shopdb-flask/plugins/endpoints/api/routes.py

"""Endpoints plugin API endpoints (read + dashboard).

Ingest is NOT here: agent reports arrive on the generic collector route
POST /api/collector/endpoints (ADR-006), authorized by a collector-scoped
service token, and are handled by EndpointsPlugin.apply_collector_payload.
These routes are the human-facing read surface, gated by user JWT + RBAC.
"""

import hashlib
import hmac
import json
import os
import time
from datetime import datetime, timedelta, timezone

from flask import Blueprint, request, current_app
from flask_jwt_extended import current_user

from shopdb.api import (
    db, Asset, Setting, success_response, error_response, paginated_response,
    ErrorCodes, get_pagination_params, paginate_query, require_permission,
    require_role, caller_holds, mint_service_token, list_service_tokens,
    revoke_service_tokens,
    AuditLog, collector_authorized, User,
    Application, AppVersion, Script,
)

from ..models import EndpointState, EndpointEvent, EndpointCommand, EndpointApplication
from ..commands import COMMAND_ALLOWLIST, is_allowed_action, ALLOWED_ACTIONS_FOR, row_actions_for, requires_target, action_targetkind, elevated_permission_for
from ..eventcatalog import CRITICAL_EVENT_IDS, event_severity, event_title

endpoints_bp = Blueprint('endpoints', __name__)


def _naive_utcnow():
    return datetime.now(timezone.utc).replace(tzinfo=None)


# A watched service or process reports a STATE STRING, not a boolean, and the
# vocabulary varies by source. Both the services panel and the fleet health
# rollup have to agree on what counts as running, so the rule lives here once -
# reading it wrong makes every service look stopped.
_RUNNING_STATES = ('running', 'started', 'up', 'ok', 'active')


def _service_running(entry):
    raw = (entry.get('state') or entry.get('status') or '').strip().lower()
    return raw in _RUNNING_STATES


# Paused, plus the four transitional states Windows reports while a service is
# on its way somewhere.
_TRANSITIONAL_STATES = ('paused', 'start pending', 'stop pending',
                        'continue pending', 'pause pending', 'starting',
                        'stopping', 'pausing')


def _service_transitional(raw):
    return (raw or '').strip().lower() in _TRANSITIONAL_STATES


def _service_name(entry):
    return entry.get('name') or entry.get('service') or 'Unknown'


def _format_uptime(seconds):
    """Whole-seconds uptime -> 'Nd Nh Nm', or None."""
    if seconds is None:
        return None
    try:
        seconds = int(seconds)
    except (TypeError, ValueError):
        return None
    days, rem = divmod(seconds, 86400)
    hours, rem = divmod(rem, 3600)
    minutes = rem // 60
    parts = []
    if days:
        parts.append('{}d'.format(days))
    if hours:
        parts.append('{}h'.format(hours))
    parts.append('{}m'.format(minutes))
    return ' '.join(parts)


# =============================================================================
# Endpoint state
# =============================================================================

@endpoints_bp.route('', methods=['GET'])
@require_permission('endpoints.fleet')
def list_endpoints():
    """List reported endpoint states.

    Registered as '' so the rule is '/api/endpoints', with no trailing slash.
    CONTRIBUTING.md's API section spells endpoints that way ('/api/machines'),
    and every other plugin's list route follows it - this was the only route in
    331 paths that did not.

    It was not cosmetic. A rule ending in a slash makes Flask answer the
    unslashed URL with a 308 to an ABSOLUTE url. On dev that resolves and
    nobody notices. Behind the IIS subpath it does not: with no ProxyFix,
    Werkzeug builds that Location from the scheme and host it believes it has,
    so it emitted http:// and dropped the /shopdb prefix. The browser refuses
    to follow an https page to http, the list request failed, and the page
    rendered "No endpoints found" - indistinguishable from a fleet with nothing
    in it. The agent had been reporting correctly the whole time.

    Joins through Computer so each row carries computerid and can deep-link
    into the existing PC detail page (/pcs/{computerid}). Endpoint state only
    exists for hostnames resolved to a Computer, so the join is not lossy.

    Sorted WORST FIRST by default: never-reported, then quiet, then reporting.
    It used to sort by lastreportedat descending, which floated the healthiest
    endpoints to the top and buried the ones somebody needs to look at.

    Each row also carries what the fleet view is actually for: which MACHINE
    depends on this PC, its PC type, and a health rollup. On a shop floor a PC
    matters because a machine stops without it, so that is the column an
    operator reads first.
    """
    from plugins.computers.models import Computer, ComputerType

    page, per_page = get_pagination_params(request)
    cutoff = _naive_utcnow() - timedelta(hours=_quiet_hours())

    # Status as a sortable expression rather than a post-hoc label, so
    # worst-first survives pagination. 0 never reported, 1 quiet, 2 reporting.
    statusrank = db.case(
        (EndpointState.lastreportedat.is_(None), 0),
        (EndpointState.lastreportedat < cutoff, 1),
        else_=2,
    )

    query = (db.session.query(EndpointState, Computer.computerid,
                              Computer.computertypeid, ComputerType.computertype,
                              Asset.assetnumber, statusrank.label('statusrank'))
             .join(Asset, Asset.assetid == EndpointState.assetid)
             .join(Computer, Computer.assetid == EndpointState.assetid)
             .outerjoin(ComputerType,
                        ComputerType.computertypeid == Computer.computertypeid)
             .filter(Asset.isactive.is_(True)))

    # Search covers hostname, asset number and the machine number a PC serves.
    # Hostname alone was too narrow: people look a PC up by the bay it runs.
    if search := (request.args.get('search') or '').strip():
        like = f'%{search}%'
        machinehits = _assetids_serving_machine_like(like)
        clauses = [EndpointState.hostname.ilike(like), Asset.assetnumber.ilike(like)]
        if machinehits:
            clauses.append(EndpointState.assetid.in_(machinehits))
        query = query.filter(db.or_(*clauses))

    if status := request.args.get('status'):
        wanted = {'never': 0, 'quiet': 1, 'reporting': 2}.get(status.lower())
        if wanted is not None:
            query = query.filter(statusrank == wanted)

    if computertypeid := request.args.get('computertypeid'):
        query = query.filter(Computer.computertypeid == computertypeid)
    if locationid := request.args.get('locationid'):
        query = query.filter(Asset.locationid == locationid)
    if businessunitid := request.args.get('businessunitid'):
        query = query.filter(Asset.businessunitid == businessunitid)

    sort = (request.args.get('sort') or 'status').lower()
    sorts = {
        'status':       (statusrank.asc(), EndpointState.hostname.asc()),
        'hostname':     (EndpointState.hostname.asc(),),
        'lastreported': (EndpointState.lastreportedat.desc(),),
        'uptime':       (EndpointState.uptimeseconds.desc(),),
    }
    query = query.order_by(*sorts.get(sort, sorts['status']))

    items, total = paginate_query(query, page, per_page)

    # Batch the per-row extras for this PAGE only. Doing them inline would be
    # one query per endpoint for machines and another for events.
    assetids = [state.assetid for state, *_ in items]
    machines = _machines_for_assets(assetids)
    critical = _critical_counts_for_assets(assetids)

    data = []
    for state, computerid, computertypeid, computertype, assetnumber, rank in items:
        row = state.to_dict()
        row['computerid'] = computerid
        row['assetnumber'] = assetnumber
        row['computertypeid'] = computertypeid
        row['computertype'] = computertype
        row['uptimedisplay'] = _format_uptime(state.uptimeseconds)
        row['status'] = ('never', 'quiet', 'reporting')[rank]
        row['machines'] = machines.get(state.assetid, [])
        row['health'] = _health_for(state, critical.get(state.assetid))
        data.append(row)

    return paginated_response(data, page, per_page, total)


def _assetids_serving_machine_like(like):
    """Asset ids of PCs whose controlled machine matches a search term.

