Skip to content

Self-hosted homelab infrastructure visualizer — interactive network diagram with live status monitoring

31.6 KB
Raw Blame
import json
from urllib.parse import quote
from mcp.server import Server
from mcp.types import Tool, TextContent
from .backend_client import backend, safe_id
from .annotations import annotate
from .devices import DEVICE_TOOL_NAMES, DEVICE_TOOLS, dispatch_device
from .documents import DOC_TOOL_NAMES, DOC_TOOLS, dispatch_document
from .racks import RACK_TOOL_NAMES, RACK_TOOLS, dispatch_rack


# Kept in sync manually with frontend/src/types/index.ts NodeType (device types only —
# group/groupRect/text are canvas annotations created via dedicated UI actions, not create_node).
# Sync is enforced automatically by test_node_types_in_sync_with_frontend (mcp/tests/test_node_types_sync.py).
NODE_TYPES = [
    "isp", "router", "firewall", "switch", "server", "proxmox", "vm", "lxc", "nas", "kvm", "iot", "ap",
    "camera", "printer", "computer", "laptop", "mobile", "cpl", "docker_host", "docker_container",
    "generic", "zigbee_coordinator", "zigbee_router", "zigbee_enddevice",
    "zwave_coordinator", "zwave_router", "zwave_enddevice", "grid", "ups", "battery", "generator",
    "solar_panel", "inverter", "circuit_breaker", "contactor", "electrical_switch", "socket",
    "light", "meter", "transformer", "load",
]

# Kept in sync manually with frontend/src/types/index.ts EdgeType.
# Sync is enforced automatically by test_edge_types_in_sync_with_frontend (mcp/tests/test_edge_types_sync.py).
EDGE_TYPES = [
    "ethernet", "wifi", "iot", "zigbee_mesh", "vlan", "virtual", "cluster", "fibre", "electrical",
]

