diff --git a/.github/ISSUE_TEMPLATE/feature_request.yml b/.github/ISSUE_TEMPLATE/feature_request.yml
index 93255e9f..2ab676e3 100644
--- a/.github/ISSUE_TEMPLATE/feature_request.yml
+++ b/.github/ISSUE_TEMPLATE/feature_request.yml
@@ -1,9 +1,9 @@
 name: Suggestion / 建议
-description: Tell us what would help / 说说你希望改进什么
+description: Tell us what would help / 说说希望改进什么
 body:
   - type: textarea
     id: idea
     attributes:
-      label: What would you like to do? / 你希望实现什么？
+      label: What would you like to do? / 希望实现什么？
     validations:
       required: true
diff --git a/freevideo_engine/agent/builtin_director.py b/freevideo_engine/agent/builtin_director.py
index 6b280055..f530ad54 100644
--- a/freevideo_engine/agent/builtin_director.py
+++ b/freevideo_engine/agent/builtin_director.py
@@ -369,7 +369,7 @@ def judge(shot_view, sheets, *, ask):
         return Decision('keep', shot['id'], '保留已选定的版本。')
     if shot.get('state') in ('waiting', 'shooting'):
         return Decision('wait', shot['id'], '等待这一镜的各版完成。')
-    uncertain = Decision('ask_user', shot['id'], '拿不准，请你挑一版。')
+    uncertain = Decision('ask_user', shot['id'], '拿不准，请挑一版。')
     takes = _selectable(shot)
     if not takes:
         return uncertain
diff --git a/freevideo_engine/agent/cast.py b/freevideo_engine/agent/cast.py
index 8352d5c3..1e7d4a2e 100644
--- a/freevideo_engine/agent/cast.py
+++ b/freevideo_engine/agent/cast.py
@@ -807,7 +807,7 @@ def _land_image(self, identity, generating, hidden, rows):
                     code = (row.get('error') or {}).get('code') if row else 'not_found'
                     code = code or (row['status'] if row else 'not_found')
                     detail = errors.image_failure(code, (row.get('error') or {}).get('report_id') if row else None)
-                    failure = failure or dict(code=code, job=job, error_zh=detail['reason']['zh'], message=(
+                    failure = failure or dict(code=code, job=job, error_zh=detail['reason']['zh'], error_en=detail['reason']['en'], message=(
                         'The generation was cancelled.' if row and row['status'] == 'cancelled'
                         else detail['reason']['en']), **detail)
                     continue
@@ -843,7 +843,7 @@ def _land(self, identity, generating, hidden, rows):
                 if row is None or row['status'] != 'succeeded':
                     code = (row.get('error') or {}).get('code') if row else 'not_found'
                     failure = failure or dict(code=code or row['status'], job=job,
-                        error_zh=errors.failure_reason((row.get('error') or dict(code=code or row['status'])) if row else dict(code=code)), message=(
+                        **errors.reasons((row.get('error') or dict(code=code or row['status'])) if row else dict(code=code)), message=(
                         'The generation was cancelled.' if row and row['status'] == 'cancelled'
                         else 'The reference clip did not finish.'))
                     continue
diff --git a/freevideo_engine/agent/drama.py b/freevideo_engine/agent/drama.py
index e47f12b1..a34f83ed 100644
--- a/freevideo_engine/agent/drama.py
+++ b/freevideo_engine/agent/drama.py
@@ -45,7 +45,6 @@
          'lines': [{'who': 'lin', 'line': '再见了。', 'how': 'says softly'}],
          'sound': 'The hum of the subway, rain tapping on glass.'},
     ],
-    'music': 'A sparse, slow piano figure over soft sustained strings.',
     'review': {'min_score': 7, 'retakes': 1, 'frames': 6},
     'transition': 'crossfade', 'transition_seconds': .4,
 }
@@ -258,7 +257,8 @@ def prompt(play, shot):
         scene = ' '.join(shot.get(k) or '' for k in ('framing', 'action', 'camera')) + ' ' + (play.get('style') or '')
         parts.append('No subtitles or captions.' if SHOWN_TEXT.search(scene) else 'No subtitles, no on-screen text.')
     sound = shot.get('sound') or 'Natural ambience of the scene.'
-    music = shot.get('music') or play.get('music') or 'N/A'
+    # Only a shot's own music is written; a score for the whole film is imported and laid on the sound track.
+    music = shot.get('music') or 'N/A'
     return ('integrated_multimodal_description: ' + ' '.join(p for p in parts if p) +
             '\n\noverall_soundscape: ' + sound + '\n\nnon_diegetic_music: ' + music)
 
diff --git a/freevideo_engine/agent/errors.py b/freevideo_engine/agent/errors.py
index 9ec75f8b..8f05a205 100644
--- a/freevideo_engine/agent/errors.py
+++ b/freevideo_engine/agent/errors.py
@@ -89,6 +89,7 @@
     'needs_confirmation': (409, False, 'The user must approve this operation first.', 'Wait with freevideo_wait (no id) for the user\'s answer; once approved, repeat the same call unchanged.'),
     'declined_by_user': (409, False, 'The user declined this operation.', 'Do not repeat this call; change the plan or ask the user what they want instead.'),
     'stopped_by_user': (409, False, 'The user stopped this session.', 'Wait for the user to continue the session; do not retry or start a new one.'),
+    'turn_abandoned': (409, False, 'The user gave up on this turn.', 'Stop working on it. Wait for the next message from the user; do not retry.'),
     'reshoot_limit': (409, False, 'This shot reached its generation limit.', 'Choose the best existing take, rewrite the shot, or ask the user.'),
     'production_limit': (409, False, 'This film reached its child job limit.', 'Choose existing takes or ask the user how to reduce the film.'),
     'managed_by_production': (409, False, 'This job belongs to a film and cannot be resumed on its own.', 'This take belongs to a film; reshoot the shot or wait for the user to continue.'),
@@ -131,6 +132,8 @@
                            '显卡暂时用不了，稍后再试。'),
     'worker_exit': ('Image generation stopped unexpectedly. Try again.', '图像生成意外停止了，请再试一次。'),
     'timeout': ('This image took too long and was stopped. Try again.', '这一张用时太长，已停止。请再试一次。'),
+    'disk_full': ('Saving the image failed: the disk is full. Free some space, then generate again.',
+                  '保存图片时出错：磁盘空间不够，空出空间后再生成。'),
     'generation_failed': ('Image generation ran into an error. Try again.', '图像生成出错了，请再试一次。'),
     'invalid_media': ('The reference image could not be read. Choose another one.',
                       '参考图读不出来，请换一张。'),
@@ -148,7 +151,7 @@ def image_failure(code, report_id=None):
     valid = isinstance(report_id, str) and re.fullmatch(r'[0-9a-f]{32}', report_id)
     return dict(reason=dict(en=en, zh=zh), report_id=report_id if valid and code not in (
         'busy', 'cancelled', 'invalid_request', 'invalid_media', 'invalid_prompt', 'invalid_dimensions',
-        'invalid_seed', 'too_many_images') else None)
+        'invalid_seed', 'too_many_images', 'disk_full') else None)
 
 
 # Recovery is a function of the committed task and phase, not just an error code.
@@ -171,6 +174,7 @@ def image_failure(code, report_id=None):
     'timeout': IMAGE_REASONS['timeout'][0],
     'generation_failed': IMAGE_REASONS['generation_failed'][0],
     'worker_exit': 'The image worker stopped before finishing this picture.',
+    'disk_full': IMAGE_REASONS['disk_full'][0],
     'cancelled': 'The picture job was cancelled.',
 }
 for _code, _message in IMAGE_ERRORS.items():
@@ -214,6 +218,8 @@ def recovery(code, *, kind=None, stage='request', resumable=True):
         return SESSION_GONE[1]
     if kind == 'image' and code == 'pack_missing':
         return ERRORS[code][3]
+    if kind == 'image' and code == 'disk_full':
+        return 'Tell the user to free disk space, then submit the picture again with a new ref.'
     if kind == 'image' and code == 'no_memory':
         return 'Lower the picture dimensions or use fewer reference pictures, then submit with a new ref.'
     if kind in RECOVERY and stage == 'terminal':
@@ -282,6 +288,7 @@ def safe_details(value):
     'join_failed': '片段已生成，但拼接成片时出错',
     'invalid_output': '生成的结果无法读取或保存',
     'storage_unavailable': '任务记录或素材无法写入磁盘',
+    'disk_full': '磁盘空间不够，图片没能保存',
     'submit_interrupted': '提交生成时服务中断，尚未确认任务是否开始',
     'timeout': '生成等了太久没有结果',
     'invalid_media': '参考素材无法读取',
@@ -299,6 +306,54 @@ def safe_details(value):
     'generation_failed': '生成没有完成',
     'internal_error': '没有完成',
 }
+FAILURE_REASONS_EN = {
+    'lan_unavailable': 'Phone access requires FreeVideo installed with the launcher',
+    'out_of_memory': 'Graphics or system memory ran out during generation',
+    'insufficient_memory': 'There is not enough free graphics or system memory for this generation',
+    'no_memory': 'There is not enough free graphics or system memory for this picture',
+    'ram_space': 'System memory ran out while loading the model',
+    'model_missing': 'A required model is not installed yet',
+    'pack_missing': 'The image model is not installed yet',
+    'not_ready': 'The generation environment is not installed yet',
+    'cancelled': 'Generation was stopped',
+    'stopped_by_user': 'Generation was stopped',
+    'interrupted': 'Generation was interrupted',
+    'worker_exit': 'Picture generation was interrupted',
+    'closed': 'The picture generation service has closed',
+    'image_unavailable': 'The picture generation service has not started yet',
+    'vlm_unavailable': 'The model that reads pictures has not started yet',
+    'vlm_failed': 'The picture could not be read',
+    'vlm_invalid_json': 'The reply could not be read',
+    'context_length': 'The description and reference pictures are too long',
+    'dependency_failed': 'A task this one depends on did not finish',
+    'lost': 'The generation service no longer has the result of this task',
+    'unknown_job': 'The picture generation service cannot find this task',
+    'not_found': 'This task or media file cannot be found',
+    'source_missing': 'The previous video needed to continue is no longer available',
+    'chain_missing': 'The generation record needed to continue is no longer available',
+    'chain_unavailable': 'The part-by-part generation service has not started yet',
+    'continuation_unavailable': 'The current generation mode cannot continue a video',
+    'join_failed': 'The parts were generated, but joining them into one video failed',
+    'invalid_output': 'The result could not be read or saved',
+    'storage_unavailable': 'The task record or media could not be written to disk',
+    'disk_full': 'Saving the image failed: the disk is full',
+    'submit_interrupted': 'The service stopped during submission, so it is not confirmed whether the task started',
+    'timeout': 'Generation took too long without a result',
+    'invalid_media': 'The reference media could not be read',
+    'clip_unreadable': 'The generated character reference video could not be read',
+    'invalid_request': 'The generation settings are outside the supported range',
+    'invalid_drama': 'The short film settings are invalid',
+    'reshoot_limit': 'This shot has reached its generation limit',
+    'production_limit': 'This short film has reached its task limit',
+    'export_failed': 'Exporting the video failed',
+    'invalid_prompt': 'The picture description is empty or too long',
+    'invalid_dimensions': 'The picture size is outside the supported range',
+    'too_many_images': 'There are more reference pictures than supported',
+    'chain_too_long': 'The continued video has more parts than supported',
+    'chain_rejected': 'The generation service cannot use these settings',
+    'generation_failed': 'Generation did not finish',
+    'internal_error': 'Did not finish',
+}
 FAILURE_MESSAGES_ZH = {
     'The previous video is too short to continue; it needs to be at least 1.7 s long.': '接续的上一段视频太短，至少需要 1.7 秒',
     'The previous video has no usable continuation data. Generate it again with this version of FreeVideo.': '上一段视频缺少可用的接续数据',
@@ -309,14 +364,18 @@ def safe_details(value):
 }
 
 
-def failure_reason(error):
-    if isinstance(error.get('error_zh'), str) and error['error_zh'].strip():
-        return error['error_zh']
+def _failure_line(error):
     lines = str(error.get('message') or '').strip().splitlines()
     line = lines[0] if lines else ''
     if line.startswith('FreeVideo generation failed') and len(lines) > 1:
         line = lines[1].strip()
-    line = re.sub(r'^(?:[\w.]*Error|[\w.]*Exception):\s*', '', line)
+    return re.sub(r'^(?:[\w.]*Error|[\w.]*Exception):\s*', '', line)
+
+
+def failure_reason(error):
+    if isinstance(error.get('error_zh'), str) and error['error_zh'].strip():
+        return error['error_zh']
+    line = _failure_line(error)
     if line in FAILURE_MESSAGES_ZH:
         return FAILURE_MESSAGES_ZH[line]
     if line.startswith(('CUDA out of memory', 'MPS backend out of memory')):
@@ -324,8 +383,24 @@ def failure_reason(error):
     return FAILURE_REASONS_ZH.get(error.get('code'), '没有完成')
 
 
+# The same lookup in English for the page: the specific sentences are already written for people.
+def failure_reason_en(error):
+    if isinstance(error.get('error_en'), str) and error['error_en'].strip():
+        return error['error_en']
+    line = _failure_line(error)
+    if line in FAILURE_MESSAGES_ZH:
+        return line
+    if line.startswith(('CUDA out of memory', 'MPS backend out of memory')):
+        return FAILURE_REASONS_EN['out_of_memory']
+    return FAILURE_REASONS_EN.get(error.get('code'), 'Did not finish')
+
+
+def reasons(error):
+    return dict(error_zh=failure_reason(error), error_en=failure_reason_en(error))
+
+
 def with_reason(error):
-    return dict(error, error_zh=failure_reason(error)) if error else error
+    return dict(error, **reasons(error)) if error else error
 
 
 def build(code, *, details=None, request=None, validation=False, message=None, next_action=None,
diff --git a/freevideo_engine/agent/guide.py b/freevideo_engine/agent/guide.py
index e5dabc94..1ac71bdb 100644
--- a/freevideo_engine/agent/guide.py
+++ b/freevideo_engine/agent/guide.py
@@ -52,7 +52,21 @@
     ],
     'sound': [
         'overall_soundscape: 1-4 sentences of ambience, action sounds and breaths; N/A only for silence.',
-        'non_diegetic_music: 1-3 sentences of instruments, tempo and dynamics, or N/A.',
+        'non_diegetic_music: 1-3 sentences of instruments, tempo and dynamics, or N/A. By default a long video or a '
+        'film of several shots has no music: write N/A in every shot, so each has only ambience, action sounds and '
+        'dialogue, and do not add music on your own. Whatever the case, never write a different non_diegetic_music '
+        'for each shot: the shots are generated separately, so the music would not be continuous and would jump at '
+        'every cut. Write music in a shot only when the user asks for it there.',
+        'When the user wants music over the film, keep every shot at N/A and lay one score on one sound track after '
+        'the cut is joined. If the user gives a music file, import it. If there is none, compose it yourself: write '
+        'a program that synthesizes one audio file (WAV, using the standard wave module and numpy or whatever is '
+        'already installed, nothing to install) as long as the finished cut, with its sections, changes and ending '
+        'on the cut points. Import it with POST /imports (MCP freevideo_import_media, CLI imports add), wait until it '
+        'is ready, and insert import:<id> at frame 0 of the project. There is no volume or fade operation, so build '
+        'them into the file: keep the music clearly quieter than the dialogue and ambience, lower it further while '
+        'someone speaks (take the line times from your own screenplay and shot lengths, or run a listen job in align '
+        'mode to time them), and fade it in at the start and out at the end. The audio goes onto a free sound track by '
+        'itself; read the project afterwards and check that one clip covers the whole film.',
     ],
     'references': [
         'Tags are numbered per kind in reference order: <Picture 1>, <Picture 2>, <Video 1>, <Audio 1>. '
@@ -355,6 +369,7 @@
     ('POST', '/sessions/{id}/events'): '向用户发送消息或计划；操作步骤由服务端记录',
     ('POST', '/sessions/{id}/stop'): '只有页面用户可以停止工作会话',
     ('POST', '/sessions/{id}/resume'): '只有页面用户可以继续工作会话',
+    ('POST', '/sessions/{id}/stage'): '说明你正在做回答的哪个阶段；页面用自己的话显示，你开口或开始任务后自动清除',
     ('GET', '/confirmations'): '查看哪些 Agent 操作需要等待用户确认',
     ('PUT', '/confirmations'): '只有页面用户可以修改需要确认的操作',
 }
@@ -480,7 +495,7 @@ def connection_guide(root, base, lang=None, offers=None, resumed=False):
                      '连接后先调一次等待接口 GET ' + base + PREFIX + '/inbox?wait=50（MCP 用不带 id 的 freevideo_wait），'
                      '回应里的 session 就是用户现在所在的对话（用户新开或换到另一段对话时它会变，话总是发到最近一次等待给的那个 session）；往 POST ' + base + PREFIX + '/sessions/{session}/events 发 '
                      '{"type":"message","text":"…"}（MCP 用 freevideo_session_post）。在这里先只说一句话，不超过 20 个字，'
-                     '不列条目、不举例（页面上已经有例句），例如“连上了。告诉我想做什么视频。”'
+                     '不列条目、不举例（页面上已经有例句），例如“连上了。请说明想做什么视频。”'
                      '你自己窗口里的那段介绍照常说，但它不是这一轮的结尾：说完紧接着去调等待接口。'
                      '然后守在这段对话里：再调等待接口，它最多约 50 秒返回；"again": true 表示用户还没说话，'
                      '马上再调下一次，一次接一次，不限次数。没有消息不是结束的理由：不要结束这一轮回复，'
@@ -498,7 +513,7 @@ def connection_guide(root, base, lang=None, offers=None, resumed=False):
                      'without id). Its "session" is the conversation the user is in (it changes when the user opens or returns to another one; always post to the session the latest wait named); POST ' + base + PREFIX + '/sessions/{session}/events '
                      'with {"type":"message","text":"..."} (MCP: freevideo_session_post). Say only one short sentence '
                      'there first, at most 12 words, no list and no examples (the page already shows examples), for example '
-                     '"Connected. Tell me what video you want.", in the language the user writes to you in. Give the introduction in your own window as usual, but it '
+                     '"Connected. Describe the video you want.", in the language the user writes to you in. Give the introduction in your own window as usual, but it '
                      'does not end your turn: right after it, call the wait. Then stay in this conversation: call the wait again; it answers within about 50 seconds, and '
                      '"again": true means the user has not spoken yet, so call it again at once, time after time, with no '
                      'limit. Nothing new is not a reason to finish: do not end your reply and do not stop to ask the '
@@ -515,7 +530,7 @@ def connection_guide(root, base, lang=None, offers=None, resumed=False):
                      'Stop only when the user clearly tells you to.',
         rules=RULES_ZH if zh else RULES,
         fallback='如果你无法执行命令、读取本地文件或连接到这台电脑，请如实告诉用户：'
-                 '“我连接不到你电脑上的 FreeVideo，请改用安装在这台电脑上的 AI 助手。”'
+                 '“连接不到您电脑上的 FreeVideo，请改用安装在这台电脑上的 AI 助手。”'
                  '用中文说明；如果用户改用其他语言，就跟随用户。不要另装程序来绕过这个限制。' if zh else
                  'If you cannot run commands, read local files, or reach this computer, tell the user honestly '
                  'in their language: I cannot connect to FreeVideo on your computer; please use an AI assistant '
diff --git a/freevideo_engine/agent/jobs.py b/freevideo_engine/agent/jobs.py
index 8328b1c6..aad1a92e 100644
--- a/freevideo_engine/agent/jobs.py
+++ b/freevideo_engine/agent/jobs.py
@@ -633,16 +633,16 @@ def finish(self, row, status, *, result=None, error=None, save=True):
             report_id = error.get('report_id')
             if code == 'invalid_json' and row['kind'] in ('vlm', 'enhance'):
                 code = 'vlm_invalid_json'
-            reason_zh = errors.failure_reason(dict(error, code=code))
+            reason_zh, reason_en = errors.failure_reason(dict(error, code=code)), errors.failure_reason_en(dict(error, code=code))
             if code != error.get('code') or not error.get('request_id') or row['kind'] == 'image':
                 error = errors.build(code, details=error.get('details'),
                     request=row.get('request_id'), action_required=code != 'cancelled',
                     kind=row['kind'], stage='terminal', resumable=not row['spec'].get('refine_from'),
                     report_id=error.get('report_id'))
                 if row['kind'] == 'image':
-                    reason_zh = error['error_zh']
+                    reason_zh, reason_en = error['error_zh'], error['error_en']
         if error is not None:
-            error = dict(error, request_id=row.get('request_id') or error['request_id'], error_zh=reason_zh)
+            error = dict(error, request_id=row.get('request_id') or error['request_id'], error_zh=reason_zh, error_en=reason_en)
             if row['kind'] != 'image' and isinstance(report_id, str) and re.fullmatch(r'[a-f0-9]{32}', report_id):
                 error['report_id'] = report_id
             if error.get('next'):
@@ -1123,7 +1123,7 @@ def view(self, row, *, detail=True):
                     details=result['error'].get('details'), request=row.get('request_id'),
                     action_required=result['error']['code'] != 'cancelled', report_id=row['error'].get('report_id'))
             else:
-                result['error'] = dict(result['error'], error_zh=errors.failure_reason(row['error']))
+                result['error'] = dict(result['error'], **errors.reasons(row['error']))
                 report_id = row['error'].get('report_id')
                 if isinstance(report_id, str) and re.fullmatch(r'[a-f0-9]{32}', report_id):
                     result['error']['report_id'] = report_id
@@ -1808,7 +1808,7 @@ def generation_failure(self, row, chain):
             next_action = 'Use the alternatives in details to lower size or duration, then POST /freevideo/agent/v1/jobs.'
         failure = dict(chain_error(code, errors.ERRORS[code][2], next_action, details=[detail],
                            retryable=code != 'insufficient_memory').view(),
-                    error_zh=errors.failure_reason(dict(code=code, message=message)))
+                    **errors.reasons(dict(code=code, message=message)))
         if isinstance(report_id, str) and re.fullmatch(r'[a-f0-9]{32}', report_id):
             failure['report_id'] = report_id
         return failure
diff --git a/freevideo_engine/agent/production.py b/freevideo_engine/agent/production.py
index 37b8b207..ca20154c 100644
--- a/freevideo_engine/agent/production.py
+++ b/freevideo_engine/agent/production.py
@@ -151,8 +151,12 @@ def plan(request, *, fields=True):
             preview.append(dict(id=shot['id'], continues=shot.get('continue', False),
                                 takes=shot.get('takes', count), prompt=text))
     _invalid(problems)
+    notice = {}
+    if play_ok and play.get('music'):
+        notice['notice'] = ('screenplay.music is not written into the shots. Music for the whole film is composed or '
+                            'provided as an audio file, imported, and laid on one sound track after the cut is joined.')
     return dict(request=dict(screenplay=play, takes_per_shot=count, director=director),
-                preview=dict(shots=preview, total=total))
+                preview=dict(shots=preview, total=total, **notice))
 
 
 def corrected_prompt(play, shot, note=None):
diff --git a/freevideo_engine/agent/productions.py b/freevideo_engine/agent/productions.py
index fcfb811e..6ee2c115 100644
--- a/freevideo_engine/agent/productions.py
+++ b/freevideo_engine/agent/productions.py
@@ -145,7 +145,7 @@ def _safe_error(error):
     rows = [_safe_row(row) for row in details] if isinstance(details, list) and details else []
     if rows and all(rows):
         result.update(details=rows, message='; '.join(row['message'] for row in rows))
-    return dict(result, error_zh=errors.failure_reason(error))
+    return dict(result, **errors.reasons(error))
 
 
 def failure(view):
@@ -177,7 +177,12 @@ def __init__(self, jobs):
         jobs.parent_allowed = self.allows
 
     def load(self):
-        for path in sorted(self.directory.glob('*.json')):
+        try:
+            paths = sorted(self.directory.glob('*.json'))
+        except OSError:
+            logging.getLogger(__name__).warning('Could not list saved films; starting without them')
+            return
+        for path in paths:
             try:
                 record = json.loads(path.read_text(encoding='utf-8'))
                 state = record['state']
@@ -209,8 +214,13 @@ def load(self):
                                                         record['stalled']['code'])
                 self.records[state['id']] = record
             except (ValueError, KeyError, TypeError, AttributeError, IndexError, RecursionError, OSError):
-                path.replace(path.with_name(path.name + '.bad-' + uuid.uuid4().hex))
-                logging.getLogger(__name__).warning('Isolated an invalid production receipt; other films remain available')
+                try:
+                    path.replace(path.with_name(path.name + '.bad-' + uuid.uuid4().hex))
+                except OSError:
+                    # A receipt that cannot be moved aside is left in place; the agent still starts.
+                    logging.getLogger(__name__).warning('Skipped an invalid production receipt that could not be renamed')
+                else:
+                    logging.getLogger(__name__).warning('Isolated an invalid production receipt; other films remain available')
 
     def _stalled(self, error):
         code = error.code if isinstance(getattr(error, 'code', None), str) and error.code in errors.ERRORS else 'internal_error'
@@ -404,6 +414,24 @@ def _take_estimate(self, state, shot, take, item=None):
         except Exception:  # Timing is optional; a forecast failure must not reject a film or job.
             return dict(seconds=None, basis=None, reason='unavailable')
 
