diff --git a/freevideo_engine/access_proxy.py b/freevideo_engine/access_proxy.py
index 1040bf6f..1e645478 100644
--- a/freevideo_engine/access_proxy.py
+++ b/freevideo_engine/access_proxy.py
@@ -23,6 +23,7 @@
 from aiohttp import ClientSession, ClientTimeout, DummyCookieJar, TCPConnector, WSMsgType, client_exceptions, web
 from yarl import URL
 from . import proxy_identity, lan_password
+from .origin_scheme import effective_scheme
 from .lan_auth import PasswordAuth, DEVICE_COOKIE, REMEMBER_SECONDS, LOGIN_DELAY
 
 PHONE_WEB = Path(__file__).resolve().parents[1] / 'web'
@@ -81,17 +82,18 @@ def same_origin(request):
     try:
         if len(request.headers.getall('Host', [])) != 1 or len(request.headers.getall('Origin', [])) != 1:
             return False
-        host = urlsplit(request.scheme + '://' + request.host)
         origin = urlsplit(request.headers['Origin'])
+        scheme = effective_scheme(request, origin.scheme)
+        host = urlsplit(scheme + '://' + request.host)
         socket = request.transport.get_extra_info('sockname')
         allowed = {socket[0].split('%')[0]}
         if ipaddress.ip_address(socket[0].split('%')[0]).is_loopback:
             allowed.update(('localhost', '127.0.0.1', '::1'))
-        own_port = host.port or (443 if request.scheme == 'https' else 80)
+        own_port = host.port or (443 if scheme == 'https' else 80)
         return (host.hostname in allowed and own_port == socket[1] and origin.hostname == host.hostname
                 and (origin.port or (443 if origin.scheme == 'https' else 80)) == own_port
-                and origin.scheme == request.scheme and not any((origin.path, origin.query, origin.fragment,
-                                                                origin.username, origin.password)))
+                and origin.scheme == scheme and not any((origin.path, origin.query, origin.fragment,
+                                                         origin.username, origin.password)))
     except (ValueError, TypeError, IndexError, KeyError):
         return False
 
diff --git a/freevideo_engine/agent/access.py b/freevideo_engine/agent/access.py
index 2c5f9e02..a46a3d07 100644
--- a/freevideo_engine/agent/access.py
+++ b/freevideo_engine/agent/access.py
@@ -4,6 +4,7 @@
 
 from .jobs import JobError
 from .. import proxy_identity
+from ..origin_scheme import effective_scheme
 
 
 def peer_address(request):
@@ -63,8 +64,9 @@ def check(request, *, local=False, lan=False):
         if origin is not None:
             parsed = urlsplit(origin)
             port = parsed.port or (443 if parsed.scheme == 'https' else 80)
-            own_port = host.port or (443 if scheme == 'https' else 80)
-            if (parsed.scheme != scheme or parsed.hostname != host.hostname or port != own_port
+            wanted = scheme if proxied else effective_scheme(request, parsed.scheme)
+            own_port = host.port or (443 if wanted == 'https' else 80)
+            if (parsed.scheme != wanted or parsed.hostname != host.hostname or port != own_port
                     or parsed.path or parsed.query or parsed.fragment or parsed.username or parsed.password):
                 raise ValueError
     except (ValueError, TypeError, IndexError):
diff --git a/freevideo_engine/agent/attachments.py b/freevideo_engine/agent/attachments.py
new file mode 100644
index 00000000..b6669fc3
--- /dev/null
+++ b/freevideo_engine/agent/attachments.py
@@ -0,0 +1,71 @@
+"""Chat images in the retained input assets directory, with server-owned filenames."""
+import filecmp
+import json
+import os
+from pathlib import Path
+import re
+import threading
+import uuid
+
+from . import media, schema
+
+FOLDER = 'freevideo/agent/assets'
+_LOCK = threading.RLock()
+
+
+def path(input_root, name):
+    if not isinstance(name, str) or not re.fullmatch(schema.CHAT_ATTACHMENT_NAME['pattern'], name):
+        raise ValueError('Invalid attachment name')
+    root = Path(input_root).resolve()
+    target = root / FOLDER / name
+    # Reject links, including links to another location inside the input directory.
+    for part in (target, *target.parents):
+        if part == root:
+            break
+        if part.is_symlink():
+            raise ValueError('Attachment paths cannot contain links')
+    if target.resolve() != target:
+        raise ValueError('Attachment leaves assets')
+    return target
+
+
+def _metadata(target):
+    result = target.with_name('.' + target.name + '.json')
+    if result.is_symlink():
+        raise ValueError('Invalid attachment metadata')
+    return result
+
+
+def publish(input_root, temporary, name, original_name):
+    """Publish decoded bytes atomically; the first filename for identical content wins."""
+    with _LOCK:
+        target = path(input_root, name)
+        target.parent.mkdir(parents=True, exist_ok=True)
+        metadata = _metadata(target)
+        if not metadata.exists():
+            pending = metadata.with_name('.attachment-' + uuid.uuid4().hex + '.tmp')
+            try:
+                pending.write_text(json.dumps(dict(original_name=original_name), ensure_ascii=False), encoding='utf-8')
+                os.replace(pending, metadata)
+            finally:
+                pending.unlink(missing_ok=True)
+        # Identical bytes are already in place; replacing a file the page is reading fails on Windows.
+        if target.is_file() and filecmp.cmp(temporary, target, shallow=False):
+            Path(temporary).unlink(missing_ok=True)
+        else:
+            os.replace(temporary, target)
+        return json.loads(metadata.read_text(encoding='utf-8'))['original_name']
+
+
+def read(input_root, name):
+    """Reinspect stored bytes; never use dimensions or filenames sent with a message."""
+    with _LOCK:
+        target = path(input_root, name)
+        if not target.is_file() or target.stat().st_size > schema.MAX_CHAT_ATTACHMENT_BYTES:
+            raise ValueError('No saved attachment within the upload limit')
+        info, suffix = media.validate_upload(target)
+        if suffix not in ('.png', '.jpg', '.webp') or info['kind'] != 'image':
+            raise ValueError('Not a chat image')
+        metadata = _metadata(target)
+        original_name = json.loads(metadata.read_text(encoding='utf-8'))['original_name'] if metadata.exists() else name
+        return dict(name=name, original_name=original_name, width=info['width'], height=info['height'])
diff --git a/freevideo_engine/agent/guide.py b/freevideo_engine/agent/guide.py
index 09ed9a1d..2a4268da 100644
--- a/freevideo_engine/agent/guide.py
+++ b/freevideo_engine/agent/guide.py
@@ -69,6 +69,7 @@
         'itself; read the project afterwards and check that one clip covers the whole film.',
     ],
     'references': [
+        "User message attachments are reference images by default. Use each name directly in media.references, media.first or media.last according to the user's words; if the intended role is unclear, ask one question. Text can be empty when a message has attachments.",
         'Tags are numbered per kind in reference order: <Picture 1>, <Picture 2>, <Video 1>, <Audio 1>. '
         'For keyframes the first frame is <Picture 1> and the last frame the next picture.',
         'Name subjects <Subject 1>, <Subject 2> in subject_definitions and reuse those labels in every section.',
@@ -211,6 +212,7 @@
 
 # How an agent is expected to work once a user has asked for something. Only what this version can honour.
 RULES = [
+    "User message attachments are reference images by default. Use each name directly in media.references, media.first or media.last according to the user's words; if the intended role is unclear, ask one question. Text can be empty when a message has attachments.",
     'Finish the job. When the user asks for a video, keep going until the finished video exists. Decide ordinary '
     'things yourself: how to split shots, how many takes, which take to keep. Tell the user what you decided in '
     'one or two sentences at the end, not along the way. Name every video you submit: label is a short name of what '
@@ -259,6 +261,7 @@
 
 
 RULES_ZH = [
+    '用户消息里的 attachments 默认当参考图。按用户的话把每张图的 name 直接用在 media.references、media.first 或 media.last；拿不准用途时问一句。有附件的消息 text 可以为空串。',
     '把任务做完。用户要求生成视频时，持续推进，直到成片生成。如何拆分镜头、拍几条、选哪一条等普通决定，'
     '由你自己做。最后用一两句话告诉用户你做了哪些决定，不要每做一步都请用户决定。'
     '提交的每条视频都起个名字：label 写拍的是什么，用用户的语言，2 到 8 个字，例如“清晨海边”。',
diff --git a/freevideo_engine/agent/routes.py b/freevideo_engine/agent/routes.py
index 3be38e26..450dc543 100644
--- a/freevideo_engine/agent/routes.py
+++ b/freevideo_engine/agent/routes.py
@@ -22,7 +22,7 @@
 import uuid
 
 from . import API_VERSION, enabled
-from . import assets
+from . import assets, attachments
 from . import private_file
 from . import spec as specs
 from .. import tiers as quality_tiers
@@ -433,6 +433,7 @@ def reply(value, status=200):
     connection = dict(connected=False, client=None, via=None, last_seen_at=None)
     # The conversation in the Autopilot view: the page and the connected agent talk in the agent's latest session.
     chat = dict(client=None, waiting=0, folds={}, told={})
+    chat_writes = asyncio.Lock()
     server._freevideo_agent_connection = connection
 
     def problem(code, message=None, status=None, **details):
@@ -473,7 +474,12 @@ async def respond(request):
                         valid = definition.get('lan') and proxy_identity.check_csrf(request)
                     else:
                         valid = valid_page_session(request)
-                    if not valid:
+                elif definition.get('page_cookie'):
+                    session = page_sessions.get(request.cookies.get('freevideo_agent_page', ''))
+                    valid = session is not None and session[1] >= time.monotonic()
+                else:
+                    valid = True
+                if not valid:
                         raise JobError('page_session_required', 'Page session required')
                 header = request.headers.get('Authorization', '')
                 given = header[7:].strip() if header[:7].lower() == 'bearer ' else request.headers.get('X-FreeVideo-Agent', '')
@@ -522,7 +528,7 @@ async def respond(request):
                 return await handler(request)
             except specs.SpecError as error:
                 value = errors.validation(error.errors, getattr(error, 'code', 'invalid_request'))
-                return reply({'error': value}, errors.ERRORS[value['code']][0])
+                return reply({'error': value}, getattr(error, 'status', errors.ERRORS[value['code']][0]))
             except JobError as error:
                 value = error.view()
                 if value.get('next') and request.match_info.get('id'):
@@ -888,7 +894,8 @@ def chat_fold(store, sid):
                 fold['seq'] = row['seq']
                 base = dict(seq=row['seq'], at=row['at'], type=row['type'])
                 if row['type'] == 'message':
-                    fold['items'].append(dict(base, role='user' if row['by'] == 'user' else 'agent', text=row['text']))
+                    fold['items'].append(dict(base, role='user' if row['by'] == 'user' else 'agent', text=row['text'],
+                                              **({'attachments': row['attachments']} if row.get('attachments') else {})))
                     if row['by'] == 'user':
                         fold['said'] = row['text']
                         if row.get('mid'):
@@ -1017,6 +1024,75 @@ def read():
                         active=sum(1 for r in rows if r['status'] in ACTIVE))
         return reply(await asyncio.to_thread(read))
 
