → createJiraIssue.
+If there is only one option at each step, use it silently. If multiple, present
+friendly names.
+
+Chain discovery when needed — e.g. for Airtable records: list_bases → pick
+base → list_tables_for_base → pick table → list_records_for_table.
+
+MULTI-ACCOUNT TOOL NAMING: When the user has multiple accounts connected for
+the same service, tool names are prefixed to avoid collisions — e.g.
+linear_25_list_issues and linear_30_list_issues instead of two list_issues.
+Each prefixed tool's description starts with [Account: <display_name>] so you
+know which account it targets. Use get_connected_accounts("<service>") to see
+the full list of accounts with their connector IDs and display names.
+When only one account is connected, tools have their normal unprefixed names.
+</parameter_resolution>

surfsense_backend/app/agents/new_chat/prompts/base/tool_routing_private.md

@@ -0,0 +1,16 @@
+<tool_routing>
+CRITICAL — You have direct tools for these services: Linear, ClickUp, Jira, Slack, Airtable.
+Their data is NEVER in the knowledge base. You MUST call their tools immediately — never
+say "I don't see it in the knowledge base" or ask the user if they want you to check.
+Ignore any knowledge base results for these services.
+
+When to use which tool:
+- Linear (issues) → list_issues, get_issue, save_issue (create/update)
+- ClickUp (tasks) → clickup_search, clickup_get_task
+- Jira (issues) → getAccessibleAtlassianResources (cloudId discovery), getVisibleJiraProjects (project discovery), getJiraProjectIssueTypesMetadata (issue type discovery), searchJiraIssuesUsingJql, createJiraIssue, editJiraIssue
+- Slack (messages, channels) → slack_search_channels, slack_read_channel, slack_read_thread
+- Airtable (bases, tables, records) → list_bases, list_tables_for_base, list_records_for_table
+- Knowledge base content (Notion, GitHub, files, notes) → automatically searched
+- Real-time public web data → call web_search
+- Reading a specific webpage → call scrape_webpage
+</tool_routing>

surfsense_backend/app/agents/new_chat/prompts/base/tool_routing_team.md

@@ -0,0 +1,16 @@
+<tool_routing>
+CRITICAL — You have direct tools for these services: Linear, ClickUp, Jira, Slack, Airtable.
+Their data is NEVER in the knowledge base. You MUST call their tools immediately — never
+say "I don't see it in the knowledge base" or ask if they want you to check.
+Ignore any knowledge base results for these services.
+
+When to use which tool:
+- Linear (issues) → list_issues, get_issue, save_issue (create/update)
+- ClickUp (tasks) → clickup_search, clickup_get_task
+- Jira (issues) → getAccessibleAtlassianResources (cloudId discovery), getVisibleJiraProjects (project discovery), getJiraProjectIssueTypesMetadata (issue type discovery), searchJiraIssuesUsingJql, createJiraIssue, editJiraIssue
+- Slack (messages, channels) → slack_search_channels, slack_read_channel, slack_read_thread
+- Airtable (bases, tables, records) → list_bases, list_tables_for_base, list_records_for_table
+- Knowledge base content (Notion, GitHub, files, notes) → automatically searched
+- Real-time public web data → call web_search
+- Reading a specific webpage → call scrape_webpage
+</tool_routing>