# Shared field schemas mirroring backend NodeBase / NodeUpdate (backend/app/schemas/nodes.py).
# create_node and update_node both expose these so the MCP is symmetric with what the
# backend already validates and stores. _dispatch forwards args verbatim, so any field
# advertised here is accepted by the backend.
_NODE_FIELDS = {
    "label":         {"type": "string"},
    "ip":            {"type": "string"},
    "hostname":      {"type": "string"},
    "mac":           {"type": "string", "description": "MAC address."},
    "os":            {"type": "string", "description": "Operating system / distribution."},
    # Live observation, not a user setting: the inventory row keeps the freshest
    # one rather than the last writer, so an update only lands while the device's
    # status is still unknown (see link_facts in inventory_sync.py). Set it on
    # create; to change it later, point check_method/check_target at the device
    # and let the status checker observe it.
    "status":        {"type": "string", "enum": ["online", "offline", "unknown", "pending"], "description": "Live status. Honoured on create; on update it is ignored unless the device's status is still unknown — the status checker owns it."},
    "check_method":  {"type": "string", "description": "Status check method (ping, http, https, ssh, prometheus, tcp)."},
    "check_target":  {"type": "string", "description": "Target host/URL used by the status check."},
    "services":      {"type": "array", "items": {"type": "object"}, "description": "Running services detected or documented on the node."},
    "notes":         {"type": "string", "description": "Free-text notes / documentation for the node."},
    "pos_x":         {"type": "number", "description": "X position on the canvas. Omit on create to auto-place (root nodes only)."},
    "pos_y":         {"type": "number", "description": "Y position on the canvas. Omit on create to auto-place (root nodes only). For child nodes this is relative to the parent container."},
    "width":         {"type": "number", "description": "Width of the node card in pixels. Mainly useful for container nodes."},
    "height":        {"type": "number", "description": "Height of the node card in pixels. Mainly useful for container nodes."},
    "parent_id":     {"type": "string", "description": "ID of the parent node (e.g. Proxmox host for a VM/LXC). Pass null to detach."},
    "container_mode": {"type": "boolean", "description": "Render this node as a container/group that can hold children."},
    "custom_icon":   {"type": "string", "description": "Override icon name for the node."},
    "cpu_count":     {"type": "integer", "description": "Number of CPU cores/threads."},
    "cpu_model":     {"type": "string", "description": "CPU model name."},
    "ram_gb":        {"type": "number", "description": "RAM in gigabytes."},
    "disk_gb":       {"type": "number", "description": "Disk capacity in gigabytes."},
    "show_hardware": {"type": "boolean", "description": "Display hardware specs on the node card."},
    # The field is `key`, not `name`: the backend keys properties on it, both to
    # merge them into the inventory row and to order them in the node's view
    # (`merge_properties` / `apply_view` in backend/app/services/inventory_sync.py).
    # A property with no `key` collapses with every other keyless one — send two
    # and the node draws one.
    # Connection points. The count is per side; the IDs an edge references are
    # derived from it — slot 0 is the bare side name, slot N >= 1 is
    # '{side}-{N + 1}' (frontend/src/utils/handleUtils.ts). Lowering a count
    # moves the edges that used the dropped handles back to the side's slot 0.
    "top_handles":    {"type": "integer", "minimum": 0, "maximum": 64, "description": "Connection points on the top side (default 1). Their IDs are 'top', 'top-2', 'top-3', ..."},
    "bottom_handles": {"type": "integer", "minimum": 0, "maximum": 64, "description": "Connection points on the bottom side (default 1). Their IDs are 'bottom', 'bottom-2', 'bottom-3', ..."},
    "left_handles":   {"type": "integer", "minimum": 0, "maximum": 64, "description": "Connection points on the left side (default 0). Their IDs are 'left', 'left-2', 'left-3', ..."},
    "right_handles":  {"type": "integer", "minimum": 0, "maximum": 64, "description": "Connection points on the right side (default 0). Their IDs are 'right', 'right-2', 'right-3', ..."},
    "show_port_numbers": {"type": "boolean", "description": "Number each connection point on the node card."},
    "port_labels":   {
        "type": "object",
        "description": (
            "Text shown in place of a connection point's number when show_port_numbers is on, "
            "keyed by handle ID — e.g. {\"top\": \"8\"} for an uplink printed as port 8. "
            "At most 8 characters; a point without a label keeps its number."
        ),
        "additionalProperties": {"type": "string", "maxLength": 8},
    },
    "properties":    {
        "type": "array",
        "description": "Key/value metadata shown on the node. `key` is the identity: two properties sharing one are the same property.",
        "items": {
            "type": "object",
            "required": ["key", "value"],
            "properties": {
                "key":     {"type": "string"},
                "value":   {"type": "string"},
                "icon":    {"type": "string", "description": "Lucide icon name shown beside the value."},
                "visible": {"type": "boolean", "description": "Draw it on the node card. Default true."},
            },
        },
    },
}

# A zone is a groupRect node: canvas furniture that visually groups the nodes
# parented to it (a VLAN, a site, a rack room). It is deliberately kept out of
# NODE_TYPES — create_node makes devices, these tools make and fill zones.
ZONE_TYPE = "groupRect"

_ZONE_COLOR_FIELDS = {
    "border":       {"type": "string", "description": "Border/label colour, e.g. '#00d4ff'."},
    "border_style": {"type": "string", "enum": ["solid", "dashed", "dotted"]},
    "border_width": {"type": "number"},
    "text_color":   {"type": "string", "description": "Label colour, e.g. '#e6edf3'."},
}

# Which connection point each end of an edge attaches to. The IDs are the node's
# per-side handle IDs (see the *_handles fields above): 'bottom', 'bottom-2',
# 'left-3', ... Omit both on create and the backend picks them from the two nodes'
# relative positions (_auto_handles in backend/app/api/routes/edges.py).
#
# The node must already have the connection point — left/right default to none —
# or the backend answers 422 rather than storing an edge the canvas cannot draw.
# Presentation, mirroring the backend EdgeBase (backend/app/schemas/edges.py) and
# the EdgeData types the canvas draws from (frontend/src/types/index.ts). Every one
# is optional: an omitted field keeps the edge type's own preset.
_EDGE_STYLE_FIELDS = {
    "animated":     {"type": "string", "enum": ["none", "basic", "snake", "flow"], "description": "Line animation. 'basic' is the dashed march, 'snake' a travelling segment, 'flow' a continuous drift."},
    "custom_color": {"type": "string", "description": "Line colour override, e.g. '#00d4ff'. Omit to use the edge type's colour."},
    "path_style":   {"type": "string", "enum": ["bezier", "smooth"], "description": "How the line is routed between the two connection points."},
    "line_style":   {"type": "string", "enum": ["solid", "dashed", "dotted"], "description": "How the line itself is stroked."},
    "width_mult":   {"type": "number", "minimum": 1, "maximum": 4, "description": "Stroke-width multiplier (1-4x) on the edge type's base width."},
    "marker_start": {"type": "string", "enum": ["none", "arrow", "arrow-open", "circle", "diamond", "square"], "description": "Marker drawn at the source end."},
    "marker_end":   {"type": "string", "enum": ["none", "arrow", "arrow-open", "circle", "diamond", "square"], "description": "Marker drawn at the target end."},
    "vlan_id":      {"type": "integer", "description": "VLAN tag carried by the link. Shown on the edge for a 'vlan' type."},
    "speed":        {"type": "string", "description": "Link speed as it should read on the canvas, e.g. '10G' or '2.5 Gbps'."},
}