+    def chat_not_connected():
+        return reply({'error': dict(code='not_connected', message='No agent is connected.', retryable=True,
+                                    error_zh='尚未连接 Agent', error_en='No agent is connected.',
+                                    details=[], next='Connect an agent in the Autopilot view, then say it again.',
+                                    request_id=errors.request_id())}, 409)
+
+    @routes.post('/freevideo/agent/chat/attach')
+    @guarded
+    async def chat_attach(request):
+        from . import media
+        if work_sessions() is None or chat['client'] is None:
+            return chat_not_connected()
+        limit = contract.MAX_CHAT_ATTACHMENT_BYTES
+        if request.content_length is not None and request.content_length > limit:
+            raise JobError('too_large', 'Upload exceeds limit', limit_bytes=limit)
+        with tempfile.TemporaryDirectory(prefix='.chat-attach-', dir=port.input_root) as temporary:
+            # The suffix selects the image decoder; Pillow recognizes the actual content format.
+            target = Path(temporary) / 'image.png'
+            size, digest = 0, hashlib.sha256()
+            with target.open('xb') as output:
+                while True:
+                    chunk = await request.content.read(min(256 * 1024, limit - size + 1))
+                    if not chunk:
+                        break
+                    size += len(chunk)
+                    if size > limit:
+                        raise JobError('too_large', 'Upload exceeds limit', limit_bytes=limit)
+                    output.write(chunk)
+                    digest.update(chunk)
+            async with upload_decoders:
+                try:
+                    info, suffix = await asyncio.to_thread(media.validate_upload, target)
+                    if suffix not in ('.png', '.jpg', '.webp'):
+                        raise ValueError('Only PNG, JPEG and WebP chat images are supported')
+                except (OSError, ValueError, SyntaxError):
+                    raise JobError('invalid_media', 'Image cannot be decoded within the upload limits') from None
+            name = digest.hexdigest() + suffix
+            original = await asyncio.to_thread(attachments.publish, port.input_root, target, name, request[query_key]['name'])
+            return reply(dict(name=name, original_name=original, width=info['width'], height=info['height'], bytes=size), 201)
+
+    @routes.get('/freevideo/agent/chat/attachment')
+    @guarded
+    async def chat_attachment(request):
+        name = request[query_key]['name']
+        try:
+            path = attachments.path(port.input_root, name)
+            async with upload_decoders:
+                await asyncio.to_thread(attachments.read, port.input_root, name)
+        except (OSError, ValueError, SyntaxError):
+            raise JobError('not_found', 'No saved chat image') from None
+        kind = {'.png': 'image/png', '.jpg': 'image/jpeg', '.jpeg': 'image/jpeg', '.webp': 'image/webp'}[path.suffix]
+        return web.FileResponse(path, headers={'Content-Type': kind, 'Cache-Control': 'private, no-store',
+                                               'X-Content-Type-Options': 'nosniff'})
+
+    def chat_attachments(values):
+        result, seen = [], set()
+        for i, item in enumerate(values):
+            field = '/attachments/%d/name' % i
+            if item['name'] in seen:
+                raise specs.SpecError(contract.diagnostic(field, 'a distinct attachment name', item['name']))
+            seen.add(item['name'])
+            try:
+                result.append(attachments.read(port.input_root, item['name']))
+            except (OSError, ValueError, SyntaxError):
+                error = specs.SpecError(contract.diagnostic(field, 'an existing uploaded image', item['name']))
+                error.status = 400
+                raise error from None
+        return result
+
     @routes.post('/freevideo/agent/chat')
     @guarded
     async def chat_say(request):
@@ -1026,17 +1102,22 @@ def write():
             if store is None or client is None:
                 return None
             before = store.current_for(client)
-            sid = chat_session(store, value.get('session')) or store.create(client=client, by='user')['id']
-            fold = chat_fold(store, sid)
-            if value['id'] in fold['sent']:
+            sid = chat_session(store, value.get('session'))
+            fold = chat_fold(store, sid) if sid else None
+            if fold is not None and value['id'] in fold['sent']:
                 seq = fold['sent'][value['id']]
                 first = next(i for i in fold['items'] if i['seq'] == seq)
                 return 200, dict(seq=seq, id=value['id'], at=first['at'], session=sid,
                                  read=seq <= store.read_until(sid, client), agent=agent_state())
+            pictures = chat_attachments(value.get('attachments', []))
+            sid = sid or store.create(client=client, by='user')['id']
+            fold = fold or chat_fold(store, sid)
             if store.stopped(sid):
                 # Speaking again continues a conversation the user stopped earlier.
                 store.append(sid, dict(type='control', action='resumed'), by='user')
             message = dict(type='message', text=value['text'], mid=value['id'])
+            if pictures:
+                message['attachments'] = pictures
             if value.get('whitebox'):
                 message['whitebox'] = True
             row = store.append(sid, message, by='user')
@@ -1046,13 +1127,11 @@ def write():
                 fold['ids'].pop(fold['sent'].pop(next(iter(fold['sent']))), None)
             chat_moved(store, before)
             return 201, dict(seq=row['seq'], id=value['id'], at=row['at'], session=sid, read=False, agent=agent_state())
-        done = await asyncio.to_thread(write)
+        async with chat_writes, upload_decoders:
+            done = await asyncio.to_thread(write)
         if done is None:
             # Nothing is kept: no agent has connected since FreeVideo started, so nobody would read it.
-            return reply({'error': dict(code='not_connected', message='No agent is connected.', retryable=True,
-                                        error_zh='尚未连接 Agent', error_en='No agent is connected.',
-                                        details=[], next='Connect an agent in the Autopilot view, then say it again.',
-                                        request_id=errors.request_id())}, 409)
+            return chat_not_connected()
         return reply(done[1], done[0])
 
     @routes.post('/freevideo/agent/chat/abandon')
@@ -1520,6 +1599,8 @@ async def session_post(request):
             value['text'] = chat_text(value['text'])
         client = session_client(request, by)
         def write():
+            if value.get('attachments'):
+                value['attachments'] = chat_attachments(value['attachments'])
             row = sessions.append(sid, value, by=by, client=client)
             here = sessions.current_for(client) if client and value['type'] == 'message' else None
             if here not in (None, sid):
diff --git a/freevideo_engine/agent/schema.py b/freevideo_engine/agent/schema.py
index 2ee30dc4..ad76ce1b 100644
--- a/freevideo_engine/agent/schema.py
+++ b/freevideo_engine/agent/schema.py
@@ -24,6 +24,8 @@
 MAX_UPLOAD_BYTES = 2 * 1024 ** 3
 MAX_JSON_UPLOAD_BYTES = 96 * 1024 ** 2
 MAX_BODY_BYTES = 1024 * 1024
+MAX_CHAT_ATTACHMENT_BYTES = 10 * 1024 ** 2
+MAX_CHAT_ATTACHMENTS = 4
 
 
 def obj(properties=None, required=(), **kw):
@@ -70,6 +72,22 @@ def enum(values, **kw):
 # Waiting for the user's words is longer: a person types slowly, and each wait is one tool call.
 INBOX_WAIT = number(0, default=0, description='x-clamp caps accepted wait values at 50 seconds without rejecting larger values.', **{'x-clamp': 50})
 CHAT_ID = string(64, pattern=r'^[A-Za-z0-9_-]{1,64}$')
+CHAT_ATTACHMENT_NAME = string(69, pattern=r'^[0-9a-f]{64}\.(png|jpg|jpeg|webp)$')
+CHAT_ATTACHMENT = obj(dict(name=CHAT_ATTACHMENT_NAME, original_name=string(255),
+                           width=number(1, integer=True), height=number(1, integer=True)),
+                      ('name', 'original_name', 'width', 'height'))
+
+
+def chat_request():
+    # Upload metadata supplied by the page is ignored; only the stored name is authoritative.
+    attachment = obj(dict(name=CHAT_ATTACHMENT_NAME, original_name={}, width={}, height={}, bytes={}), ('name',))
+    return obj(dict(text=dict(type='string', minLength=0, maxLength=8000), id=CHAT_ID,
+                    session=SESSION, whitebox={'type': 'boolean'},
+                    attachments=array(attachment, 0, MAX_CHAT_ATTACHMENTS)), ('text', 'id'),
+               **{'if': {'required': ['attachments'], 'properties': {'attachments': {'minItems': 1}}},
+                  'else': {'properties': {'text': string(8000)}}})
+
+
 EMPTY = obj()
 
 
@@ -447,7 +465,7 @@ def route(summary, body=None, query=None, **kw):
             ('/connect/reset', 'POST', EMPTY), ('/chat', 'GET', None), ('/chat/now', 'GET', None),
             ('/chat/video', 'GET', None), ('/chat/list', 'GET', None), ('/chat/new', 'POST', EMPTY),
             ('/chat/delete', 'POST', obj({'session': SESSION, 'stop': {'type': 'boolean'}}, ('session',))),
-            ('/chat', 'POST', obj({'text': string(8000), 'id': CHAT_ID, 'session': SESSION, 'whitebox': {'type': 'boolean'}}, ('text', 'id'))),
+            ('/chat', 'POST', chat_request()),
             ('/chat/abandon', 'POST', obj({'session': SESSION})),
             ('/chat/confirm', 'POST', obj({'id': string(12, pattern=r'^cf[0-9a-f]{10}$'), 'approve': {'type': 'boolean'},
                                            'session': SESSION}, ('id', 'approve')))):
@@ -461,6 +479,13 @@ def route(summary, body=None, query=None, **kw):
             query=obj({'client_id': string()}) if suffix != '/freevideo/progress' else EMPTY,
             **({'path': obj({'id': string()}, ('id',))} if '{id}' in suffix else {}))
     routes['GET', '/freevideo/agent/activity']['query'] = obj({'cursor': string(80)})
+    routes['POST', '/freevideo/agent/chat/attach'] = route(
+        'Upload one chat reference image as raw bytes (at most 10 MiB); requires a page cookie and CSRF.',
+        query=obj({'name': string(255, pattern=r'^[^/\\:\x00-\x1f]+$')}, ('name',)),
+        auth=False, local=True, raw_body=True, status=201)
+    routes['GET', '/freevideo/agent/chat/attachment'] = route(
+        'Read a saved chat image using the page cookie.', query=obj({'name': CHAT_ATTACHMENT_NAME}, ('name',)),
+        auth=False, local=True, page_cookie=True)
     routes['GET', '/freevideo/agent/chat']['query'] = obj({'after': number(0, integer=True), 'session': SESSION})
     routes['GET', '/freevideo/agent/chat/video']['query'] = obj({'id': string(43, pattern=r'^[0-9]{4}-[0-9]{2}-[0-9]{2}/[0-9a-f]{32}$')}, ('id',))
     for method, suffix, summary, extra in (
@@ -806,7 +831,7 @@ def image_schema(name):
     'create': ('POST', '/sessions', 'Create a work conversation for messages and plans shared with the user.'),
     'list': ('GET', '/sessions', 'List work conversations, most recently updated first.'),
     'get': ('GET', '/sessions/{id}', 'Read conversation metadata and whether the user stopped it.'),
-    'events': ('GET', '/sessions/{id}/events', 'Read events after a sequence number; wait up to 25 seconds or request an SSE stream.'),
+    'events': ('GET', '/sessions/{id}/events', 'Read events, including user image attachments (name, original_name, width, height); text may be empty with attachments. Wait up to 25 seconds or request an SSE stream.'),
     'post': ('POST', '/sessions/{id}/events', 'Append a message or plan for the user. Steps are recorded by the server.'),
     'stop': ('POST', '/sessions/{id}/stop', 'Only the page user may stop a work conversation.'),
     'resume': ('POST', '/sessions/{id}/resume', 'Only the page user may continue a work conversation.'),
@@ -848,6 +873,7 @@ def session_event(*, stored=False, agent_only=False):
             1, 20000), total_seconds=number(0)), ('type', 'shots'))]
     if not agent_only:
         variants[0]['properties']['mid'] = CHAT_ID
+        variants[0]['properties']['attachments'] = array(CHAT_ATTACHMENT, 1, MAX_CHAT_ATTACHMENTS)
         variants[0]['properties']['whitebox'] = {'type': 'boolean', 'description': 'The user chose a white-model preview for this request.'}
         from .errors import ERRORS
         variants += [obj(dict(type={'const': 'step'}, id=string(12, pattern=r'^st[0-9a-f]{10}$'),
@@ -1354,6 +1380,11 @@ def openapi():
                        **{str(status): {'description': 'Agent error', 'content': {'application/json': {'schema': {
                            'type': 'object', 'properties': {'error': {'$ref': '#/components/schemas/Error'}}}}}}
                           for status in sorted({r[0] for r in ERRORS.values() if r[0] >= 400})}})
+        if contract.get('auth') is False and (method != 'GET' or contract.get('page_cookie')):
+            operation['security'] = [{'page': []}]
+        if contract.get('raw_body'):
+            operation['requestBody'] = dict(required=True, content={'application/octet-stream': {'schema': {
+                'type': 'string', 'format': 'binary', 'maxLength': MAX_CHAT_ATTACHMENT_BYTES}}})
         if contract.get('page_or_token'):
             operation['security'] = [{'page': []}] if suffix.endswith(('/stop', '/resume')) else [{'bearer': []}, {'page': []}]
         if contract['body'] is not None:
@@ -1370,6 +1401,9 @@ def openapi():
                                     for k in re.findall(r'\{(.*?)\}', path)]
         if contract.get('idempotent'):
             operation['parameters'].append(dict(name='Idempotency-Key', **{'in': 'header'}, required=False, schema=REF))
+        if suffix == '/freevideo/agent/chat/attach':
+            operation['parameters'].append(dict(name='X-FreeVideo-CSRF', **{'in': 'header'}, required=True,
+                schema=string(128), description='The page session CSRF token.'))
         if contract.get('page_or_token'):
             operation['parameters'].append(dict(name='X-FreeVideo-CSRF', **{'in': 'header'}, required=False,
                 schema=string(128), description='Required with the freevideo_agent_page cookie, including reads.'))
@@ -1405,6 +1439,11 @@ def openapi():
         elif suffix == '/imports':
             response_schema = obj({'import': import_view()}, ('import',)) if method == 'POST' else obj(
                 {'items': array(import_view())}, ('items',))
+        elif suffix == '/freevideo/agent/chat/attach':
+            response_schema = obj(dict(CHAT_ATTACHMENT['properties'], bytes=number(1, MAX_CHAT_ATTACHMENT_BYTES, integer=True)),
+                                  (*CHAT_ATTACHMENT['required'], 'bytes'))
+        elif suffix == '/freevideo/agent/chat/attachment':
+            content_type, response_schema = 'image/png', {'type': 'string', 'format': 'binary'}
         elif suffix == '/editor/sources/content':
             content_type, response_schema = 'application/octet-stream', {'type': 'string', 'format': 'binary'}
         elif suffix.endswith('/file') or suffix.endswith('/thumbnail'):
@@ -1412,6 +1451,8 @@ def openapi():
             response_schema = {'type': 'string', 'format': 'binary'}
         success = operation['responses'][str(contract.get('status', 200))]
         success['content'] = {content_type: {'schema': response_schema}}
+        if suffix == '/freevideo/agent/chat/attachment':
+            success['content'].update({kind: {'schema': response_schema} for kind in ('image/jpeg', 'image/webp')})
         if suffix == '/sessions/{id}/events' and method == 'GET':
             success['content']['text/event-stream'] = {'schema': {'type': 'string'}}
         if contract.get('idempotent'):
diff --git a/freevideo_engine/agent/spec.py b/freevideo_engine/agent/spec.py
index 72d221fd..76c7bf82 100644
--- a/freevideo_engine/agent/spec.py
+++ b/freevideo_engine/agent/spec.py
@@ -93,6 +93,8 @@ def media_item(value, path, *, image=False, frame_default=None, early=False):
     errors = contract.issues(contract.media(image, local=True, early=early), value, path)
     if errors:
         raise SpecError(errors)
+    if 'file' in value and re.fullmatch(contract.CHAT_ATTACHMENT_NAME['pattern'], value['file']):
+        value['file'] = 'freevideo/agent/assets/' + value['file']
     return value
 