    Checked in BOTH directions on purpose. The canonical edge is PC controls
    machine, but production also holds reciprocal rows, so matching only
    source->target silently drops half the fleet from a machine-number search.
    """
    from shopdb.api import AssetRelationship, RelationshipType, AssetType

    controls = (RelationshipType.query
                .filter(db.func.lower(RelationshipType.relationshiptype) == 'controls')
                .first())
    if controls is None:
        return []

    # Only MACHINE assets. A PC also 'controls' printers and other kit, and
    # matching any controlled asset put office locations in the machine column.
    other = db.aliased(Asset)
    othertype = db.aliased(AssetType)
    forward = (db.session.query(AssetRelationship.sourceassetid)
               .join(other, other.assetid == AssetRelationship.targetassetid)
               .join(othertype, othertype.assettypeid == other.assettypeid)
               .filter(AssetRelationship.relationshiptypeid == controls.relationshiptypeid,
                       db.func.lower(othertype.assettype) == 'machine',
                       other.assetnumber.ilike(like)))
    backward = (db.session.query(AssetRelationship.targetassetid)
                .join(other, other.assetid == AssetRelationship.sourceassetid)
                .join(othertype, othertype.assettypeid == other.assettypeid)
                .filter(AssetRelationship.relationshiptypeid == controls.relationshiptypeid,
                        db.func.lower(othertype.assettype) == 'machine',
                        other.assetnumber.ilike(like)))
    return [r[0] for r in forward.union(backward).all()]


def _machines_for_assets(assetids):
    """{assetid: [{assetid, assetnumber, name}]} for the machines each PC serves.

    Both edge directions again, deduplicated by machine asset id. One query for
    the whole page rather than one per row.
    """
    from shopdb.api import AssetRelationship, RelationshipType, AssetType

    if not assetids:
        return {}

    controls = (RelationshipType.query
                .filter(db.func.lower(RelationshipType.relationshiptype) == 'controls')
                .first())
    if controls is None:
        return {}

    other = db.aliased(Asset)
    othertype = db.aliased(AssetType)
    forward = (db.session.query(AssetRelationship.sourceassetid.label('pc'),
                                other.assetid, other.assetnumber, other.name)
               .join(other, other.assetid == AssetRelationship.targetassetid)
               .join(othertype, othertype.assettypeid == other.assettypeid)
               .filter(AssetRelationship.relationshiptypeid == controls.relationshiptypeid,
                       db.func.lower(othertype.assettype) == 'machine',
                       AssetRelationship.sourceassetid.in_(assetids)))
    backward = (db.session.query(AssetRelationship.targetassetid.label('pc'),
                                 other.assetid, other.assetnumber, other.name)
                .join(other, other.assetid == AssetRelationship.sourceassetid)
                .join(othertype, othertype.assettypeid == other.assettypeid)
                .filter(AssetRelationship.relationshiptypeid == controls.relationshiptypeid,
                        db.func.lower(othertype.assettype) == 'machine',
                        AssetRelationship.targetassetid.in_(assetids)))

    out = {}
    for pc, machineid, assetnumber, name in forward.union(backward).all():
        seen = out.setdefault(pc, [])
        if not any(m['assetid'] == machineid for m in seen):
            seen.append({'assetid': machineid, 'assetnumber': assetnumber, 'name': name})
    return out


def _critical_counts_for_assets(assetids):
    """{assetid: (total, dominant_eventid)} of critical events in the window.

    Broken down BY EVENT ID, not just totalled. A bare count says "418
    critical", which tells an operator nothing about what to do; naming the
    event that dominates - service timeout, application crash - is the
    difference between a number and a lead.
    """
    if not assetids:
        return {}
    cutoff = _naive_utcnow() - timedelta(hours=48)
    rows = (db.session.query(EndpointEvent.assetid, EndpointEvent.eventid,
                             db.func.count(EndpointEvent.endpointeventid))
            .filter(EndpointEvent.assetid.in_(assetids),
                    EndpointEvent.eventid.in_(CRITICAL_EVENT_IDS),
                    EndpointEvent.eventtime.isnot(None),
                    EndpointEvent.eventtime >= cutoff)
            .group_by(EndpointEvent.assetid, EndpointEvent.eventid)
            .all())

    perasset = {}
    for assetid, eventid, count in rows:
        total, dominant, best = perasset.get(assetid, (0, None, 0))
        total += count
        if count > best:
            dominant, best = eventid, count
        perasset[assetid] = (total, dominant, best)
    return {assetid: (total, dominant)
            for assetid, (total, dominant, _best) in perasset.items()}


def _health_for(state, critical):
    """Worst-of rollup: a stopped watched service, or a recent critical event.

    Deliberately a PHRASE, not a score. "eDNC stopped" tells an operator what
    to do; a percentage does not - and neither does "418 critical", which is
    why the dominant event is named rather than merely counted.
    """
    if critical:
        total, dominant = critical
        title = event_title(dominant) if dominant else 'critical'
        label = title if total == 1 else '{} x{}'.format(title, total)
        return {'level': 'critical', 'label': label}

    stopped = []
    try:
        for entry in json.loads(state.servicesjson or '[]'):
            if not _service_running(entry):
                stopped.append(_service_name(entry))
    except (TypeError, ValueError):
        # Malformed payload is not a health verdict - say nothing rather than
        # claiming the endpoint is unhealthy.
        return {'level': 'unknown', 'label': None}

    if stopped:
        first = stopped[0]
        extra = f' +{len(stopped) - 1}' if len(stopped) > 1 else ''
        return {'level': 'warning', 'label': f'{first} stopped{extra}'}

    if state.lastreportedat is None:
        return {'level': 'unknown', 'label': 'no agent'}
    return {'level': 'ok', 'label': 'ok'}


@endpoints_bp.route('/<int:asset_id>', methods=['GET'])
@require_permission('endpoints.view')
def get_endpoint(asset_id: int):
    """Get the current state for one endpoint by asset id.

    Carries the same reporting/quiet/never verdict the fleet list uses. The
    quiet window is a site setting, so the SERVER decides it: a caller that
    guessed a threshold in the browser would disagree with the fleet list on
    the same endpoint the moment a site changed the setting.
    """
    state = EndpointState.query.filter_by(assetid=asset_id).first()

    if not state:
        return error_response(
            ErrorCodes.NOT_FOUND,
            f'No endpoint state for asset {asset_id}',
            http_code=404,
        )

    cutoff = _naive_utcnow() - timedelta(hours=_quiet_hours())
    if state.lastreportedat is None:
        status = 'never'
    elif state.lastreportedat < cutoff:
        status = 'quiet'
    else:
        status = 'reporting'

    data = state.to_dict()
    data['status'] = status
    data['uptimedisplay'] = _format_uptime(state.uptimeseconds)
    return success_response(data)


@endpoints_bp.route('/<int:asset_id>/events', methods=['GET'])
@require_permission('endpoints.view')
def list_endpoint_events(asset_id: int):
    """List curated Windows events for one endpoint, newest first."""
    page, per_page = get_pagination_params(request)

    query = EndpointEvent.query.filter_by(assetid=asset_id)

    if eventid := request.args.get('eventid'):
        try:
            query = query.filter(EndpointEvent.eventid == int(eventid))
        except (TypeError, ValueError):
            return error_response(ErrorCodes.VALIDATION_ERROR,
                                  'eventid must be an integer')

    query = query.order_by(EndpointEvent.eventtime.desc())

    items, total = paginate_query(query, page, per_page)
    data = [e.to_dict() for e in items]

    return paginated_response(data, page, per_page, total)


# =============================================================================
# Asset panels (ADR-010) - shaped for the generic PluginAssetPanels renderer
# =============================================================================

@endpoints_bp.route('/asset/<int:asset_id>/status', methods=['GET'])
@require_permission('endpoints.view')
def panel_status(asset_id: int):
    """Keyvalue panel: the endpoint's current status snapshot.