+    def _shape(self, state, only=None):
+        """Size and length of what a confirmation asks to make: the first shot's canvas, every shot's seconds added up."""
+        try:
+            rows = []
+            for shot in state['shots']:
+                take = next((t for s, t in self._takes(state) if s is shot and not t['stale'] and t['state'] != 'cancelled'), None)
+                if take is None or only not in (None, shot['id']):
+                    continue
+                item = deepcopy(take['spec'])
+                if shot['continues']:
+                    item['continue_from'] = dict(from_job='j000000000000')
+                rows.append(spec.normalize(item, port=self.jobs.port))
+            if not rows:
+                return {}
+            return dict(width=rows[0]['width'], height=rows[0]['height'], seconds=round(sum(row['seconds'] for row in rows), 3))
+        except Exception:  # The card is complete without it; a failure here must not reject a film.
+            return {}
+
     def _join_estimate(self, state, previous=None):
         """Use only successful local joins with the same input geometry and transition."""
         shapes = []
@@ -580,6 +608,7 @@ def start(self, request, *, by='agent', ref=None, local=True, client=None):
             self._prepare_estimates(record)
             if self.gate is not None:
                 self.gate('start', by=by, estimate_seconds=self._progress(record)['estimate_seconds'],
+                          shape=self._shape(state),
                           refs={},  # the film gets its ID only when the approved call repeats
                           question='Start shooting %s (%d shots)?' % (_quoted(state['title']), len(state['shots'])))
             result = self._view(record)
@@ -628,7 +657,7 @@ def contextual(error, join=False):
         value['progress'] = self._progress(record)
         value['over_budget'] = self._elapsed_budget(value['progress'])
         error, _ = failure(value)
-        value['error_zh'] = errors.failure_reason(error) if error else None
+        value['error_zh'], value['error_en'] = (errors.failure_reason(error), errors.failure_reason_en(error)) if error else (None, None)
         return value
 
     def get(self, identity, *, since=None, wait=0, compact=True):
@@ -654,7 +683,7 @@ def _reply(self, identity, since, compact):
             join = next((j for j in state['joins'] if j['ref'] == state['current_join']), None)
             error, _ = failure(dict(stalled=record.get('stalled'), join=join, shots=state['shots']))
             if error:
-                value['error_zh'] = errors.failure_reason(error)
+                value.update(errors.reasons(error))
             return value
         return dict(self._view(record), changed=changed)
 
@@ -704,7 +733,7 @@ def list(self, *, status='all', limit=50):
                 shot = next((s for s in view['shots'] if s['chosen'] is None), view['shots'][-1])
                 result.append(dict(id=view['id'], title=view['title'], status=view['status'], revision=view['revision'],
                     shot=shot['index'], shots=len(view['shots']), next=view['todo'][0] if view['todo'] else None,
-                    paused=view['paused'], stalled=view['stalled'], session=record.get('session'), error_zh=view['error_zh']))
+                    paused=view['paused'], stalled=view['stalled'], session=record.get('session'), error_zh=view['error_zh'], error_en=view['error_en']))
             return result
 
     def _mutate(self, action, identity, *, by, ref=None, client=None, **values):
@@ -734,7 +763,7 @@ def _mutate(self, action, identity, *, by, ref=None, client=None, **values):
                 if self.gate is not None:
                     shot = next(s for s in state['shots'] if s['id'] == values['shot'])
                     self.gate('reshoot', by=by, estimate_seconds=sum(estimates) if known else None,
-                              refs=dict(drama=identity, shot=shot['id']),
+                              refs=dict(drama=identity, shot=shot['id']), shape=self._shape(state, shot['id']),
                               question='Reshoot shot %d of %s?' % (shot['index'], _quoted(state['title'])))
             if ref:
                 draft['receipts'].append(dict(key=ref, digest=digest, result=result,
diff --git a/freevideo_engine/agent/routes.py b/freevideo_engine/agent/routes.py
index ca03c42d..a7766712 100644
--- a/freevideo_engine/agent/routes.py
+++ b/freevideo_engine/agent/routes.py
@@ -912,7 +912,8 @@ def chat_fold(store, sid):
                 elif row['type'] == 'confirm':
                     if row['state'] == 'asked':
                         item = dict(base, id=row['id'], question=row.get('question'),
-                                    estimate_seconds=row.get('estimate_seconds'), state='asked')
+                                    estimate_seconds=row.get('estimate_seconds'), state='asked',
+                                    **{key: row[key] for key in ('width', 'height', 'seconds') if row.get(key) is not None})
                         fold['asked'][row['id']] = item
                         fold['items'].append(item)
                     elif row['id'] in fold['asked']:
@@ -995,6 +996,8 @@ def read():
                                  or (task_view.get('result') or {}).get('report_id'))
                     value['failures'][identity] = dict(id=identity, kind=kind, status='failed',
                         label=entity.get('title') or entity.get('name'), error=errors.with_reason(error), report_id=report_id)
+            value['abandoned'] = store.abandoned(sid)
+            value['stage'] = store.stage(sid)
             value['current'] = current['id'] if current else None
             if current and current['id'] not in value['jobs']:
                 value['jobs'][current['id']] = chat_card(current, fold)
@@ -1047,10 +1050,25 @@ def 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 reply(done[1], done[0])
 
+    @routes.post('/freevideo/agent/chat/abandon')
+    @guarded
+    async def chat_abandon(request):
+        value = await body(request)
+        def write():
+            store = work_sessions()
+            sid = chat_session(store, value.get('session'))
+            if sid is None:
+                raise JobError('not_found', 'No conversation')
+            if not store.abandoned(sid):
+                store.append(sid, dict(type='control', action='abandoned'), by='user')
+            return dict(abandoned=True)
+        return reply(await asyncio.to_thread(write))
+
     @routes.post('/freevideo/agent/chat/confirm')
     @guarded
     async def chat_confirm(request):
@@ -1116,6 +1134,7 @@ def write():
         done = await asyncio.to_thread(write)
         if done is None:
             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 try again.',
                                         request_id=errors.request_id())}, 409)
         return reply(done[1], done[0])
@@ -1493,6 +1512,8 @@ async def session_post(request):
                 raise JobError('forbidden', 'Only the page user may stop or continue')
             if value['action'] == 'stopped':
                 return reply(await stop_session(sid), 201)
+        if by == 'agent' and value['type'] in ('message', 'plan') and sessions.abandoned(sid):
+            raise JobError('turn_abandoned', 'The user gave up on this turn')
         if by == 'agent' and value['type'] == 'message' and isinstance(value.get('text'), str):
             value['text'] = chat_text(value['text'])
         client = session_client(request, by)
@@ -1533,6 +1554,19 @@ async def session_resume(request):
         return reply(await asyncio.to_thread(sessions.append, request.match_info['id'],
             dict(type='control', action='resumed'), by='user'))
 
+    @routes.post(PREFIX + '/sessions/{id}/stage')
+    @guarded
+    async def session_stage(request):
+        if session_caller(request) == 'user':
+            raise JobError('forbidden', 'Only the agent reports a stage')
+        sid, value = request.match_info['id'], await body(request)
+        def write():
+            if sessions.abandoned(sid):
+                raise JobError('turn_abandoned', 'The user gave up on this turn')
+            sessions.report_stage(sid, value['stage'])
+            return dict(stage=value['stage'])
+        return reply(await asyncio.to_thread(write))
+
     def publish_media(work, suffix):
         from . import media
         with tempfile.TemporaryDirectory(prefix='.agent-stage-', dir=port.input_root) as temporary:
diff --git a/freevideo_engine/agent/schema.py b/freevideo_engine/agent/schema.py
index 5a88c66d..eb07103d 100644
--- a/freevideo_engine/agent/schema.py
+++ b/freevideo_engine/agent/schema.py
@@ -50,6 +50,9 @@ def enum(values, **kw):
 
 JOB = string(13, pattern=r'^j[0-9a-f]{12}$')
 SESSION = string(13, pattern=r'^s[0-9a-f]{12}$')
+# What an agent may say it is doing before it answers. The page owns the words for each key.
+STAGES = ('reading', 'scripting', 'storyboarding', 'prompting', 'checking', 'assembling')
+STAGE = enum(STAGES, description='The stage of your answer. The page shows its own wording for it.')
 PRODUCTION = string(13, pattern=r'^d[0-9a-f]{12}$')
 CAST = string(13, pattern=r'^c[0-9a-f]{12}$')
 PROJECT = string(34, pattern=r'^e-[0-9a-f]{32}$')
@@ -439,10 +442,11 @@ def route(summary, body=None, query=None, **kw):
             ('/chat/video', 'GET', None), ('/chat/list', 'GET', None), ('/chat/new', 'POST', EMPTY),
             ('/chat/delete', 'POST', obj({'session': SESSION, 'stop': {'type': 'boolean'}}, ('session',))),
             ('/chat', 'POST', obj({'text': string(8000), 'id': CHAT_ID, 'session': SESSION, 'whitebox': {'type': 'boolean'}}, ('text', 'id'))),
+            ('/chat/abandon', 'POST', obj({'session': SESSION})),
             ('/chat/confirm', 'POST', obj({'id': string(12, pattern=r'^cf[0-9a-f]{10}$'), 'approve': {'type': 'boolean'},
                                            'session': SESSION}, ('id', 'approve')))):
         routes[(method, '/freevideo/agent' + suffix)] = route('Studio Agent controls', body=body, auth=False, local=True,
-            lan=suffix in ('/chat', '/chat/now', '/chat/confirm', '/chat/list', '/chat/new', '/chat/video',
+            lan=suffix in ('/chat', '/chat/now', '/chat/confirm', '/chat/abandon', '/chat/list', '/chat/new', '/chat/video',
                            '/activity', '/activity/cancel'))
     # These existing ComfyUI handlers are registered outside the Agent dispatcher.
     # /chain is a creation POST; monitoring reads /chains and /chain/{id}.
@@ -500,9 +504,9 @@ def route(summary, body=None, query=None, **kw):
 # The fields are checked when planning; the shape is told here so an agent does not guess it.
 SCREENPLAY = {'type': 'object', 'description': (
     'characters: {key: {card: cast card ID, or name, look, voice, language}}; '
-    'shots: [{id, seconds, action or prompt, camera, framing, sound, takes, seed, '
+    'shots: [{id, seconds, action or prompt, camera, framing, sound, music (N/A unless the user asks for music in this shot), takes, seed, '
     'cast: [character keys], lines: [{who: character key, line, how}], '
-    'continue: true to carry on from the previous shot}]; also title, style, aspect, quality, music.')}
+    'continue: true to carry on from the previous shot}]; also title, style, aspect, quality; a whole-film score is not written here but imported onto a sound track.')}
 
 
 def drama_schema(name):
@@ -580,7 +584,7 @@ def cast_view():
         description=nullable(string(1000)), revision=number(1, integer=True),
         images=array(picture, 0, 6), candidates=array(picture, 0, 6),
         generating=nullable(generating), generated=nullable(generated),
-        generate_error=nullable(obj(dict(code=string(), job=JOB, message=string(), error_zh=string()), ('code', 'job', 'message')))),
+        generate_error=nullable(obj(dict(code=string(), job=JOB, message=string(), error_zh=string(), error_en=string()), ('code', 'job', 'message')))),
         ('id', 'kind', 'name', 'revision', 'images', 'candidates', 'generating')), additionalProperties=True)
 
 
@@ -800,9 +804,10 @@ def image_schema(name):
     '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.'),
+    'stage': ('POST', '/sessions/{id}/stage', 'Say which stage of the answer you are working on; the page shows its own words for it. Clears when you speak or start a task.'),
 }
 SESSION_TOOLS = dict(create='freevideo_session_create', list='freevideo_sessions',
-                     events='freevideo_session_events', post='freevideo_session_post')
+                     events='freevideo_session_events', post='freevideo_session_post', stage='freevideo_report_stage')
 
 
 CONFIRM_DEFAULTS = dict(start=True, reshoot_seconds=300)
@@ -846,13 +851,14 @@ def session_event(*, stored=False, agent_only=False):
             why=text, error=enum(ERRORS)), ('type', 'tool', 'summary', 'state', *(['id'] if stored else []))),
             obj(dict(type={'const': 'confirm'}, id=string(12, pattern=r'^cf[0-9a-f]{10}$'),
                 question=text, estimate_seconds=number(0), state=enum(('asked', 'approved', 'declined')),
+                width=number(0, integer=True), height=number(0, integer=True), seconds=number(0),
                 action=enum(('start', 'reshoot')), refs=obj(dict(drama=PRODUCTION, shot=SLUG)),
                 whitebox=obj(dict(work=SAVED, seconds=number(0), summary=text), ('work',),
                              description='The rendered white-model preview the user is asked to approve.'),
                 digest=string(64, pattern=r'^[0-9a-f]{64}$')),
                 ('type', 'id', 'state'), **{'if': {'properties': {'state': {'const': 'asked'}}},
                                         'then': {'required': ['question']}}),
-            obj(dict(type={'const': 'control'}, action=enum(('stopped', 'resumed', 'renamed')),
+            obj(dict(type={'const': 'control'}, action=enum(('stopped', 'resumed', 'renamed', 'abandoned')),
                 title=dict(type='string', maxLength=120)), ('type', 'action'),
                 **{'if': {'properties': {'action': {'const': 'renamed'}}},
                    'then': {'required': ['title']}, 'else': {'not': {'required': ['title']}}})]
@@ -874,6 +880,8 @@ def session_schema(name):
                         limit=number(1, 500, integer=True, default=500)))
     if name == 'post':
         return session_event()
+    if name == 'stage':
+        return obj(dict(stage=STAGE), ('stage',))
     return EMPTY
 
 
@@ -1432,7 +1440,7 @@ def openapi():
                            'description': 'Image diagnostic report ID, or null when no report exists.'},
                 request_id=string(), docs=string(), retry_after_seconds=number(0), card={'type': 'object'}, job=JOB,
                 project={'type': 'object'}, current_revision=number(1, integer=True),
-                error_zh=string()),
+                error_zh=string(), error_en=string()),
                 ('code', 'message', 'details', 'retryable', 'request_id', 'docs'))}))
 
 
diff --git a/freevideo_engine/agent/sessions.py b/freevideo_engine/agent/sessions.py
index adf49467..064402b5 100644
--- a/freevideo_engine/agent/sessions.py
+++ b/freevideo_engine/agent/sessions.py
@@ -5,6 +5,7 @@
 from copy import deepcopy
 from datetime import datetime, timezone
 import json
+import logging
 import os
 from pathlib import Path
 import secrets
@@ -47,6 +48,9 @@ def _line(value):
     return json.dumps(value, ensure_ascii=False, allow_nan=False, separators=(',', ':')).encode('utf-8')
 
 
+STAGE_SECONDS = 120
+
+
 class Sessions:
     def __init__(self, root: Path, *, clock=time.time, max_events=20000):
         self.root = Path(root)
@@ -60,6 +64,8 @@ def __init__(self, root: Path, *, clock=time.time, max_events=20000):
         self._initial_seq = {}
         self._readers = {}
         self._user_seq = {}  # Last event the user wrote; unread() skips the file when nothing is newer.
+        self._turn = {}  # [last message the user wrote, the one they gave up on]: a turn is given up until they speak again.
+        self._stage = {}  # What the agent last said it was doing: (key, when). Not kept across restarts.
         # When each session began or its user last acted in it (spoke, answered), and the user's first words
         # there: the conversation the user is in, and what the page calls each one in its history.
         self._acted = {}
@@ -82,44 +88,65 @@ def _get(self, sid):
         return self._meta[sid]
 
     def _load(self):
-        for path in sorted(self.root.glob('s*.jsonl')):
+        try:
+            paths = sorted(self.root.glob('s*.jsonl'))
+        except OSError:
+            logging.getLogger(__name__).warning('Could not list saved sessions; starting without them')
+            return
+        for path in paths:
             sid = path.stem
             if schema.issues(schema.SESSION, sid) or path.is_symlink():
                 continue
-            with path.open('rb') as stream:
-                first = stream.readline()
+            try:
+                self._load_file(path, sid)
+            except OSError:
+                # One unreadable file must not stop the agent from starting.
+                for table in (self._meta, self._ids, self._clients, self._acted, self._said, self._initial_seq, self._user_seq):
+                    table.pop(sid, None)
+                self._chosen.discard(sid)
+                self._partial.discard(sid)
+                logging.getLogger(__name__).warning('Skipped an unreadable session file')
+
+    def _load_file(self, path, sid):
+        with path.open('rb') as stream:
+            first = stream.readline()
+            try:
+                meta = json.loads(first)
+                if not first.endswith(b'\n') or meta.get('session') != sid:
+                    return
+                _check(schema.session_metadata(), meta)
+            except (ValueError, UnicodeDecodeError, SessionError, AttributeError):
+                return
+            self._meta[sid] = dict(id=sid, title=meta['title'], created_at=meta['created_at'],
+                updated_at=meta['created_at'], client=meta['client'], last_seq=0, stopped=False)
+            self._ids[sid] = {}
+            self._clients[sid] = {meta['client']} if meta['client'] else set()
+            self._acted[sid] = (meta['created_at'], 1)
+            if meta.get('by') == 'user':
+                self._chosen.add(sid)
+            for row in stream:
+                if not row.endswith(b'\n'):
+                    self._partial.add(sid)
+                    break
                 try:
-                    meta = json.loads(first)
-                    if not first.endswith(b'\n') or meta.get('session') != sid:
+                    event = json.loads(row)
+                    _check(schema.session_event(stored=True), event)
+                    if event['seq'] != self._meta[sid]['last_seq'] + 1:
                         continue
-                    _check(schema.session_metadata(), meta)
-                except (ValueError, UnicodeDecodeError, SessionError, AttributeError):
+                except (ValueError, UnicodeDecodeError, SessionError):
+                    # Also skips an interrupted tail separated by a later append.
                     continue
-                self._meta[sid] = dict(id=sid, title=meta['title'], created_at=meta['created_at'],
-                    updated_at=meta['created_at'], client=meta['client'], last_seq=0, stopped=False)
-                self._ids[sid] = {}
-                self._clients[sid] = {meta['client']} if meta['client'] else set()
-                self._acted[sid] = (meta['created_at'], 1)
-                if meta.get('by') == 'user':
-                    self._chosen.add(sid)
-                for row in stream:
-                    if not row.endswith(b'\n'):
-                        self._partial.add(sid)
-                        break
-                    try:
-                        event = json.loads(row)
-                        _check(schema.session_event(stored=True), event)
-                        if event['seq'] != self._meta[sid]['last_seq'] + 1:
-                            continue
-                    except (ValueError, UnicodeDecodeError, SessionError):
-                        # Also skips an interrupted tail separated by a later append.
-                        continue
-                    self._fold(sid, event)
-                self._initial_seq[sid] = self._meta[sid]['last_seq']
+                self._fold(sid, event)
+            self._initial_seq[sid] = self._meta[sid]['last_seq']
 
     def _fold(self, sid, event):
         meta = self._meta[sid]
         meta.update(last_seq=event['seq'], updated_at=event['at'])
+        turn = self._turn.setdefault(sid, [0, 0])
+        if event['by'] == 'user' and event['type'] == 'message':
+            turn[0] = event['seq']
+        elif event['type'] == 'control' and event['action'] == 'abandoned':
+            turn[1] = turn[0]
         if event['by'] == 'user':
             self._user_seq[sid] = event['seq']
             # No counter here: the order must come out the same when the files are read again after a
@@ -195,7 +222,7 @@ def delete(self, sid) -> None:
             self._path(sid).unlink(missing_ok=True)
             self._closing.discard(sid)
             waiters = self._waiters.pop(sid, ())
-            for kept in (self._meta, self._ids, self._clients, self._initial_seq, self._user_seq, self._acted, self._said):
+            for kept in (self._meta, self._ids, self._clients, self._initial_seq, self._user_seq, self._turn, self._stage, self._acted, self._said):
                 kept.pop(sid, None)
             self._chosen.discard(sid)
             self._partial.discard(sid)
@@ -265,6 +292,8 @@ def append(self, sid, event: dict, *, by, client=None) -> dict:
                 raise SessionError('session_full', 'Session event limit reached')
             self._write(sid, result)
             self._fold(sid, result)
+            if kind in ('message', 'plan', 'step'):
+                self._stage.pop(sid, None)
             for loop, signal in self._waiters.get(sid, ()):
                 try:
                     loop.call_soon_threadsafe(signal.set)
@@ -428,6 +457,32 @@ def stopped(self, sid) -> bool:
         with self.lock:
             return self._get(sid)['stopped']
 
+    def abandoned(self, sid) -> bool:
+        """The user gave up on the turn they last spoke in, and has not spoken since."""
+        with self.lock:
+            self._get(sid)
+            said, gave_up = self._turn.get(sid, (0, 0))
+            return gave_up > 0 and gave_up == said
+
+    def report_stage(self, sid, key) -> None:
+        _check(schema.STAGE, key)
+        with self.lock:
+            self._get(sid)
+            self._stage[sid] = (key, self.clock())
+
+    def stage(self, sid, *, within=STAGE_SECONDS) -> str | None:
+        """The stage the agent reported, while it is recent and the turn it belongs to is still open."""
+        with self.lock:
+            self._get(sid)
+            known = self._stage.get(sid)
+            if known is None or self.clock() - known[1] > within:
+                return None
+            return known[0]
+
+    def clear_stage(self, sid) -> None:
+        with self.lock:
+            self._stage.pop(sid, None)
+
     def unread(self, sid, reader: str) -> list[dict]:
         _check(schema.string(80), reader)
         with self.lock:
@@ -439,7 +494,8 @@ def unread(self, sid, reader: str) -> list[dict]:
                 return []
             rows = self._events(sid, after, self.max_events)
             return [row for row in rows if row['by'] == 'user' and (row['type'] == 'message'
-                    or row['type'] == 'confirm' and row['state'] in ('approved', 'declined'))]
+                    or row['type'] == 'confirm' and row['state'] in ('approved', 'declined')
+                    or row['type'] == 'control' and row['action'] == 'abandoned')]
 
 
 class Confirmations:
@@ -472,7 +528,7 @@ def update(self, values: dict) -> dict:
             os.replace(temporary, self.path)
         return value
 
-    def gate(self, action, *, by, estimate_seconds, refs, question):
+    def gate(self, action, *, by, estimate_seconds, refs, question, shape=None):
         sid, digest = errors.SESSION.get(), errors.CONFIRM.get()
         if by != 'agent' or sid is None or digest is None:
             return
@@ -493,6 +549,8 @@ def gate(self, action, *, by, estimate_seconds, refs, question):
                              refs=refs, digest=digest, state='asked')
                 if estimate_seconds is not None:
                     event['estimate_seconds'] = round(estimate_seconds)
+                # What is about to be made, for the card the user answers; only what is known is written.
+                event.update({key: shape[key] for key in ('width', 'height', 'seconds') if shape and shape.get(key) is not None})
                 row = self.sessions.append(sid, event, by='system')
         raise SessionError('needs_confirmation', 'Waiting for the user',
                            details=[dict(confirm=row['id'], seconds=row.get('estimate_seconds'))])
diff --git a/freevideo_engine/bootstrap.py b/freevideo_engine/bootstrap.py
index e16cc7ca..bf049977 100644
--- a/freevideo_engine/bootstrap.py
+++ b/freevideo_engine/bootstrap.py
@@ -31,7 +31,7 @@
 from .terminal_ui import TerminalUI, LogProgress
 from .locking import runtime_lock, LOCK_ENV
 from .environments import (ENVIRONMENTS, environment_names, select_layout, role_pythons, constraints_file,
-                           bootstrap_versions, uv_file_arguments)
+                           bootstrap_versions, uv_file_arguments, uv_build_environment)
 from . import disk_space
 from . import network
 from . import processes
@@ -86,6 +86,13 @@ def setup_memory_status(row, verbose=False, *, system=None):
     return result
 
 
+# Windows setup downloads and unpacks prebuilt wheels: a real install peaked
+# near 0.93 GiB of tree private working set. The planned budget is the memory
+# free when planning minus the 2 GiB reserve, so on a machine with 2 GiB or
+# less free it was 0 and the first sample stopped the install.
+WINDOWS_SETUP_MIN_PROCESS_BUDGET = 2 * GiB
+
+
 def setup_memory_pressure(row, budget, system):
     """Keep Windows planning headroom separate from an exhaustion stop.
 