_EDGE_HANDLE_FIELDS = {
    "source_handle": {"type": "string", "description": "Connection point on the source node, e.g. 'bottom' or 'right-2'. The node must have it already — raise that side's *_handles first, or the call is rejected. Omit to let the server pick."},
    "target_handle": {"type": "string", "description": "Connection point on the target node, e.g. 'top' or 'left-3'. The node must have it already — raise that side's *_handles first, or the call is rejected. Omit to let the server pick."},
}

# Optional design/canvas selector. The backend attaches nodes/edges to the first
# design when design_id is omitted (see backend nodes.py / edges.py), so these
# tools stay backward compatible; pass design_id to target a specific canvas.
# Use list_designs to discover the available IDs.
_DESIGN_ID_FIELD = {
    "design_id": {"type": "string", "description": "Target design/canvas ID. Omit to use the default (first) canvas; call list_designs to discover IDs."},
}


def _build_tools() -> list[Tool]:
    create_node_props = {
        "type": {"type": "string", "enum": NODE_TYPES},
        **_NODE_FIELDS,
        **_DESIGN_ID_FIELD,
    }
    create_node_props["status"] = {**_NODE_FIELDS["status"], "default": "unknown"}

    update_node_props = {
        "id":   {"type": "string"},
        "type": {"type": "string", "enum": NODE_TYPES},
        **_NODE_FIELDS,
    }