    Single indexed lookup on endpointstates.assetid. Returns the {fields:[...]}
    shape the keyvalue renderer expects; an absent state yields no fields so the
    panel hides itself.
    """
    state = EndpointState.query.filter_by(assetid=asset_id).first()
    if not state:
        return success_response({'fields': []})

    def field(label, value, mono=False):
        return {'label': label, 'value': value, 'mono': mono}

    fields = [
        field('Logged in user', state.loggedinuser),
        field('Uptime', _format_uptime(state.uptimeseconds)),
        field('Idle', _format_uptime(state.idleseconds)),
        field('Last boot',
              state.lastboottime.isoformat() + 'Z' if state.lastboottime else None),
        field('CPU %', state.cpupercent),
        field('Memory %', state.memorypercent),
        field('Agent', state.agentversion, mono=True),
        field('Last reported',
              state.lastreportedat.isoformat() + 'Z' if state.lastreportedat else None),
    ]
    fields = [f for f in fields if f['value'] not in (None, '')]
    return success_response({'fields': fields})


@endpoints_bp.route('/asset/<int:asset_id>/telemetrymeta', methods=['GET'])
@require_permission('endpoints.view')
def panel_telemetry_meta(asset_id: int):
    """Lightweight freshness probe for the panels.

    Returns just the endpoint's lastreportedat so the WebUI can tell whether a
    genuinely new telemetry snapshot has landed after it forces a collect. The
    UI captures this before enqueuing collect_now (or a row Start/Stop), then
    polls until the stamp advances - a command reaching 'succeeded' only means
    the agent ran it, not that it has yet posted the fresh services snapshot, so
    gating the refresh on this timestamp is what avoids re-rendering stale rows.
    """
    state = EndpointState.query.filter_by(assetid=asset_id).first()
    stamp = (state.lastreportedat.isoformat() + 'Z'
             if state and state.lastreportedat else None)
    return success_response({'lastreportedat': stamp})


@endpoints_bp.route('/asset/<int:asset_id>/services', methods=['GET'])
@require_permission('endpoints.view')
def panel_services(asset_id: int):
    """List panel: monitored services/processes with a run-state badge per row.

    Reshapes the latest servicesjson snapshot into rows the list renderer maps
    to name + status badge. Tolerant of a missing or unexpected snapshot shape:
    an unknown state renders as an Unknown badge rather than failing.
    """
    state = EndpointState.query.filter_by(assetid=asset_id).first()
    if not state or not state.servicesjson:
        return success_response([])

    maycommand = caller_holds('endpoints.command')

    try:
        services = json.loads(state.servicesjson)
    except (TypeError, ValueError):
        return success_response([])

    rows = []
    for entry in (services or []):
        if isinstance(entry, dict):
            name = entry.get('name') or entry.get('service') or 'Unknown'
            raw = (entry.get('state') or entry.get('status') or '').strip().lower()
            app = entry.get('app')
            kind = entry.get('kind')
        else:
            name, raw, app, kind = str(entry), '', None, None
        running = _service_running(entry)
        row = {
            'name': name,
            'kind': (kind or '').lower() or None,
            'statelabel': 'Running' if running else (raw.title() or 'Unknown'),
            # Amber for paused and the four Windows transitional states. A
            # service stuck in Start Pending is failing to start on a retry
            # cycle - which is what raises event 7011 - and a paused one was
            # paused deliberately. Painting either the same red as Stopped
            # loses the distinction that tells you which to chase.
            'statecolor': ('green' if running
                           else 'yellow' if _service_transitional(raw)
                           else 'red' if raw else 'gray'),
        }
        # Only offer Start/Stop to a caller who could actually run it. The
        # emission never consulted the caller, so a read-only viewer was shown
        # controls on a shop-floor PC that 403 on click - wrong signal, and it
        # gets worse the moment guests can read this panel.
        targets = row_actions_for(app, kind, name) if maycommand else None
        if targets:
            row['action'] = {
                'running': running,
                'start': targets['start'],
                'stop': targets['stop'],
                'status': targets['status'],
                'target': targets['target'],
                'targetkind': targets['targetkind'],
            }
        rows.append(row)
    return success_response(rows)


@endpoints_bp.route('/asset/<int:asset_id>/recentevents', methods=['GET'])
@require_permission('endpoints.view')
def panel_recent_events(asset_id: int):
    """List panel: the endpoint's most recent curated events, newest first.

    Indexed range scan on (assetid, eventtime), capped small - this is the
    at-a-glance history card, not the full report (that is the events route
    with pagination).
    """
    events = (EndpointEvent.query
              .filter_by(assetid=asset_id)
              .order_by(EndpointEvent.eventtime.desc())
              .limit(10)
              .all())

    rows = [{
        'title': event_title(e.eventid),
        'severity': event_severity(e.eventid),
        'eventtime': e.eventtime.isoformat() + 'Z' if e.eventtime else None,
    } for e in events]
    return success_response(rows)


@endpoints_bp.route('/asset/<int:asset_id>/activity', methods=['GET'])
@require_permission('endpoints.view')
def panel_activity(asset_id: int):
    """List panel: one chronological feed of what happened to this endpoint.

    Curated Windows events and operator commands merged and sorted newest
    first. They answer the same question - what happened to this box - and
    reading them as two separate cards means interleaving them by eye.

    Both sources are already keyed by assetid, so this is two indexed reads and
    a merge sort, not a join. The core AuditLog is deliberately NOT a third
    source: its rows carry the action as the entity name and the asset only
    inside a details string, so pulling it in would mean parsing prose.

    The requester's name is withheld from callers who cannot themselves issue
    commands. A read-only viewer - a guest, on a site that grants this panel -
    has no need for the roster of who restarted what, and the feed is just as
    useful without it.
    """
    naming = caller_holds('endpoints.command')

    events = (EndpointEvent.query
              .filter_by(assetid=asset_id)
              .order_by(EndpointEvent.eventtime.desc())
              .limit(15)
              .all())

    # The newest 15 of ANY severity can be entirely routine logons while the
    # health rollup is shouting about a burst of crashes underneath them - the
    # page then contradicts itself for no visible reason. Pull the newest
    # criticals separately and merge, so what health is reporting is always
    # present in the feed that is supposed to explain it.
    criticals = (EndpointEvent.query
                 .filter(EndpointEvent.assetid == asset_id,
                         EndpointEvent.eventid.in_(CRITICAL_EVENT_IDS))
                 .order_by(EndpointEvent.eventtime.desc())
                 .limit(5)
                 .all())
    seen = {e.endpointeventid for e in events}
    events.extend(e for e in criticals if e.endpointeventid not in seen)

    commands = (EndpointCommand.query
                .filter_by(assetid=asset_id)
                .order_by(EndpointCommand.createdat.desc())
                .limit(15)
                .all())