@@ -94,9 +101,18 @@ def setup_memory_pressure(row, budget, system):
     that same threshold prevented the low-memory path from continuing.
     Linux uses inference's clean-file-cache headroom, retaining the 1 GiB
     floor and the separate process budget. Raw cgroup charges remain recorded.
+    On Windows the process budget never falls below what setup itself needs;
+    a machine with little free RAM is stopped by the commit and system floors
+    when it truly runs out, not by a budget that was planned as zero.
     """
+    planned_budget = budget
+    raised = system == 'Windows' and budget < WINDOWS_SETUP_MIN_PROCESS_BUDGET
+    if raised:
+        budget = WINDOWS_SETUP_MIN_PROCESS_BUDGET
     reserve = (2 if system == 'Windows' else 1) * GiB
-    floor = min(256 * 2**20, int(budget * .05)) if system == 'Windows' else reserve
+    # A raised budget keeps the full 256 MiB stop point: 5% of the raised
+    # budget would stop a machine with little free RAM later than any other.
+    floor = (256 * 2**20 if raised else min(256 * 2**20, int(budget * .05))) if system == 'Windows' else reserve
     available = (inference_headroom(row) if system == 'Linux'
                  else row.get('effective_available_bytes', row['system_available_bytes']))
     commit = row.get('system_commit_available_bytes')
@@ -108,7 +124,7 @@ def setup_memory_pressure(row, budget, system):
         reasons.append('system_headroom')
     if used is not None and used > budget:
         reasons.append('process_budget')
-    return dict(reasons=reasons, working_bytes=used, budget_bytes=budget,
+    return dict(reasons=reasons, working_bytes=used, budget_bytes=budget, planned_budget_bytes=planned_budget,
                 guard_metric=row.get('guard_metric', 'process working memory'),
                 available_bytes=available, physical_available_bytes=row.get('system_physical_available_bytes'),
                 commit_available_bytes=commit, emergency_floor_bytes=floor,
@@ -124,8 +140,14 @@ def setup_memory_failure(pressure):
         detail = 'Installation paused: system RAM is nearly exhausted (%.2f GiB available; %.2f GiB minimum).' % (
             pressure['available_bytes'] / GiB, pressure['emergency_floor_bytes'] / GiB)
     else:
-        detail = 'Installation crossed its RAM budget (%.2f GiB used; %.2f GiB budget; %s).' % (
-            pressure['working_bytes'] / GiB, pressure['budget_bytes'] / GiB, pressure['guard_metric'])
+        # The budget is the memory free at planning minus the reserve, and setup itself
+        # peaked near 1 GiB; so the free memory needed is the reserve plus that, rounded up to
+        # the next half GiB above it, which leaves some room.
+        needed = (pressure.get('planning_reserve_bytes') or 0) + max(pressure['working_bytes'], GiB)
+        detail = ('Installation crossed its RAM budget (%.2f GiB used; %.2f GiB budget; %s). '
+                  'At least %.1f GiB of RAM must be free when it starts.') % (
+            pressure['working_bytes'] / GiB, pressure['budget_bytes'] / GiB, pressure['guard_metric'],
+            (needed * 2 // GiB + 1) / 2)
     return detail + ' Downloaded files are retained.'
 
 
@@ -509,9 +531,15 @@ def plan(args, *, local_progress=None, stage=None):
             errors.append('Native Windows uses the unified environment. Rerun with --environment unified.')
             layout = 'unified'
         if windows_target and getattr(args, 'rebuild_sage', False):
-            errors.append('Windows uses the Sage2 wheel FreeVideo tests; --rebuild-sage is a Linux source-build option.')
-        if int(snapshot['selected_gpu']['driver_version'].split('.')[0]) < 580:
-            errors.append('NVIDIA driver 580 or newer is required by the CUDA 13 encoder FreeVideo tests; update the driver first.')
+            errors.append('Windows uses the tested Sage2 wheel; --rebuild-sage is a Linux source-build option.')
+        driver = str(snapshot['selected_gpu']['driver_version']).strip()
+        # An unreadable version is left to the later checks instead of blocking setup.
+        major = re.match(r'\d+', driver)
+        if major and int(major.group()) < 580:
+            # The card names the graphics card too; brackets would end its pattern early.
+            model = re.sub(r'[()\r\n]', '', str(snapshot['selected_gpu'].get('name') or '')).strip()
+            errors.append('NVIDIA driver %s%s is older than 580, the minimum for the CUDA 13 encoder; update the driver first.'
+                          % (driver, ' (%s)' % model if model else ''))
         compatibility = hardware.cuda_compatibility()
         if compatibility['error']:
             errors.append(compatibility['error'])
@@ -783,9 +811,9 @@ def plan(args, *, local_progress=None, stage=None):
             'first_generation_note': 'Portable AdaLN tables follow model weights and the exact schedule, across GPUs and Torch/CUDA versions. New schedules or modulation-changing LoRAs require preparation. Existing environments stay on disk when switching.',
             'estimates_are_not_capacity_guarantees': True, 'errors': errors, 'warnings': warnings,
             'steps': ['Install isolated uv/Python and ' + ('one shared CUDA environment' if layout == 'unified' else 'two CUDA environments'),
-                      'Clone the VDN, patched Diffusers and native H3 text-encoder library versions FreeVideo tests',
+                      'Clone the tested versions of VDN, patched Diffusers and the native H3 text-encoder library',
                       'Install verified Windows Triton / Sage2 wheels' if windows_target else 'Install a local CUDA compiler and build Sage2 for the detected GPU only',
-                      'Download and verify the base, eight-step checkpoint and encoder weights FreeVideo tests',
+                      'Download and verify the tested base, eight-step checkpoint and encoder weights',
                       'Prepare or verify the architecture-compatible FP8 cache',
                       'Execute small attention/linear kernel probes and save machine configuration'],
             'licenses': ['https://huggingface.co/OpenVDN/vdn-minimax-h3/blob/751739ee5b9e3ac802dca5d5111075fdaeb47885/LICENSE',
@@ -1009,6 +1037,7 @@ def download(url, path, sha, progress=None, *, networking=None, env=None):
 
 
 def unpack(archive, destination):
+    from .long_paths import io_path
     destination.mkdir(parents=True, exist_ok=True)
     if zipfile.is_zipfile(archive):
         with zipfile.ZipFile(archive) as zipped:
@@ -1018,7 +1047,7 @@ def unpack(archive, destination):
                 if (not path.is_relative_to(destination.resolve()) or ':' in name or
                         (member.external_attr >> 16) & 0o170000 == 0o120000):
                     raise ValueError('Unsafe ZIP member: ' + member.filename)
-            zipped.extractall(destination)
+            zipped.extractall(io_path(destination))
         return
     with tarfile.open(archive) as tar:
         # Python 3.9-compatible containment checks, including link targets.
@@ -1030,7 +1059,7 @@ def unpack(archive, destination):
                 target = (path.parent if member.issym() else destination) / member.linkname
                 if not target.resolve().is_relative_to(destination.resolve()):
                     raise ValueError('Unsafe archive link: ' + member.name)
-        tar.extractall(destination)
+        tar.extractall(io_path(destination))
 
 
 def unpack_folder(archive, destination, name):
@@ -1390,6 +1419,7 @@ def command(self, label, args, env=None, cwd=None):
         env = temporary_environment(self.root, env or self.env)
         env = dict(env, FREEVIDEO_INSTALL_PROGRESS='1')
         args, cwd = uv_file_arguments(args, cwd)
+        env = uv_build_environment(args, self.root, env)
         with self.state_lock:
             index = len(self.state['steps'])
             log = self.run_dir / ('%02d-%s.log' % (index, label))
@@ -1458,6 +1488,12 @@ def update_progress(*, final=False):
             row['error'] = repr(error)
             if child is not None and child.poll() is None:
                 processes.stop(child)
+            # A step stopped from outside leaves nothing else in its log to explain why.
+            try:
+                with log.open('a', encoding='utf-8') as stream:
+                    stream.write('\nStopped: %s\n' % error)
+            except OSError:
+                pass
             raise
         finally:
             update_progress(final=True)
@@ -2236,15 +2272,15 @@ def run_main(argv, shown):
                         help='With --prepared-format: download and verify that model beside the installed one, which stays '
                              'in use. No other setup step runs and machine.json is unchanged')
     parser.add_argument('--model-source', choices=('prepared', 'source'), default='prepared',
-                        help='Default: download a slim model FreeVideo tests matching the GPU format. source: explicitly download original weights and convert locally')
+                        help='Default: download the tested slim model matching the GPU format. source: explicitly download original weights and convert locally')
     parser.add_argument('--encoder-models', type=Path, help='Directory containing text_encoders/')
-    parser.add_argument('--reuse-models', type=Path, help='Scan an existing model library, verify content against the version FreeVideo tests and import matching files; only missing models are downloaded')
+    parser.add_argument('--reuse-models', type=Path, help='Scan an existing model library, verify content against the tested version and import matching files; only missing models are downloaded')
     parser.add_argument('--reuse-models-manifest', type=Path, help='Versioned JSON containing model library roots, including ComfyUI extra model paths')
     parser.add_argument('--copy-existing-models', action='store_true', help='Copy verified local models instead of using same-disk hardlinks; source files are kept')
     parser.add_argument('--allow-model-restart', action='store_true', help='Permit a fresh download when an incomplete model cannot resume; retains old data and may use extra download/disk space')
     cache_options = parser.add_mutually_exclusive_group()
     cache_options.add_argument('--cache', type=Path, help='Verify and reuse an existing FP8 cache')
-    cache_options.add_argument('--rebuild-cache', action='store_true', help='Recreate FP8 from the source weights FreeVideo tests, downloading missing sources; preserve failed caches')
+    cache_options.add_argument('--rebuild-cache', action='store_true', help='Recreate FP8 from the tested source weights, downloading missing sources; preserve failed caches')
     parser.add_argument('--storage', choices=('compact', 'retain'), help='Default compact: stream FP8 and remove installer-owned conversion sources after verification; retain keeps sources for re-quantization')
     parser.add_argument('--frontend-root', type=Path, help=argparse.SUPPRESS)
     parser.add_argument('--frontend-separate', action='store_true', help=argparse.SUPPRESS)
diff --git a/freevideo_engine/capabilities.py b/freevideo_engine/capabilities.py
index 4dee6234..e7323281 100644
--- a/freevideo_engine/capabilities.py
+++ b/freevideo_engine/capabilities.py
@@ -215,7 +215,8 @@ def _machine(machine, in_use=None):
                 vram_free=max(0, int(vram) - idle['vram']),
                 ram_available=max(0, int(ram) - idle['ram']), idle_in_use=idle,
                 cap=tuple(cap) if not mac else None, precision=precision,
-                metal=int(metal) if metal is not None else None)
+                metal=int(metal) if metal is not None else None,
+                compile_owed=machine.get('compile_owed') is True and not mac)
 
 
 def _request(width, height, frames, steps, two_pass, task='t2va', refine=3,
@@ -290,7 +291,8 @@ def _admit(machine, width, height, frames, steps, two_pass, task='t2va', refine=
         disk_risk = any(e['stream_weights'] and e['pin_host_gb'] * 1e9 + 1e6
                         < (50 - e['resident_blocks']) * BLOCK_BYTES for e in engines)
         transfers = any(e['residual_offload'] or e['attention_cpu_outputs'] for e in engines)
-        return request, int(disk_risk or transfers), None
+        # The third value names what made a slow plan slow: host RAM (weights stream from disk) or GPU memory.
+        return request, int(disk_risk or transfers), 'memory' if disk_risk else 'gpu_memory' if transfers else None
     except ResourceBudgetError as error:
         # One field cannot encode both failures: prefer the system RAM limit.
         return None, -1, 'memory' if 'ram' in error.details['insufficient'] else 'gpu_memory'
@@ -423,7 +425,7 @@ def estimate(request):
     tiers = {}
     for name, tier in quality_tiers.table(machine['mac']).items():
         steps, two_pass, step_cache, upscale = tier['steps'], tier['two_pass'], tier['step_cache'], tier['upscale']
-        times, codes, maxima, frame_maxima, limits = [], [], [], [], []
+        times, codes, slow_by, maxima, frame_maxima, limits = [], [], [], [], [], []
         for size_row in sizes:
             admitted = {f: _admit(machine, size_row['width'], size_row['height'], f, steps, two_pass, task,
                                   references=references, step_cache=step_cache, upscale=upscale)
@@ -441,23 +443,27 @@ def estimate(request):
             maxima.append(max((l['seconds'] for l in lengths if admitted[l['frames']][0] is not None),
                               default=None))
             next_frame = maximum + 17 if maximum is not None else 39
-            limits.append('panel' if next_frame > lengths[-1]['frames'] else admitted[next_frame][2])
-            row_times, row_codes = [], []
+            limits.append('panel' if next_frame > lengths[-1]['frames']
+                          else admitted[next_frame][2] if admitted[next_frame][0] is None else None)
+            row_times, row_codes, row_slow = [], [], []
             for length in lengths:
-                request, code, _ = admitted[length['frames']]
+                request, code, why = admitted[length['frames']]
                 seconds = None
                 if request is not None:
                     seconds = max(1, math.floor(estimates[length['frames']] + .5))
                 row_times.append(seconds)
                 row_codes.append(code)
+                row_slow.append(why if code > 0 else None)
             times.append(row_times)
             codes.append(row_codes)
+            slow_by.append(row_slow)
         tiers[name] = dict(steps=steps, two_pass=two_pass, step_cache=step_cache, upscale=upscale, estimate_seconds=times,
-                           max_seconds=maxima, max_frames=frame_maxima, limit=limits, plan=codes)
+                           max_seconds=maxima, max_frames=frame_maxima, limit=limits, plan=codes, slow_by=slow_by)
     default = next((i for i, s in enumerate(sizes) if s['pixels'] == .98), 0)
     result = dict(version=1, status='ready',
                 machine=dict(gpu_name=machine['name'], vram_gib=machine['vram'] / GiB,
-                             ram_gib=machine['ram'] / GiB, system=machine['system']),
+                             ram_gib=machine['ram'] / GiB, system=machine['system'],
+                             compile_owed=machine['compile_owed']),
                 aspect=aspect, sizes=sizes, lengths=lengths, tiers=tiers,
                 summary=dict(width=sizes[default]['width'], height=sizes[default]['height'], tier='light',
                              max_seconds=tiers['light']['max_seconds'][default],
diff --git a/freevideo_engine/chain_runner.py b/freevideo_engine/chain_runner.py
index bf37a1f3..4d400c05 100644
--- a/freevideo_engine/chain_runner.py
+++ b/freevideo_engine/chain_runner.py
@@ -195,8 +195,26 @@ def _view(self, record, detailed=True, client_id=None):
             value['mine'] = record['client_id'] == client_id
         if record['state'] in store.ACTIVE and block is not None:
             value['current'] = block['id']
+        waiting = self._waiting(record)
+        if waiting is not None:
+            value['waiting'] = waiting
         return value
 
+    def _waiting(self, record):
+        """What a chain that has not started waits for: how many chains are ahead and, for a single
+        one, the time it has left (None while that is unknown)."""
+        if record['state'] != 'queued' or record['id'] == self.active or record['id'] not in self.queue:
+            return None
+        ahead = [] if self.active is None else [self.active]
+        for identity in self.queue[:self.queue.index(record['id'])]:
+            row = self.records.get(identity)
+            if row is not None and row['state'] == 'queued' and identity not in ahead:
+                ahead.append(identity)
+        if not ahead:
+            return None
+        remaining = _number(self.progress_values.get(ahead[0], {}).get('remaining_seconds')) if len(ahead) == 1 and ahead[0] == self.active else None
+        return {'ahead': len(ahead), 'remaining_seconds': None if remaining is None or remaining < 0 else remaining}
+
     def _depths(self, record, seen=()):
         """Parts each block of `record` continues through, itself included (a block with no source is the
         first part). A source that is an existing video counts the parts that video was made of, as far
diff --git a/freevideo_engine/comfy_capabilities.py b/freevideo_engine/comfy_capabilities.py
index a2daa4d6..4ad9c07a 100644
--- a/freevideo_engine/comfy_capabilities.py
+++ b/freevideo_engine/comfy_capabilities.py
@@ -28,8 +28,9 @@ def _hash(value):
 
 
 def _context(output_directory):
+    from .background_compile import compile_owed
     from .comfy_bridge import installation
-    _, machine = installation()
+    root, machine = installation()
     machine = dict(machine)
     device = machine.get('device_identity') if machine.get('device_backend') == 'mps' else machine.get('gpu_uuid')
     # Older machine.json stores totals only in its setup plan. Match the saved
@@ -49,6 +50,8 @@ def _context(output_directory):
     if isinstance(machine.get('hardware'), dict):
         observed.update(machine['hardware'])
     machine['hardware'] = observed
+    # Receipt read only: tells the page the first run still compiles.
+    machine['compile_owed'] = compile_owed(root)
     # Do not inherit the sampling-estimate route's 15-second history cache:
     # reopening the grid after a generation must notice the new timing record.
     rows = local_records(output_directory, device, system=machine.get('system'), stage_details=True, refresh=True)
diff --git a/freevideo_engine/comfy_editor.py b/freevideo_engine/comfy_editor.py
index 5150f7a2..55287b00 100644
--- a/freevideo_engine/comfy_editor.py
+++ b/freevideo_engine/comfy_editor.py
@@ -844,6 +844,7 @@ def list_imports(root):
                 fields = ('id', 'name', 'kind', 'state', 'fraction', 'reason', 'frames', 'fps', 'width', 'height',
                           'preview_width', 'preview_height', 'has_audio')
                 row = {k: value.get(k) for k in fields}
+                row['started'] = bool(value.get('started'))
                 for side in ('width', 'height'):
                     row['preview_' + side] = value.get('preview_' + side, value.get(side))
                 row['preview'] = '/freevideo/editor/media?' + urlencode(dict(id=child.name, file='preview')) if value['state'] == 'ready' else None
@@ -929,7 +930,7 @@ def transcode(root, identity):
             if vin is None and ain is None:
                 raise ValueError('No media')
             kind = 'video' if vin is not None else 'audio'
-            _import_update(root, identity, kind=kind)
+            _import_update(root, identity, kind=kind, started=True)
             seconds = max((clip_seconds(reader, s) or 0 for s in (vin, ain) if s is not None), default=0)
             origin = Fraction(vin.start_time or 0) * vin.time_base if vin is not None else Fraction(ain.start_time or 0) * ain.time_base
             with av.open(str(partial), 'w', format='mp4' if vin is not None else 'ipod', options={'movflags': 'faststart'}) as writer:
@@ -1055,9 +1056,11 @@ def submit(self, identity):
     def prepare(self, identity):
         try:
             transcode(self.root, identity)
-        except Exception:
+        except Exception as error:
+            # A full disk is not the file's fault, so the page says so instead of listing formats.
+            reason = _disk_full(error).get('reason', 'unreadable')
             try:
-                _import_update(self.root, identity, state='failed', reason='unreadable')
+                _import_update(self.root, identity, state='failed', reason=reason)
             except (OSError, ValueError):
                 pass  # The terminal snapshot is retained and retried by the worker/reads.
 
diff --git a/freevideo_engine/comfy_launcher.py b/freevideo_engine/comfy_launcher.py
index d3111c38..439615b8 100644
--- a/freevideo_engine/comfy_launcher.py
+++ b/freevideo_engine/comfy_launcher.py
@@ -694,9 +694,10 @@ def copy_error(self):
 
     def report_error(self):
         if self.error_text:
-            from .launcher_issue import issue_url
+            from .launcher_issue import environment_line, issue_url, product_version
             from .windows_ux import open_browser
-            open_browser(issue_url(self.error_text, __version__, None, zh=self.zh))
+            version = product_version(__version__)
+            open_browser(issue_url(self.error_text, version, None, zh=self.zh, environment=environment_line(version)))
 
     def show_details(self):
         dialog = self.tk.Toplevel(self.window); dialog.title(self.t('FreeVideo · Details', 'FreeVideo · 详情'))
diff --git a/freevideo_engine/comfy_library.py b/freevideo_engine/comfy_library.py
index 805b5dc2..3f2ece89 100644
--- a/freevideo_engine/comfy_library.py
+++ b/freevideo_engine/comfy_library.py
@@ -31,6 +31,10 @@ def _video(root, identity):
     return path
 
 
+class _Unfinished(ValueError):
+    """A run whose summary says it did not finish: not a damaged record."""
+
+
 def _report(path):
     # Read only the small request summary, never tensors, prompts or media.
     with path.with_suffix('.request.json').open('rb') as stream:
@@ -38,8 +42,10 @@ def _report(path):
     if len(data) > 4 * 1024 * 1024:
         raise ValueError('Request summary is too large')
     report = json.loads(data)
-    if not isinstance(report, dict) or report.get('success') is not True:
-        raise ValueError('Video is not complete')
+    if not isinstance(report, dict):
+        raise ValueError('Request summary is not an object')
+    if report.get('success') is not True:
+        raise _Unfinished('Video is not complete')
     return report
 
 
@@ -82,12 +88,20 @@ def list_videos(output_directory, *, before=None, limit=24):
                 candidates.append((key, path, info))
         except (OSError, ValueError):
             continue
-    rows = []
+    rows, unreadable = [], []
     for key, path, info in sorted(candidates, key=lambda item: item[0], reverse=True):
+        relative = Path('FreeVideo') / key[1] / 'video.mp4'
+        if relative.as_posix() in hidden:
+            continue
         try:
-            report = _report(path)
-            relative = Path('FreeVideo') / key[1] / 'video.mp4'
-            if relative.as_posix() in hidden:
+            try:
+                report = _report(path)
+            except (FileNotFoundError, _Unfinished):
+                # A generation still writing its video has no summary yet; a failed run is not listed.
+                continue
+            except (OSError, ValueError):
+                # Summaries are replaced atomically, so one that cannot be read is damaged.
+                unreadable.append(key)
                 continue
             summary = output_summary(report, relative)
             # Geometry may gain input metadata in future reports. Keep this API
@@ -123,7 +137,9 @@ def list_videos(output_directory, *, before=None, limit=24):
     next_cursor = rows[-1]['cursor'] if more else None
     for row in rows:
         row.pop('cursor')
-    return dict(items=rows, next=next_cursor)
+    # Each page counts the damaged summaries in its own range, so the pages add up.
+    last = (int(next_cursor.split(':', 1)[0]), next_cursor.split(':', 1)[1]) if next_cursor else None
+    return dict(items=rows, next=next_cursor, unreadable=sum(last is None or key > last for key in unreadable))
 
 
 def thumbnail(output_directory, identity):
diff --git a/freevideo_engine/comfy_metadata.py b/freevideo_engine/comfy_metadata.py
index 00e052b5..0abafdf8 100644
--- a/freevideo_engine/comfy_metadata.py
+++ b/freevideo_engine/comfy_metadata.py
@@ -19,6 +19,11 @@
 BACKUP_NAME = 'video.before-workflow.mp4'
 
 
+def _room(path, size):
+    """A rewrite needs the copy and the backup beside the original, plus some headroom."""
+    return disk_space.free_bytes(path.parent) >= 2 * size + 512 * 1024 * 1024
+
+
 def embed_comfy_metadata(path, *, prompt=None, workflow=None, backup=None):
     """Rewrite the MP4 at `path` with the graph in its metadata, copying the
     streams. The new file replaces the original only after every audio and video
@@ -38,7 +43,7 @@ def embed_comfy_metadata(path, *, prompt=None, workflow=None, backup=None):
         if backup is not None and backup.exists():
             logging.info('FreeVideo left %s unchanged: %s already exists.', path, backup.name)
             return False
-        if disk_space.free_bytes(path.parent) < 2 * size + 512 * 1024 * 1024:
+        if not _room(path, size):
             logging.info('FreeVideo left %s unchanged: not enough free disk space.', path)
             return False
         # faststart puts moov, and with it the metadata, before the media data.
@@ -147,7 +152,7 @@ def backfill(output_directory, *, stop=None):
             rows.setdefault(json.loads(row.read_text(encoding='utf-8'))['id'], []).append(row)
         except (OSError, ValueError, KeyError, TypeError):
             continue
-    counts = dict(embedded=0, skipped=0, pending=0, failed=0)
+    counts = dict(embedded=0, skipped=0, pending=0, failed=0, no_space=0)
     for video in sorted(root.glob('*/*/video.mp4')):
         if stop is not None and stop.is_set():
             return counts
@@ -172,6 +177,11 @@ def backfill(output_directory, *, stop=None):
             continue
         if not embed_comfy_metadata(video, workflow=workflow, backup=run / BACKUP_NAME):
             counts['failed'] += 1
+            # Space is checked after the attempt too: a write that ran out has already removed its copy.
+            try:
+                counts['no_space'] += not _room(video, video.stat().st_size)
+            except OSError:
+                pass
             continue
         counts['embedded'] += 1
         # Keep saved results reusable: the index records each output's size and hash.
@@ -270,12 +280,12 @@ def register():
 
     def run(output):
         try:
-            result = backfill(output) or dict(embedded=0, skipped=0, pending=0, failed=0)
+            result = backfill(output) or dict(embedded=0, skipped=0, pending=0, failed=0, no_space=0)
             if result['embedded']:
                 logging.info('FreeVideo stored the workflow in %d earlier videos.', result['embedded'])
-        except Exception:
+        except Exception as error:
             logging.warning('FreeVideo could not store the workflow in earlier videos.', exc_info=True)
-            result = dict(embedded=0, skipped=0, pending=0, failed=1)
+            result = dict(embedded=0, skipped=0, pending=0, failed=1, no_space=int(disk_space.ran_out(error)))
         with _job_lock:
             _job.update(state='done', result=result)
 
diff --git a/freevideo_engine/comfy_share.py b/freevideo_engine/comfy_share.py
index c8878784..ac1b2eca 100644
--- a/freevideo_engine/comfy_share.py
+++ b/freevideo_engine/comfy_share.py
@@ -10,6 +10,7 @@
 import hashlib
 import io
 import json
+import logging
 import math
 from pathlib import Path
 import threading
@@ -212,7 +213,12 @@ async def export(request):
                 try: await asyncio.shield(task)
                 except (InterruptedError, OSError): pass
                 raise
-            except (ValueError, TypeError, OSError, InterruptedError):
+            except (ValueError, TypeError, OSError, InterruptedError) as error:
+                from .disk_space import ran_out
+                if ran_out(error):
+                    raise web.HTTPInsufficientStorage(text='disk_full') from None
+                if not isinstance(error, InterruptedError):
+                    logging.warning('FreeVideo sharing export failed.', exc_info=error)
                 raise web.HTTPBadRequest(text='Sharing export could not complete') from None
 