    return [
        Tool(name="create_node", description="Add a new node to the homelab canvas", inputSchema={
            "type": "object",
            "required": ["type", "label"],
            "properties": create_node_props,
        }),
        Tool(name="update_node", description="Update an existing node", inputSchema={
            "type": "object",
            "required": ["id"],
            "properties": update_node_props,
        }),
        Tool(name="delete_node", description="Delete a node from the canvas", inputSchema={
            "type": "object",
            "required": ["id"],
            "properties": {"id": {"type": "string"}},
        }),
        Tool(name="create_edge", description="Create a network link between two nodes. Pass source_handle/target_handle to choose which connection points it attaches to; omit them and the server picks from the nodes' relative positions.", inputSchema={
            "type": "object",
            "required": ["source", "target"],
            "properties": {
                "source": {"type": "string"},
                "target": {"type": "string"},
                "type":   {"type": "string", "enum": EDGE_TYPES, "default": "ethernet"},
                "label":  {"type": "string"},
                **_EDGE_STYLE_FIELDS,
                **_EDGE_HANDLE_FIELDS,
                **_DESIGN_ID_FIELD,
            },
        }),
        Tool(name="update_edge", description="Update an existing link: its type, label, styling (animation, colour, line/path style, width, endpoint markers), or which connection points it attaches to. Call get_canvas to discover edge ids.", inputSchema={
            "type": "object",
            "required": ["id"],
            "properties": {
                "id":    {"type": "string", "description": "Edge id."},
                "type":  {"type": "string", "enum": EDGE_TYPES},
                "label": {"type": "string"},
                **_EDGE_STYLE_FIELDS,
                **_EDGE_HANDLE_FIELDS,
            },
        }),
        Tool(name="list_edges", description="List every link with its full detail — type, label, waypoints, and the source_handle/target_handle each end attaches to. get_canvas slims edges down to id/source/target/type/label; use this when the connection points or styling matter.", inputSchema={
            "type": "object",
            "properties": {},
        }),
        Tool(name="delete_edge", description="Delete a network link", inputSchema={
            "type": "object",
            "required": ["id"],
            "properties": {"id": {"type": "string"}},
        }),
        Tool(name="trigger_scan", description="Trigger a network discovery scan", inputSchema={
            "type": "object",
            "properties": {
                "ranges": {"type": "array", "items": {"type": "string"}, "description": "CIDR ranges to scan (uses configured defaults if omitted)"},
            },
        }),
        Tool(name="approve_device", description="Approve a pending discovered device and create a node. A device created from a rack canvas (discovery_source 'rack') is refused with a 409 — passive rack gear never goes on a logical canvas.", inputSchema={
            "type": "object",
            "required": ["id"],
            "properties": {
                "id":    {"type": "string"},
                "type":  {"type": "string", "enum": NODE_TYPES, "default": "generic"},
                "label": {"type": "string"},
                **_DESIGN_ID_FIELD,
            },
        }),
        Tool(name="hide_device", description="Hide a pending discovered device", inputSchema={
            "type": "object",
            "required": ["id"],
            "properties": {"id": {"type": "string"}},
        }),
        Tool(name="get_canvas", description="Get the full canvas: all nodes and edges in the homelab topology", inputSchema={
            "type": "object",
            "properties": {
                "design_id": {"type": "string", "description": "Only list zones on this design/canvas. Omit to list every design's zones; call list_designs to discover IDs."},
            },
        }),
        Tool(name="list_nodes", description="List all nodes (devices) in the homelab", inputSchema={
            "type": "object",
            "properties": {},
        }),
        Tool(name="list_node_summaries", description="List all nodes with only id/label/type/status — a lighter, lower-token alternative to list_nodes or get_canvas for when full node detail (ip, services, notes, hardware, properties, ...) isn't needed.", inputSchema={
            "type": "object",
            "properties": {},
        }),
        Tool(name="get_node", description="Get node(s) by id or by label. Provide exactly one of the two. Lookup by 'id' returns a single full node object. Lookup by 'label' is a case-insensitive substring match and returns a list of full node objects, since labels are not unique.", inputSchema={
            "type": "object",
            "properties": {
                "id":    {"type": "string", "description": "Node id. Returns a single node object."},
                "label": {"type": "string", "description": "Case-insensitive substring match on label. Returns a list of matching node objects."},
            },
        }),
        Tool(name="list_pending_devices", description="List devices discovered by scan but not yet approved or hidden", inputSchema={
            "type": "object",
            "properties": {},
        }),
        Tool(name="list_inventory", description="List the full device inventory (everything scanned except user-hidden devices): both pending devices awaiting triage and already-approved devices. Each row carries a 'status' field. Use the optional 'status' filter to narrow the result. Gear created from a rack canvas is listed here too, tagged discovery_source 'rack'.", inputSchema={
            "type": "object",
            "properties": {
                "status": {"type": "string", "enum": ["all", "pending", "approved"], "default": "all", "description": "Filter inventory by status. 'all' returns pending + approved."},
            },
        }),
        Tool(name="list_hidden_devices", description="List devices the user has hidden from the inventory. Hidden devices are excluded from list_pending_devices and list_inventory; use restore_device to bring one back to pending.", inputSchema={
            "type": "object",
            "properties": {},
        }),
        Tool(name="restore_device", description="Restore (un-hide) a previously hidden device, returning it to pending status so it reappears in the triage list. Use this to undo a hide_device action.", inputSchema={
            "type": "object",
            "required": ["id"],
            "properties": {"id": {"type": "string"}},
        }),
        Tool(name="create_zone", description="Create a zone (a groupRect area) on the canvas. A zone visually groups the nodes parented to it — a VLAN, a site, a room. Use add_to_zone to fill it.", inputSchema={
            "type": "object",
            "required": ["label"],
            "properties": {
                "label":  {"type": "string", "description": "Zone name, shown in its top-left corner."},
                "pos_x":  {"type": "number", "description": "X position on the canvas. Omit to auto-place."},
                "pos_y":  {"type": "number", "description": "Y position on the canvas. Omit to auto-place."},
                "width":  {"type": "number", "description": "Zone width in pixels (default 360)."},
                "height": {"type": "number", "description": "Zone height in pixels (default 240)."},
                **_ZONE_COLOR_FIELDS,
                **_DESIGN_ID_FIELD,
            },
        }),
        Tool(name="list_zones", description="List the zones (groupRect areas) on the canvas with their id, label, position, size and the nodes each one contains. Lists every design's zones unless design_id narrows it.", inputSchema={
            "type": "object",
            "properties": {
                "design_id": {"type": "string", "description": "Only list zones on this design/canvas. Omit to list every design's zones; call list_designs to discover IDs."},
            },
        }),
        Tool(name="add_to_zone", description="Move one or more nodes into a zone. Positions are rebased on the zone so the nodes keep their place on screen. A node already in the zone, the zone itself, a node on another design, and any node the zone is nested in are skipped.", inputSchema={
            "type": "object",
            "required": ["zone_id", "node_ids"],
            "properties": {
                "zone_id":  {"type": "string", "description": "Target zone id — call list_zones to discover it."},
                "node_ids": {"type": "array", "items": {"type": "string"}, "description": "Ids of the nodes to move in."},
            },
        }),
        Tool(name="remove_from_zone", description="Take nodes back out of the zone they are in, restoring their absolute canvas position. Nodes that are not in a zone are left alone.", inputSchema={
            "type": "object",
            "required": ["node_ids"],
            "properties": {
                "node_ids": {"type": "array", "items": {"type": "string"}, "description": "Ids of the nodes to detach."},
            },
        }),
        Tool(name="list_designs", description="List all designs (canvases) with their IDs and node/group/text counts", inputSchema={
            "type": "object",
            "properties": {},
        }),
        Tool(name="create_design", description="Create a new design (canvas) and return it, including its id for use as design_id", inputSchema={
            "type": "object",
            "required": ["name"],
            "properties": {
                "name":        {"type": "string", "description": "Name of the new canvas."},
                "icon":        {"type": "string", "description": "Icon name for the canvas (default: dashboard)."},
                "design_type": {"type": "string", "description": "Design type (default: network)."},
            },
        }),
        Tool(name="delete_design", description="Delete a design (canvas) and all its nodes and edges. The last remaining design cannot be deleted.", inputSchema={
            "type": "object",
            "required": ["design_id"],
            "properties": {
                "design_id": {"type": "string", "description": "ID of the design to delete. Call list_designs to discover IDs."},
            },
        }),
    ]