 
diff --git a/freevideo_engine/bootstrap.py b/freevideo_engine/bootstrap.py
index b2637fef..b3975267 100644
--- a/freevideo_engine/bootstrap.py
+++ b/freevideo_engine/bootstrap.py
@@ -5,6 +5,7 @@
 from contextlib import ExitStack, contextmanager
 import hashlib
 import json
+import shlex
 import logging
 import os
 from pathlib import Path
@@ -798,12 +799,16 @@ def plan(args, *, local_progress=None, stage=None):
     if reuse_cache and not cache_compatible(reuse_cache, **cache_format):
         errors.append('--cache must contain a prepared FP8 cache compatible with this GPU scale format.')
     from .download_settings import read as download_preferences
+    def native_git_available():
+        from .macos_bootstrap import available_git
+        return available_git() is not None
     if stage:
         stage('Testing download sources')
     networking = network.plan(json.loads((PACKAGE / 'dependencies.json').read_text(encoding='utf-8')),
         bootstrap_versions(json.loads((PACKAGE / 'bootstrap_versions.json').read_text(encoding='utf-8')), system), layout,
         mode=getattr(args, 'network', 'auto'), timeout=getattr(args, 'network_timeout', 5),
         offline=bool(args.hardware_json or errors),
+        git_available=not mac_target or bool(args.hardware_json) or native_git_available(),
         env=dict(os.environ, FREEVIDEO_HOME=str(root)),
         proxy_mode=download_preferences(root / 'download-settings.json')['proxy_mode'],
         progress=(lambda message: stage('Testing download sources', message['done'], message['total'])) if stage else None)
@@ -1916,19 +1921,22 @@ def install_git(self):
 
     def install_git_linux(self):
         """Git from conda-forge, pinned by package hashes, when the system has none new enough."""
-        spec = json.loads((PACKAGE / 'git_linux.json').read_text(encoding='utf-8'))
+        self.install_git_conda('git_linux.json', 'git-linux', self.fetch_git_linux)
+
+    def install_git_conda(self, filename, folder, fetch, wrapper=False):
+        spec = json.loads((PACKAGE / filename).read_text(encoding='utf-8'))
         prefix = self.root / 'tools' / ('git-' + spec['version'])
         executable = prefix / 'bin' / 'git'
         receipt, content = prefix / '.freevideo-git.json', {k: spec[k] for k in ('version', 'packages')}
         if not (receipt.is_file() and json.loads(receipt.read_text(encoding='utf-8')) == content and executable.is_file()):
-            micromamba, packages = self.fetch_git_linux(spec)
+            micromamba, packages = fetch(spec)
             # conda writes the prefix into the installed files, so the
             # environment is created in place; an interrupted one is replaced.
             if prefix.is_symlink():
                 prefix.unlink()
             elif prefix.exists():
                 shutil.rmtree(prefix)
-            explicit = self.root / 'downloads' / 'git-linux' / 'explicit.txt'
+            explicit = self.root / 'downloads' / folder / 'explicit.txt'
             explicit.write_text('@EXPLICIT\n' + ''.join(path.resolve().as_uri() + '\n' for path in packages), encoding='utf-8')
             cache = self.root / 'tools' / 'micromamba-root'
             env = {k: v for k, v in self.env.items() if not k.upper().startswith(('CONDA', 'MAMBA'))}
@@ -1943,7 +1951,15 @@ def install_git_linux(self):
         # the other tools of the environment.
         links = self.root / 'tools' / 'git-bin'
         links.mkdir(parents=True, exist_ok=True)
-        if not (links / 'git').is_symlink() or Path(os.readlink(links / 'git')) != executable:
+        if wrapper:
+            # macOS resolves @loader_path from the path it was started by, so a script
+            # starts the real executable instead of a link to it.
+            text = '#!/bin/sh\nexec %s "$@"\n' % shlex.quote(str(executable))
+            if not (links / 'git').is_file() or (links / 'git').is_symlink() or (links / 'git').read_text(encoding='utf-8') != text:
+                (links / 'git').unlink(missing_ok=True)
+                (links / 'git').write_text(text, encoding='utf-8')
+                (links / 'git').chmod(0o755)
+        elif not (links / 'git').is_symlink() or Path(os.readlink(links / 'git')) != executable:
             (links / 'git').unlink(missing_ok=True)
             (links / 'git').symlink_to(executable)
         self.env['PATH'] = str(links) + os.pathsep + self.env['PATH']
diff --git a/freevideo_engine/comfy_editor.py b/freevideo_engine/comfy_editor.py
index 281fc84e..3c6b1995 100644
--- a/freevideo_engine/comfy_editor.py
+++ b/freevideo_engine/comfy_editor.py
@@ -23,6 +23,7 @@
 from pathlib import Path, PurePosixPath, PureWindowsPath
 from urllib.parse import unquote, urlencode, urlsplit
 
+from .origin_scheme import effective_scheme
 from .comfy_library import _ID, _plain, _video
 from .system import windows
 from .editor_frames import frame_not_after, last_frame_end
@@ -1198,10 +1199,11 @@ def _local_page(request):
     try:
         parsed = urlsplit(origin)
         host = urlsplit('//' + request.host)
-        return (parsed.scheme in ('http', 'https') and parsed.scheme == request.scheme and not parsed.username
+        scheme = effective_scheme(request, parsed.scheme)
+        return (parsed.scheme in ('http', 'https') and parsed.scheme == scheme and not parsed.username
                 and not parsed.password and not parsed.path and not parsed.query and not parsed.fragment
                 and parsed.hostname == host.hostname
-                and (parsed.port or (443 if parsed.scheme == 'https' else 80)) == (host.port or (443 if request.scheme == 'https' else 80)))
+                and (parsed.port or (443 if parsed.scheme == 'https' else 80)) == (host.port or (443 if scheme == 'https' else 80)))
     except ValueError:
         return False
 
@@ -1215,7 +1217,7 @@ def _same_origin_audio(request):
     if referer:
         try:
             page = urlsplit(referer)
-            origin = urlsplit(request.scheme + '://' + request.host)
+            origin = urlsplit(effective_scheme(request, page.scheme) + '://' + request.host)
             return (page.scheme == origin.scheme and page.hostname == origin.hostname
                     and (page.port or (443 if page.scheme == 'https' else 80)) ==
                         (origin.port or (443 if origin.scheme == 'https' else 80))
diff --git a/freevideo_engine/comfy_launcher_runtime.py b/freevideo_engine/comfy_launcher_runtime.py
index 5a9dbed4..fd1b6740 100644
--- a/freevideo_engine/comfy_launcher_runtime.py
+++ b/freevideo_engine/comfy_launcher_runtime.py
@@ -169,7 +169,7 @@ def nearby_comfy(selected):
     """
     try:
         for parent in (selected, *list(selected.parents)[:3]):
-            if _is_comfy_root(parent):
+            if _is_comfy_root(parent) or _is_comfy_root(parent / 'ComfyUI'):
                 return parent.name
         for child in sorted(p for p in selected.iterdir() if p.is_dir())[:200]:
             if _is_comfy_root(child) or _is_comfy_root(child / 'ComfyUI'):
diff --git a/freevideo_engine/comfy_source.py b/freevideo_engine/comfy_source.py
index f57fb014..74f95587 100644
--- a/freevideo_engine/comfy_source.py
+++ b/freevideo_engine/comfy_source.py
@@ -52,8 +52,7 @@ def download(target, *, plan, env, run):
     if target.exists():
         return network.clone(target, spec['url'], spec['commit'], network=plan, env=env, run=run)
     stage = target.with_name(target.name + '.freevideo-download')
-    # A finished checkout survives cancellation before publication. Git fetch's
-    # own incomplete attempts are retained beside this directory by network.clone.
+    # A finished checkout survives cancellation before publication.
     network.clone(stage, spec['url'], spec['commit'], network=plan, env=env, run=run)
     if not all((stage / name).is_file() for name in ('main.py', 'folder_paths.py', 'requirements.txt')) or not (stage / 'comfy_api/latest').is_dir():
         raise ValueError('The downloaded ComfyUI application is incomplete; files retained at ' + str(stage))
diff --git a/freevideo_engine/diagnostic_resources.py b/freevideo_engine/diagnostic_resources.py
index a4614597..c53a7657 100644
--- a/freevideo_engine/diagnostic_resources.py
+++ b/freevideo_engine/diagnostic_resources.py
@@ -751,6 +751,10 @@ def planning(rows):
             item[key] = numbers(row.get(key), HARDWARE_MEMORY)
         item['system_memory'] = numbers(row.get('system_memory'),
             ('total_bytes', 'available_bytes', 'physical_available_bytes', 'commit_available_bytes'))
+        from .exit_diagnostics import OWN_MEMORY_KEYS
+        own = numbers(row.get('own_memory'), OWN_MEMORY_KEYS)
+        if own:
+            item['own_memory'] = own
         for key in ('budget_bytes', 'reserve_bytes'):
             item[key] = numbers(row.get(key), ('gpu', 'ram'))
         item['resource_error'] = numbers(row.get('resource_error'), tuple(resource + '_' + field
diff --git a/freevideo_engine/exit_diagnostics.py b/freevideo_engine/exit_diagnostics.py
index a7be152e..644582b6 100644
--- a/freevideo_engine/exit_diagnostics.py
+++ b/freevideo_engine/exit_diagnostics.py
@@ -19,6 +19,8 @@
 }
 _CRASH_PREFIXES = (0xC0000000, 0x80000000)
 _FIELDS = ('physical_available_bytes', 'commit_available_bytes', 'commit_limit_bytes', 'total_bytes')
+OWN_MEMORY_KEYS = ('own_private_working_set_bytes', 'own_processes', 'own_unreadable_processes',
+                   'host_private_working_set_bytes', 'resident_private_working_set_bytes')
 
 
 def _count(value):
@@ -58,6 +60,56 @@ def memory_sample(pid=None):
     return {key: sample[key] for key in sample if sample[key] is not None}
 
 
+def own_memory(pid=None):
+    """Physical RAM held by this process tree and by the largest process above it (Windows).
+
+    A refusal says how much memory the system had left; this says how much of
+    what was taken is FreeVideo's. The numbers are private working sets, which
+    exclude mapped files, so they can be compared with free physical RAM. The
+    tree is the generation process and its children; the resident worker is
+    started by the host and is added by the caller from its own status. The
+    host figure is the largest of up to four ancestors (a launcher gate is
+    small, the program that started it is not). Processes that cannot be read
+    are counted, never taken as zero; if the process itself cannot be read
+    nothing is returned. Other platforms give nothing; the function never raises.
+    """
+    if not windows():
+        return {}
+    pid = os.getpid() if pid is None else pid
+    result = {}
+    try:
+        from .win32 import ProcessMemory, processes, process_memory
+        reading = ProcessMemory().sample(pid)
+        unreadable = reading.get('unreadable_memory_pids') or []
+        if not reading.get('processes') or pid in unreadable:
+            return {}
+        result.update(own_private_working_set_bytes=reading['private_working_set_bytes'],
+                      own_processes=reading['processes'])
+        if unreadable:
+            result['own_unreadable_processes'] = len(unreadable)
+    except Exception:
+        return {}
+    try:
+        parents = processes()
+        seen, node, largest = {pid}, parents.get(pid), None
+        for _ in range(4):
+            if not node or node in seen:
+                break
+            seen.add(node)
+            try:
+                row = process_memory(node)
+            except Exception:
+                row = None
+            if row and _count(row.get('private_working_set_bytes')) is not None:
+                largest = max(largest or 0, row['private_working_set_bytes'])
+            node = parents.get(node)
+        if largest is not None:
+            result['host_private_working_set_bytes'] = largest
+    except Exception:
+        pass
+    return {key: value for key, value in result.items() if _count(value) is not None}
+
+
 def exit_reading(code, *, requested_stop=False, system=None):
     """The exit code as given, a fixed-word class and whether the system probably ended it.
 
diff --git a/freevideo_engine/failure_details.py b/freevideo_engine/failure_details.py
index b7826938..6797aeee 100644
--- a/freevideo_engine/failure_details.py
+++ b/freevideo_engine/failure_details.py
@@ -221,6 +221,10 @@ def environment_python_removed(args, pythons):
 