    # requestedby holds a user ID, not a name. One batched lookup rather than a
    # query per row, and only when the caller is allowed to see the names at all.
    names = {}
    if naming:
        userids = {c.requestedby for c in commands if c.requestedby}
        if userids:
            names = {u.userid: (u.username or '')
                     for u in User.query.filter(User.userid.in_(userids)).all()}

    rows = []
    for event in events:
        rows.append({
            'title': event_title(event.eventid),
            'severity': event_severity(event.eventid),
            'detail': 'Windows event {}'.format(event.eventid),
            'eventtime': event.eventtime,
        })

    for command in commands:
        label = COMMAND_ALLOWLIST.get(command.action, {}).get('label') or command.action
        target = ' on {}'.format(command.target) if command.target else ''
        who = ''
        if naming and command.requestedby:
            username = names.get(command.requestedby)
            if username:
                who = ' by {}'.format(username)
        rows.append({
            'title': '{}{}'.format(label, target),
            # A failed command is the one an operator needs to notice.
            'severity': 'critical' if command.status == EndpointCommand.STATUS_FAILED else 'info',
            'detail': '{}{}'.format((command.status or 'queued').title(), who),
            'eventtime': command.createdat,
        })

    # Undated rows sort last rather than crashing the comparison.
    rows.sort(key=lambda r: (r['eventtime'] is not None, r['eventtime']), reverse=True)

    # Trim to the cap without dropping the criticals that were fetched on
    # purpose: they keep their place in time, but they survive the cut.
    if len(rows) > 15:
        keep = [r for r in rows if r['severity'] == 'critical'][:5]
        rest = [r for r in rows if r not in keep]
        rows = sorted(keep + rest[:15 - len(keep)],
                      key=lambda r: (r['eventtime'] is not None, r['eventtime']),
                      reverse=True)

    # The list renderer reads a badge colour from a ROW FIELD, so without this
    # every badge falls back to grey and the feed loses the one signal it is
    # there to carry.
    colours = {'critical': 'red', 'warning': 'yellow', 'info': 'gray'}

    return success_response([{
        'title': r['title'],
        'severity': r['severity'],
        'severitycolor': colours.get(r['severity'], 'gray'),
        'detail': r['detail'],
        'eventtime': r['eventtime'].isoformat() + 'Z' if r['eventtime'] else None,
    } for r in rows[:15]])


@endpoints_bp.route('/asset/<int:asset_id>/interfaces', methods=['GET'])
@require_permission('endpoints.view')
def panel_interfaces(asset_id: int):
    """Table panel: every hardware NIC the agent last reported.

    Reshapes the latest interfacesjson snapshot into one row per adapter with
    the {alias, ipv4, mac, gateway, subnet, dhcp} shape the collector emits. A
    gatewayless adapter is flagged as the likely machine-IPS link (see
    DESIGN-DECISIONS-agent.md 9.3): those NICs speak to the machine, not the
    corporate LAN. Tolerant of a missing or unexpected snapshot: an unreadable
    payload yields no rows so the panel hides itself.
    """
    state = EndpointState.query.filter_by(assetid=asset_id).first()
    if not state or not state.interfacesjson:
        return success_response([])

    try:
        interfaces = json.loads(state.interfacesjson)
    except (TypeError, ValueError):
        return success_response([])

    rows = []
    for entry in (interfaces or []):
        if not isinstance(entry, dict):
            continue
        gateway = (entry.get('gateway') or '').strip()
        ipv4 = (entry.get('ipv4') or '').strip()
        rows.append({
            'alias': entry.get('alias') or 'Unknown',
            'ipv4': ipv4 or '-',
            'mac': (entry.get('mac') or '').strip() or '-',
            'gateway': gateway or '-',
            'subnet': (entry.get('subnet') or '').strip() or '-',
            'dhcp': 'DHCP' if entry.get('dhcp') else 'Static',
            # A gatewayless NIC with an address is the machine link; the gateway
            # NIC is the corporate LAN.
            'role': 'Machine link' if (ipv4 and not gateway) else 'LAN',
        })
    return success_response(rows)


@endpoints_bp.route('/asset/<int:asset_id>/applications', methods=['GET'])
@require_permission('endpoints.view')
def panel_applications(asset_id: int):
    """Table panel: the installed-software (SBOM) snapshot for this endpoint.

    Reads the endpointapplications rows the collector last wrote for this asset
    (a per-report snapshot, not history) and returns them sorted by name. This
    is the raw discovered inventory: appid is carried through but not yet linked
    to the applications catalog ("Managed Applications"), so no managed status
    is surfaced here. An endpoint with no reported SBOM yields no rows and the
    panel hides itself.
    """
    rows = (EndpointApplication.query
            .filter_by(assetid=asset_id)
            .order_by(db.func.lower(EndpointApplication.displayname))
            .all())
    return success_response([
        {
            'displayname': row.displayname,
            'displayversion': row.displayversion or '-',
            'publisher': row.publisher or '-',
        }
        for row in rows
    ])


# =============================================================================
# Dashboard
# =============================================================================

@endpoints_bp.route('/dashboard/summary', methods=['GET'])
@require_permission('endpoints.fleet')
def dashboard_summary():
    """High-level fleet counts for the endpoints dashboard."""
    total = db.session.query(EndpointState).join(
        Asset, Asset.assetid == EndpointState.assetid
    ).filter(Asset.isactive.is_(True)).count()

    hours = _quiet_hours()
    cutoff = _naive_utcnow() - timedelta(hours=hours)

    reporting = db.session.query(EndpointState).join(
        Asset, Asset.assetid == EndpointState.assetid
    ).filter(
        Asset.isactive.is_(True),
        EndpointState.lastreportedat.isnot(None),
        EndpointState.lastreportedat >= cutoff,
    ).count()

    # Counted separately from quiet. An endpoint that has NEVER reported is a
    # different problem from one that reported and stopped: the first is an
    # enrolment or key failure, the second is a machine that went down. Folding
    # them together (quiet = total - reporting) hid the enrolment failures.
    never = db.session.query(EndpointState).join(
        Asset, Asset.assetid == EndpointState.assetid
    ).filter(
        Asset.isactive.is_(True),
        EndpointState.lastreportedat.is_(None),
    ).count()

    return success_response({
        'total': total,
        'reporting': reporting,
        'quiet': total - reporting - never,
        'never': never,
        'quiethours': hours,
    })


@endpoints_bp.route('/dashboard/quiet', methods=['GET'])
@require_permission('endpoints.fleet')
def dashboard_quiet():
    """Endpoints that reported before, then went quiet past the window.

    Never-reported endpoints are excluded: those are records without an agent,
    not agents that stopped. Silence is the only signal - there is no heartbeat
    separate from the report.
    """
    hours = _quiet_hours()
    cutoff = _naive_utcnow() - timedelta(hours=hours)
    now = _naive_utcnow()

    from plugins.computers.models import Computer

    rows = (db.session.query(EndpointState, Computer.computerid)
            .join(Asset, Asset.assetid == EndpointState.assetid)
            .join(Computer, Computer.assetid == EndpointState.assetid)
            .filter(Asset.isactive.is_(True),
                    EndpointState.lastreportedat.isnot(None),
                    EndpointState.lastreportedat < cutoff)
            .all())

    out = [{
        'computerid': computerid,
        'assetid': s.assetid,
        'hostname': s.hostname,
        'lastreported': s.lastreportedat.isoformat() + 'Z',
        'quietdays': (now - s.lastreportedat).days,
    } for s, computerid in rows]

    out.sort(key=lambda r: -r['quietdays'])
    return success_response(out[:50])


@endpoints_bp.route('/dashboard/stability', methods=['GET'])
@require_permission('endpoints.fleet')
def dashboard_stability():
    """Endpoints with recent crash/shutdown/corruption events.