surfsense_backend/app/agents/new_chat/prompts/composer.py

@@ -0,0 +1,405 @@
+"""
+Prompt composer for the SurfSense ``new_chat`` agent.
+
+This module assembles the agent's system prompt from the markdown fragments
+under :mod:`app.agents.new_chat.prompts`. It replaces the monolithic
+``system_prompt.py`` with a clean, fragment-based composition:
+
+::
+
+    prompts/
+      base/                  # agent identity, KB policy, tool routing, …
+      providers/             # provider-specific tweaks (anthropic, gpt5, …)
+      tools/                 # one ``<name>.md`` per tool
+      examples/              # one ``<name>.md`` per tool with call examples
+      routing/               # connector-specific routing notes (linear, slack, …)
+
+The model-family dispatch step (see :func:`detect_provider_variant`)
+mirrors OpenCode's ``packages/opencode/src/session/system.ts`` — different
+model families respond best to differently-styled prompts (Claude likes
+XML/narrative, GPT-5 wants channel-aware pragmatic, Codex needs
+terse/file:line, Gemini wants formal numbered steps, etc.). LangChain's
+``dynamic_prompt`` helper supports per-call prompt swaps but ships no
+out-of-the-box family classifier, so we keep our own.
+
+Backwards compatibility
+=======================
+
+``system_prompt.py`` re-exports :func:`compose_system_prompt` and wraps it
+in functions with the same signatures as the legacy
+``build_surfsense_system_prompt`` / ``build_configurable_system_prompt`` so
+existing call sites do not change.
+"""
+
+from __future__ import annotations
+
+import re
+from collections.abc import Iterable
+from datetime import UTC, datetime
+from importlib import resources
+
+from app.db import ChatVisibility
+
+# -----------------------------------------------------------------------------
+# Provider variant detection
+# -----------------------------------------------------------------------------
+
+# String literal alias for the supported provider-specific prompt variants.
+# When adding a new variant, also drop a matching ``providers/<variant>.md``
+# file in this package and (if appropriate) extend the regex matchers below.
+#
+# Stylistic clusters: each variant is a focused style nudge, NOT a full
+# system prompt — the main prompt is already assembled from base/ +
+# tools/ + routing/. The clustering itself (which models map to which
+# style) follows OpenCode's ``system.ts`` family table; see the module
+# docstring for credits.
+ProviderVariant = str
+# Known values:
+#   "anthropic"        — Claude family (XML-friendly, narrative todos)
+#   "openai_reasoning" — GPT-5 / o-series (channel-aware pragmatic)
+#   "openai_classic"   — GPT-4 family (autonomous persistence)
+#   "openai_codex"     — gpt-*-codex (code-purist, terse, file:line refs)
+#   "google"           — Gemini (formal, <3-line, numbered workflow)
+#   "kimi"             — Moonshot Kimi-K* (action-bias, parallel tools)
+#   "grok"             — xAI Grok (extreme-terse, one-word ok)
+#   "deepseek"         — DeepSeek V3 / R1 (terse, R1-aware reasoning)
+#   "default"          — fallback, no provider-specific block emitted
+
+# IMPORTANT: order of evaluation matters in :func:`detect_provider_variant`.
+# More specific patterns must come first (e.g. ``codex`` before
+# ``openai_reasoning`` because codex model ids contain ``gpt``).
+
+_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:
+    """Pick a provider-specific prompt variant from a model id string.
+
+    Heuristic match on the model id; returns ``"default"`` when nothing
+    matches so the composer can fall back to the empty placeholder file.
+
+    Order is significant: more-specific patterns are tried first so
+    ``gpt-5-codex`` routes to ``"openai_codex"`` rather than
+    ``"openai_reasoning"`` — same dispatch order as OpenCode's
+    ``packages/opencode/src/session/system.ts``.
+    """
+    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"
+
+
+# -----------------------------------------------------------------------------
+# Fragment loading
+# -----------------------------------------------------------------------------
+
+
+_PROMPTS_PACKAGE = "app.agents.new_chat.prompts"
+
+
+def _read_fragment(subpath: str) -> str:
+    """Read a fragment file from the ``prompts/`` resource tree.
+
+    Returns the raw contents stripped of any single trailing newline so
+    composition can append explicit separators without compounding blank
+    lines. Missing files return an empty string so optional fragments
+    (e.g. provider hints) act as no-ops.
+    """
+    parts = subpath.split("/")
+    try:
+        ref = resources.files(_PROMPTS_PACKAGE).joinpath(*parts)
+        if not ref.is_file():
+            return ""
+        text = ref.read_text(encoding="utf-8")
+    except (FileNotFoundError, ModuleNotFoundError):
+        return ""
+    if text.endswith("\n"):
+        text = text[:-1]
+    return text
+
+
+# -----------------------------------------------------------------------------
+# Tool ordering + memory variant resolution
+# -----------------------------------------------------------------------------
+
+
+# Ordered for reading flow: fundamentals first, then artifact generators,
+# then memory at the end (mirrors the legacy ``_ALL_TOOL_NAMES_ORDERED``).
+ALL_TOOL_NAMES_ORDERED: tuple[str, ...] = (
+    "search_surfsense_docs",
+    "web_search",
+    "generate_podcast",
+    "generate_video_presentation",
+    "generate_report",
+    "generate_resume",
+    "generate_image",
+    "scrape_webpage",
+    "update_memory",
+)
+
+
+_MEMORY_VARIANT_TOOLS: frozenset[str] = frozenset({"update_memory"})
+
+
+def _tool_fragment_path(tool_name: str, variant: str) -> str:
+    """Resolve a tool's instruction fragment path.
+
+    Tools listed in :data:`_MEMORY_VARIANT_TOOLS` switch on the conversation
+    visibility and load ``tools/<name>_<variant>.md``; everything else
+    falls back to ``tools/<name>.md``.
+    """
+    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()
+
+
+# -----------------------------------------------------------------------------
+# Section builders
+# -----------------------------------------------------------------------------
+
+
+def _build_system_instructions(
+    *,
+    visibility: ChatVisibility,
+    resolved_today: str,
+) -> str:
+    """Reconstruct the legacy ``<system_instruction>`` block from fragments."""
+    variant = "team" if visibility == ChatVisibility.SEARCH_SPACE else "private"
+
+    sections = [
+        _read_fragment(f"base/agent_{variant}.md"),
+        _read_fragment(f"base/kb_only_policy_{variant}.md"),
+        _read_fragment(f"base/tool_routing_{variant}.md"),
+        _read_fragment("base/parameter_resolution.md"),
+        _read_fragment(f"base/memory_protocol_{variant}.md"),
+    ]
+    body = "\n\n".join(s for s in sections if s)
+    block = f"\n<system_instruction>\n{body}\n\n</system_instruction>\n"
+    return block.format(resolved_today=resolved_today)
+
+
+def _build_mcp_routing_block(
+    mcp_connector_tools: dict[str, list[str]] | None,
+) -> str:
+    """Emit the ``<mcp_tool_routing>`` block when at least one MCP server is wired."""
+    if not mcp_connector_tools:
+        return ""
+    lines: list[str] = [
+        "\n<mcp_tool_routing>",
+        "You also have direct tools from these user-connected MCP servers.",
+        "Their data is NEVER in the knowledge base — call their tools directly.",
+        "",
+    ]
+    for server_name, tool_names in mcp_connector_tools.items():
+        lines.append(f"- {server_name} → {', '.join(tool_names)}")
+    lines.append("</mcp_tool_routing>\n")
+    return "\n".join(lines)
+
+
+def _build_tools_section(
+    *,
+    visibility: ChatVisibility,
+    enabled_tool_names: set[str] | None,
+    disabled_tool_names: set[str] | None,
+) -> str:
+    """Reconstruct the ``<tools>`` block + ``<tool_call_examples>`` block."""
+    variant = "team" if visibility == ChatVisibility.SEARCH_SPACE else "private"
+
+    parts: list[str] = []
+    preamble = _read_fragment("tools/_preamble.md")
+    if preamble:
+        parts.append(preamble + "\n")
+
+    examples: list[str] = []
+
+    for tool_name in ALL_TOOL_NAMES_ORDERED:
+        if enabled_tool_names is not None and tool_name not in enabled_tool_names:
+            continue
+
+        instruction = _read_fragment(_tool_fragment_path(tool_name, variant))
+        if instruction:
+            parts.append(instruction + "\n")
+
+        example = _read_fragment(_example_fragment_path(tool_name, variant))
+        if example:
+            examples.append(example + "\n")
+
+    known_disabled = (
+        set(disabled_tool_names) & set(ALL_TOOL_NAMES_ORDERED)
+        if disabled_tool_names
+        else set()
+    )
+    if known_disabled:
+        disabled_list = ", ".join(
+            _format_tool_label(n) for n in ALL_TOOL_NAMES_ORDERED if n in known_disabled
+        )
+        parts.append(
+            "\n"
+            "DISABLED TOOLS (by user):\n"
+            f"The following tools are available in SurfSense but have been disabled by the user for this session: {disabled_list}.\n"
+            "You do NOT have access to these tools and MUST NOT claim you can use them.\n"
+            "If the user asks about a capability provided by a disabled tool, let them know the relevant tool\n"
+            "is currently disabled and they can re-enable it.\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)
+
+
+def _build_provider_block(provider_variant: ProviderVariant) -> str:
+    """Optional provider-tuned hints. Empty for ``"default"``."""
+    if not provider_variant or provider_variant == "default":
+        return ""
+    text = _read_fragment(f"providers/{provider_variant}.md")
+    return f"\n{text}\n" if text else ""
+
+
+def _build_routing_block(connector_routing: Iterable[str] | None) -> str:
+    if not connector_routing:
+        return ""
+    fragments: list[str] = []
+    for name in connector_routing:
+        text = _read_fragment(f"routing/{name}.md")
+        if text:
+            fragments.append(text)
+    if not fragments:
+        return ""
+    return "\n" + "\n\n".join(fragments) + "\n"
+
+
+def _build_citation_block(citations_enabled: bool) -> str:
+    fragment = (
+        _read_fragment("base/citations_on.md")
+        if citations_enabled
+        else _read_fragment("base/citations_off.md")
+    )
+    return f"\n{fragment}\n" if fragment else ""
+
+
+# -----------------------------------------------------------------------------
+# Public API
+# -----------------------------------------------------------------------------
+
+
+def compose_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,
+    mcp_connector_tools: dict[str, list[str]] | None = None,
+    custom_system_instructions: str | None = None,
+    use_default_system_instructions: bool = True,
+    citations_enabled: bool = True,
+    provider_variant: ProviderVariant | None = None,
+    model_name: str | None = None,
+    connector_routing: Iterable[str] | None = None,
+) -> str:
+    """Assemble the SurfSense system prompt from disk fragments.
+
+    Args:
+        today: Optional clock injection for tests.
+        thread_visibility: Private vs shared (team) — drives memory wording
+            and a few base block variants.
+        enabled_tool_names: When provided, only these tools' instructions
+            are included; ``None`` keeps the legacy "include everything"
+            behavior.
+        disabled_tool_names: User-disabled tools (note appended to prompt).
+        mcp_connector_tools: ``{server_name: [tool_names...]}`` to inject
+            an explicit MCP routing block.
+        custom_system_instructions: Free-form instructions that override
+            the default ``<system_instruction>`` block (legacy support
+            for ``NewLLMConfig.system_instructions``).
+        use_default_system_instructions: When ``custom_system_instructions``
+            is empty/None, fall back to defaults (legacy semantics).
+        citations_enabled: Include ``citations_on.md`` (true) or
+            ``citations_off.md`` (false).
+        provider_variant: Explicit provider variant override
+            (``"anthropic" | "openai_reasoning" | "openai_classic" | "google" | "default"``).
+            When ``None``, falls back to :func:`detect_provider_variant`
+            on ``model_name``.
+        model_name: Used to auto-detect ``provider_variant`` when not
+            provided explicitly.
+        connector_routing: Optional list of routing fragment names
+            (``["linear", "slack", ...]``) to include from
+            ``prompts/routing/``.
+
+    Returns:
+        The fully composed system prompt string.
+    """
+    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():
+        sys_block = custom_system_instructions.format(resolved_today=resolved_today)
+    elif use_default_system_instructions:
+        sys_block = _build_system_instructions(
+            visibility=visibility, resolved_today=resolved_today
+        )
+    else:
+        sys_block = ""
+
+    sys_block += _build_mcp_routing_block(mcp_connector_tools)
+
+    if provider_variant is None:
+        provider_variant = detect_provider_variant(model_name)
+    sys_block += _build_provider_block(provider_variant)
+    sys_block += _build_routing_block(connector_routing)
+
+    tools_block = _build_tools_section(
+        visibility=visibility,
+        enabled_tool_names=enabled_tool_names,
+        disabled_tool_names=disabled_tool_names,
+    )
+    citation_block = _build_citation_block(citations_enabled)
+
+    return sys_block + tools_block + citation_block
+
+
+__all__ = [
+    "ALL_TOOL_NAMES_ORDERED",
+    "ProviderVariant",
+    "compose_system_prompt",
+    "detect_provider_variant",
+]