# The rack canvas, the inventory write routes and the documentation space live
# in their own modules — each is big enough to bury the logical-canvas tools
# this file is about.
TOOLS = annotate(_build_tools() + RACK_TOOLS + DEVICE_TOOLS + DOC_TOOLS)


def register_tools(server: Server):

    @server.list_tools()
    async def list_tools():
        return TOOLS

    @server.call_tool()
    async def call_tool(name: str, arguments: dict):
        result = await _dispatch(name, arguments)
        return [TextContent(type="text", text=json.dumps(result, indent=2))]


def _slim_canvas(raw: dict) -> dict:
    """Strip layout/style fields — keep only semantic data for AI use.

    `GET /api/v1/canvas` reports a node flat (`NodeResponse`): the device facts
    sit beside `id` and `type`, not under a React Flow `data` key. A payload that
    does carry `data` — the shape the frontend store holds — is folded in too, so
    either form slims to the same thing.
    """
    NODE_KEEP = {
        "label", "ip", "hostname", "mac", "os", "status", "services",
        "notes", "description", "properties", "cpu_count", "cpu_model", "ram_gb",
        "disk_gb", "parent_id",
    }
    EDGE_KEEP = {"id", "source", "target", "type", "label"}

    def slim_node(n: dict) -> dict:
        fields = {**n, **n.get("data", {})}
        out = {k: v for k, v in fields.items() if k in NODE_KEEP and v not in (None, "", [])}
        out["id"] = n.get("id")
        # `type` on the wire is the node type (router, proxmox, ...); it is
        # reported under its own key so it cannot collide with a device fact.
        out["node_type"] = n.get("type")
        # Nesting is `parent_id` on the API and `parentId` in a React Flow payload.
        parent = out.pop("parent_id", None) or n.get("parentId")
        if parent:
            out["parent_id"] = parent
        return out