 # Checked before free space: freeing space cannot fix a disk that cannot hold the installation.
 PLACE_RULES = (
+    (r'Git is unavailable\. Check the network connection', 'git-unavailable',
+     ('Git could not be set up', '无法准备 Git'),
+     ('Check the network connection and click Check again, or install the Apple Command Line Tools.',
+      '请检查网络连接后点击“重新检查”，或者安装 Apple 命令行工具。')),
     (r'The installation folder is on a FAT32 disk', 'disk-format',
      ('This disk cannot hold the models', '这块硬盘放不下模型'),
      ('It is formatted as FAT32, which stores files up to 4 GB; the prompt model alone is 15.7 GB. '
diff --git a/freevideo_engine/feedback.py b/freevideo_engine/feedback.py
index d34d065d..6acb6b68 100644
--- a/freevideo_engine/feedback.py
+++ b/freevideo_engine/feedback.py
@@ -6,6 +6,8 @@
 import json
 import re
 
+from .exit_diagnostics import OWN_MEMORY_KEYS
+
 # The summary travels in a GitHub page address, which has to stay short.
 MAX_DIAGNOSTICS = 3000
 # Anything that looks like a file location or a long secret-like token is replaced.
@@ -45,6 +47,9 @@ def summary(report, machine=None):
         memory = planning[-1].get('system_memory') if isinstance(planning, list) and planning and isinstance(planning[-1], dict) else None
         if isinstance(memory, dict):
             memory = {k: memory[k] for k in MEMORY_KEYS if type(memory.get(k)) is int}
+            own = planning[-1].get('own_memory')
+            if isinstance(own, dict):
+                memory.update({k: own[k] for k in OWN_MEMORY_KEYS if type(own.get(k)) is int})
             if memory:
                 out['memory'] = memory
         errors = analysis.get('errors')
diff --git a/freevideo_engine/generate.py b/freevideo_engine/generate.py
index ff599ceb..3355497f 100644
--- a/freevideo_engine/generate.py
+++ b/freevideo_engine/generate.py
@@ -287,6 +287,15 @@ def automatic_profile(args, canvas, *, stage, evidence, descriptor=None, environ
                        mapped_bytes=mapped if type(mapped) is int else None,
                        commit_available_bytes=commit if type(commit) is int else None)
         except ResourceBudgetError as error:
+            from .exit_diagnostics import own_memory
+            own = dict(own_memory())
+            resident = ((state or {}).get('memory') or {})
+            if (resident.get('guard_metric') == 'tree private working set'
+                    and type(resident.get('guard_bytes')) is int and resident['guard_bytes'] >= 0):
+                own['resident_private_working_set_bytes'] = resident['guard_bytes']
+            if own:
+                error.add_own_memory(own)
+                row['own_memory'] = own
             row.update(status='rejected', resource_error=error.details)
             error.resource_planning = evidence
             impossible_capacity = any(error.details[name+'_capacity_bytes'] - error.details[name+'_reserve_bytes']
diff --git a/freevideo_engine/git_macos.json b/freevideo_engine/git_macos.json
new file mode 100644
index 00000000..656aa09d
--- /dev/null
+++ b/freevideo_engine/git_macos.json
@@ -0,0 +1,126 @@
+{
+  "version": "2.56.0",
+  "micromamba": {
+    "version": "2.9.0-0",
+    "url": "https://github.com/mamba-org/micromamba-releases/releases/download/2.9.0-0/micromamba-osx-arm64",
+    "sha256": "ec2a072f028e1a7cf20f3e2e74d5a8127cf5a5f27636375b5359811565f4e5be",
+    "bytes": 14596096
+  },
+  "channels": {
+    "official": "https://conda.anaconda.org/conda-forge",
+    "tuna": "https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/conda-forge",
+    "nju": "https://mirror.nju.edu.cn/anaconda/cloud/conda-forge"
+  },
+  "packages": [
+    {
+      "path": "osx-arm64/libexpat-2.8.5-h6111c0a_0.conda",
+      "sha256": "50aa761b83fc7a116ca88fe0cb4cefce5abababb10a926bf8afbafd5c298d102",
+      "bytes": 69565
+    },
+    {
+      "path": "osx-arm64/perl-5.44.0-1_hb001647_perl5.conda",
+      "sha256": "d68a8d9a8f329e46920d7750b495f613e093fa40736d00a227b727f5898804bb",
+      "bytes": 10327598
+    },
+    {
+      "path": "osx-arm64/libiconv-1.18-he4c29f2_3.conda",
+      "sha256": "689a14968267f2f97c07112fca5636e7d756036b9aee969911267c20bd3eea1f",
+      "bytes": 750816
+    },
+    {
+      "path": "osx-arm64/libzlib-1.3.2-h8088a28_3.conda",
+      "sha256": "a18fa5d5bac452401459f966cf0d872224e8080c4ff93c77e168d43ab42ef9d7",
+      "bytes": 47822
+    },
+    {
+      "path": "osx-arm64/bzip2-1.0.8-h4e30115_10.conda",
+      "sha256": "8ec22f0ba25cbfc2e64d70cf29459eccd7ffdf6436f6a6ff15bbfef799f7d4f6",
+      "bytes": 124965
+    },
+    {
+      "path": "osx-arm64/libcxx-23.1.4-h55c6f16_2.conda",
+      "sha256": "9c390b4f403db32346754ccef6b84b7e5d9b70cb6d76258fd23b05f2db2ab216",
+      "bytes": 577255
+    },
+    {
+      "path": "osx-arm64/libev-4.33-h1a92334_3.conda",
+      "sha256": "c0100064506ae8abb432c5a506d474f10af2cf48c33d62bc221fb28b6d6ff6ac",
+      "bytes": 39991
+    },
+    {
+      "path": "osx-arm64/c-ares-1.34.9-h74c22ad_0.conda",
+      "sha256": "74b52d784b3576643da3aea1240a4120c1f25ee057f14ae5a89e8a7b3fb2e69f",
+      "bytes": 201979
+    },
+    {
+      "path": "osx-arm64/ncurses-6.6-he64c551_1.conda",
+      "sha256": "7024a48c8c0d0114ed4ab53c76bf9275d50e91ba7cea367a9aead638d3c29c68",
+      "bytes": 804298
+    },
+    {
+      "path": "osx-arm64/icu-78.3-py310h579977c_2.conda",
+      "sha256": "6cdb5dee54c72e56ab189fb3ad33cb28533553d42590e7e831160248f4416a43",
+      "bytes": 14070242
+    },
+    {
+      "path": "osx-arm64/libintl-0.25.1-hae62183_2.conda",
+      "sha256": "b43c6645e668f86d04d5c06d6dc17efe921bbd9244be4b77946d56fb950a3fcc",
+      "bytes": 90024
+    },
+    {
+      "path": "osx-arm64/zstd-1.5.7-hf451053_7.conda",
+      "sha256": "da867f5092eb0cb746d353694f0098031fd9817a4ce7d5743121209ae0f406ca",
+      "bytes": 433687
+    },
+    {
+      "path": "osx-arm64/pcre2-10.47-he63d830_1.conda",
+      "sha256": "f8c415329b542e1fca1ff2917b80e687c18302dfa0353530668b90cb225b494f",
+      "bytes": 851704
+    },
+    {
+      "path": "osx-arm64/libedit-3.1.20260512-pl5321h26f1114_0.conda",
+      "sha256": "5692a969e49b200c7874d1e869d281753361980b2ecfe9191b571d7e550f8b92",
+      "bytes": 108512
+    },
+    {
+      "path": "osx-arm64/libpsl-0.23.1-hdb0c161_1.conda",
+      "sha256": "c8c3153df39abedf3413483f672151ad70f1adb0c06c192bdf7ec432c3616b07",
+      "bytes": 72673
+    },
+    {
+      "path": "noarch/ca-certificates-2026.10.23-hbd8a1cb_0.conda",
+      "sha256": "95619b9c1184be7175392eeb0ee7eeb5ccf25d280951f4225f2d62e725790be3",
+      "bytes": 130057
+    },
+    {
+      "path": "osx-arm64/openssl-3.6.5-h55eecbc_0.conda",
+      "sha256": "44fbc526456571a9fb8fa18157f6f269d4d4bd40eef20183a93b9c9296f94497",
+      "bytes": 3111573
+    },
+    {
+      "path": "osx-arm64/libssh2-1.11.1-hbf2880b_1.conda",
+      "sha256": "a056e518c3d2d09fbb1778cea80c0c95338fb24adf1a817bbf90fb484850a304",
+      "bytes": 280492
+    },
+    {
+      "path": "osx-arm64/krb5-1.22.2-h34f8a20_2.conda",
+      "sha256": "aaea1d42b07769920db2af7471ece9399d2613448863da64ad8272998db5db34",
+      "bytes": 1165740
+    },
+    {
+      "path": "osx-arm64/libnghttp2-1.68.1-h435687b_1.conda",
+      "sha256": "445440d8cccccb6585f71d1eb7c949d64a2af2ced5481d924621424ff3b6a398",
+      "bytes": 567024
+    },
+    {
+      "path": "osx-arm64/libcurl-8.23.0-h6651222_0.conda",
+      "sha256": "d9f60525b665e50da15c84fcb4e74e6b1f0a56ea4c89a02d4258f23c3fe1f02c",
+      "bytes": 425637
+    },
+    {
+      "path": "osx-arm64/git-2.56.0-pl5440h66f49b9_0.conda",
+      "sha256": "2faaa571f7da2dd0b17f54ddb8a0437c9bde290af2202257c75ba73889819975",
+      "bytes": 13480930
+    }
+  ]
+}
diff --git a/freevideo_engine/macos_bootstrap.py b/freevideo_engine/macos_bootstrap.py
index f9987bbd..2f1b0ef1 100644
--- a/freevideo_engine/macos_bootstrap.py
+++ b/freevideo_engine/macos_bootstrap.py
@@ -38,6 +38,9 @@ def versions(value):
         uv=dict(UV, url=UV['sources']['official']), github_probe_url=value['uv']['url'])
 
 
+GIT_UNAVAILABLE = 'Git is unavailable. Check the network connection, or install Git (Apple Command Line Tools), then retry.'
+
+
 def available_git():
     executable = shutil.which('git')
     if not executable:
@@ -86,8 +89,6 @@ def preflight(args, snapshot=None, *, ram_gib=None, vram_gib=None):
         errors.append('CUDA device and SageAttention build options do not apply to Mac.')
     if vram_gib is not None:
         errors.append('Mac has one unified memory pool; use --ram-gib instead of a separate VRAM limit.')
-    if not snapshot.get('git'):
-        errors.append('Git is unavailable. Install or repair Git (Apple Command Line Tools), then retry.')
     total, available = hardware['ram_total'], hardware['ram_available']
     if any(type(n) is not int or n < 0 for n in (total, available)) or not 0 < total or available > total:
         raise ValueError('Invalid unified-memory observation')
diff --git a/freevideo_engine/macos_setup.py b/freevideo_engine/macos_setup.py
index 9f6b4c55..cb010837 100644
--- a/freevideo_engine/macos_setup.py
+++ b/freevideo_engine/macos_setup.py
@@ -20,6 +20,19 @@ def device_configuration(self):
     def prepare_tools(self):
         return native.prepare_uv(self.root, self.network, self.env, self.fetch)
 
+    def install_git(self):
+        """The system Git when it runs; otherwise Git from conda-forge in FreeVideo's own folder."""
+        git = native.available_git()
+        if git:
+            return
+        try:
+            self.install_git_conda('git_macos.json', 'git-macos',
+                                   lambda spec: self.fetch_conda_lock(spec, 'git-macos', 'Git'), wrapper=True)
+        except Exception as error:
+            if self.cancel.is_set():
+                raise
+            raise RuntimeError(native.GIT_UNAVAILABLE) from error
+
     def component_tasks(self, uv):
         tasks = {}
         def task(name, run, after=(), writes=()):
@@ -27,6 +40,7 @@ def task(name, run, after=(), writes=()):
         python = self.pythons['engine']
         writer = ('environment:unified',)
         task('python', lambda: self.python_install(uv))
+        task('git', self.install_git)
         task('unified-venv', lambda: None if python.is_file() else
              self.command('unified-venv', [uv, 'venv', '--python', native.PYTHON_VERSION, self.root / 'envs/unified']),
              ('python',), writer)
@@ -36,12 +50,12 @@ def task(name, run, after=(), writes=()):
         task('engine', lambda: self.packages('unified-engine', [uv, 'pip', 'install', '--python', python,
             '--no-deps', '-e', SOURCE]), ('runtime',), writer)
         task('vdn-source', lambda: self.command('vdn-source', [python, '-c',
-            'from freevideo_engine.install import install; install()']), ('unified-venv',))
+            'from freevideo_engine.install import install; install()']), ('unified-venv', 'git'))
         task('vdn-install', lambda: self.packages('vdn', [uv, 'pip', 'install', '--python', python,
             '--no-deps', '-e', self.root / 'vendor/vdn' / self.spec['vdn']['diffusers_subdirectory']]),
             ('engine', 'vdn-source'), writer)
         task('encoder-source', lambda: self.clone('h3-text-encoder',
-            self.spec['encoder']['comfy_url'], self.spec['encoder']['comfy_commit']))
+            self.spec['encoder']['comfy_url'], self.spec['encoder']['comfy_commit']), ('git',))
         task('models', lambda: self.command('models', [python, '-m', 'freevideo_engine.provision',
             '--plan', self.run_dir / 'plan.json']), ('engine',))
         return tasks, {'unified': python}, self.root / 'vendor/h3-text-encoder'
diff --git a/freevideo_engine/network.py b/freevideo_engine/network.py
index ea3000a0..b17120c7 100644
--- a/freevideo_engine/network.py
+++ b/freevideo_engine/network.py
@@ -395,14 +395,14 @@ def probe_score(row, family):
 