surfsense_backend/app/agents/new_chat/prompts/examples/__init__.py

@@ -0,0 +1 @@
+

surfsense_backend/app/agents/new_chat/prompts/examples/generate_image.md

@@ -0,0 +1,12 @@
+
+- User: "Generate an image of a cat"
+  - Call: `generate_image(prompt="A fluffy orange tabby cat sitting on a windowsill, bathed in warm golden sunlight, soft bokeh background with green houseplants, photorealistic style, cozy atmosphere")`
+  - The generated image will automatically be displayed in the chat.
+- User: "Draw me a logo for a coffee shop called Bean Dream"
+  - Call: `generate_image(prompt="Minimalist modern logo design for a coffee shop called 'Bean Dream', featuring a stylized coffee bean with dream-like swirls of steam, clean vector style, warm brown and cream color palette, white background, professional branding")`
+  - The generated image will automatically be displayed in the chat.
+- User: "Show me this image: https://example.com/image.png"
+  - Simply include it in your response using markdown: `![Image](https://example.com/image.png)`
+- User uploads an image file and asks: "What is this image about?"
+  - The user's uploaded image is already visible in the chat.
+  - Simply analyze the image content and respond directly.

surfsense_backend/app/agents/new_chat/prompts/examples/generate_podcast.md

@@ -0,0 +1,7 @@
+
+- User: "Give me a podcast about AI trends based on what we discussed"
+  - First search for relevant content, then call: `generate_podcast(source_content="Based on our conversation and search results: [detailed summary of chat + search findings]", podcast_title="AI Trends Podcast")`
+- User: "Create a podcast summary of this conversation"
+  - Call: `generate_podcast(source_content="Complete conversation summary:\n\nUser asked about [topic 1]:\n[Your detailed response]\n\nUser then asked about [topic 2]:\n[Your detailed response]\n\n[Continue for all exchanges in the conversation]", podcast_title="Conversation Summary")`
+- User: "Make a podcast about quantum computing"
+  - First explore `/documents/` (ls/glob/grep/read_file), then: `generate_podcast(source_content="Key insights about quantum computing from retrieved files:\n\n[Comprehensive summary of findings]", podcast_title="Quantum Computing Explained")`

surfsense_backend/app/agents/new_chat/prompts/examples/generate_report.md

@@ -0,0 +1,13 @@
+
+- User: "Generate a report about AI trends"
+  - Call: `generate_report(topic="AI Trends Report", source_strategy="kb_search", search_queries=["AI trends recent developments", "artificial intelligence industry trends", "AI market growth and predictions"], report_style="detailed")`
+  - WHY: Has creation verb "generate" → call the tool. No prior discussion → use kb_search.
+- User: "Write a research report from this conversation"
+  - Call: `generate_report(topic="Research Report", source_strategy="conversation", source_content="Complete conversation summary:\n\n...", report_style="deep_research")`
+  - WHY: Has creation verb "write" → call the tool. Conversation has the content → use source_strategy="conversation".
+- User: (after a report on Climate Change was generated) "Add a section about carbon capture technologies"
+  - Call: `generate_report(topic="Climate Crisis: Causes, Impacts, and Solutions", source_strategy="conversation", source_content="[summary of conversation context if any]", parent_report_id=<previous_report_id>, user_instructions="Add a new section about carbon capture technologies")`
+  - WHY: Has modification verb "add" + specific deliverable target → call the tool with parent_report_id.
+- User: (after a report was generated) "What else could we add to have more depth?"
+  - Do NOT call generate_report. Answer in chat with suggestions.
+  - WHY: No creation/modification verb directed at producing a deliverable.

surfsense_backend/app/agents/new_chat/prompts/examples/generate_resume.md

@@ -0,0 +1,19 @@
+
+- User: "Build me a resume. I'm John Doe, engineer at Acme Corp..."
+  - Call: `generate_resume(user_info="John Doe, engineer at Acme Corp...", max_pages=1)`
+  - WHY: Has creation verb "build" + resume → call the tool.
+- User: "Create my CV with this info: [experience, education, skills]"
+  - Call: `generate_resume(user_info="[experience, education, skills]", max_pages=1)`
+- User: "Build me a resume" (and there is a resume/CV document in the conversation context)
+  - Extract the FULL content from the document in context, then call:
+    `generate_resume(user_info="Name: John Doe\nEmail: john@example.com\n\nExperience:\n- Senior Engineer at Acme Corp (2020-2024)\n  Led team of 5...\n\nEducation:\n- BS Computer Science, MIT (2016-2020)\n\nSkills: Python, TypeScript, AWS...", max_pages=1)`
+  - WHY: Document content is available in context — extract ALL of it into user_info. Do NOT ignore referenced documents.
+- User: (after resume generated) "Change my title to Senior Engineer"
+  - Call: `generate_resume(user_info="", user_instructions="Change the job title to Senior Engineer", parent_report_id=<previous_report_id>, max_pages=1)`
+  - WHY: Modification verb "change" + refers to existing resume → set parent_report_id.
+- User: (after resume generated) "Make this 2 pages and expand projects"
+  - Call: `generate_resume(user_info="", user_instructions="Expand projects and keep this to at most 2 pages", parent_report_id=<previous_report_id>, max_pages=2)`
+  - WHY: Explicit page increase request → set max_pages to 2.
+- User: "How should I structure my resume?"
+  - Do NOT call generate_resume. Answer in chat with advice.
+  - WHY: No creation/modification verb.

surfsense_backend/app/agents/new_chat/prompts/examples/generate_video_presentation.md

@@ -0,0 +1,7 @@
+
+- User: "Give me a presentation about AI trends based on what we discussed"
+  - First search for relevant content, then call: `generate_video_presentation(source_content="Based on our conversation and search results: [detailed summary of chat + search findings]", video_title="AI Trends Presentation")`
+- User: "Create slides summarizing this conversation"
+  - Call: `generate_video_presentation(source_content="Complete conversation summary:\n\nUser asked about [topic 1]:\n[Your detailed response]\n\nUser then asked about [topic 2]:\n[Your detailed response]\n\n[Continue for all exchanges in the conversation]", video_title="Conversation Summary")`
+- User: "Make a video presentation about quantum computing"
+  - First explore `/documents/` (ls/glob/grep/read_file), then: `generate_video_presentation(source_content="Key insights about quantum computing from retrieved files:\n\n[Comprehensive summary of findings]", video_title="Quantum Computing Explained")`

surfsense_backend/app/agents/new_chat/prompts/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/new_chat/prompts/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")`
+- User: "How do I use Docker to run SurfSense?"
+  - Call: `search_surfsense_docs(query="Docker installation setup")`

surfsense_backend/app/agents/new_chat/prompts/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/new_chat/prompts/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/new_chat/prompts/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")`
+  - Then answer using the returned web results with citations.
+- 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/new_chat/prompts/providers/__init__.py