     @server.routes.get('/freevideo/share/video')
diff --git a/freevideo_engine/desktop.py b/freevideo_engine/desktop.py
index a2e9dfc9..3871693d 100644
--- a/freevideo_engine/desktop.py
+++ b/freevideo_engine/desktop.py
@@ -241,7 +241,7 @@ def setup_page(self, page):
 
     def generate_page(self, page):
         _, page = self.scrolled(page, 1)
-        self.label(page, 'Describe your video', '描述你想生成的视频').pack(anchor='w', pady=(0, 10))
+        self.label(page, 'Describe your video', '描述想生成的视频').pack(anchor='w', pady=(0, 10))
         prompt_card = Card(page, padding=branding.space(3), radius=branding.RADIUS_MD)
         prompt_card.pack(fill='x')
         self.prompt = self.tk.Text(prompt_card.body, height=6, font=(branding.FAMILY, branding.STRONG),
diff --git a/freevideo_engine/disk_space.py b/freevideo_engine/disk_space.py
index ceaaf791..c9575eb9 100644
--- a/freevideo_engine/disk_space.py
+++ b/freevideo_engine/disk_space.py
@@ -7,6 +7,7 @@
 one APFS container share their space, so they are counted as one disk.
 Other systems keep shutil.disk_usage, which matches Explorer and df.
 """
+import errno
 import os
 from pathlib import Path
 import re
@@ -195,6 +196,18 @@ def free_bytes(path):
         return shutil.disk_usage(io_path(existing(path))).free
 
 
+def ran_out(error):
+    """Whether a write failed because the disk (or the user's quota) is full, through wrapper exceptions."""
+    seen = set()
+    while error is not None and id(error) not in seen:
+        seen.add(id(error))
+        if isinstance(error, OSError) and (error.errno in (errno.ENOSPC, getattr(errno, 'EDQUOT', errno.ENOSPC))
+                                           or getattr(error, 'winerror', None) in (39, 112)):
+            return True
+        error = error.__cause__ or (None if error.__suppress_context__ else error.__context__)
+    return False
+
+
 @plain_errors
 def disk_key(path):
     """Paths with the same key draw on the same free space."""
diff --git a/freevideo_engine/editor_export.py b/freevideo_engine/editor_export.py
index 01e95197..671ddf29 100644
--- a/freevideo_engine/editor_export.py
+++ b/freevideo_engine/editor_export.py
@@ -75,7 +75,8 @@ def _seam_call(operation, *args, **kwargs):
 def _reason(error):
     if isinstance(error, ExportError):
         return error.reason
-    if isinstance(error, OSError) and error.errno == errno.ENOSPC:
+    from .disk_space import ran_out
+    if ran_out(error):
         return 'disk_full'
     if isinstance(error, PermissionError) or isinstance(error, OSError) and error.errno in (errno.EACCES, errno.EROFS):
         return 'not_writable'
@@ -989,9 +990,10 @@ def __init__(self, root, *, loader=None, resolver=None, renderer=None):
         self.pending = deque()
         self.worker = None
 
-    @staticmethod
-    def _view(job):
+    def _view(self, job):
+        # queued: still waiting for its turn; state stays 'running' for older readers.
         return {**{key: job[key] for key in ('id', 'state', 'fraction', 'work', 'reason')},
+                'queued': job['state'] == 'running' and job['id'] in self.pending,
                 **job['options'].status(), 'path': job.get('path'), 'url': job.get('url'),
                 'project_id': job['project'].get('id'), 'revision': job['project'].get('revision')}
 
@@ -1100,11 +1102,15 @@ def progress(fraction):
                         _checked_target(folder, self.root).rmdir()
                 except (OSError, ExportError):
                     pass
+                reason = _reason(error)
+                if reason == 'failed' and not isinstance(error, ExportCancelled):
+                    # The page only says the export failed; the log keeps why for the diagnostic report.
+                    logging.getLogger(__name__).error('Editor export failed.', exc_info=error)
                 with self.lock:
                     if isinstance(error, ExportCancelled) or job['stop'].is_set():
                         job.update(state='cancelled', reason=None)
                     else:
-                        job.update(state='failed', reason=_reason(error))
+                        job.update(state='failed', reason=reason)
 
 
 def register():
diff --git a/freevideo_engine/encode_worker.py b/freevideo_engine/encode_worker.py
index af346347..a8c0c37e 100644
--- a/freevideo_engine/encode_worker.py
+++ b/freevideo_engine/encode_worker.py
@@ -293,6 +293,10 @@ def phase(name, **metrics):
     import torch
     torch.set_num_threads(8)
     torch.set_grad_enabled(False)
+    if resident is not None and not idle:
+        # The guard counts the whole session, retained engine weights included,
+        # from the first sample; free host room before CUDA setup, not after.
+        resident.make_room('encoder', 0, ram_need=8*2**30)
     phase('encoder_cuda_setup')
     if idle:
         from .idle_encoder import gpu_preload_budget
@@ -392,7 +396,9 @@ def phase(name, **metrics):
         clip = resident.take('encoder', encoder_key)
         estimated = Path(path).stat().st_size * 1.4 + 2*2**30
         if not idle and clip is None:
-            resident.make_room('encoder', min(resident.gpu_budget, estimated), model=clip, ram_need=2*2**30)
+            # Measured: a cold encoder load peaks near 8 GiB of guarded host memory
+            # (7.7 GiB PSS); retained engine weights must make room for that.
+            resident.make_room('encoder', min(resident.gpu_budget, estimated), model=clip, ram_need=8*2**30)
         elif idle and (Path(request['cancel']).exists() or not resident.ram_fits() or resident.available() < 2*2**30):
             return dict(state='skipped', reason='No spare live memory, or a foreground request arrived')
         import comfy.model_management
diff --git a/freevideo_engine/environments.py b/freevideo_engine/environments.py
index 6fdb6a3f..26753c31 100644
--- a/freevideo_engine/environments.py
+++ b/freevideo_engine/environments.py
@@ -1,4 +1,5 @@
 """Pinned installation layouts; model lifetimes are independent of venv count."""
+import logging
 import os
 from pathlib import Path
 import platform
@@ -28,6 +29,65 @@ def constraints_file(name, system=None):
                                   '--build-constraints', '--override', '--overrides'))
 
 
+# uv 0.9.0 adds sdists-v9/<index>/<name>/<version>/<revision>/src and
+# setuptools adds build/bdist.<platform>/wheel/... beneath the cache. Reserve
+# 200 characters below the engine root, including downloads/uv-cache. This
+# allows headroom above measured staging paths, not arbitrary sdists.
+UV_BUILD_PATH_CHARS = 200
+
+
+PYTHON_NAMES = ('python.exe', 'python', 'pythonw.exe')
+
+
+def uv_build_environment(command, root, environ):
+    """Windows uv pip builds use a short view of the existing cache and scratch.
+
+    TEMP alone does not move uv's sdist source tree or isolated build venv.
+    Mark only this child's view: processes.popen translates its paths, while
+    installation settings, cache ownership and the parent's environment stay put.
+    """
+    # `venv` and `python` also write into the cache (interpreter records, downloads) below the same root.
+    # A Python child also needs it: it runs `uv` itself (VDN and ComfyUI packages) and the loader
+    # refuses a DLL such as torch\\lib\\c10.dll whose full path exceeds the limit (WinError 206).
+    runs_uv = list(command[1:2]) in (['pip'], ['venv'], ['python']) and Path(command[0]).name.lower() in ('uv', 'uv.exe')
+    if not command or not (runs_uv or Path(str(command[0])).name.lower() in PYTHON_NAMES):
+        return environ
+    from . import launch_view, runtime_alias
+    from .long_paths import io_path, path_chars, plain
+    from .package_paths import WINDOWS_PATH_LIMIT
+    from .triton_compat import TEMP_KEYS, TEMP_ROOT_ENV, _writable_temporary
+    if not runtime_alias._backend.windows():
+        return environ
+    root = Path(os.path.abspath(plain(root)))
+    view = launch_view.current(environ)
+    if (view is None or not root.is_relative_to(view.real)
+            or path_chars(launch_view.value(root, view)) + 1 + UV_BUILD_PATH_CHARS > WINDOWS_PATH_LIMIT):
+        try:
+            view = launch_view.View(root, runtime_alias.runtime_view(root, longest=UV_BUILD_PATH_CHARS))
+        except OSError as error:
+            if (path_chars(root) + 1 + UV_BUILD_PATH_CHARS <= WINDOWS_PATH_LIMIT
+                    and runtime_alias._backend.ansi_path(root)):
+                view = launch_view.View(root, root)
+            else:
+                return environ  # No short entry: keep the old behaviour; packages without source builds still install.
+    if path_chars(launch_view.value(root, view)) + 1 + UV_BUILD_PATH_CHARS > WINDOWS_PATH_LIMIT:
+        return environ
+    temporary = Path(launch_view.value(root / 'tmp' / 'uv', view))
+    try:
+        for path in (temporary.parent, temporary):
+            folder = io_path(path)
+            folder.mkdir(parents=True, exist_ok=True)
+            if folder.is_symlink() or runtime_alias.is_junction(path):
+                raise OSError('The package build temporary folder is a link')
+        _writable_temporary(temporary)
+    except OSError as error:
+        # Like a missing short entry, an unusable scratch folder must not stop installs that worked before.
+        logging.getLogger(__name__).warning('Package builds keep the previous environment: %s', error)
+        return environ
+    return dict(launch_view.environment_view(environ, view),
+                **{key: str(temporary) for key in (*TEMP_KEYS, TEMP_ROOT_ENV)})
+
+
 def uv_file_arguments(command, cwd=None):
     """Return (command, cwd) for which uv reads constraint files in folders with spaces.
 
diff --git a/freevideo_engine/failure_details.py b/freevideo_engine/failure_details.py
index d2a09cad..5ebbfda2 100644
--- a/freevideo_engine/failure_details.py
+++ b/freevideo_engine/failure_details.py
@@ -6,6 +6,7 @@
 import sys
 
 from .long_paths import io_path, plain_text
+from .locking import ENGINE_BUSY
 from .diagnostics import Redactor
 from .tiers import STEP_CACHES, TIERS, on_mac, table_level
 from .two_pass import SCALES
@@ -13,6 +14,11 @@
 
 STEP_OUTPUT = '\n\nFull step output:\n'
 QUALITY_LEVELS = (('Light', '轻量'), ('Medium', '标准'), ('High', '精细'), ('Max', '极致'))  # tiers.TIERS order
+ENGINE_BUSY_TITLE = ('FreeVideo is busy with another task', 'FreeVideo 正在处理另一个任务')
+ENGINE_BUSY_DETAIL = (
+    'Try again when it finishes. If nothing else is running, close duplicate FreeVideo instances, then try again.',
+    '请在它完成后重试。若没有其他任务在进行，请关闭重复运行的 FreeVideo 后重试。',
+)
 
 
 IMAGE_PACK_FAILURES = {
@@ -354,6 +360,12 @@ def launcher_failure(value, *, zh=False, other_disk=False, disk_name='', can_rep
     if not value:
         return dict(title='', detail='', action='', kind='')
     text = _without_ready_curl_diagnostics(str(value))
+    # Match only the cause position, never an echoed input or a log line below it.
+    causes = [line.strip() for line in text.splitlines() if line.strip()]
+    index = 1 if causes and re.fullmatch(r'FreeVideo generation failed(?: \(exit -?\d+\))?', causes[0]) else 0
+    cause = re.sub(r'^(?:[A-Za-z_][\w.]*(?:Error|Exception)):\s*', '', causes[index]) if len(causes) > index else ''
+    if cause == ENGINE_BUSY:
+        return dict(title=ENGINE_BUSY_TITLE[zh], detail=ENGINE_BUSY_DETAIL[zh], action='', kind='engine-busy')
     from .offline_errors import alias_failure, guidance
     if alias_failure(text):
         return dict(title=guidance('title', zh), detail=guidance('runtime-alias', zh),
@@ -531,6 +543,10 @@ def launcher_failure(value, *, zh=False, other_disk=False, disk_name='', can_rep
          ('If this computer has an NVIDIA graphics card, install NVIDIA driver 580 or newer, '
           'restart the computer, then click Retry installation.',
           '如果这台电脑有 NVIDIA 显卡，请安装 580 或更新版本的 NVIDIA 显卡驱动，重启电脑后点击“重试安装”。')),
+        (r'NVIDIA driver [\w.]+(?: \([^)]*\))? is older than 580', 'gpu-driver-old',
+         ('The NVIDIA driver is too old', 'NVIDIA 显卡驱动版本过低'),
+         ('Update the NVIDIA driver to version 580 or newer, restart the computer, then click "Retry installation".',
+          '请将 NVIDIA 显卡驱动更新到 580 或更新版本，重启电脑后点击“重试安装”。')),
         (r'CUDA_ARCH_UNSUPPORTED', 'gpu-architecture',
          ('This GPU lacks the required compute support', '这张显卡不满足当前计算要求'),
          ('The detected GPU and compute capability are shown below. The CUDA runtime requires native BF16 computation (SM80 or newer). Use a compatible GPU; changing the driver does not change its hardware capability.',
@@ -555,6 +571,10 @@ def launcher_failure(value, *, zh=False, other_disk=False, disk_name='', can_rep
          ('A video is still generating', '还有视频正在生成'),
          ('The update is installed. Click Connect again after the job finishes to restart ComfyUI with it.',
           '更新已装好。等当前任务完成后点“重新连接”，ComfyUI 会重启并使用新版。')),
+        (r'WinError 206|The filename or extension is too long', 'path-too-long',
+         ("The installation folder's path is too long", '安装文件夹的路径太长'),
+         ('Install FreeVideo in a folder with a shorter path, such as D:\\FreeVideo, then click "Retry installation".',
+          '请把 FreeVideo 安装到路径较短的文件夹，例如 D:\\FreeVideo，然后点击“重试安装”。')),
         (r'Installation paused: Windows commit headroom', 'memory-commit',
          ('Windows is low on available memory', 'Windows 可用内存不足'),
          ('Close memory-heavy applications, then retry. Downloaded files are retained.',
@@ -570,6 +590,10 @@ def launcher_failure(value, *, zh=False, other_disk=False, disk_name='', can_rep
         (r'No space left on device|WinError 112|disk (?:is )?full|Not enough disk space|Insufficient disk space', 'disk',
          ('Not enough disk space', '磁盘空间不足'),
          ('Free up space on the installation drive, then retry.', '清理安装盘空间后，点击重试。')),
+        (r'Could not reach the pinned prepared model sources', 'model-source',
+         ('The model download sources cannot be reached', '无法连接模型下载源'),
+         ('Check your network or proxy, or change the source or connection mode, then retry.',
+          '请检查网络或代理，或切换下载源、连接模式后重试。')),
         (r'Model download paused|unexpected-transfer-size|ConnectionError|ConnectTimeout|ReadTimeout|HTTP probe failed|Could not resolve host|SSL certificate|Every download source failed|All Git sources failed|All package sources failed|Python download failed on every route', 'download',
          ('Download interrupted', '下载中断了'),
          ('Check your network or change the source or connection mode in Settings → Downloads, then retry.', '检查网络，或在“设置 → 下载”中切换下载源、连接模式后重试。')),
@@ -606,11 +630,33 @@ def launcher_failure(value, *, zh=False, other_disk=False, disk_name='', can_rep
          ('Memory is unavailable for this step', '这一步可用内存不足'),
          ('Close other memory-heavy applications and retry. Copy the details if it continues.', '关闭占用内存较多的程序后重试；仍失败时可复制详情反馈。')),
     )
+    if re.search(r'Installation crossed its RAM budget', text):
+        # Only the installer's memory monitor says this. The sentence carries the amount; an
+        # older retained error does not, and then the card names none.
+        free = re.search(r'At least ([\d.]+) GiB of RAM must be free', text)
+        if free:
+            action = ('Close other applications until at least %s GiB of RAM is free, then click "Retry installation". Downloaded files are retained.' % free.group(1),
+                      '请关闭其他程序，直到至少有 %s GiB 内存空闲，然后点击“重试安装”。已下载的文件会保留。' % free.group(1))
+        else:
+            action = ('Close other applications to free memory, then click "Retry installation". Downloaded files are retained.',
+                      '请关闭其他程序以空出内存，然后点击“重试安装”。已下载的文件会保留。')
+        return dict(title=('There is not enough free memory to continue installing', '可用内存不足，无法继续安装')[zh],
+                    detail='', action=action[zh], kind='memory-budget')
     for pattern, kind, title, action in rules:
         if re.search(pattern, text, re.I):
             if kind in WITHOUT_REPAIR and not can_repair:
                 action = WITHOUT_REPAIR[kind].get(unavailable) or WITHOUT_REPAIR[kind].get('none', WITHOUT_REPAIR_NONE)
-            return dict(title=title[zh], detail='', action=action[zh], kind=kind)
+            detail = ''
+            if kind == 'gpu-driver-old':
+                version, model = re.search(r'NVIDIA driver ([\w.]+)(?: \(([^)]*)\))? is older than 580', text, re.I).groups()
+                if model:
+                    detail = ('检测到：%s，驱动 %s' if zh else 'Detected: %s, driver %s') % (model, version)
+                else:
+                    detail = ('当前驱动版本：' if zh else 'Installed driver: ') + version
+                # The launcher's failure card shows the action only when there is no detail, so the
+                # instruction goes under the detected driver instead of being hidden by it.
+                return dict(title=title[zh], detail=detail + '\n' + action[zh], action='', kind=kind)
+            return dict(title=title[zh], detail=detail, action=action[zh], kind=kind)
     # Validation messages are already actionable. Show the cause, not a stack
     # trace or a guessed hardware diagnosis, and leave long details expandable.
     lines = [line.strip() for line in text.splitlines() if line.strip()]
@@ -619,8 +665,38 @@ def launcher_failure(value, *, zh=False, other_disk=False, disk_name='', can_rep
         if not line.startswith(('File ', 'Traceback', '^', 'All files retained.', 'Rerun the same setup command'))), '')
     detail = re.sub(r'^(?:ValueError|RuntimeError|OSError):\s*', '', detail)
     from .launcher_copy import lan_error
-    return dict(title='暂时未能完成' if zh else 'Couldn’t complete this step',
-                detail=lan_error(detail, zh), action='', kind='unknown')
+    title = '暂时未能完成' if zh else 'Couldn’t complete this step'
+    # The cause line is the last error line, which a traceback or pip's "ERROR:" can displace; the
+    # failed step's own line is found anywhere in the text.
+    step = re.search(r'^Failed step: ([\w.-]+) \(exit code (-?\w+)\)', text, re.M)
+    if step:
+        # The internal step id never reaches the card; the exit code stays as the detail.
+        name = _step_name(step.group(1), zh)
+        if name:
+            title = ('“%s”这一步没有完成' if zh else 'The step "%s" did not finish') % name
+        detail = ('退出码 %s' if zh else 'Exit code %s') % step.group(2)
+    # A message the launcher raised itself (a ValueError) is the next step already.
+    unexpected = 'Traceback' in text and not re.match(r'(?:[\w.]+\.)?ValueError:', causes[-1] if causes else '')
+    detail = lan_error(detail, zh)
+    if step or unexpected:
+        # The launcher's failure card shows the action only when there is no detail, so the next
+        # step goes under the detail rather than being hidden by it.
+        return dict(title=title, detail=(detail + '\n' if detail else '') + UNKNOWN_NEXT_STEP[zh], action='', kind='unknown')
+    return dict(title=title, detail=detail, action='', kind='unknown')
+
+
+UNKNOWN_NEXT_STEP = ('Try again. If it fails again, export the report, then use "Report a problem" and attach it.',
+                     '请重试。仍然失败时，请先导出报告，再点“提交问题”并附上报告。')
+
+
+def _step_name(label, zh):
+    """The name the progress card showed for a setup step, in the interface language, or ''."""
+    try:
+        from .bootstrap import setup_task
+        from .launcher_copy import display
+        return display(setup_task(label)[0], zh)
+    except Exception:
+        return ''
 
 
 def _read(path, limit, *, tail=False):
@@ -666,6 +742,20 @@ def _checkpoint_change(report):
     return ''
 
 
+# Labels of the resource-planning snapshots (generate.py), never stages of the run.
+SNAPSHOT_LABEL = re.compile(r'(?:before|after)-encoding|retry-\d+')
+
+
+def _attempt_phase(report):
+    """The stage the last failed video attempt stopped in, or ''."""
+    attempts = report.get('resource_attempts')
+    for row in reversed(attempts) if isinstance(attempts, list) else ():
+        phase = row.get('phase') if isinstance(row, dict) and row.get('failure') else None
+        if isinstance(phase, str) and re.fullmatch(r'[a-z_]{1,32}', phase) and phase not in ('unknown', 'complete'):
+            return phase
+    return ''
+
+
 # Stands in for the repair sentence until the report text is redacted: the
 # redactor reads "Install / repair" as a path.
 REPAIR_MARK = 'FREEVIDEO_REPAIR_SENTENCE'
@@ -687,6 +777,13 @@ def runtime_explanation(run, repair=None):
     return _runtime_explanation(_report(run), repair)
 
 
+# A traceback's final line names the error at the start of the line; the one a parent
+# process raises about its failed child only repeats the exit code.
+CHILD_ERROR = re.compile(r'[\w.]*(?:Error|Exception|OutOfMemory\w*)\b')
+WRAPPING_ERROR = re.compile(r'(?:subprocess\.)?CalledProcessError\b')
+ANSI = re.compile(r'\x1b(?:\[[0-?]*[ -/]*[@-~]|\][^\x07\x1b]*(?:\x07|\x1b\\))')
+
+
 def generation_failure(run, exit_code=None, *, repair=None):
     """Surface the cause, including a child traceback, even without a report.
 
@@ -700,6 +797,11 @@ def generation_failure(run, exit_code=None, *, repair=None):
     lines = ['FreeVideo generation failed' + (' (exit %s)' % exit_code if exit_code is not None else '')]
     report = _report(run)
     redactor.structured(report)  # Collect prompt values without showing the request.
+    # Read the whole bounded log, not a tail that could hide other output.
+    # A legacy managed failure has exactly this one nonempty line and exit 1.
+    if exit_code == 1 and [line for line in _read(run / 'generate.log', 8192).splitlines()
+                           if line.strip()] == [ENGINE_BUSY]:
+        lines.append(ENGINE_BUSY)
     # What the worker found when cuDNN refused its own files, ahead of the trace.
     explained = _runtime_explanation(report, REPAIR_MARK)
     if explained:
@@ -714,13 +816,21 @@ def generation_failure(run, exit_code=None, *, repair=None):
     encoding_failure = report.get('encoding_failure')
     phase = encoding_failure.get('phase') if isinstance(encoding_failure, dict) else None
     phase = phase or report.get('phase')
+    snapshot = None
+    if isinstance(planning, list) and planning and isinstance(planning[-1], dict):
+        snapshot = planning[-1].get('stage')
+    if isinstance(phase, str) and SNAPSHOT_LABEL.fullmatch(phase):
+        # A planning label is where the resources were last read, not where the run stopped.
+        snapshot, phase = snapshot or phase, None
+    phase = phase or _attempt_phase(report)
+    if not phase and (snapshot or report.get('resource_attempts')):
+        phase = 'generation'
     if phase:
         lines.append('Stage: ' + str(phase))
     if isinstance(planning, list) and planning and isinstance(planning[-1], dict):
         last = planning[-1]
-        if not phase:
-            lines.append('Stage: ' + str(last.get('stage', 'unknown')))
-        lines.append('\nLast resource-planning snapshot (before this failure; not the stopping sample):')
+        lines.append('\nLast resource-planning snapshot (%s, before this failure; not the stopping sample):'
+                     % (re.sub(r'[^a-z0-9-]', '', str(snapshot))[:32] or 'unlabeled'))
         state = last.get('idle_cache') or {}
         accounting = state.get('ram_accounting', {}) if isinstance(state, dict) else {}
         if not isinstance(accounting, dict):
@@ -744,14 +854,25 @@ def generation_failure(run, exit_code=None, *, repair=None):
             lines.append('Idle cache release: ' + str(recovery.get('reason', recovery) if isinstance(recovery, dict) else recovery))
     # Child logs carry the actual CUDA/encoder error when the parent only knows
     # an exit code. Read bounded tails; do not read model data or entire logs.
+    causes, known = [], redactor.text('\n'.join(lines))[:300]
     for name in ('generate.log', 'video.engine.log', 'video.encoding.log'):
         value = _read(run / name, 8192, tail=True)
         if not value:
             continue
+        # Redact the whole tail first: a prompt or credential is only recognized in full.
+        value = redactor.text(value)
         rows = [line for line in value.splitlines() if line.strip() and not line.lstrip().startswith('{"event":')]
         trace = '\n'.join(rows[-45:])
+        causes += [line for line in (ANSI.sub('', row) for row in rows)
+                   if CHILD_ERROR.match(line) and not WRAPPING_ERROR.match(line)]
         if trace:
             lines.extend(['\n' + name + ':', trace])
+    # A long child traceback pushes its own last line out of a shortened copy of this text
+    # (the report's error already holds the log's end), so the error it ends with also
+    # stands near the top unless the text starts with it. It is taken from the redacted tails.
+    promoted = bool(causes) and causes[-1] not in known
+    if promoted:
+        lines.insert(1, 'Cause: ' + causes[-1][:300])
     if len(lines) == 1:
         lines.append('The worker ended without a readable error report.')
     lines.append('\nRetained outputs: ' + str(run))
@@ -769,7 +890,8 @@ def generation_failure(run, exit_code=None, *, repair=None):
         from .runtime_libraries import REINSTALL
         value = value.replace(REPAIR_MARK, repair or REINSTALL, 1)
     value = re.sub(r'\x1b(?:\[[0-?]*[ -/]*[@-~]|\][^\x07\x1b]*(?:\x07|\x1b\\))', '', value)
-    return re.sub(r'[\x00-\x08\x0b-\x1f\x7f]', '', value)
+    value = re.sub(r'[\x00-\x08\x0b-\x1f\x7f]', '', value)
+    return value
 
 
 def startup_failure(message, log, *, exit_code=None, context=None):
diff --git a/freevideo_engine/image_gen/failure.py b/freevideo_engine/image_gen/failure.py
index c7cde206..9055e01b 100644
--- a/freevideo_engine/image_gen/failure.py
+++ b/freevideo_engine/image_gen/failure.py
@@ -9,14 +9,16 @@
 import traceback
 
 from .. import __version__
+from ..disk_space import ran_out
 from ..exit_diagnostics import clean_reading, clean_sample
 from ..failure_details import Redactor
 