 
 def plan(spec, versions, layout='unified', *, mode='auto', timeout=5, offline=False, env=None,
-         model_only=False, measure_speed=True, progress=None, proxy_mode='auto'):
+         model_only=False, measure_speed=True, progress=None, proxy_mode='auto', git_available=True):
     from .environments import ENVIRONMENTS, environment_names
     env = proxy_environment(env)
     started = time.monotonic()
     if proxy_mode not in proxy.MODES:
         raise ValueError('Unknown download connection mode')
     proxy_configured = proxy.configured(env)
-    git_keys = [] if offline or model_only else proxy.git_proxy_keys(env)
+    git_keys = [] if offline or model_only or not git_available else proxy.git_proxy_keys(env)
     families = ['pypi', 'models', 'vdn-models', 'image-models', 'github', 'git', 'cuda']
     if spec['models'].get('edge_modelscope'):
         families.insert(3, 'edge-models')
@@ -422,8 +422,15 @@ def plan(spec, versions, layout='unified', *, mode='auto', timeout=5, offline=Fa
         if mode != 'official' and configured_source(group, env):
             names.insert(0, 'user')
         selections[family] = []
-        routes = proxy.routes(env, git=family == 'git', mode=proxy_mode)
+        routes = proxy.routes(env, git=family == 'git' and git_available, mode=proxy_mode)
         jobs += [(family, name, route) for name in names for route in routes]
+    # Without a usable Git, any `git` on PATH may be Apple's tool launcher,
+    # which opens an installation dialog when run. Leave those sources unmeasured.
+    skipped = [job for job in jobs if job[0] == 'git' and not git_available and not offline]
+    if skipped:
+        jobs = [job for job in jobs if job not in skipped]
+        for family, name, route in skipped:
+            selections[family].append({'id': name, 'route': route, 'ok': None, 'score': 0, 'scope': 'Git not available yet; not measured'})
     if offline:
         for family, name, route in jobs:
             selections[family].append({'id': name, 'route': route, 'ok': None, 'scope': 'Offline fixture; not measured'})
@@ -1392,6 +1399,12 @@ def execute(label, command):
     io_path(target.parent).mkdir(parents=True, exist_ok=True)
     candidates = urls(dict(network, sources=dict(network.get('sources', {}), github=network.get('sources', {}).get('git', [{'id': 'official'}]))), 'github', url, env)
     configured_env = env
+    # A depth-1 fetch cannot continue an earlier attempt, so what a past run left beside the target is only waste.
+    for stale in io_path(target.parent).glob(target.name + '.partial-*'):
+        try:
+            remove_tree(stale)
+        except OSError:
+            pass  # Held by another program; the download goes on and a later run tries again.
     for name, endpoint, route in [(name, endpoint, route) for name, endpoint in candidates
                                  for route in route_order(network, 'git', name, configured_env)]:
         env = route_environment(network, 'git', name, configured_env, route)
@@ -1439,6 +1452,12 @@ def execute(label, command):
             source_health(network, 'git', name, False)
             event(network, category='git', source=name, route=route, file=target.name, action='fallback', reason='network-fetch-failed')
         finally:
-            if scratch and scratch.exists():
-                remove_tree(scratch)
-    raise RuntimeError('All Git sources failed. Check your network or connection mode in Downloads. Incomplete checkouts retained beside ' + str(target))
+            # Best effort: a file still held by antivirus or git.exe must not replace the
+            # error being handled or stop the fallback to the next source.
+            for leftover in (scratch, io_path(temporary)):
+                try:
+                    if leftover and leftover.exists():
+                        remove_tree(leftover)
+                except OSError:
+                    pass
+    raise RuntimeError('All Git sources failed. Check your network or connection mode in Downloads.')
diff --git a/freevideo_engine/origin_scheme.py b/freevideo_engine/origin_scheme.py
new file mode 100644
index 00000000..6ec5f846
--- /dev/null
+++ b/freevideo_engine/origin_scheme.py
@@ -0,0 +1,16 @@
+"""The one rule for comparing a page's Origin protocol with the protocol the server saw."""
+
+
+def effective_scheme(request, origin_scheme):
+    """The protocol to compare an Origin with.
+
+    A TLS-terminating proxy turns https into http before the request arrives. Only when the request carries a single
+    X-Forwarded-Proto equal to the Origin's own protocol is that difference accepted; host and port are still compared
+    by the caller, and no other forwarding header is read.
+    """
+    headers = request.headers
+    forwarded = headers.getall('X-Forwarded-Proto', []) if hasattr(headers, 'getall') else [headers['X-Forwarded-Proto']] if 'X-Forwarded-Proto' in headers else []
+    if (origin_scheme in ('http', 'https') and len(forwarded) == 1
+            and forwarded[0].strip().lower() == origin_scheme):
+        return origin_scheme
+    return request.scheme
diff --git a/freevideo_engine/policy.py b/freevideo_engine/policy.py
index 01d4b231..d7ea6248 100644
--- a/freevideo_engine/policy.py
+++ b/freevideo_engine/policy.py
@@ -622,6 +622,22 @@ def __init__(self, hardware, gpu_total, ram_total, free_gpu, free_ram, reserve_g
         lines.append('Available memory includes validated idle-cache credit, if present. Resolution, frames and steps were not reduced.')
         super().__init__('\n'.join(lines))
 
+    def add_own_memory(self, own):
+        """Say how much of the physical RAM already taken is FreeVideo's own."""
+        self.details.update(own)
+        parts = []
+        if 'own_private_working_set_bytes' in own:
+            partial = own.get('own_unreadable_processes', 0)
+            parts.append('FreeVideo\'s generation processes hold %s%.2f GiB of private physical RAM (%d processes%s)' %
+                         ('at least ' if partial else '', own['own_private_working_set_bytes'] / GiB,
+                          own.get('own_processes', 0), ', %d unreadable' % partial if partial else ''))
+        if 'resident_private_working_set_bytes' in own:
+            parts.append('FreeVideo\'s resident worker holds %.2f GiB' % (own['resident_private_working_set_bytes'] / GiB))
+        if 'host_private_working_set_bytes' in own:
+            parts.append('the largest process above the generation process holds %.2f GiB' % (own['host_private_working_set_bytes'] / GiB))
+        if parts:
+            parts[0] = parts[0][0].upper() + parts[0][1:]
+            self.args = ('%s\n%s.' % (self.args[0], '; '.join(parts)),)
 
 
 def feasible_alternatives(hardware, canvas, **options):
diff --git a/freevideo_engine/release_notes.json b/freevideo_engine/release_notes.json
index 3888e749..91ea03f2 100644
--- a/freevideo_engine/release_notes.json
+++ b/freevideo_engine/release_notes.json
@@ -1,21 +1,42 @@
 {
   "schema": 1,
-  "product_version": "0.4.21",
+  "product_version": "0.4.22",
   "en": {
-    "summary": "Agent image tasks can place characters in a scene by position, size and facing; the page check no longer reports a covered toolbar behind ComfyUI dialogs.",
+    "summary": "Autopilot chat can attach reference pictures; the launcher and the editor explain more precisely why a step failed; Mac setup no longer needs Git installed.",
     "changes": [
-      "Agent images: an image task can take a scene picture and one to three characters (character card or picture) with a position, a height, a full-body or half-body framing and a facing direction. The picture is generated with the scene as the first reference instead of a pasted collage, and the result can be used directly as a first frame.",
-      "Page check: the FreeVideo toolbar is no longer reported as covered when ComfyUI shows a dimming or click-away layer over it, such as the Templates dialog on a first start or the Help panel; the details now name what covers the toolbar."
+      "Autopilot: a chat message can carry reference pictures, added with the Picture button, by pasting or by dragging, up to 4 per message and 10 MB each. A picture that fails to upload stays in place and can be clicked to retry.",
+      "Launcher and setup: choosing a folder inside a portable ComfyUI, such as its python_embeded folder, now names the portable folder to select. Mac setup downloads its own Git when the system has none. A failed Git download no longer leaves an incomplete folder behind, and leftovers from earlier attempts are removed. On Windows, when security software briefly holds the newly downloaded installer tool, setup retries for a few seconds instead of stopping with \"This folder is not writable\".",
+      "Edit: a failed save now states its cause, and a lost connection is reported as a lost connection. Edit also works on an https page served through a proxy that terminates the secure connection, as on some cloud GPU hosts.",
+      "Diagnostic reports: the page check states why an open creative panel cannot be seen. When a request is refused for lack of memory on Windows, the report records how much of the used memory FreeVideo itself holds."
     ]
   },
   "zh": {
-    "summary": "Agent 图片任务可按位置、大小和朝向把人物放进场景；页面检查不再因 ComfyUI 的对话框遮罩报告工具栏被遮挡。",
+    "summary": "AI 全自动聊天可以附参考图；启动器和编辑器更准确地说明失败原因；Mac 安装不再需要预先装好 Git。",
     "changes": [
-      "Agent 图片：图片任务可以指定一张场景图和一至三个人物（角色卡或图片），并设定位置、人物高度、全身或半身景别和朝向。图片以场景为第一参考图生成，不再拼贴粘图，结果可直接作首帧。",
-      "页面检查：ComfyUI 的调暗或点击遮罩（首次启动的模板对话框、帮助面板等）盖在 FreeVideo 工具栏上时，不再报告工具栏被遮挡；诊断详情会写出遮挡工具栏的元素。"
+      "AI 全自动：聊天消息可以附参考图，用“图片”按钮、粘贴或拖入添加，每条最多 4 张、每张不超过 10 MB。上传失败的图片会保留，点击即可重试。",
+      "启动器和安装：选中便携版 ComfyUI 内部的文件夹（例如 python_embeded）时，提示会写出应选择的便携版文件夹。Mac 系统没有 Git 时，安装会自行下载 Git。Git 下载失败后不再留下不完整的文件夹，以前残留的也会清除。Windows 上安全软件短暂占用刚下载的安装工具时，安装会重试数秒，不再以“此文件夹不可写”停止。",
+      "编辑：保存失败时会写明原因，连接断开时显示为连接已断开。通过代理终止安全连接的 https 页面（部分云端显卡主机即是如此）上也可以正常使用编辑。",
+      "诊断报告：页面检查会写明创作面板已打开却看不到的原因。Windows 上请求因内存不足被拒绝时，报告会写出已用内存中 FreeVideo 自己占用多少。"
     ]
   },
   "history": [
+    {
+      "product_version": "0.4.21",
+      "en": {
+        "summary": "Agent image tasks can place characters in a scene by position, size and facing; the page check no longer reports a covered toolbar behind ComfyUI dialogs.",
+        "changes": [
+          "Agent images: an image task can take a scene picture and one to three characters (character card or picture) with a position, a height, a full-body or half-body framing and a facing direction. The picture is generated with the scene as the first reference instead of a pasted collage, and the result can be used directly as a first frame.",
+          "Page check: the FreeVideo toolbar is no longer reported as covered when ComfyUI shows a dimming or click-away layer over it, such as the Templates dialog on a first start or the Help panel; the details now name what covers the toolbar."
+        ]
+      },
+      "zh": {
+        "summary": "Agent 图片任务可按位置、大小和朝向把人物放进场景；页面检查不再因 ComfyUI 的对话框遮罩报告工具栏被遮挡。",
+        "changes": [
+          "Agent 图片：图片任务可以指定一张场景图和一至三个人物（角色卡或图片），并设定位置、人物高度、全身或半身景别和朝向。图片以场景为第一参考图生成，不再拼贴粘图，结果可直接作首帧。",
+          "页面检查：ComfyUI 的调暗或点击遮罩（首次启动的模板对话框、帮助面板等）盖在 FreeVideo 工具栏上时，不再报告工具栏被遮挡；诊断详情会写出遮挡工具栏的元素。"
+        ]
+      }
+    },
     {
       "product_version": "0.4.20",
       "en": {
@@ -238,35 +259,6 @@
           "安装：Windows 上的库文件检查也支持层级很深的文件夹。"
         ]
       }
-    },
-    {
-      "product_version": "0.4.13",
-      "en": {
-        "summary": "High is faster than Max again, videos at another size of the same shape can be continued, LoRAs on int8 models apply what they can, Settings adds step-by-step how-tos and an offline tutorial, and the installers are smaller.",
-        "changes": [
-          "High: the step cache is back with added checks, so High runs faster than Max again.",
-          "Long videos: parts of a long video and continued parts use 8 + 3 steps, so a second part is no longer refused.",
-          "Continue: a video at another size of the same shape now continues from its own ending. Videos that are too short or have too low a frame rate show a clear message.",
-          "LoRA: on int8 models, the parts of a LoRA that cannot be applied are skipped and the rest takes effect, instead of the whole LoRA being refused.",
-          "Help: \"User guide\" in Settings shows three how-tos on the real interface, and the new \"Tutorial\" opens an offline illustrated manual in English and Chinese.",
-          "Upgrade: a FreeVideo tab left open from an older version switches to the current workflow and keeps the prompt, size, length, seed, assets and LoRAs.",
-          "Agent: assembled videos are saved to \"Your creations\", and their task cards can be opened.",
-          "Install: the Windows and Linux installers are smaller, the prompt enhancement model can also download from ModelScope, and the GPU acceleration step shows one steady name and description while it builds."
-        ]
-      },
-      "zh": {
-        "summary": "精细档重新比极致档快；同比例、不同尺寸的视频可以接着生成；int8 模型上的 LoRA 能用的部分照常生效；设置中新增分步演示和离线教程；安装包变小。",
-        "changes": [
-          "精细档：重新启用步缓存，并加上检查，精细档重新比极致档快。",
-          "长视频：长视频的分段和接着生成的分段统一使用 8 + 3 步，第二段不再被拒绝。",
-          "接着生成：同比例、不同尺寸的视频改为从它自己的结尾接着生成；太短或帧率太低的视频会给出明确提示。",
-          "LoRA：在 int8 模型上，LoRA 中无法应用的部分会跳过，其余照常生效，不再整个拒绝。",
-          "帮助：设置中的“使用说明”改为在真实界面上演示三项操作；新增“教程”，可打开中英文离线图文手册。",
-          "升级：旧版本留在标签页中的 FreeVideo 工作流会换成当前版本，并保留提示词、尺寸、时长、种子、素材和 LoRA。",
-          "Agent：拼接的成片保存到“我的作品”，任务卡可以打开。",
-          "安装：Windows 和 Linux 安装包变小；提示词增强模型也可以从魔搭下载；GPU 加速组件构建时，名称和说明保持不变。"
-        ]
-      }
     }
   ]
 }
diff --git a/freevideo_engine/uv_bootstrap.py b/freevideo_engine/uv_bootstrap.py
index 1ea42338..6fc7c320 100644
--- a/freevideo_engine/uv_bootstrap.py
+++ b/freevideo_engine/uv_bootstrap.py
@@ -4,6 +4,7 @@
 from pathlib import Path
 import re
 import shutil
+import time
 import zipfile
 
 from . import network
@@ -27,6 +28,30 @@ def verified(path, sha, size):
     return path.is_file() and path.stat().st_size == size and network.hash_file(path) == sha
 