    def slim_edge(e: dict) -> dict:
        return {k: v for k, v in e.items() if k in EDGE_KEEP and v not in (None, "")}

    return {
        "nodes": [slim_node(n) for n in raw.get("nodes", [])],
        "edges": [slim_edge(e) for e in raw.get("edges", [])],
    }


def _slim_node_summary(n: dict) -> dict:
    """Only the fields needed to identify a node and reference it in later calls."""
    return {"id": n.get("id"), "label": n.get("label"), "type": n.get("type"), "status": n.get("status")}


def _slim_zone(z: dict, nodes: list[dict]) -> dict:
    """A zone plus the ids of the nodes parented to it."""
    return {
        "id": z.get("id"),
        "label": z.get("label"),
        "design_id": z.get("design_id"),
        "pos_x": z.get("pos_x"),
        "pos_y": z.get("pos_y"),
        "width": z.get("width"),
        "height": z.get("height"),
        "node_ids": [n["id"] for n in nodes if n.get("parent_id") == z.get("id")],
    }


def _is_ancestor(nodes_by_id: dict[str, dict], maybe_ancestor_id: str, node_id: str) -> bool:
    """True when `maybe_ancestor_id` sits on `node_id`'s parent chain. Cycle-safe."""
    seen = {node_id}
    current = nodes_by_id.get(node_id)
    while current:
        pid = current.get("parent_id")
        if not pid or pid in seen:
            return False
        if pid == maybe_ancestor_id:
            return True
        seen.add(pid)
        current = nodes_by_id.get(pid)
    return False


async def _dispatch(name: str, args: dict) -> dict:
    if name in RACK_TOOL_NAMES:
        return await dispatch_rack(name, args)

    if name in DEVICE_TOOL_NAMES:
        return await dispatch_device(name, args)

    if name in DOC_TOOL_NAMES:
        return await dispatch_document(name, args)

    if name == "create_node":
        return await backend.post("/api/v1/nodes", args)

    if name == "update_node":
        node_id = safe_id(args.pop("id"), field="node id")
        return await backend.patch(f"/api/v1/nodes/{node_id}", args)

    if name == "delete_node":
        return await backend.delete(f"/api/v1/nodes/{safe_id(args['id'], field='node id')}")

    if name == "create_edge":
        return await backend.post("/api/v1/edges", args)

    if name == "update_edge":
        edge_id = safe_id(args.pop("id"), field="edge id")
        return await backend.patch(f"/api/v1/edges/{edge_id}", args)

    if name == "list_edges":
        return await backend.get("/api/v1/edges")

    if name == "delete_edge":
        return await backend.delete(f"/api/v1/edges/{safe_id(args['id'], field='edge id')}")

    if name == "trigger_scan":
        body = {"ranges": args["ranges"]} if "ranges" in args else {}
        return await backend.post("/api/v1/scan/trigger", body)

    if name == "approve_device":
        device_id = safe_id(args.pop("id"), field="device id")
        return await backend.post(f"/api/v1/scan/pending/{device_id}/approve", args)

    if name == "hide_device":
        return await backend.post(f"/api/v1/scan/pending/{safe_id(args['id'], field='device id')}/hide", {})

    if name == "get_canvas":
        design_id = args.get("design_id")
        path = f"/api/v1/canvas?design_id={quote(safe_id(design_id, field='design id'))}" if design_id else "/api/v1/canvas"
        raw = await backend.get(path)
        return _slim_canvas(raw)

    if name == "list_nodes":
        return await backend.get("/api/v1/nodes")

    if name == "list_node_summaries":
        nodes = await backend.get("/api/v1/nodes")
        return [_slim_node_summary(n) for n in nodes]

    if name == "get_node":
        node_id = args.get("id")
        label = args.get("label")
        if node_id:
            return await backend.get(f"/api/v1/nodes/{safe_id(node_id, field='node id')}")
        if label:
            return await backend.get(f"/api/v1/nodes?label={quote(label)}")
        raise ValueError("get_node requires either 'id' or 'label'")