 
 ERRORS = frozenset(('no_memory', 'small_memory', 'busy_video', 'busy', 'pack_missing', 'model_unreadable',
-                    'device_unavailable', 'worker_exit', 'timeout', 'cancelled',
+                    'device_unavailable', 'worker_exit', 'timeout', 'cancelled', 'disk_full',
                     'invalid_request', 'invalid_media', 'invalid_points', 'segment_not_installed', 'generation_failed'))
-NO_REPORT = frozenset(('small_memory', 'busy', 'invalid_request', 'invalid_media', 'invalid_points', 'segment_not_installed', 'cancelled'))
+NO_REPORT = frozenset(('small_memory', 'busy', 'invalid_request', 'invalid_media', 'invalid_points', 'segment_not_installed', 'cancelled',
+                       'disk_full'))
 STAGES = ('start', 'load', 'text', 'encode', 'denoise', 'decode')
 KINDS = ('generate', 'edit', 'cast', 'compose', 'download', 'install')
 MESSAGE_ZH = {
@@ -29,6 +31,7 @@
     'device_unavailable': '显卡暂时用不了，稍后再试。',
     'worker_exit': '图像生成意外停止了，请再试一次。',
     'timeout': '这一张用时太长，已停止。请再试一次。',
+    'disk_full': '保存图片时出错：磁盘空间不够，空出空间后再生成。',
     'generation_failed': '图像生成出错了，请再试一次。',
     'invalid_request': '图片设置有误，请调整后重试。',
     'invalid_media': '图片无效，请重新选择输入目录中的图片。点选支持不超过 4096×4096 像素总数的图片。',
@@ -76,6 +79,8 @@ def classify_exception(error, *, stage='start', cancelled=False):
         names = {cls.__name__ for cls in type(item).__mro__}
         if isinstance(item, MemoryError) or 'OutOfMemoryError' in names or _MPS_MEMORY.search(str(item)):
             return 'no_memory'
+    if chain and ran_out(chain[0]):
+        return 'disk_full'
     for item in chain:
         cuda_init = any(frame.f_globals.get('__name__') == 'torch.cuda'
                         and frame.f_code.co_name in ('_lazy_init', '_check_driver')
@@ -282,7 +287,10 @@ def report(self, code, *, row=None, error=None, video_busy=False, worker_restart
         hardware = {key: clean(value, self.redactor) for key, value in hardware.items()}
         geometry = {key: self.request[key] for key in ('width', 'height') if key in self.request}
         plan = safe_plan(row.get('plan') if row.get('plan') is not None else self.plan, self.redactor)
-        return dict(schema='freevideo.image-report', schema_version=1,
+        # Macs only: which attention kernel ran and why it left MLX (mlx_attention.summary).
+        attention = row.get('attention')
+        extra = dict(attention=clean(attention, self.redactor)) if isinstance(attention, dict) else {}
+        return dict(schema='freevideo.image-report', schema_version=1, **extra,
                     created_at=datetime.now(timezone.utc).isoformat(), kind=self.kind,
                     analysis=dict(version=__version__, hardware=hardware, geometry=geometry),
                     summary=dict(request_seconds=self.timing.seconds()), request=dict(self.request),
diff --git a/freevideo_engine/image_gen/mlx_attention.py b/freevideo_engine/image_gen/mlx_attention.py
index ad4573d1..31a2fed8 100644
--- a/freevideo_engine/image_gen/mlx_attention.py
+++ b/freevideo_engine/image_gen/mlx_attention.py
@@ -10,6 +10,12 @@
 
 from ..backends.mps_mlx_attention import to_mlx_many, to_torch
 
+# The first MLX call with and without a mask is compared with the fallback on this many
+# queries; each query attends independently, so the slice is exact. Measured MLX-to-Torch
+# differences on M-series Macs were 0.004-0.0098; a wrong kernel is off by far more.
+CHECK_QUERIES = 64
+TOLERANCE = .05
+
 
 def available():
     try:
@@ -22,16 +28,26 @@ def available():
 def routed(fallback, *, min_queries=256):
     """``attention`` for image-sized calls; short text segments, where the hand-over costs more, keep ``fallback``.
 
-    A call MLX cannot run goes to ``fallback`` instead of failing the picture, and so does
-    every later call of the process.
+    A call MLX cannot run, or whose first result disagrees with ``fallback``, goes to
+    ``fallback`` instead of failing or spoiling the picture, and so does every later call of
+    the process. ``failed`` keeps why, for the image report.
     """
     failed = []
+    checked = set()
 
     def run(query, key, value, attn_mask=None):
         if failed or query.shape[1] < min_queries:
             return fallback(query, key, value, attn_mask=attn_mask)
         try:
-            return attention(query, key, value, attn_mask=attn_mask)
+            out = attention(query, key, value, attn_mask=attn_mask)
+            kind = attn_mask is not None
+            if kind not in checked:
+                reason = differs(query, key, value, attn_mask, out, fallback)
+                checked.add(kind)
+                if reason:
+                    failed.append(reason)
+                    return fallback(query, key, value, attn_mask=attn_mask)
+            return out
         except Exception as error:
             failed.append(f'{type(error).__name__}: {error}'[:300])
             return fallback(query, key, value, attn_mask=attn_mask)
@@ -39,6 +55,31 @@ def run(query, key, value, attn_mask=None):
     return run
 
 
+def differs(query, key, value, attn_mask, out, fallback):
+    """Why ``out`` cannot be trusted, or None when its first queries match ``fallback``."""
+    count = min(CHECK_QUERIES, query.shape[1])
+    mask = attn_mask
+    if mask is not None and mask.dim() >= 2 and mask.shape[-2] == query.shape[1]:
+        mask = mask[..., :count, :]
+    reference = fallback(query[:, :count], key, value, attn_mask=mask).float()
+    got = out[:, :count].float()
+    if not torch.isfinite(got).all():
+        # Non-finite values the fallback also gives come from the inputs, not from MLX.
+        return 'mismatch: non-finite output' if torch.isfinite(reference).all() else None
+    difference = (got - reference).abs().max().item()
+    if difference > TOLERANCE * max(1., reference.abs().max().item()):
+        return 'mismatch: max difference %.3g over %d queries' % (difference, count)
+    return None
+
+
+def summary(function):
+    """What the image report says about attention: None unless ``function`` came from routed()."""
+    failed = getattr(function, 'failed', None)
+    if failed is None:
+        return None
+    return dict(kernel='torch' if failed else 'mlx', **({'fallback_reason': failed[0]} if failed else {}))
+
+
 def attention(query, key, value, attn_mask=None):
     """Same contract as ``qwen_dit.attention``: (batch, sequence, heads, dim), boolean mask or None."""
     import mlx.core as mx
diff --git a/freevideo_engine/image_gen/qwen_pipeline.py b/freevideo_engine/image_gen/qwen_pipeline.py
index b0eeb89e..475cfa6b 100644
--- a/freevideo_engine/image_gen/qwen_pipeline.py
+++ b/freevideo_engine/image_gen/qwen_pipeline.py
@@ -306,8 +306,8 @@ def report(stage, fraction):
         ref_latents = []
         if vae_images:
             with self._stage("encode"):
-                for image in vae_images:
-                    check_cancel()
+                for index, image in enumerate(vae_images):
+                    report("encode", index / len(vae_images))
                     # Always tiled, whatever the memory plan: the latents (and so the picture)
                     # must not depend on the card, and the tiled pass needs far less memory.
                     encoded = _encoded_latents(
@@ -315,7 +315,7 @@ def report(stage, fraction):
                     mean, std = self._stats(encoded)
                     encoded = (encoded - mean) / std
                     ref_latents.append(self._pack_latents(encoded, 1, channels, *encoded.shape[-2:]).to(embeds.dtype))
-                    check_cancel()
+                report("encode", 1.0)
         keep_latents = cell_mask = None
         if edit_mask is not None:
             from . import region_edit
diff --git a/freevideo_engine/image_gen/service.py b/freevideo_engine/image_gen/service.py
index 6a0dcb6f..1b4d5cd7 100644
--- a/freevideo_engine/image_gen/service.py
+++ b/freevideo_engine/image_gen/service.py
@@ -606,8 +606,10 @@ def wait_queued(self):
                         except FailureError as error:
                             self.discard_queued(error.status)
                         except Exception as error:
-                            # A previously valid reference may disappear while waiting.
-                            code = ('generation_failed' if str(error) in INVALID_ERRORS
+                            # A previously valid reference may disappear while waiting; that keeps its
+                            # own reason. Other settings were already checked, so a change there is a fault.
+                            code = (str(error) if str(error) == 'invalid_media'
+                                    else 'generation_failed' if str(error) in INVALID_ERRORS
                                     else classify_exception(error))
                             self.discard_queued(self.failed(code, pending['context'], error=error))
                         else:
@@ -1087,6 +1089,8 @@ def infer(self, session, value):
                         raise RuntimeError('worker_exit')
                     result = dict(phase='complete', file='freevideo-image/' + path.name,
                                   seconds=seconds, plan=row.get('plan'))
+                    if isinstance(row.get('attention'), dict):
+                        result['attention'] = row['attention']
                 with self.mutex:
                     if 'fitted' in self.state:
                         result['fitted'] = self.state['fitted']
diff --git a/freevideo_engine/image_gen/worker.py b/freevideo_engine/image_gen/worker.py
index f45022d8..469acf6f 100644
--- a/freevideo_engine/image_gen/worker.py
+++ b/freevideo_engine/image_gen/worker.py
@@ -152,6 +152,11 @@ def generate(seed):
                     if back_check is not None:
                         plan = dict(plan or {}, back_check=back_check)
                 result = dict(phase='result', path=value['path'], seconds=time.monotonic() - started, plan=plan)
+                try:
+                    result.update(attention_report(runner, context.redactor))
+                except Exception:
+                    # Diagnostics must not turn a saved picture into an error.
+                    pass
             except Exception as error:
                 code = classify_exception(error, stage=context.timing.stage, cancelled=inbox.cancel.is_set())
                 plan = None
@@ -165,6 +170,10 @@ def generate(seed):
                 result = dict(phase='error', code=code, stage=context.timing.stage, reloaded=reload,
                               seconds=context.timing.seconds(), stages=context.timing.snapshot(), plan=plan,
                               error=exception_details(error, context.redactor, code))
+                try:
+                    result.update(attention_report(runner, context.redactor))
+                except Exception:
+                    pass
             finally:
                 if result is not None and result.get('reloaded'):
                     # A failure inside a layer pass leaves the weights streamer part-way through it and the
@@ -192,6 +201,19 @@ def generate(seed):
                 runner.close()
 
 
+def attention_report(runner, redactor=None):
+    """{'attention': ...} on Macs where MLX attention was installed (why it fell back, if it did); {} elsewhere."""
+    if runner is None:
+        return {}
+    from . import mlx_attention, qwen_dit
+    value = mlx_attention.summary(qwen_dit.attention)
+    if value is None:
+        return {}
+    if redactor is not None and 'fallback_reason' in value:
+        value['fallback_reason'] = redactor.text(value['fallback_reason'])[:300]
+    return dict(attention=value)
+
+
 def main():
     parser = argparse.ArgumentParser(description=__doc__)
     parser.add_argument('pack_dir', type=Path)
diff --git a/freevideo_engine/install.py b/freevideo_engine/install.py
index 4d4b00da..ce74e48f 100644
--- a/freevideo_engine/install.py
+++ b/freevideo_engine/install.py
@@ -7,6 +7,7 @@
 import sys
 import time
 from . import execution_paths
+from .long_paths import io_path
 from .paths import vdn_root
 from . import network, subprocess_output
 
@@ -16,8 +17,7 @@ def install(root=None, *, packages=False):
     spec = dependencies['vdn']
     root = (root or vdn_root()).expanduser().resolve()
     network.clone(root, spec['url'], spec['commit'])
-    head = subprocess_output.output_external(['git', 'rev-parse', 'HEAD'], cwd=root).strip()
-    dirty = subprocess_output.output_external(['git', 'status', '--porcelain', '--untracked-files=no'], cwd=root)
+    head, dirty = network.repository_state(root, 'git', None)
     if head != spec['commit'] or dirty:
         raise RuntimeError('Existing VDN checkout does not match the clean pinned dependency; use another --vdn-root.')
     diffusers = prepare_diffusers(root, dependencies)
@@ -76,22 +76,36 @@ def prepare_diffusers(root, dependencies):
     """Build the official two-patch tree atomically; never reset a user checkout."""
     spec = dependencies['diffusers']
     target = root / dependencies['vdn']['diffusers_subdirectory']
-    if not target.exists():
-        checkout = root / ('engine-diffusers-' + str(time.time_ns()))
-        network.clone(checkout, spec['url'], spec['base'])
-        patches = sorted((root / 'diffusers_patches').glob('*.patch'))
-        if len(patches) != 2:
-            raise RuntimeError('Expected the two pinned VDN Diffusers patches')
-        for path in patches:
-            execution_paths.run(['git', 'apply', '--index', str(path)], cwd=checkout, check=True)
-        tree = subprocess_output.output_external(['git', 'write-tree'], cwd=checkout).strip()
-        if tree != spec['patched_tree']:
-            raise RuntimeError('Patched Diffusers source tree does not match the tested dependency')
-        execution_paths.run(['git', '-c', 'user.name=VDN Engine dependency setup', '-c', 'user.email=engine@localhost',
-                        'commit', '--quiet', '-m', 'Apply pinned OpenVDN MiniMax H3 patches'], cwd=checkout, check=True)
-        checkout.rename(target)
-    tree = subprocess_output.output_external(['git', 'rev-parse', 'HEAD^{tree}'], cwd=target).strip()
-    dirty = subprocess_output.output_external(['git', 'status', '--porcelain', '--untracked-files=no'], cwd=target)
-    if tree != spec['patched_tree'] or dirty:
-        raise RuntimeError('Existing Diffusers differs from the clean tested tree; use another VDN dependency directory.')
+    built = False
+    if not io_path(target).exists():
+        # Where the folder is too deep for Git to work in, build in a short real one.
+        scratch = network.git_scratch_folder() if network.git_folder_too_deep(root / 'engine-diffusers-0') else None
+        checkout = scratch / 'repo' if scratch else root / ('engine-diffusers-' + str(time.time_ns()))
+        try:
+            network.clone(checkout, spec['url'], spec['base'])
+            patches = sorted(io_path(root / 'diffusers_patches').glob('*.patch'))
+            if len(patches) != 2:
+                raise RuntimeError('Expected the two pinned VDN Diffusers patches')
+            for path in patches:
+                # The patch path carries the long-path prefix, which Git cannot open; pass the content instead.
+                execution_paths.run(['git', '-c', 'core.longpaths=true', 'apply', '--index', '-'],
+                                    cwd=checkout, input=path.read_bytes(), check=True)
+            tree = subprocess_output.output_external(['git', '-c', 'core.longpaths=true', 'write-tree'], cwd=checkout).strip()
+            if tree != spec['patched_tree']:
+                raise RuntimeError('Patched Diffusers source tree does not match the tested dependency')
+            execution_paths.run(['git', '-c', 'core.longpaths=true', '-c', 'user.name=VDN Engine dependency setup', '-c', 'user.email=engine@localhost',
+                            'commit', '--quiet', '-m', 'Apply pinned OpenVDN MiniMax H3 patches'], cwd=checkout, check=True)
+            if scratch:
+                partial = root / ('engine-diffusers-' + str(time.time_ns()))
+                network.relocate(checkout, partial)
+                checkout = partial
+            io_path(checkout).rename(io_path(target))
+            built = True
+        finally:
+            if scratch and scratch.exists():
+                network.remove_tree(scratch)
+    if not built:
+        tree, dirty = network.repository_state(target, 'git', None, 'HEAD^{tree}')
+        if tree != spec['patched_tree'] or dirty:
+            raise RuntimeError('Existing Diffusers differs from the clean tested tree; use another VDN dependency directory.')
     return target
diff --git a/freevideo_engine/launch_view.py b/freevideo_engine/launch_view.py
index f9c020bf..e97ef240 100644
--- a/freevideo_engine/launch_view.py
+++ b/freevideo_engine/launch_view.py
@@ -67,24 +67,30 @@ def view_for(real_root) -> View:
 def for_engine(engine_root) -> View:
     """The view of the offline runtime around this engine folder, if it needs one.
 
-    An installation made online is its engine folder: it gets a view only where the
-    system code page cannot spell its folder, and one that cannot be made is not an error.
+    An installation made online is its engine folder: it gets a view only where the system
+    code page cannot spell its folder or its deepest library is beyond the loader's limit,
+    and one that cannot be made is not an error.
     """
     from . import runtime_alias
     engine = Path(plain(engine_root))
-    if engine.name.casefold() == ONLINE_ENGINE:
-        try:
-            return view_for(engine) if runtime_alias.needs_view(engine, spelling_only=True) else View(engine, engine)
-        except OSError:  # Includes a junction that cannot be made: the checks after it name what fails.
-            return View(engine, engine)
     root = engine.parent
     # The first two tests read nothing from disk wherever no junction is needed.
     # The third asks by the extended path: where a junction is needed, the plain
     # one is not found beyond 259 characters.
-    if (engine.name.casefold() != 'engine' or not runtime_alias.needs_view(root)
-            or not io_path(root / 'portable.json').is_file()):
+    if (engine.name.casefold() == 'engine' and runtime_alias.needs_view(root)
+            and io_path(root / 'portable.json').is_file()):
+        return view_for(root)
+    if engine.name.casefold() == 'engine':
         return View(engine, engine)
-    return view_for(root)
+    try:
+        # An installation made online is its engine folder. It gets a view where the system code
+        # page cannot spell the folder or the loader cannot open its deepest library; one that
+        # cannot be made is not an error.
+        if runtime_alias.needs_view(engine, spelling_only=True) or runtime_alias.online_loaded_too_long(engine):
+            return view_for(engine)
+    except OSError:  # Includes a junction that cannot be made: the checks after it name what fails.
+        pass
+    return View(engine, engine)
 
 
 def for_file(path):
diff --git a/freevideo_engine/launcher/Main.qml b/freevideo_engine/launcher/Main.qml
index 5b26d445..8f724186 100644
--- a/freevideo_engine/launcher/Main.qml
+++ b/freevideo_engine/launcher/Main.qml
@@ -36,7 +36,7 @@ ApplicationWindow {
     property bool compact: width < 1000
     property bool narrow: width < 640
     property bool shortWindow: height < 700
-    onModelInfoOpenChanged: if (modelInfoOpen) backend.prepareDownloadList()
+    onModelInfoOpenChanged: if (modelInfoOpen) backend.prepareDownloadList(); else backend.setDownloadScope("video")
     readonly property bool manualEnvironment: s.offline.runtime_supported !== false && s.form.new_comfy && s.form.environment_method === "manual"
     readonly property bool usingRuntime: manualEnvironment && s.offline.runtime
     readonly property bool needsRuntime: s.page === "comfy" && manualEnvironment && !s.offline.runtime
@@ -403,7 +403,7 @@ ApplicationWindow {
                 Item { height: shortWindow ? 0 : 16 }
                 ColumnLayout {
                     Layout.fillWidth: true; spacing: 8; Layout.bottomMargin: shortWindow ? 0 : 10
-                    FText { objectName: "pageHeading"; horizontalAlignment: Text.AlignHCenter; text: s.page === "launcher" ? t("Your workspace", "你的工作空间") : s.page === "comfy" ? t("Set up FreeVideo", "安装 FreeVideo") : s.page === "models" ? t("Set up your models", "准备模型") : s.status === "review" ? t("Ready to install", "确认安装") : s.busy ? t("Setting things up", "正在准备 FreeVideo") : t("Continue your setup", "继续安装"); font.pixelSize: win.compact ? 26 : theme.hero; font.weight: Font.DemiBold; font.letterSpacing: -0.6; Layout.fillWidth: true }
+                    FText { objectName: "pageHeading"; horizontalAlignment: Text.AlignHCenter; text: s.page === "launcher" ? t("Your workspace", "工作空间") : s.page === "comfy" ? t("Set up FreeVideo", "安装 FreeVideo") : s.page === "models" ? t("Set up your models", "准备模型") : s.status === "review" ? t("Ready to install", "确认安装") : s.busy ? t("Setting things up", "正在准备 FreeVideo") : t("Continue your setup", "继续安装"); font.pixelSize: win.compact ? 26 : theme.hero; font.weight: Font.DemiBold; font.letterSpacing: -0.6; Layout.fillWidth: true }
                     FText { objectName: "actionNotice"; visible: !!s.action_notice; text: s.action_notice || ""; color: theme.muted; Layout.fillWidth: true; horizontalAlignment: Text.AlignHCenter }
                 }
 