 
+def publish(temporary, destination, delays=(0.25, 0.5, 1, 2, 2, 2, 2, 2)):
+    """Replace the tool with the extracted copy. Windows refuses the replacement while a scanner or a running
+    tool holds the file, so it is retried for a few seconds; an identical tool already in place is kept."""
+    error = None
+    for delay in (*delays, None):
+        try:
+            os.replace(temporary, destination)
+            return destination
+        except PermissionError as denied:
+            error = denied
+            if delay is None:
+                break
+            time.sleep(delay)
+    try:
+        same = destination.is_file() and destination.stat().st_size == temporary.stat().st_size \
+            and network.hash_file(destination) == network.hash_file(temporary)
+    except OSError:
+        same = False
+    if not same:
+        raise error
+    temporary.unlink(missing_ok=True)
+    return destination
+
+
 def extract(archive, destination, member, size=None):
     """Only extract the pinned executable, with bounded memory and atomic replacement."""
     destination.parent.mkdir(parents=True, exist_ok=True)
@@ -37,8 +62,7 @@ def extract(archive, destination, member, size=None):
             raise ValueError('Unexpected uv executable size in the verified archive')
         with source.open(info) as incoming, temporary.open('wb') as outgoing:
             shutil.copyfileobj(incoming, outgoing, 1024 * 1024)
-    os.replace(temporary, destination)
-    return destination
+    return publish(temporary, destination)
 
 
 def prepare_windows(root, uv, networking, env, fetch):
diff --git a/pyproject.toml b/pyproject.toml
index 2e898f61..0fbe352f 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -36,4 +36,4 @@ include = ["freevideo_engine*"]
 [tool.setuptools.package-data]
 "freevideo_engine.prompt_vlm" = ["model.json"]
 "freevideo_engine.image_gen" = ["image_models.json", "licenses/NOTICE.txt"]
-freevideo_engine = ["stage_estimate_params.json", "build-version.txt", "build-identity.json", "release_notes.json", "dependencies.json", "model_files.json", "model_shares.json", "prepared_models.json", "macos_reference_catalog.json", "superseded_tables.json", "bootstrap_versions.json", "curl_windows.json", "git_linux.json", "gcc_linux.json", "uv_downloads.json", "torch_downloads.json", "test_prompts.json", "windows_bootstrap.ps1", "assets/*.png", "assets/*.ico", "launcher/*.qml", "launcher/licenses/*.txt", "licenses/*.txt", "licenses/runtime/*.txt", "subtitle_pack.json", "frame_preview_weights.json"]
+freevideo_engine = ["stage_estimate_params.json", "build-version.txt", "build-identity.json", "release_notes.json", "dependencies.json", "model_files.json", "model_shares.json", "prepared_models.json", "macos_reference_catalog.json", "superseded_tables.json", "bootstrap_versions.json", "curl_windows.json", "git_linux.json", "git_macos.json", "gcc_linux.json", "uv_downloads.json", "torch_downloads.json", "test_prompts.json", "windows_bootstrap.ps1", "assets/*.png", "assets/*.ico", "launcher/*.qml", "launcher/licenses/*.txt", "licenses/*.txt", "licenses/runtime/*.txt", "subtitle_pack.json", "frame_preview_weights.json"]
diff --git a/web/agent_session.js b/web/agent_session.js
index 9eb20b7c..d4a49989 100644
--- a/web/agent_session.js
+++ b/web/agent_session.js
@@ -45,6 +45,18 @@ export async function agentSend(api, method, path, body, headers = {}) {
 
 // UI actions retry once, including HTTP refusals. Callers keep the same
 // idempotency id/body and show the result where the user clicked.
+// Raw bytes (one picture) with the same page session and one retry after a restart.
+export async function agentUpload(api, path, body, signal) {
+    for (let attempt = 0; ; attempt++) {
+        const response = await api.fetchApi(path, {method: 'POST', credentials: 'same-origin', signal, body,
+            headers: {'Content-Type': 'application/octet-stream', ...await agentHeaders(api)}});
+        if (response.status !== 403 || attempt) return response;
+        const value = await response.clone().json().catch(() => null);
+        if (value?.error?.code !== 'page_session_required') return response;
+        sessions.delete(api);
+    }
+}
+
 export async function retryAgentPost(api, path, body, {retry = async () => body, accept = () => true} = {}) {
     for (let attempt = 0; attempt < 2; attempt++) {
         let response = null;
diff --git a/web/autopilot_chat.css b/web/autopilot_chat.css
index 532c7ac7..593ca5b3 100644
--- a/web/autopilot_chat.css
+++ b/web/autopilot_chat.css
@@ -128,7 +128,7 @@
 @keyframes fv-chat-breathe{0%,100%{opacity:.25}40%{opacity:1}}
 .fv-studio .fv-chat-down{position:absolute;right:max(40px,calc(50% - 360px));bottom:96px;width:34px;height:34px;min-height:0;padding:0;border-radius:50%;background:var(--fv-raised);box-shadow:0 6px 18px #0008}
 .fv-chat svg{width:16px;height:16px;fill:none;stroke:currentColor;stroke-linecap:round;stroke-linejoin:round}
-.fv-chat-input{box-sizing:border-box;justify-self:center;width:100%;max-width:720px;min-height:54px;display:grid;grid-template-columns:minmax(0,1fr) auto auto auto;gap:10px;align-items:end;border:1px solid var(--fv-border);background:#0e151e;border-radius:14px;padding:8px 8px 8px 16px;transition:border-color var(--fv-quick),box-shadow var(--fv-quick)}
+.fv-chat-input{box-sizing:border-box;justify-self:center;width:100%;max-width:720px;min-height:54px;display:grid;grid-template-columns:minmax(0,1fr) auto auto auto auto;gap:10px;align-items:end;border:1px solid var(--fv-border);background:#0e151e;border-radius:14px;padding:8px 8px 8px 16px;transition:border-color var(--fv-quick),box-shadow var(--fv-quick)}
 .fv-chat-notice-row{justify-self:center;width:100%;max-width:720px;display:flex;align-items:center;flex-wrap:wrap;gap:4px 12px}.fv-chat-notice-row>.fv-issue-button{margin:6px 0}
 .fv-chat-notice{flex:0 1 auto;margin:6px 0;font-size:14px;color:var(--fv-ink)}
 .fv-chat-preview[data-loading=true] .fv-chat-play::before{border-color:var(--fv-accent);animation:fv-chat-breathe 1s infinite}
@@ -141,11 +141,11 @@
 .fv-studio.fv-autopilot .fv-chat-input textarea,.fv-chat-input textarea{box-sizing:border-box;width:100%;height:36px;min-height:36px;max-height:none;padding:7px 0;border:0;border-radius:0;background:none;color:var(--fv-ink);font-size:14.5px;line-height:22px;resize:none;outline:none;overflow-y:auto}
 .fv-chat-input textarea::placeholder{color:var(--fv-disabled)}
 .fv-chat-input textarea{min-width:0}
-.fv-studio :is(.fv-chat-whitebox,.fv-chat-scenes){height:36px;min-height:36px;padding:0 8px;gap:5px;border:1px solid var(--fv-border);border-radius:9px;background:transparent;color:var(--fv-muted);font-size:12px;white-space:nowrap}
-:is(.fv-chat-whitebox,.fv-chat-scenes) svg{flex:none}
+.fv-studio :is(.fv-chat-whitebox,.fv-chat-scenes,.fv-chat-attach){height:36px;min-height:36px;padding:0 8px;gap:5px;border:1px solid var(--fv-border);border-radius:9px;background:transparent;color:var(--fv-muted);font-size:12px;white-space:nowrap}
+:is(.fv-chat-whitebox,.fv-chat-scenes,.fv-chat-attach) svg{flex:none}
 .fv-studio .fv-chat-whitebox[aria-pressed=true]{background:var(--fv-selected);border-color:var(--fv-accent);color:var(--fv-accent)}
-.fv-studio :is(.fv-chat-whitebox,.fv-chat-scenes):disabled{opacity:.45;cursor:not-allowed}
-.fv-studio :is(.fv-chat-whitebox,.fv-chat-scenes):focus-visible{outline:2px solid var(--fv-accent);outline-offset:2px}
+.fv-studio :is(.fv-chat-whitebox,.fv-chat-scenes,.fv-chat-attach):disabled{opacity:.45;cursor:not-allowed}
+.fv-studio :is(.fv-chat-whitebox,.fv-chat-scenes,.fv-chat-attach):focus-visible{outline:2px solid var(--fv-accent);outline-offset:2px}
 .fv-studio .fv-chat-send{width:36px;height:36px;min-height:0;padding:0;border:0;border-radius:9px;background:var(--fv-accent);color:var(--fv-bg)}
 .fv-studio .fv-chat-send:hover:not(:disabled){background:var(--fv-accent-hover)}
 .fv-studio .fv-chat-send:disabled{opacity:1;background:#22344a;color:#9fb3c8}
@@ -163,7 +163,7 @@
  .fv-chat[data-chats-open=true] .fv-chat-conversations:not([hidden]){display:flex}
 }
 @media(max-width:480px){.fv-chat-main{padding:0 14px 14px}.fv-chat-scroll{margin:0 -14px;padding:16px 14px 2px}.fv-chat-hello{padding-top:32px}.fv-chat-job{max-width:none;gap:12px}.fv-chat-confirm{max-width:none}
-@media(max-width:560px){.fv-studio .fv-chat-scenes{width:36px;padding:0;justify-content:center}.fv-chat-scenes>span{display:none}}
+@media(max-width:560px){.fv-studio :is(.fv-chat-scenes,.fv-chat-whitebox){width:36px;padding:0;justify-content:center}:is(.fv-chat-scenes,.fv-chat-whitebox)>span{display:none}}
  .fv-chat-preview,.fv-chat-preview[data-shape=wide]{width:98px}.fv-chat-length{display:none}.fv-chat-percent{font-size:22px}.fv-studio.fv-autopilot .fv-chat-input textarea{font-size:16px}.fv-studio .fv-chat-down{right:14px;bottom:84px}}
 @media(prefers-reduced-motion:reduce){.fv-chat-water,.fv-chat-percent,.fv-chat-parts>i::after{transition:none}.fv-chat-water::after,.fv-chat-dots>i,.fv-chat-preview>img,.fv-chat-menu{animation:none}}
 