    if name == "list_pending_devices":
        # Backend /scan/pending returns the whole inventory: approved rows stay
        # listed so the frontend can show a canvas-presence badge. This tool
        # promises only devices "not yet approved or hidden", so filter to
        # actual pending rows (legacy rows without a status count as pending).
        devices = await backend.get("/api/v1/scan/pending")
        return [d for d in devices if d.get("status", "pending") == "pending"]

    if name == "list_inventory":
        # /scan/pending returns the whole inventory minus hidden rows (pending +
        # approved). Legacy rows without a status field count as pending.
        devices = await backend.get("/api/v1/scan/pending")
        wanted = args.get("status", "all")
        if wanted == "all":
            return devices
        return [d for d in devices if d.get("status", "pending") == wanted]

    if name == "list_hidden_devices":
        return await backend.get("/api/v1/scan/hidden")

    if name == "restore_device":
        return await backend.post(f"/api/v1/scan/pending/{safe_id(args['id'], field='device id')}/restore", {})

    if name == "create_zone":
        colors = {k: args.pop(k) for k in list(_ZONE_COLOR_FIELDS) if k in args}
        body = {
            "type": ZONE_TYPE,
            "status": "unknown",
            "width": args.pop("width", 360),
            "height": args.pop("height", 240),
            **args,
        }
        if colors:
            body["custom_colors"] = colors
        return await backend.post("/api/v1/nodes", body)

    if name == "list_zones":
        # /api/v1/nodes has no design filter — it returns every design's nodes,
        # so narrow here when the caller named one.
        design_id = args.get("design_id")
        nodes = await backend.get("/api/v1/nodes")
        return [
            _slim_zone(n, nodes)
            for n in nodes
            if n.get("type") == ZONE_TYPE and (design_id is None or n.get("design_id") == design_id)
        ]

    if name == "add_to_zone":
        zone_id = args["zone_id"]
        nodes = await backend.get("/api/v1/nodes")
        by_id = {n["id"]: n for n in nodes}
        zone = by_id.get(zone_id)
        if zone is None or zone.get("type") != ZONE_TYPE:
            raise ValueError(f"add_to_zone: {zone_id} is not a zone")
        moved, skipped = [], []
        for node_id in args["node_ids"]:
            node = by_id.get(node_id)
            if (
                node is None
                or node_id == zone_id
                or node.get("parent_id") == zone_id
                # A zone only groups its own canvas: parenting across designs
                # would hide the node on the design it belongs to.
                or node.get("design_id") != zone.get("design_id")
                or _is_ancestor(by_id, node_id, zone_id)
            ):
                skipped.append(node_id)
                continue
            # The canvas stores a child's position relative to its parent, so
            # rebase or the node jumps by the zone's offset on the next load.
            await backend.patch(f"/api/v1/nodes/{safe_id(node_id, field='node id')}", {
                "parent_id": zone_id,
                "pos_x": (node.get("pos_x") or 0) - (zone.get("pos_x") or 0),
                "pos_y": (node.get("pos_y") or 0) - (zone.get("pos_y") or 0),
            })
            moved.append(node_id)
        return {"zone_id": zone_id, "moved": moved, "skipped": skipped}

    if name == "remove_from_zone":
        nodes = await backend.get("/api/v1/nodes")
        by_id = {n["id"]: n for n in nodes}
        detached, skipped = [], []
        for node_id in args["node_ids"]:
            node = by_id.get(node_id)
            zone = by_id.get(node.get("parent_id") or "") if node else None
            if node is None or zone is None or zone.get("type") != ZONE_TYPE:
                skipped.append(node_id)
                continue
            # Zone-relative → absolute, so the node stays where it looks.
            await backend.patch(f"/api/v1/nodes/{safe_id(node_id, field='node id')}", {
                "parent_id": None,
                "pos_x": (node.get("pos_x") or 0) + (zone.get("pos_x") or 0),
                "pos_y": (node.get("pos_y") or 0) + (zone.get("pos_y") or 0),
            })
            detached.append(node_id)
        return {"detached": detached, "skipped": skipped}

    if name == "list_designs":
        return await backend.get("/api/v1/designs")

    if name == "create_design":
        return await backend.post("/api/v1/designs", args)

    if name == "delete_design":
        return await backend.delete(f"/api/v1/designs/{safe_id(args['design_id'], field='design id')}")

    raise ValueError(f"Unknown tool: {name}")