@@ -443,7 +443,7 @@ ApplicationWindow {
                             FButton { objectName: "chooseDisk"; visible: diskFailure && s.form.new_comfy && diskChoices.length > (bestDisk ? 1 : 0); enabled: !s.busy
                                       text: t("Other disks…", "其他硬盘…"); onClicked: diskDialogOpen = true }
                             FButton { objectName: "diskCleanupButton"; visible: s.failure.kind === "disk"; enabled: s.can_cleanup; text: t("Clean download cache", "清理缓存"); onClicked: { settingsTab = "general"; settingsOpen = true; backend.cleanupDownloads(false) } }
-                            FButton { visible: s.failure.kind === "download"; text: t("Change source", "切换下载源"); onClicked: { settingsTab = "downloads"; settingsOpen = true } }
+                            FButton { objectName: "changeSource"; visible: s.failure.kind === "download" || s.failure.kind === "model-source"; text: t("Change source", "切换下载源"); onClicked: { settingsTab = "downloads"; settingsOpen = true } }
                             FButton { objectName: "repairCardButton"; primary: true; visible: !!s.repair_offered; text: t("Repair", "修复"); onClicked: backend.action("repair", false) }
                             FButton { objectName: "latestLauncherLink"; visible: s.failure.kind === "older-launcher"; flat: true
                                       text: t("Download the latest ↗", "下载最新版 ↗"); onClicked: backend.link("https://github.com/FlashML-org/FreeVideo/releases/latest") }
@@ -634,7 +634,7 @@ ApplicationWindow {
                                     delegate: FButton { required property var modelData; required property int index; objectName: "offlineShare-" + index; text: t("Download from Quark ↗", "打开夸克网盘 ↗"); onClicked: backend.link(modelData.url) }
                                 }
                                 FText { objectName: "offlinePackageGuide"; text: s.offline.guide; color: theme.muted; font.pixelSize: theme.micro; Layout.fillWidth: true }
-                                FButton { objectName: "otherModelLinks"; text: downloads.copy.title + "  ›"; flat: true; enabled: !s.busy; implicitHeight: theme.heightSm; leftPadding: 0; font.pixelSize: theme.micro + 1; onClicked: modelInfoOpen = true }
+                                FButton { objectName: "otherModelLinks"; text: downloads.copy.title + "  ›"; flat: true; enabled: !s.busy; implicitHeight: theme.heightSm; leftPadding: 0; font.pixelSize: theme.micro + 1; onClicked: { backend.setDownloadScope("video"); modelInfoOpen = true } }
                             }
                             ColumnLayout {
                                 Layout.fillWidth: true; Layout.preferredWidth: 1; Layout.alignment: Qt.AlignTop; spacing: 10
@@ -675,7 +675,7 @@ ApplicationWindow {
                         }
                         RowLayout {
                             objectName: "promptModelRow"
-                            visible: !!s.prompt_model && !s.prompt_model.ready && !s.prompt_model.checking
+                            visible: !!s.prompt_model && !s.prompt_model.ready && !s.prompt_model.checking && !s.prompt_model.unavailable
                             Layout.fillWidth: true; spacing: 14
                             FCheck {
                                 objectName: "promptModelSwitch"; Layout.fillWidth: true
@@ -800,7 +800,7 @@ ApplicationWindow {
                                 }
                                 // What this group downloads, as in the plan; nothing for one already here.
                                 FText { objectName: "modelSize-" + modelData.id; visible: modelData.download > 0; text: bytes(modelData.download).replace(" ", "\u00a0"); font.pixelSize: theme.micro; color: theme.muted; font.features: { "tnum": 1 } }
-                                FButton { objectName: "modelFiles-" + modelData.id; text: downloads.copy.title + "  ›"; implicitHeight: theme.heightSm; flat: true; font.pixelSize: theme.micro + 1; visible: s.form.model_method === "manual"; onClicked: modelInfoOpen = true }
+                                FButton { objectName: "modelFiles-" + modelData.id; text: downloads.copy.title + "  ›"; implicitHeight: theme.heightSm; flat: true; font.pixelSize: theme.micro + 1; visible: s.form.model_method === "manual"; onClicked: { backend.setDownloadScope(modelData.id === "image" ? "image" : "video"); modelInfoOpen = true } }
                             }
                         }
                     }
@@ -890,7 +890,7 @@ ApplicationWindow {
                                 Layout.fillWidth: true; spacing: 6
                                 // A failed start stays failed while another task (an import, a cleanup) runs: the controller's own tasks clear it first.
                                 FText { objectName: "statusTitle"; text: s.status === "open" ? t("FreeVideo is ready", "FreeVideo 已就绪") : s.status === "failed" ? t("Startup interrupted", "启动未完成") : s.busy ? t("Starting ComfyUI", "正在启动 ComfyUI") : s.status === "restart-required" ? t("Restart ComfyUI", "请重启 ComfyUI") : t("Ready to launch", "准备就绪"); font.pixelSize: theme.section + 2; font.weight: Font.DemiBold; Layout.fillWidth: true }
-                                FText { objectName: "statusDetail"; text: s.status === "open" ? s.url : s.status === "restart-required" ? t("Restart ComfyUI after the update, then select \"Connect again\".", "更新后需重启 ComfyUI，再点“重新连接”。") : s.status === "failed" ? t("ComfyUI did not start. See the message above.", "ComfyUI 未能启动，原因见上方。") : t("Your models and environment are connected.", "已连接你的模型与运行环境。"); color: theme.muted; Layout.fillWidth: true }
+                                FText { objectName: "statusDetail"; text: s.status === "open" ? s.url : s.status === "restart-required" ? t("Restart ComfyUI after the update, then select \"Connect again\".", "更新后需重启 ComfyUI，再点“重新连接”。") : s.status === "failed" ? t("ComfyUI did not start. See the message above.", "ComfyUI 未能启动，原因见上方。") : t("Your models and environment are connected.", "已连接模型与运行环境。"); color: theme.muted; Layout.fillWidth: true }
                                 FText { objectName: "portMovedNote"; visible: s.status === "open" && s.port_from > 0; text: t("Port %1 is in use by another program, so FreeVideo now uses port %2 and will keep using this address.", "端口 %1 已被其他程序占用，FreeVideo 改用端口 %2，以后也使用这个地址。").arg(s.port_from).arg(s.port_to); font.pixelSize: 13; color: theme.muted; wrapMode: Text.WordWrap; Layout.fillWidth: true; Layout.topMargin: -2 }
                                 FText { objectName: "pluginScopedNote"; visible: s.status === "open" && s.plugin_scoped; text: t("Some installed plugins are not compatible with FreeVideo, so only FreeVideo was loaded this time.", "部分已装的插件与 FreeVideo 不兼容，这次只载入了 FreeVideo。"); font.pixelSize: 13; color: theme.muted; wrapMode: Text.WordWrap; Layout.fillWidth: true; Layout.topMargin: -2 }
                             }
@@ -1617,13 +1617,13 @@ ApplicationWindow {
             RowLayout {
                 objectName: "downloadListActions"; Layout.fillWidth: true; Layout.minimumWidth: 0; spacing: 8
                 // The actions share what the Close button leaves, so a long label wraps instead of leaving the dialog.
-                readonly property int actionCount: checkDownloadList.visible ? 3 : 2
+                readonly property int actionCount: (checkDownloadList.visible ? 1 : 0) + (downloads.scope === "image" ? 1 : 2)
                 readonly property real actionWidth: Math.max(80, Math.floor((width - closeDownloadList.implicitWidth - (actionCount + 1) * spacing) / actionCount))
                 FButton { id: closeDownloadList; text: downloads.copy.close; flat: true; onClicked: modelInfoOpen = false }
                 Item { Layout.fillWidth: true }
-                FButton { id: checkDownloadList; objectName: "checkDownloadList"; Layout.maximumWidth: parent.actionWidth; visible: !downloads.available || s.download_list.needs_check; text: downloads.copy.check; enabled: !s.busy; onClicked: backend.prepareDownloadList() }
-                FButton { objectName: "useDownloadedFolder"; Layout.maximumWidth: parent.actionWidth; text: downloads.copy.use_folder; enabled: s.can_manage; onClicked: backend.useDownloadedFolder() }
-                FButton { objectName: "exportDownloadList"; Layout.maximumWidth: parent.actionWidth; text: downloads.copy.export; primary: true; enabled: downloads.available && !s.busy; onClicked: backend.exportDownloadList() }
+                FButton { id: checkDownloadList; objectName: "checkDownloadList"; Layout.maximumWidth: parent.actionWidth; visible: downloads.scope !== "image" && (!downloads.available || s.download_list.needs_check); text: downloads.copy.check; enabled: !s.busy; onClicked: backend.prepareDownloadList() }
+                FButton { objectName: "useDownloadedFolder"; visible: downloads.scope !== "image"; Layout.maximumWidth: parent.actionWidth; text: downloads.copy.use_folder; enabled: s.can_manage; onClicked: backend.useDownloadedFolder() }
+                FButton { objectName: "exportDownloadList"; Layout.maximumWidth: parent.actionWidth; text: downloads.copy.export; primary: true; enabled: downloads.available && (downloads.scope === "image" || !s.busy); onClicked: backend.exportDownloadList() }
             }
         }
     }
@@ -1706,7 +1706,7 @@ ApplicationWindow {
             spacing: 14
             FText { objectName: "updateActionNotice"; visible: !!s.action_notice; text: s.action_notice || ""; color: theme.muted; Layout.fillWidth: true }
             FText { text: s.update.candidate ? t("Update available", "有可用更新") : s.update.engine ? t("Engine update", "引擎更新") : t("Updates", "更新"); font.pixelSize: theme.section + 2; font.weight: Font.DemiBold }
-            FText { objectName: "updateDialogText"; text: s.update.phase === "waiting" ? t("A video is still generating. FreeVideo restarts and updates as soon as it finishes.", "还有视频正在生成，完成后会自动重启并更新。") : s.update.status === "ready" && s.update.candidate ? t("Downloaded. Restart to finish; models and settings are kept.", "下载完成，重启即可完成更新，模型和设置都会保留。") : s.update.candidate && s.update.manual ? t("FreeVideo ", "FreeVideo ") + releaseVersion(s.update.candidate) + t(" is available (current ", " 已发布（当前 ") + releaseVersion(currentRelease) + t("). Download it and replace this version; models and settings are kept.", "）。下载新版后替换当前版本，模型和设置会保留。") : s.update.candidate ? t("FreeVideo ", "FreeVideo ") + releaseVersion(s.update.candidate) + t(" is available (current ", " 已发布（当前 ") + releaseVersion(currentRelease) + t("). Updating keeps your models and settings. FreeVideo restarts after the download; running videos finish first.", "）。更新会保留模型和设置，下载完成后自动重启；正在生成的视频会先完成。") : s.update.engine ? t("This launcher already includes engine ", "启动器已带有新版引擎 ") + releaseVersion(currentRelease) + t(" (installed: ", "（已安装 ") + engineVersion + (s.status === "open" ? t("). Updating takes about a minute, restarts ComfyUI once and keeps your models and settings.", "）。更新约需 1 分钟，会重启一次 ComfyUI，模型和设置都会保留。") : t("). Updating takes about a minute, then FreeVideo starts; models and settings are kept.", "）。更新约需 1 分钟，完成后自动启动，模型和设置都会保留。")) : s.update.status === "current" ? t("You're up to date.", "已是最新版本。") : s.update.status === "development" ? t("Running from source. Update with Git.", "当前从源码运行，请通过 Git 更新。") : s.update.status === "error" ? t("Could not check for updates.", "暂时无法检查更新。") : t("Checking the latest release…", "正在检查最新版本…"); color: theme.muted; Layout.fillWidth: true }
+            FText { objectName: "updateDialogText"; text: s.update.phase === "waiting" ? t("A video is still generating. FreeVideo restarts and updates as soon as it finishes.", "还有视频正在生成，完成后会自动重启并更新。") : s.update.status === "ready" && s.update.candidate ? t("Downloaded. Restart to finish; models and settings are kept.", "下载完成，重启可完成更新，模型和设置都会保留。") : s.update.candidate && s.update.manual ? t("FreeVideo ", "FreeVideo ") + releaseVersion(s.update.candidate) + t(" is available (current ", " 已发布（当前 ") + releaseVersion(currentRelease) + t("). Download it and replace this version; models and settings are kept.", "）。下载新版后替换当前版本，模型和设置会保留。") : s.update.candidate ? t("FreeVideo ", "FreeVideo ") + releaseVersion(s.update.candidate) + t(" is available (current ", " 已发布（当前 ") + releaseVersion(currentRelease) + t("). Updating keeps your models and settings. FreeVideo restarts after the download; running videos finish first.", "）。更新会保留模型和设置，下载完成后自动重启；正在生成的视频会先完成。") : s.update.engine ? t("This launcher already includes engine ", "启动器已带有新版引擎 ") + releaseVersion(currentRelease) + t(" (installed: ", "（已安装 ") + engineVersion + (s.status === "open" ? t("). Updating takes about a minute, restarts ComfyUI once and keeps your models and settings.", "）。更新约需 1 分钟，会重启一次 ComfyUI，模型和设置都会保留。") : t("). Updating takes about a minute, then FreeVideo starts; models and settings are kept.", "）。更新约需 1 分钟，完成后自动启动，模型和设置都会保留。")) : s.update.status === "current" ? t("You're up to date.", "已是最新版本。") : s.update.status === "development" ? t("Running from source. Update with Git.", "当前从源码运行，请通过 Git 更新。") : s.update.status === "error" ? t("Could not check for updates.", "暂时无法检查更新。") : t("Checking the latest release…", "正在检查最新版本…"); color: theme.muted; Layout.fillWidth: true }
             FText { visible: !!s.update.error; text: s.update.error || ""; color: theme.danger; Layout.fillWidth: true; font.pixelSize: theme.micro }
             FButton { objectName: "updateManualDownloadButton"; text: t("Download from the release page ↗", "去发布页手动下载 ↗"); flat: true; Layout.fillWidth: true; visible: s.update.status === "error" && !!s.update.release_page; onClicked: backend.link(s.update.release_page) }
             FButton { objectName: "updateDialogReport"; text: t("Report a problem", "提交问题"); flat: true; Layout.fillWidth: true; visible: s.update.status === "error" && !!s.update.error; onClicked: backend.reportProblem(s.update.error, t("Update", "更新")) }
diff --git a/freevideo_engine/launcher_copy.py b/freevideo_engine/launcher_copy.py
index d5436a91..65e54d43 100644
--- a/freevideo_engine/launcher_copy.py
+++ b/freevideo_engine/launcher_copy.py
@@ -11,7 +11,11 @@
                          '下载目录已不存在，请重新导出清单'),
     'installation_missing': ('There is no installation information yet. Complete installation first.',
                              '还没有安装信息，请先完成安装'),
+    'installation_unavailable': ('The saved ComfyUI folder could not be opened. It may have been moved or renamed, or its drive may be disconnected. Choose its current location below, or install ComfyUI.',
+                                 '无法打开已保存的 ComfyUI 文件夹，它可能已被移动、改名，或所在磁盘未连接。请在下方选择它现在的位置，或安装 ComfyUI。'),
     'checking_update': ('Checking for updates. Please wait.', '正在检查更新，请稍候'),
+    'update_install_failed': ('The update could not be installed. You can keep using the current version.',
+                              '无法安装更新。当前版本可以照常使用。'),
     'update_page_failed': ('Couldn\'t open the download page. Please try again later.', '无法打开下载页，请稍后再试。'),
     'busy': ('The background task is busy. You can do this when it finishes.',
              '后台正在忙，完成后可以操作'),
diff --git a/freevideo_engine/launcher_issue.py b/freevideo_engine/launcher_issue.py
index f0e393ed..222f7f79 100644
--- a/freevideo_engine/launcher_issue.py
+++ b/freevideo_engine/launcher_issue.py
@@ -18,15 +18,55 @@ def _lines(error):
     return [line for line in scrub(error).splitlines() if line.strip()]
 
 
-def issue_url(error, version=None, title=None, zh=False):
+def product_version(fallback=None):
+    """The 0.4.x release of this launcher; the fallback (the build number) only when no release says."""
+    try:
+        from pathlib import Path
+        from .release_notes import installed_details
+        return installed_details(Path(__file__).parent, fallback).get('product_version') or fallback
+    except Exception:
+        return fallback
+
+
+def machine_hardware(engine):
+    """The hardware record of the installation at engine (machine.json), or {} when it is not readable."""
+    try:
+        import json
+        from pathlib import Path
+        from .long_paths import io_path
+        hardware = json.loads(io_path(Path(engine) / 'machine.json').read_text(encoding='utf-8'))['hardware']
+        return hardware if isinstance(hardware, dict) else {}
+    except Exception:
+        return {}
+
+
+def environment_line(version=None, hardware=None):
+    """One line about this computer: release, system, GPU, driver and memory. No paths, user names or build numbers."""
+    hardware = hardware if isinstance(hardware, dict) else {}
+
+    def gib(key):
+        value = hardware.get(key)
+        return '%.1f GiB' % (value / 2**30) if type(value) is int and value > 0 else ''
+
+    gpu = ', '.join(part for part in (str(hardware.get('gpu_name') or ''), gib('vram_total') + ' VRAM' if gib('vram_total') else '') if part)
+    parts = ['FreeVideo ' + str(version).lstrip('v') if version else '', str(hardware.get('system') or ''), gpu,
+             'driver ' + str(hardware['driver_version']) if hardware.get('driver_version') else '',
+             gib('ram_total') + ' RAM' if gib('ram_total') else '']
+    return scrub(' · '.join(part for part in parts if part))[:300]
+
+
+def issue_url(error, version=None, title=None, zh=False, environment=''):
     lines = _lines(error)
     first = (lines[-1] if lines and lines[0].startswith('Traceback') else (lines[0] if lines else '')).strip()
     place = '启动器' if zh else 'Launcher'
     problem = ('FreeVideo 启动器显示了一个错误。\n' if zh else 'The FreeVideo launcher showed an error.\n') + first[:300]
     report = '\n'.join(lines)
+    # The environment line is for triage, so it stays English in either language and is kept whole;
+    # only the error text gives way to fit the address.
+    head = (('Environment: ' + scrub(environment).strip() + '\n\n') if str(environment).strip() else '')
 
     def fit(limit):
-        return report if len(report) <= limit else report[:limit // 3] + '\n...\n' + report[-(limit - limit // 3 - 5):]
+        return head + (report if len(report) <= limit else report[:limit // 3] + '\n...\n' + report[-(limit - limit // 3 - 5):])
 
     def build(body):
         params = {'template': 'bug_launcher.yml', 'title': f'[{place}] ' + (str(title).strip() if title else place),
diff --git a/freevideo_engine/launcher_session.py b/freevideo_engine/launcher_session.py
index a83a49fa..3d7b7188 100644
--- a/freevideo_engine/launcher_session.py
+++ b/freevideo_engine/launcher_session.py
@@ -21,7 +21,8 @@
 from .launcher_terminal import Tail
 from .model_status import FAMILIES, NAMES
 from .model_guidance import (package_instructions, video_instructions, runtime_packages_supported, download_view,
-                             export_download_list, DownloadListError, DOWNLOAD_COPY)
+                             export_download_list, DownloadListError, DOWNLOAD_COPY,
+                             image_download_view, export_image_download_list, IMAGE_COPY)
 from .setup_progress import LineHold, ModelLine, progress_text
 from .terminal_ui import clean, duration
 from .model_upgrade import TARGET
@@ -46,12 +47,12 @@
 
 
 
-@lru_cache(maxsize=4)
 def _lan_password():
     # Resolved per call: the access module is optional and may be absent.
     return importlib.import_module('.lan_password', __package__)
 
 
+@lru_cache(maxsize=4)
 def _qr_rows(text):
     from .qr_code import encode
     try:
@@ -145,6 +146,9 @@ def __init__(self, source=None, *, controller=None, store=None, updater=None, sm
         self.image_installer = self.prompt_installer = None
         self.download_export = self.download_error = self.download_error_key = ''
         self._download_view = None
+        # 'image' opens the image model's own list, which needs no file check; any other value is the video list.
+        self.download_scope = 'video'
+        self._image_download_view = None
         self.browser_attempted = False
         self.browser_url = None
         self.port_notice_url = None
@@ -346,7 +350,15 @@ def add_folder(self, folder):
         kept = [p for p in self.form['model_dirs'] if io_path(Path(p).expanduser()).is_dir()]
         self.edit('model_dirs', library_roots(kept + [folder]))
 
+    def set_download_scope(self, scope):
+        scope = 'image' if scope == 'image' else 'video'
+        if scope != self.download_scope:
+            self.download_scope = scope
+            self.download_export = self.download_error = self.download_error_key = ''
+
     def prepare_download_list(self):
+        if self.download_scope == 'image':
+            return
         if self.controller.busy or self.importer.busy or self.cleaner.busy:
             return
         if (self.controller.state.get('plan', {}).get('download_files_verified')
@@ -361,12 +373,14 @@ def prepare_download_list(self):
     def export_downloads(self, folder):
         """The dialog shows a worded error; nothing raises into a traceback."""
         self.download_export = self.download_error = self.download_error_key = ''
-        if self.controller.busy or self.controller.state.get('status') != 'review':
+        image = self.download_scope == 'image'
+        if not image and (self.controller.busy or self.controller.state.get('status') != 'review'):
             self.download_error = self.t('Wait for the file check to finish.', '请等待文件检查完成。')
             return
         preferences = download_preferences(self.engine_root() / 'download-settings.json')
         try:
-            self.download_export = export_download_list(self.controller.state.get('plan', {}), folder, preferences['source'])
+            self.download_export = (export_image_download_list(folder, preferences['source'], self._image_location()[0]) if image else
+                                    export_download_list(self.controller.state.get('plan', {}), folder, preferences['source']))
         except DownloadListError as error:
             self.download_error_key = error.key
             reason = ''
@@ -376,7 +390,8 @@ def export_downloads(self, folder):
                 from .failure_details import redacted_launcher_error
                 print('Download list export failed (%s): %s' % (error.reason, redacted_launcher_error(error.original_error)),
                       file=sys.stderr)
-            self.download_error = self.t(*DOWNLOAD_COPY['error_' + error.key]).format(reason=reason)
+            texts = dict(DOWNLOAD_COPY, error_changed=IMAGE_COPY['error_changed']) if image else DOWNLOAD_COPY
+            self.download_error = self.t(*texts['error_' + error.key]).format(reason=reason)
 
     def use_downloaded_folder(self, folder):
         if not self.can_manage():
@@ -388,6 +403,12 @@ def use_downloaded_folder(self, folder):
         self.prepare_download_list()
 
     def downloads_view(self):
+        if self.download_scope == 'image':
+            zh, decimal_sizes = self.language.startswith('zh'), sys.platform == 'darwin'
+            key = (zh, decimal_sizes)
+            if self._image_download_view is None or self._image_download_view[0] != key:
+                self._image_download_view = (key, image_download_view(zh, decimal_sizes=decimal_sizes))
+            return self._image_download_view[1]
         row = self.controller.state
         plan = row.get('plan', {}) if row.get('status') == 'review' else {}
         if not plan.get('download_files_verified', True):
@@ -479,6 +500,8 @@ def activate_offline(self, root):
         self.persist()
 
     def action(self, name, accepted=False, *, retry=None):
+        if self.action_notice == 'installation_unavailable':
+            self.action_notice = ''
         if name == 'stop':
             self.pending_action = None
             if self.action_notice == 'setup_queued':
@@ -557,7 +580,11 @@ def action(self, name, accepted=False, *, retry=None):
             self.page = 'comfy'; return
         if name == 'launcher' and self.selected:
             if not self.controller.restore({'installation': self.selected}):
-                raise ValueError(self.t('Locate or repair this installation in Setup.', '请在安装设置中重新定位或修复这份安装。'))
+                # A moved, renamed or unplugged folder is expected: go to Setup
+                # with the saved paths filled in instead of a failure card.
+                self.action_notice = 'installation_unavailable'
+                self.page = 'comfy'
+                return
             self.page = 'launcher'
             return
         if name == 'update-engine' and self.selected:
@@ -795,7 +822,8 @@ def start_image_model_after_install(self):
         if self.form['prompt_model'] and self.installation_engine() and self.controller.state.get('status') == 'open' \
                 and not self.controller.busy and not (self.prompt_installer and self.prompt_installer.busy):
             from .prompt_vlm import assets
-            if not assets.ready(self._image_location()[1]):
+            # Without its model list the download cannot start; the tick stays for a later start.
+            if assets.available() and not assets.ready(self._image_location()[1]):
                 self.prompt_model_action('install')
             if self.prompt_installer and self.prompt_installer.busy or assets.ready(self._image_location()[1]):
                 self.form['prompt_model'] = False
@@ -827,7 +855,7 @@ def prompt_model_action(self, action):
         if self.prompt_installer and self.prompt_installer.busy:
             return
         from .prompt_vlm import assets
-        if assets.checking(engine):
+        if assets.checking(engine) or not assets.available():
             return
         if not (self.prompt_installer and self.prompt_installer.root == engine):
             self.prompt_installer = prompt_installer.Installer(engine, busy=self.generating)
@@ -836,7 +864,13 @@ def prompt_model_action(self, action):
     def prompt_model_view(self):
         """The prompt model as a row of the download card and as the suggestion on the home page."""
         from .prompt_vlm import assets
-        spec = assets.catalog()
+        try:
+            spec = assets.catalog()
+        except (OSError, ValueError):
+            # Without its model list the row can neither be described nor offered, but
+            # a missing resource must not take the whole launcher page down with it.
+            return dict(ready=False, checking=False, unavailable=True, total_bytes=0, enough_disk=True, busy=False,
+                        install={}, failure='', downloaded_bytes=0, title=self.t(*NAMES['text']), suggest=False)
         total = sum(row['bytes'] for row in spec['files'])
         engine = self._image_location()[1] if self.installation_engine() else None
         ready = bool(engine) and assets.ready(engine)
@@ -1519,7 +1553,7 @@ def update_view(self):
         view['release_page'] = release_page(self.updater.current) if self.updater else ''
         if self.updater and self.updater.install_failed and view.get('candidate'):
             # The new app did not go in and the old one is back: the reminder says so.
-            view['status'], view['error'] = 'error', 'The update could not be installed.'
+            view['status'], view['error'] = 'error', self.t(*ACTION_COPY['update_install_failed'])
         key = self.update_key()
         view.update(engine=self.engine_update_pending(), current=__version__,
                     current_release=public_details(self.updater.current) if self.updater else self.release_details,
diff --git a/freevideo_engine/launcher_settings.py b/freevideo_engine/launcher_settings.py
index 12314a70..c502e7d1 100644
--- a/freevideo_engine/launcher_settings.py
+++ b/freevideo_engine/launcher_settings.py
@@ -3,6 +3,7 @@
 import json
 import os
 from pathlib import Path
+import shutil
 import sys
 import time
 
@@ -91,6 +92,34 @@ def read(path):
         return {}
 
 
+KEPT_COPIES = 3
+
+
+def keep_unreadable(path):
+    """Copy a settings file that exists but cannot be parsed aside before a save replaces it."""
+    try:
+        text = io_path(path).read_text(encoding='utf-8')
+        if not text.strip() or isinstance(json.loads(text), dict):
+            return None
+    except FileNotFoundError:
+        return None
+    except (OSError, ValueError):
+        pass
+    kept = Path(str(path) + '.bad-' + time.strftime('%Y%m%d-%H%M%S'))
+    try:
+        shutil.copyfile(io_path(path), io_path(kept))
+    except OSError:
+        return None
+    # A save that keeps failing leaves the file unreadable, so every retry copies it again.
+    older = sorted(kept.parent.glob(Path(path).name + '.bad-*'))[:-KEPT_COPIES]
+    for stale in older:
+        try:
+            io_path(stale).unlink()
+        except OSError:
+            pass
+    return kept
+
+
 def write(path, values, language, selection=None):
     payload = dict(values, language=language, installation=selection)
     save(path, sanitize(payload))
@@ -174,6 +203,8 @@ def write_locked(self, values, language, selection=None, setup=None):
                 previous = read(path)
                 if previous:
                     save(path.with_suffix('.previous.json'), previous)
+                else:
+                    keep_unreadable(path)
                 save(path, payload)
                 written.append(path)
             except OSError as error:
diff --git a/freevideo_engine/launcher_view.py b/freevideo_engine/launcher_view.py
index 9ce17936..5342ca54 100644
--- a/freevideo_engine/launcher_view.py
+++ b/freevideo_engine/launcher_view.py
@@ -257,7 +257,7 @@ def wheel(event):
     choices.columnconfigure((0, 1), weight=1, uniform='choices')
     app.mode_buttons = []
     for index, (value, title, detail) in enumerate((
-        ('existing', ('I have ComfyUI', '我已安装 ComfyUI'), ('Connect your installation.', '接入现有安装。')),
+        ('existing', ('Use existing ComfyUI', '使用已有 ComfyUI'), ('Connect your installation.', '接入现有安装。')),
         ('new', ('Install ComfyUI for me', '帮我安装 ComfyUI'), ('ComfyUI, Python and FreeVideo together.', '一并准备 ComfyUI、Python 和 FreeVideo。')))):
         choice = Choice(choices, app, value, title, detail)
         choice.grid(row=0, column=index, sticky='nsew', padx=(0, b.space(2)) if index == 0 else (b.space(2), 0))
diff --git a/freevideo_engine/locking.py b/freevideo_engine/locking.py
index 9110dd81..9524bc6f 100644
--- a/freevideo_engine/locking.py
+++ b/freevideo_engine/locking.py
@@ -6,10 +6,12 @@
 occupied an entire multi-GPU host.
 """
 from contextlib import contextmanager
+import math
 import os
 from pathlib import Path
 import re
 import sys
+import time
 
 
 from .paths import data_root
@@ -17,6 +19,7 @@
 from .processes import inherited_fd
 LOCK_PATH = Path(os.environ.get('FREEVIDEO_LOCK_PATH', data_root() / 'engine.lock'))
 LOCK_ENV = 'FREEVIDEO_RUNTIME_LOCK_FD'
+ENGINE_BUSY = 'Setup, testing or generation is already using this installation/GPU lock. Retry after it finishes.'
 