@@ -184,3 +184,18 @@
 .fv-studio .fv-chat-scene-item:focus-visible{outline:2px solid var(--fv-accent);outline-offset:-2px}
 .fv-chat-main > .fv-request-assets{font-size:12px;line-height:1.5;margin:0 0 6px;padding:0 4px}
 .fv-chat-main > .fv-request-assets[hidden]{display:none}
+
+.fv-studio .fv-chat-attach{width:36px;padding:0;justify-content:center}
+.fv-chat-attached{grid-column:1/-1;display:flex;flex-wrap:wrap;gap:8px;margin:0 0 2px;padding:0;list-style:none}
+.fv-chat-attached[hidden]{display:none}
+.fv-chat-attached>li{position:relative;width:52px;height:52px;border-radius:9px;overflow:hidden;border:1px solid var(--fv-border);transition:opacity var(--fv-quick)}
+.fv-chat-attached>li[data-state=uploading]{opacity:.55}
+.fv-chat-attached>li[data-state=failed]{border:2px solid var(--fv-danger);cursor:pointer}
+.fv-chat-attached>li:focus-visible{outline:2px solid var(--fv-accent);outline-offset:2px}
+.fv-chat-thumb{display:grid;place-items:center;width:100%;height:100%;background:var(--fv-hover)}
+.fv-chat-thumb img{display:block;width:100%;height:100%;object-fit:cover}
+.fv-chat-thumb[data-broken] svg{display:block;width:24px;height:24px;fill:none;stroke:var(--fv-muted);stroke-width:2;stroke-linecap:round;stroke-linejoin:round}
+.fv-studio .fv-chat-attached .fv-chat-attached-x{position:absolute;top:2px;right:2px;width:20px;height:20px;min-height:0;padding:0;border:1px solid rgba(255,255,255,.22);border-radius:50%;background:rgba(8,12,18,.9)!important;color:var(--fv-ink);font-size:14px;line-height:1}
+.fv-chat-refs{width:max-content;max-width:100%;display:flex;flex-wrap:wrap;gap:6px;justify-content:flex-end;margin:0 0 6px auto;padding:0;list-style:none}
+.fv-chat-refs>li{width:96px;height:96px;border-radius:10px;overflow:hidden}
+@media(max-width:560px){.fv-chat-refs>li{width:64px;height:64px}}
diff --git a/web/autopilot_chat.js b/web/autopilot_chat.js
index 6308a9a2..a1fdeca9 100644
--- a/web/autopilot_chat.js
+++ b/web/autopilot_chat.js
@@ -1,10 +1,11 @@
 // The conversation with the connected agent: what the user says, what the agent answers, and the agent's jobs as
 // cards at the place they happened. Reads and sends use the page session; the agent's token never enters the page.
 import {agentGet, agentPost, retryAgentPost} from './agent_session.js';
+import {ATTACH, attachmentURL, pickPictures, thumbnail, uploadPicture} from './chat_attachments.js';
 import {SCENE_TEXT, sceneList, startScene, watchScene} from './whitebox_scenes.js';
 import {chatData, chatDelay, chatReadRequest, emptyChat, chatMerge, chatEntries, canSend, newId, SEND_ATTEMPTS, WAIT_FOR_AGENT, sendRefused, sendError} from './chat_data.js';
 export {emptyChat, chatMerge, chatEntries, canSend} from './chat_data.js';
-import {isolateDialogKeys, routeDialogKey} from './dialog_keys.js';
+import {isolateDialogKeys, listenDialogKey, routeDialogKey} from './dialog_keys.js';
 import {closeDialog} from './motion.js';
 import { estimatedFraction, remainingAbove, remainingShown, createDiagnosticReportControl } from './generation_progress.js';
 import {createErrorReport} from './error_panel.js';
@@ -259,7 +260,7 @@ export function agentChat(t, api, {open, edit, shown = () => true, changed = ()
         sendButton.classList.toggle('fv-chat-halt', stopMode);
         sendButton.setAttribute('aria-label', label); sendButton.title = stopMode ? label : '';
         sendButton.querySelector('path').setAttribute('d', stopMode ? STOP_GLYPH : SEND_GLYPH);
-        sendButton.disabled = stopMode ? false : !composeAllowed() || draftSending() || !input.value.trim();
+        sendButton.disabled = stopMode ? false : !composeAllowed() || draftSending() || !readyPictures().length && !input.value.trim() || stagedList().some(item => !item.attachment);
     }
     async function abandon() {
         const session = state.session;
@@ -319,7 +320,72 @@ export function agentChat(t, api, {open, edit, shown = () => true, changed = ()
             }
         } catch (failure) { sceneWaiting = false; say(t(...SCENE_TEXT.failed) + reason(failure)); }
     }
-    form.append(input, scenesButton, whiteboxButton, sendButton); scroller.append(log); main.append(scroller, down, sceneMenu, form); element.append(main);
+    // Pictures waiting to be sent, per conversation draft: {key, url, attachment|null}. A picture is uploaded the moment it is added.
+    const attachButton = press('fv-chat-attach', null, () => fileInput.click()), fileInput = el('input'), strip = el('ul', 'fv-chat-attached');
+    attachButton.append(glyph(ATTACH.icon)); attachButton.title = t(...ATTACH.add); attachButton.setAttribute('aria-label', attachButton.title);
+    fileInput.type = 'file'; fileInput.accept = ATTACH.types.join(','); fileInput.multiple = true; fileInput.hidden = true;
+    strip.hidden = true; strip.setAttribute('aria-label', t('Attached pictures', '已添加的图片'));
+    const staged = new Map(), uploads = new Set();
+    const stagedList = (session = draftSession) => { if (!staged.has(session)) staged.set(session, []); return staged.get(session); };
+    const readyPictures = () => stagedList().filter(item => item.attachment).map(item => ({...item.attachment, preview: item.url}));
+    function renderStrip() {
+        const list = stagedList();
+        strip.hidden = !list.length;
+        strip.replaceChildren(...list.map(item => {
+            const li = el('li'), drop = press('fv-chat-attached-x', '×', () => removePicture(item));
+            li.dataset.state = item.failed ? 'failed' : item.attachment ? 'ready' : 'uploading';
+            drop.title = t(...ATTACH.remove); drop.setAttribute('aria-label', drop.title);
+            li.append(thumbnail(item.url), drop);
+            // A picture that did not upload is tried again by clicking it.
+            if (item.failed === 'failed') {
+                li.tabIndex = 0; li.setAttribute('role', 'button'); li.title = t(...ATTACH.retry); li.setAttribute('aria-label', li.title);
+                li.addEventListener('click', event => { if (event.target === drop) return; void upload(item, list); });
+                li.addEventListener('keydown', event => { if (event.key === 'Enter' || event.key === ' ') { event.preventDefault(); void upload(item, list); } });
+            }
+            return li;
+        }));
+        grow();
+    }
+    function removePicture(item) {
+        item.controller?.abort();
+        for (const list of staged.values()) { const at = list.indexOf(item); if (at >= 0) list.splice(at, 1); }
+        URL.revokeObjectURL?.(item.url); renderStrip(); input.focus();
+    }
+    async function upload(item, list, session = draftSession) {
+        item.controller = new AbortController(); item.failed = null; uploads.add(item); renderStrip();
+        try {
+            item.attachment = await uploadPicture(api, item.file, item.controller.signal);
+            // The same picture twice is one stored picture; the copy is dropped.
+            if (list.some(other => other !== item && other.attachment?.name === item.attachment.name)) removePicture(item);
+        } catch (failure) {
+            if (!failure?.problem) return;
+            // Unreadable pictures are taken out; an upload that failed stays, marked, to be tried again.
+            if (failure.problem === 'unreadable') { const at = list.indexOf(item); if (at >= 0) list.splice(at, 1); URL.revokeObjectURL?.(item.url); }
+            else item.failed = 'failed';
+            if (session === draftSession) say(t(...ATTACH[failure.problem]));
+        } finally { uploads.delete(item); if (!stopped && session === draftSession) renderStrip(); }
+    }
+    function addPictures(files) {
+        files = [...files || []];
+        if (!files.length || !composeAllowed()) return;
+        const session = draftSession, list = stagedList(session), {accepted, problem} = pickPictures(files, ATTACH.count - list.length);
+        say(problem ? t(...ATTACH[problem]) : '');
+        for (const file of accepted) {
+            const item = {name: file.name, file, url: URL.createObjectURL(file), attachment: null, failed: null, controller: null};
+            list.push(item); void upload(item, list, session);
+        }
+        renderStrip();
+    }
+    fileInput.addEventListener('change', () => { addPictures(fileInput.files); fileInput.value = ''; });
+    listenDialogKey(input, 'paste', event => {
+        const files = [...event.clipboardData?.files || []].filter(file => file.type.startsWith('image/'));
+        if (files.length) { event.preventDefault(); addPictures(files); }
+    });
+    for (const type of ['dragover', 'drop']) main.addEventListener(type, event => {
+        if (![...event.dataTransfer?.types || []].includes('Files')) return;
+        event.preventDefault(); if (type === 'drop') addPictures(event.dataTransfer.files);
+    });
+    form.append(strip, input, attachButton, scenesButton, whiteboxButton, sendButton, fileInput); scroller.append(log); main.append(scroller, down, sceneMenu, form); element.append(main);
     const notice = el('p', 'fv-chat-notice'); notice.hidden = true; notice.setAttribute('role', 'status');
     const noticeRow = el('div', 'fv-chat-notice-row'); noticeRow.append(notice); main.insertBefore(noticeRow, form);
     // What the next video would download first, the studio's line, above the box the request is typed in.
@@ -364,7 +430,7 @@ export function agentChat(t, api, {open, edit, shown = () => true, changed = ()
             saveDraft(); grow(); return;
         }
         if (draftSession === session) return;
-        saveDraft(); draftSession = session; input.value = drafts.get(session)?.text || ''; grow();
+        saveDraft(); draftSession = session; input.value = drafts.get(session)?.text || ''; renderStrip();
     }
     const reason = error => error?.message || t('Connection lost. Try again.', '连接中断，请重试。');
     async function requestError(response) {
@@ -711,9 +777,18 @@ export function agentChat(t, api, {open, edit, shown = () => true, changed = ()
     // ---- one entry of the conversation each ----
     function message(role) {
         const item = el('li', role === 'user' ? 'fv-chat-me' : 'fv-chat-ai'), text = el('p'); item.append(text);
-        let status = null, error = null, retry = null, report = null;
+        let status = null, error = null, retry = null, report = null, shown = '';
         return {element: item, update(entry) {
-            text.textContent = entry.text;
+            text.textContent = entry.text; text.hidden = !entry.text && !!entry.attachments?.length;
+            const names = (entry.attachments || []).map(picture => picture.name).join('|');
+            if (names !== shown) {
+                shown = names; item.querySelector('.fv-chat-refs')?.remove();
+                if (names) {
+                    const row = el('ul', 'fv-chat-refs');
+                    for (const picture of entry.attachments) { const li = el('li'); li.append(thumbnail(picture.preview || attachmentURL(api, picture.name))); row.append(li); }
+                    item.prepend(row);
+                }
+            }
             if (role !== 'user') return;
             item.dataset.taken = String(entry.taken !== false);
             item.dataset.failed = String(!!entry.error);
@@ -1159,19 +1234,24 @@ export function agentChat(t, api, {open, edit, shown = () => true, changed = ()
             drafts.delete(entry.session);
             if (draftSession === entry.session) { input.value = ''; grow(); }
         }
+        if (entry.attachments?.length) {
+            const list = stagedList(entry.session), sent = new Set(entry.attachments.map(item => item.name));
+            staged.set(entry.session, list.filter(item => !sent.has(item.attachment?.name)));
+            if (draftSession === entry.session) renderStrip();
+        }
     }
     const waitingNotice = code => (code ? code === 'not_listening' : state.agent.connected === true)
         ? t('Will send when Agent is back', '等 Agent 回来后自动发送')
         : t('Will send when Agent connects', '等 Agent 连上后自动发送');
-    async function send(text, original = null, draft = null, automatic = false) {
+    async function send(text, original = null, draft = null, automatic = false, attachments = null) {
         text = String(text).trim();
         original ||= pending.find(entry => entry.session === (state.session ?? null) && entry.text === text && (entry.error || entry.sending || entry.automatic));
-        if (!text || original?.sending || original && original.session !== state.session) return false;
+        if (!text && !(original?.attachments || attachments)?.length || original?.sending || original && original.session !== state.session) return false;
         if (!inSelectedChat()) {
             say(creating ? t('Creating a new chat…', '正在新建对话…') : t('Loading this chat. Please send again shortly.', '正在读取对话，请稍后再发一次'));
             return false;
         }
-        const entry = original || {id: newId(), text, session: state.session ?? null, draft, whiteboxPreview: whiteboxPreview && state.agent?.builtin !== true};
+        const entry = original || {id: newId(), text, attachments: attachments || [], session: state.session ?? null, draft, whiteboxPreview: whiteboxPreview && state.agent?.builtin !== true};
         if (draft) entry.draft = draft;
         if (!original) pending.push(entry);
         if (!automatic) entry.attempts = 0;
@@ -1243,9 +1323,10 @@ export function agentChat(t, api, {open, edit, shown = () => true, changed = ()
         if (stopMode) { void abandon(); return; }
         const text = input.value;
         const draft = saveDraft();
-        if (!text.trim() || !composeAllowed() || draftSending()) return;
+        const pictures = readyPictures();
+        if (!text.trim() && !pictures.length || !composeAllowed() || draftSending() || stagedList().some(item => !item.attachment)) return;
         const failed = pending.find(entry => entry.session === state.session && entry.draft === draft && entry.error);
-        void send(text, failed, draft);
+        void send(text, failed, draft, false, pictures);
     });
 