    Bounded to a fixed 48h window and the critical event catalog, capped at 50
    rows: cost is a range scan on eventtime filtered to a short id list, not a
    scan of all history. Joins through Computer so each row links to the PC
    detail page.
    """
    from plugins.computers.models import Computer

    if not CRITICAL_EVENT_IDS:
        return success_response([])

    cutoff = _naive_utcnow() - timedelta(hours=48)

    rows = (db.session.query(
                Computer.computerid,
                EndpointState.hostname,
                db.func.count(EndpointEvent.endpointeventid),
                db.func.max(EndpointEvent.eventtime),
            )
            .select_from(EndpointEvent)
            .join(Asset, Asset.assetid == EndpointEvent.assetid)
            .join(Computer, Computer.assetid == EndpointEvent.assetid)
            .outerjoin(EndpointState,
                       EndpointState.assetid == EndpointEvent.assetid)
            .filter(Asset.isactive.is_(True),
                    EndpointEvent.eventid.in_(CRITICAL_EVENT_IDS),
                    EndpointEvent.eventtime.isnot(None),
                    EndpointEvent.eventtime >= cutoff)
            .group_by(Computer.computerid, EndpointState.hostname)
            .all())

    out = [{
        'computerid': computerid,
        'hostname': hostname,
        'eventcount': count,
        'lastevent': lasttime.isoformat() + 'Z' if lasttime else None,
    } for computerid, hostname, count, lasttime in rows]

    out.sort(key=lambda r: -r['eventcount'])
    return success_response(out[:50])


def _quiet_hours() -> int:
    """Quiet-report window in hours, from a site setting, default 24."""
    hours = 24
    setting = Setting.query.filter_by(key='endpoints_quietreporthours').first()
    if setting and (setting.value or '').strip():
        try:
            hours = int(setting.value)
        except (TypeError, ValueError):
            pass
    return hours


# =============================================================================
# Endpoint agent key (UI-managed collector.ingest.endpoints kill switch)
# =============================================================================

# The single permission the endpoint agent key authorizes. A managed token
# scoped to ONLY this authorizes the endpoints collector and nothing else, so
# rotating/revoking it stops endpoint-agent traffic without touching the
# computers/geenforce collectors.
AGENT_KEY_SCOPE = 'collector.ingest.endpoints'
AGENT_KEY_NAME = 'Endpoint agent key'


def _agentkey_status_payload():
    """Current state of the endpoint agent key for the settings page.

    Reports whether an active managed key exists (and its non-secret metadata)
    plus whether the bootstrap .env fallback is still accepted. Never exposes a
    secret; the full secret is only ever returned once, at regenerate time.
    """
    tokens = list_service_tokens(AGENT_KEY_SCOPE)
    active = tokens[0] if tokens else None
    envfallback = Setting.get('plugin.endpoints.envkeyfallback', default=True)
    provisioned = Setting.get('plugin.endpoints.agentkeyprovisioned',
                              default=False)
    return {
        'provisioned': bool(provisioned) or active is not None,
        'envkeyfallback': bool(envfallback),
        'key': active.to_dict() if active is not None else None,
    }


@endpoints_bp.route('/agentkey', methods=['GET'])
@require_role('admin')
def get_agentkey():
    """Report the endpoint agent key state (no secret)."""
    return success_response(_agentkey_status_payload())


@endpoints_bp.route('/agentkey/regenerate', methods=['POST'])
@require_role('admin')
def regenerate_agentkey():
    """Mint a fresh endpoint agent key, revoking any previous one.

    Hard cutover: the old key is deactivated in the same call, and the .env
    fallback is switched off, so the new managed token becomes the sole
    credential and revoking it is a true kill switch. The full secret is
    returned ONCE here and never again.
    """
    try:
        token, secret = mint_service_token(
            AGENT_KEY_SCOPE, AGENT_KEY_NAME, current_user.userid)
    except PermissionError as error:
        return error_response(ErrorCodes.FORBIDDEN, str(error), http_code=403)

    Setting.set('plugin.endpoints.agentkeyprovisioned', True,
                valuetype='boolean', category='plugin.endpoints')
    Setting.set('plugin.endpoints.envkeyfallback', False,
                valuetype='boolean', category='plugin.endpoints')

    result = _agentkey_status_payload()
    # The secret appears here and NOWHERE else, ever. Not stored, not logged.
    result['secret'] = secret
    result['warning'] = ('Save this key now. It will not be shown again. '
                         'Configure it as the agent X-API-Key.')
    return success_response(result, message='Endpoint agent key generated',
                            http_code=201)


@endpoints_bp.route('/agentkey', methods=['DELETE'])
@require_role('admin')
def revoke_agentkey():
    """Revoke the endpoint agent key: stop all managed-token endpoint traffic.

    Leaves envkeyfallback off (the kill switch stays armed). To restore
    reporting an operator regenerates a new key.
    """
    revoked = revoke_service_tokens(AGENT_KEY_SCOPE)
    Setting.set('plugin.endpoints.agentkeyprovisioned', False,
                valuetype='boolean', category='plugin.endpoints')
    result = _agentkey_status_payload()
    result['revoked'] = revoked
    return success_response(result, message='Endpoint agent key revoked')


# =============================================================================
# Endpoint commands (WebUI -> agent, allowlisted actions only)
# =============================================================================
#
# Operators enqueue a named, server-allowlisted action for one endpoint (JWT +
# endpoints.command). The agent long-polls /commands/next with its collector
# token, claims the single next pending command, runs it locally, and reports
# the outcome to /commands/<id>/result. The wire never carries a command string
# - only an allowlisted action name - so a compromised WebUI cannot ask an
# endpoint to run something arbitrary.

# How long the agent claim poll is held open server-side before returning empty.
# Kept below the agent HttpClient timeout and typical proxy idle timeouts so a
# quiet fleet re-polls cleanly instead of tripping a read timeout.
_CLAIM_HOLD_SECONDS = 25
_CLAIM_POLL_INTERVAL = 1.0


def _pin_holds(command, descriptor):
    """Does the descriptor still carry the digest this command was queued for?

    True when nothing was pinned (actions that deliver no content), or when the
    descriptor is missing entirely - a vanished script or version already sends
    the command through with no descriptor so the agent fails closed, and that
    path is not a content swap.
    """
    if not command.pinnedsha256:
        return True
    if descriptor is None:
        return True
    return hmac.compare_digest(
        str(descriptor.get('sha256') or ''), command.pinnedsha256)


def _fail_on_pin_drift(command, action):
    """Refuse a claim whose content changed after the operator approved it."""
    command.status = EndpointCommand.STATUS_FAILED
    command.resultmessage = (
        'Refused: the {} content changed after this command was queued'
        .format(action))
    command.completedat = _naive_utcnow()
    db.session.commit()
    current_app.logger.warning(
        'Endpoint command %s (%s, asset %s) refused: pinned sha256 %s no '
        'longer matches the resolved descriptor',
        command.endpointcommandid, action, command.assetid,
        command.pinnedsha256)
    AuditLog.log('refused', 'EndpointCommand',
                 entityid=command.endpointcommandid,
                 entityname=action,
                 details='content digest changed after enqueue')
    db.session.commit()
    return ('', 204)


def _reap_stale_commands(asset_id):
    """Expire in-flight commands an agent will never finish for this endpoint.