@@ -0,0 +1 @@
+

surfsense_backend/app/agents/new_chat/prompts/providers/anthropic.md

@@ -0,0 +1,20 @@
+<provider_hints>
+You are running on an Anthropic Claude model.
+
+Structured reasoning:
+- Use XML tags liberally to organise intermediate reasoning when a task is non-trivial. `<thinking>...</thinking>` blocks are encouraged before tool calls or before producing a complex final answer.
+- For multi-step requests, briefly outline a plan inside a `<plan>` block before issuing the first tool call.
+
+Professional objectivity:
+- Prioritise technical accuracy over validating the user's beliefs. Provide direct, factual guidance without unnecessary superlatives, praise, or emotional validation.
+- When uncertain, investigate (search the KB, fetch the page) rather than confirming the user's assumption.
+- Disagree with the user when the evidence warrants it; respectful correction beats false agreement.
+
+Task management:
+- For tasks with 3+ distinct steps use the todo / planning tool aggressively. Mark items in_progress before starting, completed immediately when finished — do not batch completions.
+- Narrate progress through the todo list itself, not through chatty status lines.
+
+Tool calls:
+- Run independent tool calls in parallel within one response. Sequence them only when a later call genuinely needs an earlier one's output.
+- Never chain bash-like commands with `;` or `&&` to "narrate" — use prose between tool calls instead.
+</provider_hints>

surfsense_backend/app/agents/new_chat/prompts/providers/deepseek.md

@@ -0,0 +1,18 @@
+<provider_hints>
+You are running on a DeepSeek model (DeepSeek-V3 chat / DeepSeek-R1 reasoning).
+
+Reasoning hygiene (R1-aware):
+- If the model surfaces explicit `<think>` blocks, keep that internal scratch focused — do NOT restate the user's question inside it; jump straight to the analysis.
+- Never paste the contents of `<think>` into your final answer. Final answer should reflect only the conclusion, citations, and any user-facing rationale.
+- Do not let chain-of-thought leak into tool-call arguments — keep tool inputs minimal and structural.
+
+Output style:
+- Be concise. Default to a one-paragraph answer; expand only when the user asks for detail.
+- Don't open with sycophantic phrasing ("Great question", "Sure, here you go"). Lead with the answer or the next action.
+- For factual answers, cite once with `[citation:chunk_id]` and stop.
+
+Tool calls:
+- Issue independent tool calls in parallel within a single turn.
+- Prefer the knowledge-base search tools before any web-search; this model has strong recall but stale training data.
+- Don't fabricate file paths, chunk ids, or URLs — only use values returned by tools or provided by the user.
+</provider_hints>

surfsense_backend/app/agents/new_chat/prompts/providers/default.md

@@ -0,0 +1 @@
+

surfsense_backend/app/agents/new_chat/prompts/providers/google.md

@@ -0,0 +1,20 @@
+<provider_hints>
+You are running on a Google Gemini model.
+
+Output style:
+- Concise & direct. Aim for fewer than 3 lines of prose (excluding tool output, citations, and code/snippets) when the task allows.
+- No conversational filler — skip openers like "Okay, I will now…" and closers like "I have finished the changes…". Get straight to the action or answer.
+- Format with GitHub-flavoured Markdown; assume monospace rendering.
+- For one-line factual answers, just answer. No headers, no bullets.
+
+Workflow for non-trivial tasks (Understand → Plan → Act → Verify):
+1. **Understand:** read the user's request and the relevant KB / connector context. Use search and read tools (in parallel when independent) before assuming anything.
+2. **Plan:** when the task touches multiple steps, share an extremely concise plan first.
+3. **Act:** call the appropriate tools, strictly adhering to the prompts/routing already established for this agent.
+4. **Verify:** confirm with a follow-up read or search where it materially de-risks the answer.
+
+Discipline:
+- Do not take significant actions beyond the clear scope of the user's request without confirming first.
+- Do not assume a connector / tool / file exists — check (e.g. via `get_connected_accounts`) before referencing it.
+- Path arguments must be the exact strings returned by tools; do not synthesise file paths.
+</provider_hints>

surfsense_backend/app/agents/new_chat/prompts/providers/grok.md

@@ -0,0 +1,17 @@
+<provider_hints>
+You are running on an xAI Grok model.
+
+Maximum terseness:
+- Answer in fewer than 4 lines unless the user asks for detail. One-word answers are best when they suffice.
+- No preamble ("The answer is", "Here's what I'll do"), no postamble ("Hope that helps", "Let me know"). Get straight to the answer.
+- Avoid restating the user's question.
+- For factual lookups inside the knowledge base, give the answer with a single `[citation:chunk_id]` and stop.
+
+Tool discipline:
+- Use exactly ONE tool per assistant turn when investigating; wait for the result before deciding the next call. Do not loop on the same tool with the same arguments — pick a result and act.
+- For obviously parallelizable read-only batches (multiple independent searches), one turn with several tool calls is fine — but never chain into a fishing expedition.
+
+Style:
+- No emojis unless the user asked. No nested bullets, no headers for short answers.
+- If you can't help, say so in 1-2 sentences without explaining "why this could lead to…".
+</provider_hints>

surfsense_backend/app/agents/new_chat/prompts/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+).
+
+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 before stating anything as factual; don't fabricate citations.
+- Keep it stupidly simple. Don't overcomplicate.
+</provider_hints>

surfsense_backend/app/agents/new_chat/prompts/providers/openai_classic.md

@@ -0,0 +1,21 @@
+<provider_hints>
+You are running on a classic OpenAI chat model (GPT-4 family).
+
+Persistence:
+- Keep going until the user's query is completely resolved before yielding back. Don't end the turn at "I would do X" — actually do X.
+- When you say "Next I will…" or "Now I will…", you MUST actually take that action in the same turn.
+- If a tool call fails, diagnose and try again with corrected arguments; do not surface the raw error and stop.
+
+Planning:
+- Plan extensively before each tool call and reflect briefly on the result of the previous call. For tasks with 3+ steps, use the todo / planning tool and mark items as `in_progress` / `completed` as you go.
+- Always announce the next action in ONE concise sentence before making a non-trivial tool call ("I'll search the KB for the migration spec.").
+
+Output style:
+- Conversational but professional. Plain prose for explanations, bullet points for findings, fenced code blocks (with language tags) for code.
+- Don't dump tool output verbatim — summarise the relevant lines.
+- Don't add a closing recap unless the user asked for one. After completing the work, just stop.
+
+Tool calls:
+- Issue independent tool calls in parallel within one response.
+- Use specialised tools over generic ones (e.g. KB search before web search; named connectors over MCP fallback).
+</provider_hints>