 
 def lock_holders(paths):
@@ -56,7 +59,10 @@ def device_lock_path(device, path=None):
 
 
 @contextmanager
-def runtime_lock(path=None, *, inherit=True, shared=False):
+def runtime_lock(path=None, *, inherit=True, shared=False, wait_seconds=0):
+    """Take a lease, optionally waiting a bounded time for acquisition only."""
+    if not math.isfinite(wait_seconds) or wait_seconds < 0:
+        raise ValueError('Runtime lock wait must be finite and nonnegative')
     path = Path(path or os.environ.get('FREEVIDEO_LOCK_PATH', data_root() / 'engine.lock'))
     path.parent.mkdir(parents=True, exist_ok=True)
     inherited = inherited_fd() if inherit else None
@@ -71,23 +77,36 @@ def runtime_lock(path=None, *, inherit=True, shared=False):
     try:
         # An inherited descriptor shares the parent's open-file description.
         # Close our descriptor on exit; do not unlock the parent's live work.
-        if windows():
-            if inherited is None:
-                import msvcrt
-                if path.stat().st_size == 0:
-                    stream.write(b'\0')
-                    stream.flush()
-                stream.seek(0)
-                try:
-                    msvcrt.locking(stream.fileno(),
-                                   msvcrt.LK_NBRLCK if shared else msvcrt.LK_NBLCK, 1)
-                except OSError as error:
-                    raise BlockingIOError('Another process holds the installation lease') from error
-            # The parent holds the byte-range lock; its Job Object owns this
-            # worker tree. Closing a duplicated child handle must not unlock it.
-        else:
-            import fcntl
-            fcntl.flock(stream, (fcntl.LOCK_SH if shared else fcntl.LOCK_EX) | fcntl.LOCK_NB)
+        deadline = time.monotonic() + wait_seconds
+        while True:
+            try:
+                if windows():
+                    if inherited is None:
+                        import msvcrt
+                        if path.stat().st_size == 0:
+                            stream.write(b'\0')
+                            stream.flush()
+                        stream.seek(0)
+                        try:
+                            msvcrt.locking(stream.fileno(),
+                                           msvcrt.LK_NBRLCK if shared else msvcrt.LK_NBLCK, 1)
+                        except OSError as error:
+                            raise BlockingIOError('Another process holds the installation lease') from error
+                    # The parent holds the byte-range lock; its Job Object owns this
+                    # worker tree. Closing a duplicated child handle must not unlock it.
+                else:
+                    import fcntl
+                    fcntl.flock(stream, (fcntl.LOCK_SH if shared else fcntl.LOCK_EX) | fcntl.LOCK_NB)
+                break
+            except BlockingIOError:
+                remaining = deadline - time.monotonic()
+                if remaining <= 0:
+                    raise
+                time.sleep(min(0.075, remaining))
+                # A delayed wakeup must not start another attempt after budget.
+                if time.monotonic() >= deadline:
+                    raise
+        # The body is outside the retry loop: a failed request must never rerun.
         yield stream.fileno()
     finally:
         # Releasing the lease must not discard completed work. Taking it still
diff --git a/freevideo_engine/lora_check.py b/freevideo_engine/lora_check.py
index c031fac5..05321610 100644
--- a/freevideo_engine/lora_check.py
+++ b/freevideo_engine/lora_check.py
@@ -11,7 +11,7 @@
 import struct
 import threading
 
-from .lora_names import INT8_UNMERGEABLE, _targets, adapter_keys, int8_skip_reason, resolve_target
+from .lora_names import INT8_UNMERGEABLE, _targets, adapter_keys, int8_skip_reason, resolve_target, supported
 
 MAX_HEADER_BYTES = 64 * 1024 * 1024
 BLOCKS, REFINER_BLOCKS = 50, 2
@@ -58,6 +58,11 @@ def _layout():
 
 LAYOUT = _layout()
 UNUSABLE = 'This LoRA was not trained for the FreeVideo model and cannot be used.'
+ACCELERATION = ('This acceleration LoRA was made for the original model and cannot be used. '
+                'The FreeVideo model is already accelerated and does not need one.')
+# Step-distillation adapters (for example parallel-decoding ones) ship a bank of
+# whole replacement output heads, one per stretch of the sampling schedule.
+HEADS = ('proj_out', 'audio_proj_out')
 _CACHE, _LOCK = {}, threading.Lock()
 
 
@@ -115,6 +120,17 @@ def _bad(reason, keys, **detail):
     return dict(reason=reason, family=family(keys) if reason == 'other-model' else None, **detail)
 
 
+def _head_bank(shapes):
+    """The first tensor that is a stack of whole output-head weights, or None."""
+    for key, shape in shapes.items():
+        if not key.endswith('.weight') or len(shape) != 3:
+            continue
+        name = _targets(key[:-7])[0][0]
+        if name in HEADS and list(shape[1:]) == list(LAYOUT[name]):
+            return key
+    return None
+
+
 def _block(target):
     found = re.search(r'blocks\.(\d+)\.', target)
     return int(found[1]) + 1 if found else None
@@ -122,15 +138,22 @@ def _block(target):
 
 def assess(shapes, *, int8_targets=None, skipped=None):
     """None when the adapter fits, else why not: reason 'other-model' (names of another model), 'shape' (our names,
-    a different size) or 'format' (not a LoRA we can read). 'layer' and the sizes say where. A layer name the table
-    does not list, in a file that matches no other known model, is let through.
+    a different size), 'acceleration' (a step-distillation adapter that replaces the output heads and needs its own
+    sampler) or 'format' (not a LoRA we can read; 'layer' names the first tensor that is not a LoRA factor).
+    'layer' and the sizes say where. A layer name the table does not list, in a file that matches no other known
+    model, is let through.
     With int8_targets, use the installed linear shapes and collect the runtime's
     skipped patch categories; 'int8-unmergeable' means every patch was skipped."""
     keys = list(shapes)
+    heads = _head_bank(shapes)
+    if heads:
+        return _bad('acceleration', keys, layer=heads)
     try:
         pairs, full = adapter_keys(keys)
     except ValueError:
-        return _bad('other-model' if family(keys) else 'format', keys, layer=None)
+        if family(keys):
+            return _bad('other-model', keys, layer=None)
+        return _bad('format', keys, layer=next((key for key in keys if not supported(key)), None))
     if not any('A' in pair for pair in pairs.values()) and not full:
         return _bad('format', keys, layer=None)
     wrong, applied = [], set()
@@ -263,4 +286,4 @@ def require_usable(path):
     """Stop a request before generation starts when a LoRA cannot work with the model."""
     result = check_installed_file(path)
     if not result['usable']:
-        raise ValueError(INT8_UNMERGEABLE if result['reason'] == 'int8-unmergeable' else UNUSABLE)
+        raise ValueError({'int8-unmergeable': INT8_UNMERGEABLE, 'acceleration': ACCELERATION}.get(result['reason'], UNUSABLE))
diff --git a/freevideo_engine/lora_names.py b/freevideo_engine/lora_names.py
index 16a2ca5c..41d64837 100644
--- a/freevideo_engine/lora_names.py
+++ b/freevideo_engine/lora_names.py
@@ -8,17 +8,33 @@
                     'so it cannot be used for now. LoRAs that change only linear layers work.')
 
 
+# Factor names trainers write: PEFT lora_A/lora_B, Kohya and diffusers
+# lora_down/lora_up, each with or without the trailing ".weight".
+_FACTOR = re.compile(r'(.+)\.lora_(A|B|down|up)(?:\.weight)?')
+_SIDE = {'A': 'A', 'down': 'A', 'B': 'B', 'up': 'B'}
+
+
+def _canonical(key):
+    return key.replace('.default.', '.')
+
+
+def supported(key):
+    """Whether adapter_keys can index this tensor name."""
+    canonical = _canonical(key)
+    return bool(_FACTOR.fullmatch(canonical)) or canonical.endswith(('.alpha', '.diff', '.diff_b'))
+
+
 def adapter_keys(keys):
     """Index factor/difference names without reading any tensors."""
     pairs, full = {}, []
     for key in keys:
-        canonical = key.replace('.default.', '.').replace('.lora_down.weight', '.lora_A.weight').replace('.lora_up.weight', '.lora_B.weight')
-        match = re.fullmatch(r'(.+)\.lora_([AB])\.weight', canonical)
+        canonical = _canonical(key)
+        match = _FACTOR.fullmatch(canonical)
         if match:
-            pair = pairs.setdefault(match[1], {})
-            if match[2] in pair:
+            pair, side = pairs.setdefault(match[1], {}), _SIDE[match[2]]
+            if side in pair:
                 raise ValueError('Duplicate LoRA factor: ' + key)
-            pair[match[2]] = key
+            pair[side] = key
         elif canonical.endswith('.alpha'):
             pairs.setdefault(canonical[:-6], {})['alpha'] = key
         elif canonical.endswith(('.diff', '.diff_b')):
diff --git a/freevideo_engine/lora_online_cache.py b/freevideo_engine/lora_online_cache.py
index 92b2c7da..c00096fa 100644
--- a/freevideo_engine/lora_online_cache.py
+++ b/freevideo_engine/lora_online_cache.py
@@ -21,7 +21,7 @@ def _supported(adapters, linears, skip_modulation=False):
     for adapter in adapters:
         with open_tensors(adapter['path']) as source:
             for key in source.keys():
-                name = re.sub(r'\.(?:lora_[AB]\.weight|lora_(?:down|up)\.weight|alpha)$',
+                name = re.sub(r'\.(?:lora_(?:A|B|down|up)(?:\.weight)?|alpha)$',
                               '', key.replace('.default.', '.'))
                 for target, _, _ in _targets(name):
                     if target not in linears and '.attn.' in target:
diff --git a/freevideo_engine/mac_app_update.py b/freevideo_engine/mac_app_update.py
index 28d7b3b4..6447145f 100644
--- a/freevideo_engine/mac_app_update.py
+++ b/freevideo_engine/mac_app_update.py
@@ -50,12 +50,23 @@
 pattern=$(printf '%s\n' "$app/Contents/MacOS/" | /usr/bin/sed 's/[][\.*^$+?(){}|]/\\&/g')
 alive() { "${FREEVIDEO_PGREP:-/usr/bin/pgrep}" -f -- "$pattern" >/dev/null 2>&1; }
 stable=0; seen=0; tries=0
-while [ "$stable" -lt "$stable_need" ]; do
-    if alive; then seen=1; stable=$((stable + 1)); elif [ "$seen" = 1 ]; then break; fi
-    tries=$((tries + 1))
-    if [ "$tries" -gt $((appear_max + stable_need)) ]; then break; fi
-    sleep "$tick"
-done
+watch() {
+    while [ "$stable" -lt "$stable_need" ]; do
+        if alive; then
+            seen=1; stable=$((stable + 1))
+        elif [ "$seen" = 1 ]; then
+            return
+        else
+            # Only the wait for the app to appear is limited; once it is up it is never given up on early.
+            tries=$((tries + 1))
+            if [ "$tries" -gt "$appear_max" ]; then return; fi
+        fi
+        sleep "$tick"
+    done
+}
+watch
+# An app that shows up just as the wait ends is running: never delete it to bring the old one back.
+if [ "$stable" -lt "$stable_need" ] && alive; then stable=0; seen=1; watch; fi
 if [ "$stable" -lt "$stable_need" ]; then
     note "the new app did not stay open"; /bin/mv "$app" "$staged" && restore; exit 1
 fi
diff --git a/freevideo_engine/macos_encoder.py b/freevideo_engine/macos_encoder.py
index 3bdc43f4..20bc20c1 100644
--- a/freevideo_engine/macos_encoder.py
+++ b/freevideo_engine/macos_encoder.py
@@ -45,7 +45,7 @@ def nvfp4_weight(weight, dtype, device, *, rows=128, row_range=None):
     return result[:count if row_range is not None else params.orig_shape[0], :params.orig_shape[1]]
 
 
-def install_weight_reader(clip, backend, checkpoint, *, decoder=nvfp4_weight):
+def install_weight_reader(clip, backend, checkpoint, *, decoder=nvfp4_weight, pool_bytes=None):
     """Adapt only this encoder instance; never patch the shared Comfy library."""
     import torch
     from comfy.quant_ops import QuantizedTensor, get_layout_class
@@ -119,7 +119,8 @@ def progress(module, inputs, output, index=index):
     stats['projection_slices'], slice_handles = install(
         [module for module, function in selected if function is linear], model.layers,
         decoder=decoder, quantized_type=QuantizedTensor, backend=backend,
-        guard=guard, release=lambda: release_mapped_pages([checkpoint]), reader_stats=stats)
+        guard=guard, release=lambda: release_mapped_pages([checkpoint]), reader_stats=stats,
+        pool_bytes=pool_bytes)
     handles.extend(slice_handles)
     return stats, restores, handles
 
@@ -198,9 +199,11 @@ def release_encoder():
             clip = load_clip([checkpoint], clip_type=comfy.sd.CLIPType.MINIMAX,
                 model_options=dict(initial_device=torch.device('cpu'), load_device=torch.device('mps'),
                     offload_device=torch.device('cpu'), dtype=torch.bfloat16), disable_dynamic=True)
-            stats, restores, handles = install_weight_reader(clip, backend, checkpoint, decoder=decoder)
+            pool_bytes = policy['effective_allocator_limit_bytes']
+            stats, restores, handles = install_weight_reader(clip, backend, checkpoint, decoder=decoder,
+                                                             pool_bytes=pool_bytes)
             from .macos_encoder_attention import install
-            stats['attention'], attention_handles = install(clip)
+            stats['attention'], attention_handles = install(clip, pool_bytes)
             handles.extend(attention_handles)
             # Placement is handled by the instance reader; native tokenization,
             # forward and output packing still execute through the CLIP API.
diff --git a/freevideo_engine/macos_encoder_attention.py b/freevideo_engine/macos_encoder_attention.py
index bb77e347..351e25d4 100644
--- a/freevideo_engine/macos_encoder_attention.py
+++ b/freevideo_engine/macos_encoder_attention.py
@@ -13,7 +13,7 @@
 SLICE_WORKING_BYTES = 256 * 2**20
 
 
-def slice_plan(tokens, heads, dim, kv_heads=None):
+def slice_plan(tokens, heads, dim, kv_heads=None, pool=1):
     """Pinned batch-one FP32 contract; include full output and both tile outputs.
 
     GQA visits one KV head's query heads at a time, with zero-stride KV views.
@@ -30,17 +30,17 @@ def slice_plan(tokens, heads, dim, kv_heads=None):
     output_bytes = tokens * heads * dim * 4
     per_row = 2 * group_heads * (tokens + dim) * 4
     # Even if a whole GQA group fits, large calls still split the query axis.
-    rows = min((tokens + 1) // 2, max(1, (SLICE_WORKING_BYTES - output_bytes) // per_row))
+    rows = min((tokens + 1) // 2, max(1, (SLICE_WORKING_BYTES * pool - output_bytes) // per_row))
     return dict(query_rows=rows, head_groups=groups,
                 slice_working_bytes=output_bytes + rows * per_row,
                 minimum_slice_bytes=output_bytes + per_row)
 
 
-def plan(tokens, heads, dim, kv_heads=None):
+def plan(tokens, heads, dim, kv_heads=None, pool=1):
     kv_heads = heads if kv_heads is None else kv_heads
-    selected = slice_plan(tokens, heads, dim, kv_heads)
+    selected = slice_plan(tokens, heads, dim, kv_heads, pool)
     full_bytes = heads * tokens * tokens * 4
-    full = full_bytes <= FULL_ATTENTION_MAX_BYTES
+    full = full_bytes <= FULL_ATTENTION_MAX_BYTES * pool
     if full:
         selected.update(query_rows=tokens, head_groups=1, slice_working_bytes=None)
     return dict(tokens=tokens, heads=heads, kv_heads=kv_heads, head_dim=dim,
@@ -99,7 +99,7 @@ def sliced_basic(q, k, v, heads, rows, basic, *, mask=None, skip_reshape=True, e
     return output
 
 
-def install(clip):
+def install(clip, pool_bytes=None):
     """Return shared dispatch statistics and removable instance hooks.
 
     Counts are attempts/planned basic invocations, not successful completions.
@@ -113,16 +113,18 @@ def install(clip):
             or any((m.num_heads, m.head_dim) != (16, 72) for m in vision)
             or any((m.num_heads, m.num_kv_heads, m.head_dim) != (64, 8, 128) for m in language)):
         raise ValueError('Expected the pinned H3 vision and language attention geometry')
-    stats = dict(full_calls=0, sliced_calls=0, slices=0, plans=[],
-                 full_attention_max_bytes=FULL_ATTENTION_MAX_BYTES,
-                 slice_working_bytes=SLICE_WORKING_BYTES,
+    from .macos_encoder_linear import pool_scale
+    pool = pool_scale(pool_bytes)
+    stats = dict(pool_scale=pool, full_calls=0, sliced_calls=0, slices=0, plans=[],
+                 full_attention_max_bytes=FULL_ATTENTION_MAX_BYTES * pool,
+                 slice_working_bytes=SLICE_WORKING_BYTES * pool,
                  kernel='Comfy attention_basic: einsum, scale, additive mask, softmax, einsum; FP32')
     seen = set()
 
     def attention(q, k, v, heads, *, mask=None, skip_reshape=False, enable_gqa=False):
         if q.device.type != 'mps' or not skip_reshape:
             raise ValueError('Bounded native encoder attention requires MPS and pre-shaped heads')
-        selected = plan(q.shape[2], heads, q.shape[3], k.shape[1])
+        selected = plan(q.shape[2], heads, q.shape[3], k.shape[1], pool)
         stats['sliced_calls'] += 1
         stats['slices'] += ((q.shape[2] + selected['query_rows'] - 1) // selected['query_rows']
                            * selected['head_groups'])
@@ -136,7 +138,7 @@ def select(module, args, kwargs):
             x = args[0] if args else kwargs['x' if kind == 'vision' else 'hidden_states']
             tokens = x.shape[0] if kind == 'vision' else x.shape[1]
             selected = plan(tokens, module.num_heads, module.head_dim,
-                            module.num_heads if kind == 'vision' else module.num_kv_heads)
+                            module.num_heads if kind == 'vision' else module.num_kv_heads, pool)
             key = (kind, tokens)
             if key not in seen:
                 seen.add(key)
diff --git a/freevideo_engine/macos_encoder_linear.py b/freevideo_engine/macos_encoder_linear.py
index 19951706..7f900532 100644
--- a/freevideo_engine/macos_encoder_linear.py
+++ b/freevideo_engine/macos_encoder_linear.py
@@ -14,14 +14,29 @@
 FULL_LINEAR_MAX_BYTES = 128 * 2**20
 LINEAR_WORKING_BYTES = 128 * 2**20
 MLP_WORKING_BYTES = 256 * 2**20
+# The limits above were sized for the 3 GiB trial pool. A larger live pool
+# scales them by the same ratio; a smaller one keeps them as they are.
+REFERENCE_POOL_BYTES = 3 * 2**30
+# Output rows of one slice never drop below this, however large the assembled
+# output is: one row per slice means one decode, fence and cache release each.
+MIN_SLICE_ROWS = 256
 
 
-def linear_plan(tokens, inputs, outputs, itemsize=4, *, packed_columns=None, packed_rows=None):
+def pool_scale(pool_bytes):
+    if pool_bytes is None:
+        return 1
+    if type(pool_bytes) is not int or pool_bytes < 1:
+        raise ValueError('Expected a positive encoder pool size')
+    return max(1, pool_bytes // REFERENCE_POOL_BYTES)
+
+
+def linear_plan(tokens, inputs, outputs, itemsize=4, *, packed_columns=None, packed_rows=None, pool=1):
     """Bound full output + decoded slab + packed/scales + slab result + bias.
 
     Packed geometry can include padding on the input axis. A result whose full
-    output alone exceeds the budget still uses one row and reports that minimum
-    explicitly; the parent MLP's token plan avoids this for the H3 workload.
+    output alone exceeds the budget still slices at least MIN_SLICE_ROWS output
+    rows (or all outputs, if fewer) and reports the over-budget estimate in
+    slice_working_bytes; minimum_slice_bytes is the theoretical one-row bound.
     """
     if min(tokens, inputs, outputs, itemsize) < 1:
         raise ValueError('Expected positive linear dimensions')
@@ -34,8 +49,9 @@ def linear_plan(tokens, inputs, outputs, itemsize=4, *, packed_columns=None, pac
     upload = 0 if packed_columns is None else columns // 2 + columns // 16 * 4
     per_row = columns * itemsize + upload + (tokens + 1) * itemsize
     scale = 4 if packed_columns is not None else 0
-    rows = min(outputs, max(1, (LINEAR_WORKING_BYTES - output - scale) // per_row))
-    full = weight + output <= FULL_LINEAR_MAX_BYTES
+    floor = min(outputs, MIN_SLICE_ROWS)
+    rows = min(outputs, max(floor, (LINEAR_WORKING_BYTES * pool - output - scale) // per_row))
+    full = weight + output <= FULL_LINEAR_MAX_BYTES * pool
     return dict(tokens=tokens, inputs=inputs, outputs=outputs, itemsize=itemsize,
                 decoded_weight_bytes=weight, output_bytes=output,
                 full_weight_output_bytes=weight + output,
@@ -44,7 +60,7 @@ def linear_plan(tokens, inputs, outputs, itemsize=4, *, packed_columns=None, pac
                 minimum_slice_bytes=output + per_row + scale)
 
 
-def mlp_plan(tokens, hidden, intermediate, itemsize=4):
+def mlp_plan(tokens, hidden, intermediate, itemsize=4, pool=1):
     """Three simultaneous intermediate arrays plus two hidden-sized row arrays.
 
     The full assembled MLP output stays live across chunks. Linear workspace is
@@ -54,8 +70,9 @@ def mlp_plan(tokens, hidden, intermediate, itemsize=4):
         raise ValueError('Expected positive MLP dimensions')
     output = tokens * hidden * itemsize
     per_row = (3 * intermediate + 2 * hidden) * itemsize
-    full = tokens * per_row <= MLP_WORKING_BYTES
-    rows = min(tokens, max(1, (MLP_WORKING_BYTES - output) // per_row))
+    budget = MLP_WORKING_BYTES * pool
+    full = tokens * per_row <= budget
+    rows = min(tokens, max(1, (budget - output) // per_row))
     return dict(tokens=tokens, hidden=hidden, intermediate=intermediate, itemsize=itemsize,
                 mode='full' if full else 'sliced', token_rows=tokens if full else rows,
                 full_activation_bytes=tokens * per_row,
@@ -121,16 +138,19 @@ def remove(self):
             setattr(self.module, self.attribute, self.original)
 
 
-def install(linears, layers, *, decoder, quantized_type, backend, guard, release, reader_stats):
+def install(linears, layers, *, decoder, quantized_type, backend, guard, release, reader_stats,
+            pool_bytes=None):
     """Select before the native forward looks up its projection/MLP callable.
 