    The queue is at-most-one-in-flight: a single PENDING or RUNNING row blocks
    every new command. If the agent is killed, uninstalled, or goes offline
    after a command is claimed (RUNNING) or before it claims one (PENDING), that
    row would pin the queue forever and every operator action returns 409. Reap
    rows past their staleness ceiling to STATUS_EXPIRED so the queue self-heals.
    Caller commits (this runs inside the enqueue transaction).
    """
    now = _naive_utcnow()
    running_cutoff = now - timedelta(seconds=EndpointCommand.RUNNING_STALE_SECONDS)
    pending_cutoff = now - timedelta(seconds=EndpointCommand.PENDING_STALE_SECONDS)

    stale = (EndpointCommand.query
             .filter(EndpointCommand.assetid == asset_id,
                     db.or_(
                         db.and_(
                             EndpointCommand.status == EndpointCommand.STATUS_RUNNING,
                             EndpointCommand.claimedat.isnot(None),
                             EndpointCommand.claimedat < running_cutoff),
                         db.and_(
                             EndpointCommand.status == EndpointCommand.STATUS_PENDING,
                             EndpointCommand.createdat < pending_cutoff)))
             .all())
    for command in stale:
        command.status = EndpointCommand.STATUS_EXPIRED
        command.completedat = now
        if not command.resultmessage:
            command.resultmessage = (
                'Expired: agent did not report a result before the staleness '
                'ceiling (agent offline, killed, or reinstalled mid-command).')
    return stale


def _collector_auth_or_error():
    """Authorize an agent-facing command route on the endpoints collector token.

    Reuses the collector auth path so the command channel shares the endpoint
    agent key (kill switch) and introduces no second secret.
    """
    return collector_authorized('endpoints')


def _resolve_asset_for_hostname(hostname):
    """Resolve a reported hostname to its asset id via the computers plugin."""
    from plugins.computers.models import Computer

    computer = (Computer.query
                .filter(db.func.lower(Computer.hostname) == hostname.lower())
                .first())
    return computer.assetid if computer is not None else None


_HOSTED_PACKAGE_PREFIX = '/api/applications/package/'


def _hosted_package_filename(application):
    """Return the stored filename for a ShopDB-hosted installer, or None.

    installpath is the core package URL (/api/applications/package/<filename>);
    a site-typed share path (not our upload) has no servable file here.
    """
    installpath = application.installpath or ''
    if not installpath.startswith(_HOSTED_PACKAGE_PREFIX):
        return None
    return installpath[len(_HOSTED_PACKAGE_PREFIX):]


def _backfill_app_installer_sha256(application):
    """Compute and persist installersha256 for a pre-migration package.

    Packages uploaded before the 7d36 migration have a null installersha256,
    which used to force a re-upload (Bug B). When the hosted package file is
    present on disk, hash it once, stamp the app, and commit so future resolves
    hit the stored value. Returns the digest, or None when no servable file.
    """
    filename = _hosted_package_filename(application)
    if not filename:
        return None
    package_dir = os.path.join(current_app.instance_path, 'applicationpackages')
    package_path = os.path.join(package_dir, filename)
    if not os.path.isfile(package_path):
        return None
    digest = hashlib.sha256()
    with open(package_path, 'rb') as package_file:
        for chunk in iter(lambda: package_file.read(65536), b''):
            digest.update(chunk)
    application.installersha256 = digest.hexdigest()
    db.session.commit()
    return application.installersha256


def _resolve_update_app(target):
    """Resolve an update_app target to a delivery descriptor.

    Accepts either an app-level or a version-level source, so a push works
    whether or not the site tracks AppVersion rows (the applications UI stores a
    single app-level installer; versioning is optional):

      - "app:<appid>"       - the app's own uploaded installer (installersha256)
      - "version:<id>"      - a specific AppVersion's pinned package
      - bare "<int>"        - legacy: treated as an appversionid

    Returns (descriptor, error_message). The descriptor carries what the agent
    needs to fetch and run the pinned installer: a collector-authenticated
    download URL, its SHA-256, and the silent-install switches. The agent
    fetches the bytes with its existing collector credential; no bytes here.
    """
    kind, _sep, raw = str(target or '').partition(':')
    if not raw:
        # No prefix: legacy bare appversionid.
        kind, raw = 'version', kind

    try:
        ident = int(raw)
    except (TypeError, ValueError):
        return None, 'update_app target must be app:<id>, version:<id>, or an appversionid'

    if kind == 'app':
        application = db.session.get(Application, ident)
        if application is None:
            return None, f'Unknown appid: {ident}'
        if not application.installpath:
            return None, (f'Application {ident} has no uploaded installer package')
        # Backfill installersha256 for packages uploaded before the 7d36 migration
        # so app-level delivery resolves without a re-upload (Bug B).
        sha256 = application.installersha256 or _backfill_app_installer_sha256(application)
        if not sha256:
            return None, (f'Application {ident} has no uploaded installer package')
        # Carry the real installer filename (with extension) on the delivery URL so
        # the agent's Path.GetExtension routes MSI through msiexec (Bug A rc2). The
        # trailing filename is cosmetic for the agent; the app_id drives the lookup.
        filename = _hosted_package_filename(application)
        installpath = '/api/endpoints/apppackage/app/{}'.format(application.appid)
        if filename:
            installpath = '{}/{}'.format(installpath, filename)
        return {
            'source': 'app',
            'appid': application.appid,
            'appversionid': None,
            'appname': application.appname,
            'version': None,
            'installpath': installpath,
            'sha256': sha256,
            'silentargs': None,
        }, None

    if kind == 'version':
        version = db.session.get(AppVersion, ident)
        if version is None:
            return None, f'Unknown appversionid: {ident}'
        application = db.session.get(Application, version.appid)
        if application is None or not application.installpath:
            return None, ('Application for appversionid {} has no installer package'
                          .format(ident))
        # PIN THE FILE THAT IS ACTUALLY SERVED.
        #
        # /api/endpoints/apppackage/<appversionid> resolves the version to its
        # APPLICATION and serves application.installpath - there is no
        # version-specific artifact behind that route. So version.sha256 pinned
        # a digest that need not describe the bytes delivered, and where it was
        # null the descriptor carried no digest at all and the agent had nothing
        # to verify against. Deriving it from the application, exactly as the
        # app branch does, makes the pin describe the delivery.
        sha256 = (application.installersha256
                  or _backfill_app_installer_sha256(application))
        if not sha256:
            return None, ('Application for appversionid {} has an installer '
                          'package with no recorded SHA-256, so the delivery '
                          'cannot be verified'.format(ident))

        # Carry the real filename so the agent's Path.GetExtension routes an MSI
        # through msiexec. Without it every MSI push failed, because this branch
        # ended the URL at the id and the extension was simply absent.
        filename = _hosted_package_filename(application)
        installpath = '/api/endpoints/apppackage/{}'.format(version.appversionid)
        if filename:
            installpath = '/api/endpoints/apppackage/app/{}/{}'.format(
                application.appid, filename)

        return {
            'source': 'version',
            'appversionid': version.appversionid,
            'appid': version.appid,
            'appname': application.appname,
            'version': version.version,
            'installpath': installpath,
            'sha256': sha256,
            'silentargs': version.silentargs,
        }, None

    return None, f'Unknown update_app target kind: {kind!r}'


def _resolve_run_script(target):
    """Resolve a run_script target to a script delivery descriptor.