surfsense_backend/app/agents/new_chat/prompts/providers/openai_codex.md

@@ -0,0 +1,19 @@
+<provider_hints>
+You are running on an OpenAI Codex-class model (gpt-codex / codex-mini / gpt-*-codex).
+
+Output style:
+- Be concise. Don't dump fetched/searched content back at the user — reference paths or chunk ids instead.
+- Reference sources as `path:line` (or `chunk:<id>`) so they're clickable. Stand-alone paths per reference, even when repeated.
+- Prefer numbered lists (`1.`, `2.`, `3.`) when offering options the user can pick by replying with a single number.
+- Skip headers and heavy formatting for simple confirmations.
+- No emojis, no em-dashes, no nested bullets. Single-level lists only.
+
+Code & structured-output tasks:
+- Lead with a one-sentence explanation of the change before context. Don't open with "Summary:" — jump in.
+- Suggest natural next steps (run tests, diff review, commit) only when they're genuinely the next move.
+- For multi-line snippets use fenced code blocks with a language tag.
+
+Tool calls:
+- Run independent tool calls in parallel; chain only when later calls need earlier results.
+- Don't ask permission ("Should I proceed?") — proceed with the most reasonable default and state what you did.
+</provider_hints>

surfsense_backend/app/agents/new_chat/prompts/providers/openai_reasoning.md

@@ -0,0 +1,21 @@
+<provider_hints>
+You are running on an OpenAI reasoning model (GPT-5+ / o-series).
+
+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.
+- 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. Don't stop at analysis when the user clearly wants the change applied.
+</provider_hints>

surfsense_backend/app/agents/new_chat/prompts/routing/__init__.py

@@ -0,0 +1 @@
+

surfsense_backend/app/agents/new_chat/prompts/routing/jira.md

@@ -0,0 +1 @@
+

surfsense_backend/app/agents/new_chat/prompts/routing/linear.md

@@ -0,0 +1 @@
+

surfsense_backend/app/agents/new_chat/prompts/routing/slack.md

@@ -0,0 +1 @@
+

surfsense_backend/app/agents/new_chat/prompts/tools/__init__.py

@@ -0,0 +1 @@
+

surfsense_backend/app/agents/new_chat/prompts/tools/_preamble.md

@@ -0,0 +1,6 @@
+<tools>
+You have access to the following tools:
+
+IMPORTANT: You can ONLY use the tools listed below. If a capability is not listed here, you do NOT have it.
+Do NOT claim you can do something if the corresponding tool is not listed.
+

surfsense_backend/app/agents/new_chat/prompts/tools/generate_image.md

@@ -0,0 +1,11 @@
+
+- generate_image: Generate images from text descriptions using AI image models.
+  - Use this when the user asks you to create, generate, draw, design, or make an image.
+  - Trigger phrases: "generate an image of", "create a picture of", "draw me", "make an image", "design a logo", "create artwork"
+  - 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 with the generated image metadata. The image will automatically be displayed in the chat.
+  - IMPORTANT: Write a detailed, descriptive prompt for best results. Don't just pass the user's words verbatim -
+    expand and improve the prompt with specific details about style, lighting, composition, and mood.
+  - If the user's request is vague (e.g., "make me an image of a cat"), enhance the prompt with artistic details.

surfsense_backend/app/agents/new_chat/prompts/tools/generate_podcast.md

@@ -0,0 +1,15 @@
+
+- generate_podcast: Generate an audio podcast from provided content.
+  - Use this when the user asks to create, generate, or make a podcast.
+  - Trigger phrases: "give me a podcast about", "create a podcast", "generate a podcast", "make a podcast", "turn this into a podcast"
+  - Args:
+    - source_content: The text content to convert into a podcast. This MUST be comprehensive and include:
+      * If discussing the current conversation: Include a detailed summary of the FULL chat history (all user questions and your responses)
+      * If based on knowledge base search: Include the key findings and insights from the search results
+      * You can combine both: conversation context + search results for richer podcasts
+      * The more detailed the source_content, the better the podcast quality
+    - podcast_title: Optional title for the podcast (default: "SurfSense Podcast")
+    - user_prompt: Optional instructions for podcast style/format (e.g., "Make it casual and fun")
+  - Returns: A task_id for tracking. The podcast will be generated in the background.
+  - IMPORTANT: Only one podcast can be generated at a time. If a podcast is already being generated, the tool will return status "already_generating".
+  - After calling this tool, inform the user that podcast generation has started and they will see the player when it's ready (takes 3-5 minutes).

surfsense_backend/app/agents/new_chat/prompts/tools/generate_report.md

@@ -0,0 +1,39 @@
+
+- generate_report: Generate or revise a structured Markdown report artifact.
+  - WHEN TO CALL THIS TOOL — the message must contain a creation or modification VERB directed at producing a deliverable:
+    * Creation verbs: write, create, generate, draft, produce, summarize into, turn into, make
+    * Modification verbs: revise, update, expand, add (a section), rewrite, make (it shorter/longer/formal)
+    * Example triggers: "generate a report about...", "write a document on...", "add a section about budget", "make the report shorter", "rewrite in formal tone"
+  - WHEN NOT TO CALL THIS TOOL (answer in chat instead):
+    * Questions or discussion about the report: "What can we add?", "What's missing?", "Is the data accurate?", "How could this be improved?"
+    * Suggestions or brainstorming: "What other topics could be covered?", "What else could be added?", "What would make this better?"
+    * Asking for explanations: "Can you explain section 2?", "Why did you include that?", "What does this part mean?"
+    * Quick follow-ups or critiques: "Is the conclusion strong enough?", "Are there any gaps?", "What about the competitors?"
+    * THE TEST: Does the message contain a creation/modification VERB (from the list above) directed at producing or changing a deliverable? If NO verb → answer conversationally in chat. Do NOT assume the user wants a revision just because a report exists in the conversation.
+  - IMPORTANT FORMAT RULE: Reports are ALWAYS generated in Markdown.
+  - Args:
+    - topic: Short title for the report (max ~8 words).
+    - source_content: The text content to base the report on.
+      * For source_strategy="conversation" or "provided": Include a comprehensive summary of the relevant content.
+      * For source_strategy="kb_search": Can be empty or minimal — the tool handles searching internally.
+      * For source_strategy="auto": Include what you have; the tool searches KB if it's not enough.
+    - source_strategy: Controls how the tool collects source material. One of:
+      * "conversation" — The conversation already contains enough context (prior Q&A, discussion, pasted text, scraped pages). Pass a thorough summary as source_content.
+      * "kb_search" — The tool will search the knowledge base internally. Provide search_queries with 1-5 targeted queries.
+      * "auto" — Use source_content if sufficient, otherwise fall back to internal KB search using search_queries.
+      * "provided" — Use only what is in source_content (default, backward-compatible).
+    - search_queries: When source_strategy is "kb_search" or "auto", provide 1-5 specific search queries for the knowledge base. These should be precise, not just the topic name repeated.
+    - report_style: Controls report depth. Options: "detailed" (DEFAULT), "deep_research", "brief".
+      Use "brief" ONLY when the user explicitly asks for a short/concise/one-page report (e.g., "one page", "keep it short", "brief report", "500 words"). Default to "detailed" for all other requests.
+    - user_instructions: Optional specific instructions (e.g., "focus on financial impacts", "include recommendations"). When revising (parent_report_id set), describe WHAT TO CHANGE. If the user mentions a length preference (e.g., "one page", "500 words", "2 pages"), include that VERBATIM here AND set report_style="brief".
+    - parent_report_id: Set this to the report_id from a previous generate_report result when the user wants to MODIFY an existing report. Do NOT set it for new reports or questions about reports.
+  - Returns: A dictionary with status "ready" or "failed", report_id, title, and word_count.
+  - The report is generated immediately in Markdown and displayed inline in the chat.
+  - Export/download formats (PDF, DOCX, HTML, LaTeX, EPUB, ODT, plain text) are produced from the generated Markdown report.
+  - SOURCE STRATEGY DECISION (HIGH PRIORITY — follow this exactly):
+    * If the conversation already has substantive Q&A / discussion on the topic → use source_strategy="conversation" with a comprehensive summary as source_content.
+    * If the user wants a report on a topic not yet discussed → use source_strategy="kb_search" with targeted search_queries.
+    * If you have some content but might need more → use source_strategy="auto" with both source_content and search_queries.
+    * When revising an existing report (parent_report_id set) and the conversation has relevant context → use source_strategy="conversation". The revision will use the previous report content plus your source_content.
+    * NEVER run a separate KB lookup step and then pass those results to generate_report. The tool handles KB search internally.
+  - AFTER CALLING THIS TOOL: Do NOT repeat, summarize, or reproduce the report content in the chat. The report is already displayed as an interactive card that the user can open, read, copy, and export. Simply confirm that the report was generated (e.g., "I've generated your report on [topic]. You can view the Markdown report now, and export it in various formats from the card."). NEVER write out the report text in the chat.