     // Both a server-side session change and a list selection reset the same reader. Job clocks stay keyed by job id.
diff --git a/web/chat_attachments.js b/web/chat_attachments.js
new file mode 100644
index 00000000..68a1375b
--- /dev/null
+++ b/web/chat_attachments.js
@@ -0,0 +1,54 @@
+// Pictures attached to a chat message: limits, checks, upload and thumbnail address.
+import {agentUpload} from './agent_session.js';
+
+export const ATTACH = {
+    count: 4, bytes: 10 * 1024 * 1024, types: ['image/png', 'image/jpeg', 'image/webp'],
+    add: ['Add pictures', '添加图片'], remove: ['Remove', '移除'],
+    tooMany: ['Up to 4 pictures per message.', '一条消息最多 4 张图片。'],
+    tooLarge: ['Each picture must be under 10 MB.', '每张图片不能超过 10 MB。'],
+    notPicture: ['Only pictures can be added.', '只能添加图片。'],
+    unreadable: ['Could not read this picture. Please choose another one.', '无法读取这张图片，请换一张。'],
+    failed: ['Could not upload the picture. Click it to try again.', '图片上传失败，请点击该图片重试。'],
+    retry: ['Upload again', '重新上传'],
+    icon: 'M4 6a2 2 0 0 1 2-2h12a2 2 0 0 1 2 2v12a2 2 0 0 1-2 2H6a2 2 0 0 1-2-2zM4 16l4-4 4 4 3-3 5 5M15 9h.01',
+};
+
+// Splits files into the ones that fit and the first reason for leaving one out: {accepted, problem}.
+export function pickPictures(files, room) {
+    const accepted = []; let problem = null;
+    for (const file of files) {
+        if (!ATTACH.types.includes(file.type)) problem ||= 'notPicture';
+        else if (file.size > ATTACH.bytes) problem ||= 'tooLarge';
+        else if (accepted.length >= room) problem ||= 'tooMany';
+        else accepted.push(file);
+    }
+    return {accepted, problem};
+}
+
+export const attachmentURL = (api, name) => {
+    const path = '/freevideo/agent/chat/attachment?' + new URLSearchParams({name});
+    return api.apiURL?.(path) || path;
+};
+
+// Resolves to the attachment object; rejects with {problem} naming a key of ATTACH.
+export async function uploadPicture(api, file, signal) {
+    let response;
+    try { response = await agentUpload(api, '/freevideo/agent/chat/attach?' + new URLSearchParams({name: file.name || 'picture'}), file, signal); }
+    catch (error) { if (signal?.aborted) throw {problem: null}; throw {problem: 'failed'}; }
+    if (response.status === 201 || response.ok) {
+        const value = await response.json().catch(() => null);
+        if (value?.name) return value;
+        throw {problem: 'failed'};
+    }
+    const code = (await response.json().catch(() => null))?.error?.code;
+    throw {problem: code === 'too_large' ? 'tooLarge' : code === 'invalid_media' ? 'unreadable' : 'failed', code};
+}
+
+// A thumbnail that turns into a grey cell with the picture glyph when the picture cannot be shown.
+export function thumbnail(src) {
+    const cell = document.createElement('span'), image = document.createElement('img');
+    cell.className = 'fv-chat-thumb'; image.alt = ''; image.draggable = false;
+    image.addEventListener('error', () => { cell.dataset.broken = 'true'; image.remove(); cell.innerHTML = `<svg viewBox="0 0 24 24" aria-hidden="true"><path d="${ATTACH.icon}"/></svg>`; });
+    image.src = src; cell.append(image);
+    return cell;
+}
diff --git a/web/chat_data.js b/web/chat_data.js
index 592c74ac..a5bb9f9e 100644
--- a/web/chat_data.js
+++ b/web/chat_data.js
@@ -63,6 +63,7 @@ export async function previewUnsupported(response, field) {
 }
 export async function sendChatMessage(api, entry, {sessionSupported = true, stopped = () => false, previewField = null} = {}) {
     const body = {id: entry.id, text: entry.text};
+    if (entry.attachments?.length) body.attachments = entry.attachments.map(({preview, ...rest}) => rest);
     if (sessionSupported && entry.session) body.session = entry.session;
     if (entry.whiteboxPreview && previewField) body[previewField] = true;
     entry.attempts = (entry.attempts || 0) + 1;
diff --git a/web/editor.js b/web/editor.js
index ddf8e52a..1c55f65b 100644
--- a/web/editor.js
+++ b/web/editor.js
@@ -91,9 +91,15 @@ function exportErrorText(value, t) {
 }
 // The save route's own messages are internal English, so only a full disk is named.
 function saveErrorText(value, t) {
+    if (isNetworkTypeError(value)) return connectionLostText(t);
     return value?.reason === 'disk_full' ? t('Saving changes failed: the disk is full. Free some space, then export again.', '保存改动时出错：磁盘空间不够，空出空间后再导出。')
         : t('Saving changes failed. Please export again.', '保存改动时出错，请再导出一次。');
 }
+// Names the failed save for the report: the server's reason, else the HTTP status, a lost connection, or no failure recorded yet.
+function saveReason(failure) {
+    if (!failure) return 'unsaved';
+    return failure.reason || (failure.status ? 'http_' + failure.status : isNetworkTypeError(failure) ? 'network' : undefined);
+}
 function restoreGeneration(record, before, prepared) {
     const same = (a, b) => JSON.stringify(a) === JSON.stringify(b);
     const entries = value => new Map(value.tracks.flatMap(track => track.clips.map(clip => [clip.id, {track: track.id, clip}])));
@@ -491,7 +497,7 @@ export function openEditor({api, app, from = null, panel = {}, components: given
             await queue.flush();
         }
         if (disposed || closing || epoch !== projectEpoch) return false;
-        if (queue.dirty) { notify(saveErrorText(queue.failure, t), EDITOR_ISSUE, failureKind('save', {reason: queue.failure?.reason})); return false; }
+        if (queue.dirty) { notify(saveErrorText(queue.failure, t), EDITOR_ISSUE, failureKind('save', {reason: saveReason(queue.failure)})); return false; }
         return !!state.record.id;
     }
     function updateTop() {
@@ -708,7 +714,7 @@ export function openEditor({api, app, from = null, panel = {}, components: given
             const words = deleteEditText(t);
             if (jobActive()) throw {reason: 'generation_active'};
             await queue.flush(); releaseStamp();
-            if (queue.dirty) { notify(saveErrorText(queue.failure, t), EDITOR_ISSUE, failureKind('save', {reason: queue.failure?.reason})); return; }
+            if (queue.dirty) { notify(saveErrorText(queue.failure, t), EDITOR_ISSUE, failureKind('save', {reason: saveReason(queue.failure)})); return; }
             // Load the replacement before deleting, so a read failure keeps the edit open.
             const rows = (await store.projects()).filter(row => row.id !== state.record.id);
             // Most recently used: the order the edits were opened, then the last save.
diff --git a/web/health.js b/web/health.js
index 55ba9eff..ad011fe8 100644
--- a/web/health.js
+++ b/web/health.js
@@ -13,7 +13,7 @@ const t = (en, zh) => cn ? zh : en;
 // launcher.js removes this flag during setup; read it before that.
 const launchVisit = query.get('freevideo') === 'launch';
 // Every script this folder ships; a test keeps the list in step with web/*.js.
-export const MODULES = ['chat_data.js', 'phone_model.js', 'phone.js', 'lan_access.js', 'qrcodegen.js', 'agent_activity.js', 'agent_connect.js', 'agent_session.js', 'autopilot.js', 'autopilot_chat.js', 'autopilot_conversations.js', 'browser_extension.js', 'caret_rect.js', 'chain_support.js', 'capability_view.js', 'cast_generate.js', 'cast_mention.js', 'cast_overlay.js', 'cast_source.js', 'compatibility.js', 'entry_showcase.js', 'error_panel.js', 'first_run_guide.js', 'fonts.js', 'freevideo.js', 'generation_progress.js',
+export const MODULES = ['chat_data.js', 'phone_model.js', 'phone.js', 'lan_access.js', 'qrcodegen.js', 'agent_activity.js', 'agent_connect.js', 'agent_session.js', 'autopilot.js', 'autopilot_chat.js', 'chat_attachments.js', 'autopilot_conversations.js', 'browser_extension.js', 'caret_rect.js', 'chain_support.js', 'capability_view.js', 'cast_generate.js', 'cast_mention.js', 'cast_overlay.js', 'cast_source.js', 'compatibility.js', 'entry_showcase.js', 'error_panel.js', 'first_run_guide.js', 'fonts.js', 'freevideo.js', 'generation_progress.js',
     'editor.js', 'editor_actions.js', 'editor_lifecycle.js', 'editor_adjust.js', 'editor_adjust_picture.js', 'editor_picture_fields.js', 'editor_picture_session.js', 'editor_preview_picture.js', 'editor_entry.js', 'editor_expand.js', 'editor_export_dialog.js', 'editor_first_frame.js', 'editor_media_references.js', 'editor_references.js', 'editor_genbar.js', 'editor_geometry.js', 'editor_library.js', 'editor_menu.js', 'editor_model.js', 'editor_frame.js', 'editor_player.js', 'editor_room.js', 'editor_select.js', 'editor_state.js', 'editor_store.js', 'editor_tracks.js', 'editor_generation.js', 'editor_transport.js', 'editor_transport_model.js', 'editor_audio.js', 'editor_audio_model.js', 'editor_audio_notes.js', 'editor_audio_replace.js', 'editor_audio_rules.js', 'editor_speed.js', 'editor_speed_model.js', 'editor_failure.js', 'editor_header_fit.js',
     'editor_loader.js', 'editor_text_coverage.js', 'editor_text.js', 'editor_toolbar_more.js', 'editor_text_model.js', 'editor_text_layout.js', 'editor_notes.js', 'editor_icons.js', 'dialog_keys.js', 'done_notify.js', 'failure_issue.js', 'feedback.js', 'frame_sizes.js', 'health.js', 'image_commands.js', 'image_editor.js', 'image_editor_engine.js', 'image_generation.js', 'launch_visit.js', 'launcher.js', 'length_plan.js', 'lora_triggers.js', 'lora_check.js', 'library.js', 'motion.js', 'native_errors.js', 'translate_guard.js', 'output_download.js', 'port_notice.js', 'preview_scene.js',
     'progress_connection.js', 'prompt_draft.js', 'prompt_enhance.js', 'prompt_references.js', 'prompt_terms.js', 'request_assets.js', 'repair_hint.js', 'report_issue.js', 'issue_button.js', 'studio_recent.js',
@@ -344,6 +344,15 @@ function onScreen(element) {
 // canvas: inside the window, sized, and invisible. Look at what is on top of
 // its view switch; a dialog, menu, toast or this card only passes over it.
 const OVERLAYS = 'dialog, [role="dialog"], [role="menu"], .p-dialog-mask, .p-overlay-mask, .p-toast, .litecontextmenu, .fv-health';
+// Why a dialog that is open does not count as on screen, for the details of the report: which test it fails
+// and the box and window it was measured against (no page content).
+function whyOffScreen(element) {
+    const style = getComputedStyle(element), box = element.getBoundingClientRect();
+    const reason = !element.isConnected ? 'not in the page' : element.hidden ? 'hidden attribute'
+        : style.display === 'none' ? 'display: none' : style.visibility === 'hidden' ? 'visibility: hidden'
+        : Number(style.opacity) === 0 ? 'opacity: 0' : !(box.width > 0 && box.height > 0) ? 'no size' : 'outside the window';
+    return `${reason}; box ${Math.round(box.left)},${Math.round(box.top)} ${Math.round(box.width)}x${Math.round(box.height)}; window ${innerWidth}x${innerHeight}`;
+}
 // A layer hides the toolbar only if it paints over it. ComfyUI's dimming
 // backdrops and click-away layers (a dialog's shade, the help panel's) leave it
 // visible underneath, so they are not a fault.
@@ -397,7 +406,8 @@ function workspaceProblem(app) {
     // The editor shares the studio shell; either view replaces the floating toolbar.
     const studio = document.querySelector('dialog.fv-editor[open]') || document.querySelector('dialog.fv-studio[open]');
     if (studio) return onScreen(studio) ? null
-        : {text: t('The creative workspace is open but cannot be seen on the page.', '创作面板已经打开，但页面上看不到。'), advice: 'reload'};
+        : {text: t('The creative workspace is open but cannot be seen on the page.', '创作面板已经打开，但页面上看不到。'), advice: 'reload',
+            state: `${nameOf(studio)}: ${whyOffScreen(studio)}`};
     const bar = document.querySelector('.fv-view-navigation');
     if (!bar) return {text: t('The FreeVideo toolbar that opens the creative workspace is missing from the page.', '页面上没有打开创作面板的 FreeVideo 工具栏。'), advice: 'reload'};
     const cover = onScreen(bar) ? coveredBy(bar) : null;
@@ -477,7 +487,7 @@ function details(problems) {
         `ComfyUI frontend: ${window.__COMFYUI_FRONTEND_VERSION__ || 'unknown'}`,
         `FreeVideo UI: ${JSON.stringify(window.FreeVideoUI || {})}`,
         `Automatic reload: ${retried ? new Date(Number(retryRecord.at)).toISOString() : retryRecord ? 'not needed' : 'unavailable (no session storage)'}`,
-        '', 'Problems:', ...problems.flatMap(p => [`- ${p.text}`, ...(p.cover ? [`  covered by: ${p.cover}`] : [])]),
+        '', 'Problems:', ...problems.flatMap(p => [`- ${p.text}`, ...(p.cover ? [`  covered by: ${p.cover}`] : []), ...(p.state ? [`  state: ${p.state}`] : [])]),
         ...(checks.length ? ['', 'Load check:', ...checks.map(row => `- ${row}`)] : []),
         ...(cache.length ? ['', 'Browser cache:', ...cache.map(row => `- ${row}`)] : []),
         ...(before.length ? ['', 'Browser cache before the automatic reload:', ...before.map(row => `- ${row}`)] : []),