    Accepts either "script:<scriptid>" or a bare script name. The descriptor
    carries a collector-authenticated download URL, the body's SHA-256, and the
    per-script run context (system or user). The agent fetches the body with
    its collector credential and runs it; the body never crosses the wire.
    """
    kind, _sep, raw = str(target or '').partition(':')
    script = None
    if kind == 'script' and raw:
        try:
            script = db.session.get(Script, int(raw))
        except (TypeError, ValueError):
            return None, 'run_script target script:<id> must be numeric'
    else:
        # Bare name (the whole target, prefix included when there was no colon).
        name = (target or '').strip()
        if name:
            script = Script.query.filter_by(scriptname=name).first()

    if script is None:
        return None, f'Unknown script: {target!r}'
    if not script.filename or not script.sha256:
        return None, f'Script {script.scriptname!r} has no uploaded body'
    if not script.isactive:
        return None, f'Script {script.scriptname!r} is inactive'

    return {
        'scriptid': script.scriptid,
        'scriptname': script.scriptname,
        'installpath': '/api/endpoints/scriptbody/{}'.format(script.scriptid),
        'sha256': script.sha256,
        'runcontext': script.runcontext,
    }, None


@endpoints_bp.route('/<int:asset_id>/commands', methods=['GET'])
@require_permission('endpoints.view')
def list_endpoint_commands(asset_id: int):
    """Recent commands for one endpoint, newest first (operator view)."""
    rows = (EndpointCommand.query
            .filter_by(assetid=asset_id)
            .order_by(EndpointCommand.createdat.desc())
            .limit(20)
            .all())
    return success_response([row.to_dict() for row in rows])


@endpoints_bp.route('/<int:asset_id>/commands', methods=['POST'])
@require_permission('endpoints.command')
def enqueue_endpoint_command(asset_id: int):
    """Queue an allowlisted command for one endpoint.

    Rejects unknown actions and refuses to stack a second pending/running
    command on the same endpoint, so the queue stays at-most-one-in-flight and
    the WebUI toggle cannot race itself.
    """
    payload = request.get_json(silent=True) or {}
    action = (payload.get('action') or '').strip()
    if not is_allowed_action(action):
        return error_response(ErrorCodes.VALIDATION_ERROR,
                              f'Unknown or disallowed action: {action!r}')

    # Destructive actions (for example system_restart) demand a stricter grant
    # ON TOP OF endpoints.command, so a remote reboot cannot ride the routine
    # toggle permission.
    elevated = elevated_permission_for(action)
    if elevated and not current_user.haspermission(elevated):
        return error_response(
            ErrorCodes.FORBIDDEN,
            f'Action {action!r} requires the {elevated} permission',
            http_code=403)

    if action not in ALLOWED_ACTIONS_FOR(assetid=asset_id):
        return error_response(
            ErrorCodes.VALIDATION_ERROR,
            f'Action not permitted by policy for this endpoint: {action!r}')

    # Generic actions (service_* / process_*) must carry the subject to act on;
    # fixed-target actions (mtconnect_*) must not. Validate both directions so a
    # caller cannot send a targetless generic action (nothing to act on) or
    # smuggle a target onto a pinned action.
    target = (payload.get('target') or '').strip() or None
    targetkind = action_targetkind(action)
    if requires_target(action):
        if not target:
            return error_response(
                ErrorCodes.VALIDATION_ERROR,
                f'Action {action!r} requires a target')
    elif target is not None:
        return error_response(
            ErrorCodes.VALIDATION_ERROR,
            f'Action {action!r} does not accept a target')

    # update_app pins a specific application version; verify it resolves to a
    # real installer now so a broken push is rejected at enqueue rather than
    # failing opaquely on the endpoint minutes later.
    # Digest of the content this push was APPROVED against. Kept on the row so
    # the claim path can refuse a descriptor that has changed since; see
    # EndpointCommand.pinnedsha256.
    pinnedsha256 = None

    if action == 'update_app':
        descriptor, resolve_error = _resolve_update_app(target)
        if resolve_error is not None:
            return error_response(ErrorCodes.VALIDATION_ERROR, resolve_error)
        pinnedsha256 = (descriptor or {}).get('sha256')

    # run_script pins a ShopDB-hosted script; verify it resolves to a real
    # uploaded body now so a broken push is rejected at enqueue rather than
    # failing opaquely on the endpoint later.
    if action == 'run_script':
        script, script_error = _resolve_run_script(target)
        if script_error is not None:
            return error_response(ErrorCodes.VALIDATION_ERROR, script_error)
        pinnedsha256 = (script or {}).get('sha256')

    state = EndpointState.query.filter_by(assetid=asset_id).first()
    if state is None:
        return error_response(ErrorCodes.NOT_FOUND,
                              f'No endpoint state for asset {asset_id}',
                              http_code=404)

    # Self-heal a queue wedged by a command the agent will never finish (killed,
    # uninstalled, or offline mid-flight) before enforcing at-most-one-in-flight,
    # so a dead in-flight row cannot permanently 409 every operator action.
    _reap_stale_commands(asset_id)

    pending = (EndpointCommand.query
               .filter(EndpointCommand.assetid == asset_id,
                       EndpointCommand.status.in_((
                           EndpointCommand.STATUS_PENDING,
                           EndpointCommand.STATUS_RUNNING)))
               .first())
    if pending is not None:
        return error_response(
            ErrorCodes.VALIDATION_ERROR,
            'A command is already in flight for this endpoint',
            http_code=409)

    command = EndpointCommand(
        assetid=asset_id,
        action=action,
        target=target,
        targetkind=targetkind if target else None,
        status=EndpointCommand.STATUS_PENDING,
        requestedby=getattr(current_user, 'userid', None),
        pinnedsha256=pinnedsha256,
    )
    db.session.add(command)

    AuditLog.log('created', 'EndpointCommand',
                 entityname=action,
                 details=f'action={action} target={target} asset={asset_id}')

    db.session.commit()
    return success_response(command.to_dict(),
                            message='Command queued', http_code=201)


@endpoints_bp.route('/commands/next', methods=['GET'])
def claim_next_command():
    """Agent long-poll: claim the next pending command for a hostname.

    Collector-token auth (not JWT). Holds the request open briefly and returns
    the single oldest pending command for the caller's endpoint, atomically
    marking it running so no other poll claims it. Returns 204 when nothing is
    queued within the hold window - the agent simply re-polls.
    """
    error = _collector_auth_or_error()
    if error is not None:
        return error

    hostname = (request.args.get('hostname') or '').strip()
    if not hostname:
        return error_response(ErrorCodes.VALIDATION_ERROR,
                              'hostname is required')

    asset_id = _resolve_asset_for_hostname(hostname)
    if asset_id is None:
        return error_response(ErrorCodes.NOT_FOUND,
                              f'Unknown hostname: {hostname!r}', http_code=404)

    deadline = time.monotonic() + _CLAIM_HOLD_SECONDS
    while True:
        command = (EndpointCommand.query
                   .filter_by(assetid=asset_id,
                              status=EndpointCommand.STATUS_PENDING)
                   .order_by(EndpointCommand.createdat.asc())
                   .first())
        if command is not None:
            command.status = EndpointCommand.STATUS_RUNNING
            command.claimedat = _naive_utcnow()
            db.session.commit()
            claim = {
                'endpointcommandid': command.endpointcommandid,
                'action': command.action,
                'target': command.target,
                'targetkind': command.targetkind,
            }
            # update_app rides a resolved delivery descriptor: the agent gets
            # the package URL, its pinned SHA-256, and the version's silent-
            # install switches, and fetches the bytes with its collector
            # credential. If the version was deleted between enqueue and claim,
            # send the command through with no descriptor and let the agent
            # fail closed rather than 500 the poll.
            if command.action == 'update_app':
                descriptor, _resolve_error = _resolve_update_app(command.target)
                if not _pin_holds(command, descriptor):
                    return _fail_on_pin_drift(command, 'update_app')
                if descriptor is not None:
                    claim['app'] = descriptor
            # run_script rides a resolved script descriptor: the agent gets a
            # collector-authenticated body URL, the pinned SHA-256, and the
            # per-script run context. Fetch happens agent-side with the
            # collector credential. If the script vanished between enqueue and
            # claim, send the command through with no descriptor and let the
            # agent fail closed rather than 500 the poll.
            if command.action == 'run_script':
                script_descriptor, _script_error = _resolve_run_script(command.target)
                if not _pin_holds(command, script_descriptor):
                    return _fail_on_pin_drift(command, 'run_script')
                if script_descriptor is not None:
                    claim['script'] = script_descriptor
            return success_response(claim)

        # Nothing queued: release the session so a concurrent enqueue is visible
        # on the next poll, wait a beat, then re-check until the hold expires.
        db.session.rollback()
        if time.monotonic() >= deadline:
            return ('', 204)
        time.sleep(_CLAIM_POLL_INTERVAL)


@endpoints_bp.route('/commands/<int:command_id>/result', methods=['POST'])
def report_command_result(command_id: int):
    """Agent reports a claimed command's outcome (collector-token auth).