surfsense_backend/app/agents/new_chat/prompts/tools/generate_resume.md

@@ -0,0 +1,30 @@
+
+- generate_resume: Generate or revise a professional resume as a Typst document.
+  - WHEN TO CALL: The user asks to create, build, generate, write, or draft a resume or CV.
+    Also when they ask to modify, update, or revise an existing resume from this conversation.
+  - WHEN NOT TO CALL: General career advice, resume tips, cover letters, or reviewing
+    a resume without making changes. For cover letters, use generate_report instead.
+  - The tool produces Typst source code that is compiled to a PDF preview automatically.
+  - PAGE POLICY:
+    - Default behavior is ONE PAGE. For new resume creation, set max_pages=1 unless the user explicitly asks for more.
+    - If the user requests a longer resume (e.g., "make it 2 pages"), set max_pages to that value.
+  - Args:
+    - user_info: The user's resume content — work experience, education, skills, contact
+      info, etc. Can be structured or unstructured text.
+      CRITICAL: user_info must be COMPREHENSIVE. Do NOT just pass the user's raw message.
+      You MUST gather and consolidate ALL available information:
+        * Content from referenced/mentioned documents (e.g., uploaded resumes, CVs, LinkedIn profiles)
+          that appear in the conversation context — extract and include their FULL content.
+        * Information the user shared across multiple messages in the conversation.
+        * Any relevant details from knowledge base search results in the context.
+      The more complete the user_info, the better the resume. Include names, contact info,
+      work experience with dates, education, skills, projects, certifications — everything available.
+    - user_instructions: Optional style or content preferences (e.g. "emphasize leadership",
+      "keep it to one page"). For revisions, describe what to change.
+    - parent_report_id: Set this when the user wants to MODIFY an existing resume from
+      this conversation. Use the report_id from a previous generate_resume result.
+    - max_pages: Maximum resume length in pages (integer 1-5). Default is 1.
+  - Returns: Dict with status, report_id, title, and content_type.
+  - After calling: Give a brief confirmation. Do NOT paste resume content in chat. Do NOT mention report_id or any internal IDs — the resume card is shown automatically.
+  - VERSIONING: Same rules as generate_report — set parent_report_id for modifications
+    of an existing resume, leave as None for new resumes.

surfsense_backend/app/agents/new_chat/prompts/tools/generate_video_presentation.md

@@ -0,0 +1,9 @@
+
+- generate_video_presentation: Generate a video presentation from provided content.
+  - Use this when the user asks to create a video, presentation, slides, or slide deck.
+  - Trigger phrases: "give me a presentation", "create slides", "generate a video", "make a slide deck", "turn this into a presentation"
+  - Args:
+    - source_content: The text content to turn into a presentation. The more detailed, the better.
+    - video_title: Optional title (default: "SurfSense Presentation")
+    - user_prompt: Optional style instructions (e.g., "Make it technical and detailed")
+  - After calling this tool, inform the user that generation has started and they will see the presentation when it's ready.

surfsense_backend/app/agents/new_chat/prompts/tools/scrape_webpage.md

@@ -0,0 +1,30 @@
+
+- scrape_webpage: Scrape and extract the main content from a webpage.
+  - Use this when the user wants you to READ and UNDERSTAND the actual content of a webpage.
+  - CRITICAL — WHEN TO USE (always attempt scraping, never refuse before trying):
+    * When a user asks to "get", "fetch", "pull", "grab", "scrape", or "read" content from a URL
+    * When the user wants live/dynamic data from a specific webpage (e.g., tables, scores, stats, prices)
+    * When a URL was mentioned earlier in the conversation and the user asks for its actual content
+    * When `/documents/` knowledge-base data is insufficient and the user wants more
+  - Trigger scenarios:
+    * "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?"
+    * "Can you analyze this article?"
+    * "Can you get the live table/data from [URL]?"
+    * "Scrape it" / "Can you scrape that?" (referring to a previously mentioned URL)
+    * "Fetch the content from [URL]"
+    * "Pull the data from that page"
+  - Args:
+    - url: The URL of the webpage to scrape (must be HTTP/HTTPS)
+    - max_length: Maximum content length to return (default: 50000 chars)
+  - Returns: The page title, description, full content (in markdown), word count, and metadata
+  - After scraping, provide a comprehensive, well-structured summary with key takeaways using headings or bullet points.
+  - Reference the source using markdown links [descriptive text](url) — never bare URLs.
+  - IMAGES: The scraped content may contain image URLs in markdown format like `![alt text](image_url)`.
+    * When you find relevant/important images in the scraped content, include them in your response using standard markdown image syntax: `![alt text](image_url)`.
+    * This makes your response more visual and engaging.
+    * Prioritize showing: diagrams, charts, infographics, key illustrations, or images that help explain the content.
+    * Don't show every image - just the most relevant 1-3 images that enhance understanding.

surfsense_backend/app/agents/new_chat/prompts/tools/search_surfsense_docs.md

@@ -0,0 +1,7 @@
+
+- search_surfsense_docs: Search the official SurfSense documentation.
+  - Use this tool when the user asks anything about SurfSense itself (the application they are using).
+  - Args:
+    - query: The search query about SurfSense
+    - top_k: Number of documentation chunks to retrieve (default: 10)
+  - Returns: Documentation content with chunk IDs for citations (prefixed with 'doc-', e.g., [citation:doc-123])

surfsense_backend/app/agents/new_chat/prompts/tools/update_memory_private.md

@@ -0,0 +1,31 @@
+
+- update_memory: Update your personal memory document about the user.
+  - Your current memory is already in <user_memory> in your context.  The `chars` and
+    `limit` attributes show your current usage and the maximum allowed size.
+  - This is your curated long-term memory — the distilled essence of what you know about
+    the user, not raw conversation logs.
+  - Call update_memory when:
+    * The user explicitly asks to remember or forget something
+    * The user shares durable facts or preferences that will matter in future conversations
+  - The user's first name is provided in <user_name>. Use it in memory entries
+    instead of "the user" (e.g. "{name} works at..." not "The user works at...").
+    Do not store the name itself as a separate memory entry.
+  - 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) [marker] text
+    Markers:
+      [fact]  — durable facts (role, background, projects, tools, expertise)
+      [pref]  — preferences (response style, languages, formats, tools)
+      [instr] — standing instructions (always/never do, response rules)
+  - Keep it concise and well under the character limit shown in <user_memory>.
+  - Every entry MUST be under a `##` heading. Keep heading names short (2-3 words) and
+    natural. Do NOT include the user's name in headings. Organize by context — e.g.
+    who they are, what they're focused on, how they prefer things. 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: [instr] > [pref] > [fact].

surfsense_backend/app/agents/new_chat/prompts/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/new_chat/prompts/tools/web_search.md

@@ -0,0 +1,18 @@
+
+- web_search: Search the web for real-time information using all configured search engines.
+  - Use this for current events, news, prices, weather, public facts, or any question requiring
+    up-to-date information from the internet.
+  - This tool dispatches to all configured search engines (SearXNG, Tavily, Linkup, Baidu) in
+    parallel and merges the results.
+  - IMPORTANT (REAL-TIME / PUBLIC WEB QUERIES): For questions that require current public web data
+    (e.g., live exchange rates, stock prices, breaking news, weather, current events), you MUST call
+    `web_search` instead of answering from memory.
+  - For these real-time/public web queries, DO NOT answer from memory and DO NOT say you lack internet
+    access before attempting a web search.
+  - If the search returns no relevant results, explain that web sources did not return enough
+    data and ask the user if they want you to retry with a refined query.
+  - Args:
+    - query: The search query - use specific, descriptive terms
+    - top_k: Number of results to retrieve (default: 10, max: 50)
+  - If search snippets are insufficient for the user's question, use `scrape_webpage` on the most relevant result URL for full content.
+  - When presenting results, reference sources as markdown links [descriptive text](url) — never bare URLs.

surfsense_backend/app/agents/new_chat/skills/__init__.py

@@ -0,0 +1,7 @@
+"""SurfSense built-in agent skills (Anthropic Skills format).
+
+Each subdirectory corresponds to one skill and contains a ``SKILL.md`` file
+with YAML frontmatter (name, description, allowed_tools) plus markdown
+instructions. The :class:`BuiltinSkillsBackend` exposes them to the
+deepagents :class:`SkillsMiddleware`.
+"""

surfsense_backend/app/agents/new_chat/skills/builtin/__init__.py

@@ -0,0 +1 @@
+

surfsense_backend/app/agents/new_chat/skills/builtin/email-drafting/SKILL.md

@@ -0,0 +1,25 @@
+---
+name: email-drafting
+description: Draft an email matching the user's voice, with structured intent and CTA
+allowed-tools: search_surfsense_docs
+---
+
+# Email drafting
+
+## When to use this skill
+"Draft an email to ...", "reply to this thread", "write a follow-up to X". Plain "summarize the email" is **not** in scope — that's a comprehension task.
+
+## Voice
+Search the KB for prior emails from the user to similar audiences (same recipient, same topic class). Mirror tone, opening style, sign-off, and length distribution. If there is no precedent, default to: warm, direct, no filler, short paragraphs, one clear ask.
+
+## Required structure
+Every draft includes, in this order:
+
+1. **Subject line** — concrete, ≤ 8 words, no clickbait, no `Re:` unless replying.
+2. **Opening (1 sentence)** — context the recipient already shares; never restate what they wrote unless the thread is long.
+3. **Body** — the actual point in one short paragraph. Bullets only if there are >3 discrete items.
+4. **Single explicit CTA** — what you want the recipient to do, with a soft deadline if relevant.
+5. **Sign-off** — match the user's prior closing style.
+
+## Always offer alternatives
+End your message with: "Want me to make it shorter, more formal, or add a different angle?" — give the user one obvious next step.

surfsense_backend/app/agents/new_chat/skills/builtin/kb-research/SKILL.md

@@ -0,0 +1,23 @@
+---
+name: kb-research
+description: Structured approach to finding and synthesizing information from the user's knowledge base
+allowed-tools: search_surfsense_docs, scrape_webpage, read_file, ls_tree, grep, web_search
+---
+
+# Knowledge-base research
+
+## When to use this skill
+- The user asks "find/look up/research" something specifically inside their knowledge base.
+- The user references documents, notes, repos, or connector data they expect to exist already.
+- A multi-document synthesis is required (e.g., "summarize what we've discussed about X across all my notes").
+
+## Plan
+1. Decompose the user's question into 2-4 specific, citation-worthy sub-questions.
+2. For each sub-question, run **one** targeted KB search (focused on terms the user would have written, not synonyms). Open the most relevant 2-3 documents fully via `read_file` if their excerpts are too short.
+3. Use `grep` to find supporting passages in long files instead of re-reading them end to end.
+4. Cite every claim with `[citation:chunk_id]` exactly as the chunk tag specifies.
+
+## What good output looks like
+- Short paragraphs with inline citations.
+- Quoted phrases when wording matters.
+- An explicit "Not found in your knowledge base" callout when a sub-question has no support — never fabricate.

surfsense_backend/app/agents/new_chat/skills/builtin/meeting-prep/SKILL.md

@@ -0,0 +1,22 @@
+---
+name: meeting-prep
+description: Pull together briefing materials before a scheduled meeting
+allowed-tools: search_surfsense_docs, web_search, scrape_webpage, read_file
+---
+
+# Meeting preparation
+
+## When to use this skill
+The user mentions an upcoming meeting, call, or interview and asks you to "prep", "brief me", "pull background", or "what do I need to know about X before tomorrow".
+
+## Output structure
+Always produce these sections (omit any with no signal — don't pad):
+
+1. **Attendees & context** — who's in the room, their roles, what they care about. Pull from KB notes about prior interactions; supplement with public profile facts via `web_search` when names or companies are unfamiliar.
+2. **Open threads** — outstanding action items, unresolved decisions, last-mentioned blockers from prior conversation history.
+3. **Recent moves** — within the last 30 days: relevant launches, hires, news. Cite KB chunks when present, otherwise external sources.
+4. **Suggested questions** — 3-5 questions the user could ask, tailored to the open threads and the attendees' likely priorities.
+
+## Source ordering
+- Always check the user's KB **first** for prior meeting notes, internal docs, or Slack threads about these attendees.
+- Only fall back to `web_search` for *publicly verifiable* facts — never to fabricate a participant's preferences or relationships.

surfsense_backend/app/agents/new_chat/skills/builtin/report-writing/SKILL.md

@@ -0,0 +1,23 @@
+---
+name: report-writing
+description: How to scope, draft, and revise a Markdown report artifact via generate_report
+allowed-tools: generate_report, search_surfsense_docs, read_file
+---
+
+# Report writing
+
+## When to use this skill
+The user explicitly requests a deliverable: "write a report on …", "draft a memo", "produce a brief", "expand the previous report". A creation or modification verb pointed at an artifact is required (see `generate_report`'s when-to-call rules).
+
+## Decision flow
+1. **Source strategy.** Decide which `source_strategy` fits:
+   - `conversation` — substantive Q&A on the topic already in chat.
+   - `kb_search` — fresh topic; supply 1–5 precise `search_queries`.
+   - `auto` — partial conversation context; let the tool fall back.
+   - `provided` — verbatim source text only.
+2. **Style.** Default to `report_style="detailed"` unless the user explicitly asks for "brief", "one page", "500 words".
+3. **Revisions.** When modifying an existing report from this conversation, set `parent_report_id` and put the change list in `user_instructions` ("add carbon-capture section", "tighten conclusion").
+4. **Never paste the report back into chat** after `generate_report` returns — confirm and let the artifact card render itself.
+
+## Hooks for KB-only mode
+If `kb_search`/`auto` returns no results, do **not** silently switch to general knowledge. Surface the gap in your confirmation message.