    Finalizes the command row and, for MTConnect actions that observed the run
    state, caches it on the EndpointState so the WebUI toggle reflects reality
    between polls. Both writes land in one commit.
    """
    error = _collector_auth_or_error()
    if error is not None:
        return error

    command = db.session.get(EndpointCommand, command_id)
    if command is None:
        return error_response(ErrorCodes.NOT_FOUND,
                              f'No command {command_id}', http_code=404)

    # An agent may only finalize ITS OWN command. The collector key is shared
    # fleet-wide, so without this any agent could post a result for any command
    # id - marking an install succeeded on a machine that never ran it, which
    # writes a false compliance record. Binding to the claimed hostname is a
    # floor, not a fix: the credential still does not prove which endpoint is
    # calling, and only a per-endpoint credential closes that.
    # Optional because the shipped agent does not send it yet; enforced when it
    # does, so an updated fleet gets the check without a flag day.
    hostname = (request.args.get('hostname') or '').strip()
    if hostname:
        caller_asset_id = _resolve_asset_for_hostname(hostname)
        if caller_asset_id != command.assetid:
            return error_response(
                ErrorCodes.FORBIDDEN,
                'This command belongs to another endpoint', http_code=403)

    # Only a claimed command can be finalized. Needs no agent change, and it
    # closes two things the hostname check cannot: reporting on a command that
    # was never claimed, and overwriting a terminal result - which would let a
    # recorded failure be replaced by a success.
    if command.status != EndpointCommand.STATUS_RUNNING:
        return error_response(
            ErrorCodes.VALIDATION_ERROR,
            'Command {} is {}, not running'.format(
                command_id, command.status),
            http_code=409)

    payload = request.get_json(silent=True) or {}
    succeeded = bool(payload.get('succeeded'))
    command.status = (EndpointCommand.STATUS_SUCCEEDED if succeeded
                      else EndpointCommand.STATUS_FAILED)
    command.resultmessage = (payload.get('message') or '')[:2000] or None
    detail = payload.get('detail')
    command.resultjson = json.dumps(detail, separators=(',', ':')) \
        if detail is not None else None
    command.completedat = _naive_utcnow()

    # Cache the observed MTConnect state so the toggle shows a real value. The
    # agent sends mtconnectrunning on any mtconnect_* result that probed it.
    if 'mtconnectrunning' in payload:
        state = EndpointState.query.filter_by(assetid=command.assetid).first()
        if state is not None:
            running = payload.get('mtconnectrunning')
            state.mtconnectrunning = bool(running) if running is not None else None
            state.mtconnectcheckedat = _naive_utcnow()

    db.session.commit()
    return success_response(command.to_dict(), message='Result recorded')


@endpoints_bp.route('/apppackage/<int:appversionid>', methods=['GET'])
def download_app_package(appversionid: int):
    """Agent-facing installer download (collector-token auth).

    The core /api/applications/package route is gated on applications.view, a
    permission the endpoint agent's collector token does not hold. This route
    authorizes on the same endpoints collector token the agent already uses for
    telemetry and command polling, so the whole app-delivery path rides one key.
    Serves the exact bytes the update_app descriptor pinned by SHA-256.
    """
    from shopdb.api import send_upload

    error = _collector_auth_or_error()
    if error is not None:
        return error

    version = db.session.get(AppVersion, appversionid)
    if version is None:
        return error_response(ErrorCodes.NOT_FOUND,
                              f'Unknown appversionid: {appversionid}', http_code=404)
    application = db.session.get(Application, version.appid)
    if application is None or not application.installpath:
        return error_response(ErrorCodes.NOT_FOUND,
                              'No installer package for this application',
                              http_code=404)

    # installpath is the core package URL (/api/applications/package/<filename>);
    # take the stored filename and serve it from the core package directory. A
    # site-typed share path (not our upload) has no servable file here.
    prefix = '/api/applications/package/'
    if not application.installpath.startswith(prefix):
        return error_response(ErrorCodes.NOT_FOUND,
                              'Application installer is not a ShopDB-hosted package',
                              http_code=404)
    filename = application.installpath[len(prefix):]
    package_dir = os.path.join(current_app.instance_path, 'applicationpackages')
    return send_upload(package_dir, filename, as_attachment=True)


@endpoints_bp.route('/apppackage/app/<int:app_id>', methods=['GET'])
@endpoints_bp.route('/apppackage/app/<int:app_id>/<path:filename>', methods=['GET'])
def download_app_package_by_app(app_id: int, filename: str = None):
    """Agent-facing installer download by APP (collector-token auth).

    The app-level counterpart to download_app_package: serves the application's
    own uploaded installer (installpath + installersha256) so update_app can
    resolve an install source by app alone, with no AppVersion row. Same
    collector-token auth and same package directory as the per-version route.

    The optional trailing <filename> segment lets the delivery URL carry the
    real installer extension (so the agent routes MSI through msiexec); it is
    validated against the stored filename and does not drive the lookup.
    """
    from shopdb.api import send_upload

    error = _collector_auth_or_error()
    if error is not None:
        return error

    application = db.session.get(Application, app_id)
    if application is None or not application.installpath:
        return error_response(ErrorCodes.NOT_FOUND,
                              'No installer package for this application',
                              http_code=404)
    prefix = '/api/applications/package/'
    if not application.installpath.startswith(prefix):
        return error_response(ErrorCodes.NOT_FOUND,
                              'Application installer is not a ShopDB-hosted package',
                              http_code=404)
    stored_filename = application.installpath[len(prefix):]
    if filename is not None and filename != stored_filename:
        return error_response(ErrorCodes.NOT_FOUND,
                              'Requested installer filename does not match this application',
                              http_code=404)
    package_dir = os.path.join(current_app.instance_path, 'applicationpackages')
    return send_upload(package_dir, stored_filename, as_attachment=True)


@endpoints_bp.route('/scriptbody/<int:script_id>', methods=['GET'])
def download_script_body(script_id: int):
    """Agent-facing script body download (collector-token auth).

    The core /api/scripts/body route is gated on applications.view, which the
    endpoint agent's collector token does not hold. This route authorizes on
    the same endpoints collector token the agent already uses, keeping the
    whole script-delivery path on one key. Serves the exact bytes the
    run_script descriptor pinned by SHA-256.
    """
    from shopdb.api import send_upload

    error = _collector_auth_or_error()
    if error is not None:
        return error

    script = db.session.get(Script, script_id)
    if script is None or not script.filename:
        return error_response(ErrorCodes.NOT_FOUND,
                              'No body for this script', http_code=404)
    script_dir = os.path.join(current_app.instance_path, 'scripts')
    return send_upload(script_dir, script.filename, as_attachment=True)