surfsense_backend/app/agents/new_chat/skills/builtin/slack-summary/SKILL.md

@@ -0,0 +1,26 @@
+---
+name: slack-summary
+description: Distill a Slack channel or thread into actionable summary
+allowed-tools: search_surfsense_docs
+---
+
+# Slack summarization
+
+## When to use this skill
+The user asks to summarize Slack ("what happened in #eng-platform this week", "what did Alice say about the launch", "catch me up on the design channel").
+
+## Required inputs
+Confirm before searching:
+- **Which channel(s) or thread(s)?** Don't guess if ambiguous.
+- **What time window?** Default to the last 7 days when not specified, but say so.
+
+## Output shape
+Produce three concise sections:
+1. **Key decisions** — explicit choices that were made, with the deciding message cited.
+2. **Open questions** — things asked but not answered, with the asking message cited.
+3. **Action items** — `@mention` who owes what by when, *only if explicitly stated*. Don't invent assignees.
+
+## What not to do
+- Never produce a chronological play-by-play of every message — distill.
+- Never quote private messages without flagging them as such.
+- If the channel was empty in the time window, say so — don't fabricate filler.

surfsense_backend/app/agents/new_chat/state_reducers.py

@@ -0,0 +1,201 @@
+"""Reducers and sentinels for SurfSense filesystem state.
+
+These reducers back the extra state fields used by the cloud-mode filesystem
+agent (`cwd`, `staged_dirs`, `pending_moves`, `dirty_paths`, `doc_id_by_path`,
+`kb_priority`, `kb_matched_chunk_ids`, `kb_anon_doc`, `tree_version`).
+
+Tools mutate these fields ONLY via `Command(update={...})` returns; the
+reducers are responsible for merging successive updates atomically and for
+honouring an explicit reset sentinel (`_CLEAR`) so that a single update can
+both reset and reseed a list (used by `move_file` / `aafter_agent`).
+
+The sentinel is intentionally a plain string constant rather than a custom
+object so that LangGraph's checkpointer (which serializes raw `Command.update`
+deltas via ``ormsgpack`` BEFORE reducers are applied) can round-trip writes
+that contain it. The token uses a NUL-bracketed form that cannot collide with
+any real virtual path, document title, or dict key produced by the agent.
+"""
+
+from __future__ import annotations
+
+from typing import Any, Final, TypeVar
+
+_CLEAR: Final[str] = "\x00__SURFSENSE_FILESYSTEM_CLEAR__\x00"
+"""Reset sentinel; pass it inside a list/dict update to request a reset.
+
+For list reducers: ``[_CLEAR, *items]`` resets the field then appends ``items``.
+For dict reducers: ``{_CLEAR: True, **items}`` resets the field then merges ``items``.
+
+Because the value is a plain string with embedded NUL bytes, it is natively
+serializable by ``ormsgpack`` (used by LangGraph's PostgreSQL checkpointer)
+yet still distinct from any real path / key produced by application code.
+"""
+
+
+T = TypeVar("T")
+
+
+def _replace_reducer[T](left: T | None, right: T | None) -> T | None:
+    """Replace `left` outright with `right`. ``None`` on the right is honored as a reset."""
+    return right
+
+
+def _is_clear(value: Any) -> bool:
+    return isinstance(value, str) and value == _CLEAR
+
+
+def _add_unique_reducer(
+    left: list[Any] | None,
+    right: list[Any] | None,
+) -> list[Any]:
+    """Append items from ``right`` to ``left`` while preserving uniqueness.
+
+    Semantics:
+    - If ``right`` is ``None`` or empty, return ``left`` unchanged.
+    - If ``right`` contains the ``_CLEAR`` sentinel anywhere, the result is
+      reseeded with only the items that appear AFTER the LAST occurrence of
+      ``_CLEAR`` (deduplicated, preserving first-seen order). This gives a
+      single-update "reset and reseed" capability.
+    - Otherwise, items from ``right`` are appended to ``left`` (order preserved
+      from first seen) while skipping values that are already present.
+    """
+    if right is None:
+        return list(left or [])
+    if not right:
+        return list(left or [])
+
+    last_clear = -1
+    for index, item in enumerate(right):
+        if _is_clear(item):
+            last_clear = index
+
+    if last_clear >= 0:
+        seed: list[Any] = []
+        seen: set[Any] = set()
+        for item in right[last_clear + 1 :]:
+            if _is_clear(item):
+                continue
+            try:
+                if item in seen:
+                    continue
+                seen.add(item)
+            except TypeError:
+                if item in seed:
+                    continue
+            seed.append(item)
+        return seed
+
+    base = list(left or [])
+    try:
+        seen: set[Any] = set(base)
+    except TypeError:
+        seen = set()
+    for item in right:
+        if _is_clear(item):
+            continue
+        try:
+            if item in seen:
+                continue
+            seen.add(item)
+        except TypeError:
+            if item in base:
+                continue
+        base.append(item)
+    return base
+
+
+def _list_append_reducer(
+    left: list[Any] | None,
+    right: list[Any] | None,
+) -> list[Any]:
+    """Append items from ``right`` to ``left`` preserving order and duplicates.
+
+    Honours the ``_CLEAR`` sentinel exactly like :func:`_add_unique_reducer`,
+    but does NOT deduplicate. Used for queues whose ordering and duplicate
+    occurrences matter (e.g. ``pending_moves``).
+    """
+    if right is None:
+        return list(left or [])
+    if not right:
+        return list(left or [])
+
+    last_clear = -1
+    for index, item in enumerate(right):
+        if _is_clear(item):
+            last_clear = index
+
+    if last_clear >= 0:
+        return [item for item in right[last_clear + 1 :] if not _is_clear(item)]
+
+    base = list(left or [])
+    base.extend(item for item in right if not _is_clear(item))
+    return base
+
+
+def _dict_merge_with_tombstones_reducer(
+    left: dict[Any, Any] | None,
+    right: dict[Any, Any] | None,
+) -> dict[Any, Any]:
+    """Merge ``right`` into ``left`` with two extra capabilities:
+
+    * Keys whose value is ``None`` are removed from the merged result
+      (tombstone semantics, matching the deepagents file-data reducer).
+    * The special key ``_CLEAR`` (with any truthy value) resets ``left`` to
+      ``{}`` before merging the remaining keys from ``right``. This makes it
+      possible to atomically clear and reseed the dictionary in a single
+      update.
+    """
+    if right is None:
+        return dict(left or {})
+
+    if _CLEAR in right or any(_is_clear(k) for k in right):
+        result: dict[Any, Any] = {}
+        for key, value in right.items():
+            if _is_clear(key):
+                continue
+            if value is None:
+                result.pop(key, None)
+                continue
+            result[key] = value
+        return result
+
+    if left is None:
+        return {key: value for key, value in right.items() if value is not None}
+
+    result = dict(left)
+    for key, value in right.items():
+        if value is None:
+            result.pop(key, None)
+        else:
+            result[key] = value
+    return result
+
+
+def _initial_filesystem_state() -> dict[str, Any]:
+    """Default empty values for SurfSense filesystem state fields.
+
+    Consumers should always treat these fields as ``state.get(key) or
+    DEFAULT`` so that fresh threads (without checkpointed state) work
+    correctly.
+    """
+    return {
+        "cwd": "/documents",
+        "staged_dirs": [],
+        "pending_moves": [],
+        "doc_id_by_path": {},
+        "dirty_paths": [],
+        "kb_priority": [],
+        "kb_matched_chunk_ids": {},
+        "kb_anon_doc": None,
+        "tree_version": 0,
+    }
+
+
+__all__ = [
+    "_CLEAR",
+    "_add_unique_reducer",
+    "_dict_merge_with_tombstones_reducer",
+    "_initial_filesystem_state",
+    "_list_append_reducer",
+    "_replace_reducer",
+]