     Below the thresholds do not assign anything: the fourth-version weight
     reader and vendor MLP function objects remain intact. MLP selection occurs
     at the parent layer because Module reads its own forward before pre-hooks.
     """
-    stats = dict(linear_sliced_calls=0, linear_slices=0, mlp_sliced_calls=0, mlp_slices=0,
-                 plans=[], full_linear_max_bytes=FULL_LINEAR_MAX_BYTES,
-                 linear_working_bytes=LINEAR_WORKING_BYTES, mlp_working_bytes=MLP_WORKING_BYTES)
+    pool = pool_scale(pool_bytes)
+    stats = dict(pool_scale=pool, linear_sliced_calls=0, linear_slices=0, mlp_sliced_calls=0, mlp_slices=0,
+                 plans=[], full_linear_max_bytes=FULL_LINEAR_MAX_BYTES * pool,
+                 linear_working_bytes=LINEAR_WORKING_BYTES * pool,
+                 mlp_working_bytes=MLP_WORKING_BYTES * pool)
     handles, seen = [], set()
 
     def record(kind, selected):
@@ -151,7 +171,8 @@ def before(module, args, kwargs):
             columns = module.weight._qdata.shape[1] * 2 if quantized else None
             selected = linear_plan(math.prod(x.shape[:-1]), module.in_features,
                                    module.out_features, x.element_size(), packed_columns=columns,
-                                   packed_rows=module.weight._qdata.shape[0] if quantized else None)
+                                   packed_rows=module.weight._qdata.shape[0] if quantized else None,
+                                   pool=pool)
             record('linear', selected)
             if selected['mode'] == 'full':
                 return None
@@ -210,7 +231,7 @@ def before(module, args, kwargs):
             nonlocal active
             x = args[0] if args else kwargs['x']
             selected = mlp_plan(math.prod(x.shape[:-1]), mlp.down_proj.out_features,
-                                mlp.down_proj.in_features, x.element_size())
+                                mlp.down_proj.in_features, x.element_size(), pool)
             record('mlp', selected)
             if selected['mode'] == 'full':
                 return None
diff --git a/freevideo_engine/managed.py b/freevideo_engine/managed.py
index 7954b116..25ef5a8a 100644
--- a/freevideo_engine/managed.py
+++ b/freevideo_engine/managed.py
@@ -9,7 +9,7 @@
 from .agent import enabled as agent_enabled
 from .bootstrap import DEFAULT_ROOT
 from . import __version__
-from .locking import runtime_lock, LOCK_ENV
+from .locking import runtime_lock, LOCK_ENV, ENGINE_BUSY
 from . import processes
 from .system import windows
 
@@ -117,7 +117,7 @@ def _main(argv=None):
                 or (command == 'doctor' and '--probe' not in tail) or '--help' in tail or '-h' in tail)
     snapshot = json.loads(config.read_text(encoding='utf-8'))
     lock_path = os.environ.get('FREEVIDEO_LOCK_PATH', str(Path(snapshot['root']) / 'engine.lock'))
-    with nullcontext() if readonly else runtime_lock(lock_path) as descriptor:
+    with nullcontext() if readonly else runtime_lock(lock_path, wait_seconds=3 if command == 'generate' else 0) as descriptor:
         machine = json.loads(config.read_text(encoding='utf-8'))
         if machine != snapshot:
             raise ValueError('Installation changed while starting. Rerun the command.')
@@ -140,7 +140,7 @@ def main(argv=None):
     try:
         return _main(argv)
     except BlockingIOError:
-        print('Setup, testing or generation is already using this installation/GPU lock. Retry after it finishes.', file=sys.stderr)
+        print(ENGINE_BUSY, file=sys.stderr)
         return 1
     except KeyboardInterrupt:
         print('Cancelled. Workers stopped; existing outputs and reports retained.', file=sys.stderr)
diff --git a/freevideo_engine/model_guidance.py b/freevideo_engine/model_guidance.py
index de39f4e3..fd93a263 100644
--- a/freevideo_engine/model_guidance.py
+++ b/freevideo_engine/model_guidance.py
@@ -13,7 +13,7 @@
 import uuid
 from urllib.parse import urlsplit
 
-from . import network
+from . import __version__, network
 from .long_paths import io_path, plain_errors
 from .model_status import FAMILIES, NAMES, family
 
@@ -129,7 +129,7 @@ def download_files(entries):
     return result
 
 
-def download_view(plan, zh=False, *, decimal_sizes=False):
+def download_view(plan, zh=False, *, decimal_sizes=False, families=LIST_FAMILIES, extra_copy=None):
     from .launcher_copy import size_text
     def size(value):
         if value < (10**6 if decimal_sizes else 2**20):
@@ -137,9 +137,10 @@ def size(value):
             return ('不到 1 ' if zh else 'under 1 ') + unit
         return size_text(value, decimal_sizes=decimal_sizes)
     copy = download_copy(zh)
+    copy.update({key: value[bool(zh)] for key, value in (extra_copy or {}).items()})
     rows = plan.get('download_files', [])
     groups = []
-    for group in LIST_FAMILIES:
+    for group in families:
         files = [dict(row, name=row['file'], size=size(row['bytes'])) for row in rows if row['group'] == group]
         amount = sum(row['bytes'] for row in files)
         groups.append(dict(id=group, title=copy[group], files=files, count=len(files), bytes=amount,
@@ -196,6 +197,10 @@ def export_download_list(plan, folder, source='auto'):
                  'SHA256SUMS': '\n'.join(sums) + ('\n' if sums else ''),
                  'urls.txt': '\n'.join(urls) + ('\n' if urls else ''),
                  'README.txt': DOWNLOAD_COPY['readme'][1] + '\n' + DOWNLOAD_COPY['readme'][0]}
+    return publish_export(folder, documents)
+
+
+def publish_export(folder, documents, folder_name=EXPORT_FOLDER, generator=GENERATOR):
     # Finish every file before anything appears or is replaced. No payload is downloaded.
     stage = None
     try:
@@ -210,7 +215,7 @@ def export_download_list(plan, folder, source='auto'):
         # Two launchers exporting into one folder take turns: the five files
         # always come from the same export.
         with export_lock(folder):
-            target, earlier = export_target(folder)
+            target, earlier = export_target(folder, folder_name, generator)
             if earlier:
                 replace_export(stage, target)
             else:
@@ -224,6 +229,109 @@ def export_download_list(plan, folder, source='auto'):
     return str(target)
 
 
+# The optional image model has its own list: files laid out as FreeVideo keeps them, one export folder of its own.
+IMAGE_FOLDER = 'image_gen'
+IMAGE_EXPORT_FOLDER = 'FreeVideo-image-model-download-list'
+IMAGE_GENERATOR = 'FreeVideo image model download list'
+IMAGE_COPY = {
+    'image': NAMES['image'],
+    'guide': ('Download the files with the exported list and place them as its README.txt describes. '
+              'Back in the launcher, choose Resume on Image generation models; FreeVideo checks every file before using it.',
+              '用导出的清单下载这些文件，按清单里 README.txt 的说明放好。'
+              '回到启动器，在“图像生成模型”处点“继续”，FreeVideo 校验每个文件后才会使用。'),
+    # The list needs no re-check, so the shared "Check again" text does not apply.
+    'error_changed': ('The download list could not be generated. Restart the launcher and try again.',
+                      '下载清单未能生成，请重启启动器后再试。'),
+    # {model_folder} is replaced with the folder that holds image_gen.
+    'readme': (
+        'FreeVideo image generation models download list\n'
+        '\n'
+        'Run this command in this folder:\n'
+        'aria2c --input-file=aria2.txt --continue=true --check-integrity=true\n'
+        '\n'
+        'You can also import urls.txt into another download tool. urls.txt uses the download source chosen in Settings (Automatic uses Hugging Face).\n'
+        '\n'
+        'When the download finishes, copy the image_gen folder inside payload into the folder below, merging it with any image_gen folder already there:\n'
+        '{model_folder}\n'
+        'Then return to the launcher and choose "Resume" on Image generation models. '
+        'FreeVideo checks every file by its checksum and downloads again any file that is missing or damaged.\n'
+        '\n'
+        'SHA256SUMS lists the checksum of every file. This list covers only the image generation models; the video model list is exported separately.\n',
+        'FreeVideo 图像生成模型下载清单\n'
+        '\n'
+        '在这个文件夹里运行下面的命令：\n'
+        'aria2c --input-file=aria2.txt --continue=true --check-integrity=true\n'
+        '\n'
+        '也可以把 urls.txt 导入其他下载工具。urls.txt 用的是设置里选中的下载源（选“自动”时用 Hugging Face）。\n'
+        '\n'
+        '下载完成后，把 payload 里的 image_gen 文件夹复制到下面这个文件夹，和其中已有的 image_gen 合并：\n'
+        '{model_folder}\n'
+        '然后回到启动器，在“图像生成模型”处点“继续”。FreeVideo 会按校验值逐个检查文件，缺少或损坏的文件会重新下载。\n'
+        '\n'
+        'SHA256SUMS 列出每个文件的校验值。这份清单只含图像生成模型，视频模型的清单另外导出。\n'),
+}
+
+
+def image_files():
+    """The image model's files as list rows; the catalog ships with the engine and is validated on load."""
+    from .image_gen import image_pack
+    result = []
+    for row in image_pack.rows():
+        ids = {source['id'] for source in row['sources']}
+        link = lambda name: network.model_url(row, name, env={}) if name in ids else ''
+        result.append(dict(group='image', repo=row['repo'], hf_revision=row['revision'],
+            modelscope_repo='', modelscope_revision='', file=row['file'], bytes=row['bytes'], sha256=row['sha256'],
+            relative_path='/'.join(('payload', IMAGE_FOLDER, row['file'])),
+            hf_url=link('official'), hf_mirror_url=link('hf-mirror'), modelscope_url=link('modelscope')))
+    return result
+
+
+def image_download_view(zh=False, *, decimal_sizes=False):
+    view = download_view(dict(download_files=image_files()), zh, decimal_sizes=decimal_sizes,
+                         families=('image',), extra_copy=IMAGE_COPY)
+    return dict(view, scope='image', reused='')
+
+
+def export_image_download_list(folder, source='auto', model_folder=None):
+    """Only the image model, with every source; nothing of the video list changes.
+
+    model_folder is where the README tells the user to put image_gen: the folder the
+    launcher looks in, or the installed model folder when none is given.
+    """
+    folder = Path(folder).expanduser().resolve()
+    if not folder.is_dir():
+        raise DownloadListError('folder')
+    if source not in ('auto', 'official', 'hf-mirror', 'modelscope'):
+        raise DownloadListError('changed')
+    if model_folder is None:
+        from .paths import installed_model_root
+        model_folder = installed_model_root()
+    rows = image_files()
+    aria, sums, urls, outputs = [], [], [], set()
+    for row in rows:
+        relative = row['relative_path']
+        path = PurePosixPath(relative)
+        if (path.is_absolute() or any(part in ('', '.', '..') for part in relative.split('/'))
+                or any(char in relative for char in '\\:\r\n\t\x00') or relative in outputs
+                or not re.fullmatch('[0-9a-f]{64}', row['sha256'] or '') or not row['hf_url']
+                or type(row['bytes']) is not int or row['bytes'] <= 0):
+            raise DownloadListError('changed')
+        outputs.add(relative)
+        aria.extend(['\t'.join(url for url in (row['hf_url'], row['hf_mirror_url'], row['modelscope_url']) if url),
+                     '  dir=' + path.parent.as_posix(), '  out=' + path.name,
+                     '  checksum=sha-256=' + row['sha256']])
+        sums.append(row['sha256'] + '  ' + relative)
+        chosen = {'official': 'hf_url', 'auto': 'hf_url', 'hf-mirror': 'hf_mirror_url', 'modelscope': 'modelscope_url'}[source]
+        urls.append(row[chosen] or row['hf_url'])
+    manifest = dict(schema_version=1, generator=IMAGE_GENERATOR, engine_version=__version__, source=source,
+                    total_bytes=sum(row['bytes'] for row in rows), files=rows)
+    documents = {'manifest.json': json.dumps(manifest, ensure_ascii=False, indent=2) + '\n',
+                 'aria2.txt': '\n'.join(aria) + '\n', 'SHA256SUMS': '\n'.join(sums) + '\n',
+                 'urls.txt': '\n'.join(urls) + '\n',
+                 'README.txt': (IMAGE_COPY['readme'][1] + '\n' + IMAGE_COPY['readme'][0]).replace('{model_folder}', str(model_folder))}
+    return publish_export(folder, documents, IMAGE_EXPORT_FOLDER, IMAGE_GENERATOR)
+
+
 def export_lock(folder, wait=30.):
     """Hold the cross-process lock of one export folder; waits for another export.
 
@@ -245,7 +353,7 @@ def export_lock(folder, wait=30.):
             time.sleep(.1)
 
 
-def earlier_export(path):
+def earlier_export(path, generator=GENERATOR):
     """A folder this launcher exported before, identified by its manifest."""
     try:
         if path.is_symlink() or not path.is_dir():
@@ -253,18 +361,18 @@ def earlier_export(path):
         value = json.loads((path / 'manifest.json').read_text(encoding='utf-8'))
     except (OSError, ValueError):
         return False
-    return isinstance(value, dict) and value.get('generator') == GENERATOR
+    return isinstance(value, dict) and value.get('generator') == generator
 
 
-def export_target(folder):
+def export_target(folder, name=EXPORT_FOLDER, generator=GENERATOR):
     """(folder, earlier): our earlier export is refreshed; anything else is left alone."""
     for index in range(1, 100):
-        path = folder / (EXPORT_FOLDER if index == 1 else '%s-%d' % (EXPORT_FOLDER, index))
+        path = folder / (name if index == 1 else '%s-%d' % (name, index))
         if not os.path.lexists(path):
             return path, False
-        if earlier_export(path):
+        if earlier_export(path, generator):
             return path, True
-    raise OSError(errno.EEXIST, 'Too many folders named ' + EXPORT_FOLDER)
+    raise OSError(errno.EEXIST, 'Too many folders named ' + name)
 
 
 def replace_export(stage, target):
diff --git a/freevideo_engine/model_repair.py b/freevideo_engine/model_repair.py
index 8e3ef9f7..c0b84f01 100644
--- a/freevideo_engine/model_repair.py
+++ b/freevideo_engine/model_repair.py
@@ -1,4 +1,4 @@
-"""Review and retain local files that differ from the model FreeVideo tests."""
+"""Review and retain local files that differ from the tested model."""
 import ctypes
 import errno
 import json
@@ -16,15 +16,15 @@
 
 # Shared copy for setup, the launcher and the browser plan.
 REVIEW = (
-    '{file} ({size}) differs from the version FreeVideo tests: it will be renamed to {keep_as} and kept, '
+    '{file} ({size}) differs from the tested version: it will be renamed to {keep_as} and kept, '
     'and the tested version ({expected_size}) will be downloaded.',
     '{file}（{size}）与 FreeVideo 测试过的版本不同：将改名为 {keep_as} 保留，'
     '再下载测试过的版本（{expected_size}）。',
 )
 PROGRESS = ('Keeping the original file: ', '正在保留原文件：')
-FAILURE_TITLE = ('An existing model differs from the version FreeVideo tests', '已有模型与 FreeVideo 测试过的版本不同')
+FAILURE_TITLE = ('An existing model differs from the tested version', '已有模型与 FreeVideo 测试过的版本不同')
 FAILURE = (
-    '{file} differs from the version FreeVideo tests and was left unchanged. Click {entry} again: '
+    '{file} differs from the tested version and was left unchanged. Click {entry} again: '
     'the confirmation page will list this file, and after you confirm, it will be renamed and kept, '
     'and the tested version will be downloaded.',
     '{file} 与 FreeVideo 测试过的版本不同，这次没有改动它。请再次点击“{entry}”：'
@@ -35,7 +35,7 @@
 MISMATCH_PATH = 'MODEL_MISMATCH_PATH'
 # Where the launcher offers no Repair: the first sentence of FAILURE, then one way out
 # for each reason (failure_details.WITHOUT_REPAIR).
-UNCHANGED = ('{file} differs from the version FreeVideo tests and was left unchanged.',
+UNCHANGED = ('{file} differs from the tested version and was left unchanged.',
              '{file} 与 FreeVideo 测试过的版本不同，这次没有改动它。')
 BUSY = ('When the current task ends, a Repair button appears here; after you click it and confirm, '
         'FreeVideo renames and keeps this file, then downloads the tested version.',
@@ -44,7 +44,7 @@
            'then import the Prompt model package (FreeVideo-文本编码器.zip) again on the Installation page.',
            '离线包安装暂时不能一键换回。请把下面这个文件移走或改名，再在“安装”页重新导入“提示词模型（FreeVideo-文本编码器.zip）”包。')
 FIRST_INSTALL = (
-    '{file} differs from the version FreeVideo tests and was left unchanged. Click Retry installation below: '
+    '{file} differs from the tested version and was left unchanged. Click Retry installation below: '
     'the confirmation page will list this file, and after you confirm, it is renamed and kept, '
     'and the tested version is downloaded.',
     '{file} 与 FreeVideo 测试过的版本不同，这次没有改动它。请点击下方的“重试安装”：'
diff --git a/freevideo_engine/modern_launcher.py b/freevideo_engine/modern_launcher.py
index 744bf27c..fac53883 100644
--- a/freevideo_engine/modern_launcher.py
+++ b/freevideo_engine/modern_launcher.py
@@ -191,13 +191,17 @@ def browse(self, key):
             if path:
                 self.invoke(session.add_folder, path) if key == 'model_dirs' else self.invoke(session.edit, key, path)
 
+        @Slot(str)
+        def setDownloadScope(self, scope):
+            self.invoke(session.set_download_scope, scope)
+
         @Slot()
         def prepareDownloadList(self):
             self.invoke(session.prepare_download_list)
 
         @Slot()
         def exportDownloadList(self):
-            if session.controller.busy:
+            if session.controller.busy and session.download_scope != 'image':
                 return
             path = QFileDialog.getExistingDirectory(None, session.downloads_view()['copy']['export'], '', QFileDialog.ShowDirsOnly)
             if path:
@@ -294,10 +298,13 @@ def link(self, value):
         @Slot(str, str)
         def reportProblem(self, error, title):
             from . import __version__
-            from .launcher_issue import issue_url
+            from .launcher_issue import environment_line, issue_url, machine_hardware, product_version
             from .windows_ux import open_browser
-            url = issue_url(error, session.installed_product_version() or __version__, title,
-                            zh=session.language.startswith('zh'))
+            # The release (0.4.x) is named, not the build number; only fields that carry no path or user name travel.
+            version = session.installed_product_version() or product_version(__version__)
+            engine = session.installation_engine()
+            url = issue_url(error, version, title, zh=session.language.startswith('zh'),
+                            environment=environment_line(version, machine_hardware(engine) if engine else {}))
             self.invoke(open_browser, url)
 
         @Slot()
diff --git a/freevideo_engine/network.py b/freevideo_engine/network.py
index 3fdd6a6a..30c379a7 100644
--- a/freevideo_engine/network.py
+++ b/freevideo_engine/network.py
@@ -15,6 +15,7 @@
 import re
 import shutil
 import subprocess
+import tempfile
 from types import SimpleNamespace
 from . import execution_paths
 from . import subprocess_output
@@ -23,7 +24,7 @@
 import time
 from urllib.parse import quote, unquote, urlencode, urljoin, urlsplit
 from . import proxy
-from .long_paths import io_path, plain_errors
+from .long_paths import io_path, is_windows, path_chars, plain, plain_errors
 from .system import curl_executable, missing_curl_message, windows
 
 
@@ -833,7 +834,44 @@ def _redirect_fields(text):
     return status.strip(), destination.strip()
 
 
-def download(candidates, path, expected, progress=None, *, network=None, env=None, size=None,
+DOWNLOAD_SIBLING_CHARS = 21  # '.partial.source.json' beside the file
+
+
+def move_file(source, destination):
+    """Move one file by the long-path entry; across volumes copy it and remove the source."""
+    source, destination = io_path(source), io_path(destination)
+    try:
+        source.replace(destination)
+    except OSError:
+        shutil.copy2(source, destination)
+        source.unlink()
+
+
+def download(candidates, path, expected, *args, **options):
+    """Fetch into `path`; where curl cannot name it (Windows, beyond 259 characters), in a short folder.
+
+    The file and what resumes it (partial bytes, source record) go to the short
+    folder first, and whatever is left there comes back, also after a failure.
+    """
+    path = Path(path)
+    scratch = None
+    if is_windows() and path_chars(plain(path)) + DOWNLOAD_SIBLING_CHARS > GIT_PATH_LIMIT:
+        scratch = git_scratch_folder()
+    if scratch is None:
+        return download_here(candidates, path, expected, *args, **options)
+    try:
+        io_path(path.parent).mkdir(parents=True, exist_ok=True)
+        for entry in io_path(path.parent).iterdir():
+            if entry.name == path.name or entry.name.startswith(path.name + '.partial'):
+                move_file(entry, scratch / entry.name)
+        return download_here(candidates, scratch / path.name, expected, *args, **options)
+    finally:
+        for entry in io_path(scratch).iterdir():
+            move_file(entry, path.parent / entry.name)
+        remove_tree(scratch)
+
+
+def download_here(candidates, path, expected, progress=None, *, network=None, env=None, size=None,
              algorithm='sha256', git_blob=False, headers_for=None, stall_seconds=None, category='download',
              low_speed_limit=1024, max_seconds=None, cycles=2, slow_seconds=None, keep_partial=False,
              validate_url=None, verify=None, fallbacks=None):
@@ -847,6 +885,13 @@ def download(candidates, path, expected, progress=None, *, network=None, env=Non
     path = Path(path)
     updating = category == 'launcher-update'
     integrity_failed = False
+    attempts = {}
+
+    def note_failure(source, route, reason, status=''):
+        # The last reason each source and route gave, for the final error.
+        attempts.pop((source, route), None)  # a repeat moves to the end, so the last 8 are the latest
+        attempts[source, route] = reason + (' (HTTP %s)' % status if str(status).isdigit() and status != '000' else '')
+
     failed_sources = set()
     transferred, high_water, last_rate = 0, 0, 0.
     transfer_samples = deque([(time.monotonic(), 0)])
@@ -1103,6 +1148,7 @@ def remember_source():
                         integrity_failed = True
                         failed_sources.add(name)
                     source_health(network, category, name, False)
+                    note_failure(name, route, 'response-size-exceeded')
                     event(network, category=category, source=name, file=path.name, action='fallback',
                           method='curl', reason='response-size-exceeded', observed_bytes=oversized,
                           expected_bytes=size, resume_bytes=offset)
@@ -1130,10 +1176,12 @@ def remember_source():
                             continue
                         failed_sources.add(name)
                     source_health(network, category, name, False)
+                    note_failure(name, route, 'integrity-mismatch')
                     event(network, category=category, source=name, file=path.name, action='fallback', reason='integrity-mismatch')
                     break
                 if offset and (process.returncode == 33 or process.returncode == 22 and status == '416') and restart == 0:
                     if model or keep_partial:
+                        note_failure(name, route, 'source-does-not-support-range')
                         event(network, category=category, source=name, file=path.name,
                               action='resume-refused', resume_bytes=offset, reason='source-does-not-support-range')
                         break
@@ -1154,6 +1202,7 @@ def remember_source():
                 if updating and stalled:
                     failed_sources.add(name)
                 reason = 'no-progress' if stalled else 'slow-transfer' if slow else 'faster-source' if faster else 'curl-' + str(process.returncode)
+                note_failure(name, route, reason, status)
                 alternatives = route_order(network, category, name, env)
                 if alternatives.index(route) + 1 < len(alternatives):
                     event(network, category=category, source=name, route=alternatives[alternatives.index(route)+1],
@@ -1169,7 +1218,12 @@ def remember_source():
             event(network, category=category, source=name, file=path.name, action='restart-approved',
                   retained_bytes=partial.stat().st_size, reason='resume-unavailable')
             retain_partial(partial, 'restart-approved')
-    error = DownloadError('Every download source failed for %s. Check your network or choose another connection mode in Downloads. Partial files retained' % path.name)
+    message = 'Every download source failed for %s. Check your network or choose another connection mode in Downloads. Partial files retained' % path.name
+    if attempts:
+        message += '. Attempts: ' + '; '.join('%s/%s: %s' % (source, route, reason)
+                                              for (source, route), reason in list(attempts.items())[-8:])
+    error = DownloadError(message)
+    error.attempts = [dict(source=source, route=route, reason=reason) for (source, route), reason in attempts.items()]
     error.integrity = integrity_failed
     raise error
 
@@ -1202,6 +1256,109 @@ def git_parameter(value):
     return "'" + value.replace("'", "'\\''").replace('!', "'\\!'") + "'"
 
 
+# Git for Windows cannot find a repository from a working folder deeper than
+# the ordinary path limit, with or without core.longpaths, and a junction does
+# not help: the folder it reports is the real one. Git therefore works in a short
+# real scratch folder and the finished checkout is moved into place.
+GIT_FOLDER_RESERVE = 140
+GIT_SCRATCH_BASE_CHARS = 80
+GIT_PATH_LIMIT = 259
+
+
+def git_folder_too_deep(path):
+    return is_windows() and path_chars(plain(path)) + GIT_FOLDER_RESERVE > GIT_PATH_LIMIT
+
+
+def git_scratch_folder():
+    """A new short real folder for Git to work in, or None where none can be made."""
+    bases = [tempfile.gettempdir()]
+    for key in ('ProgramData', 'LOCALAPPDATA'):
+        if os.environ.get(key):
+            bases.append(os.path.join(os.environ[key], 'FV'))
+    for base in bases:
+        if path_chars(plain(base)) > GIT_SCRATCH_BASE_CHARS:
+            continue
+        try:
+            os.makedirs(base, exist_ok=True)
+            return Path(tempfile.mkdtemp(prefix='fvgit-', dir=base))
+        except OSError:
+            continue
+    return None
+
+
+def remove_tree(path):
+    def writable(function, name, _):
+        os.chmod(name, 0o700)
+        function(name)
+    shutil.rmtree(io_path(path), onerror=writable)
+
+
+def relocate(source, destination):
+    """Move a finished folder; where a rename cannot (another volume) copy it."""
+    source, destination = io_path(source), io_path(destination)
+    try:
+        source.rename(destination)
+        return
+    except OSError:
+        pass
+    try:
+        shutil.copytree(source, destination, symlinks=True)
+    except BaseException:
+        if destination.exists():
+            remove_tree(destination)
+        raise
+    remove_tree(source)
+
+
+def repository_state(path, git, env, revision='HEAD', **options):
+    """What `revision` is and what local changes the checkout at `path` has."""
+    path = Path(path)
+    prefix = [git, '-c', 'core.longpaths=true']
+    hidden = dict(hidden_console(), **options)
+    if not git_folder_too_deep(path):
+        head = subprocess_output.output_external([*prefix, '-C', str(path), 'rev-parse', revision], env=env, **hidden).strip()
+        dirty = subprocess_output.output_external([*prefix, '-C', str(path), 'status', '--porcelain', '--untracked-files=no'], env=env, **hidden)
+        return head, dirty
+    # Name the repository instead of entering it. If Git cannot read it this
+    # way, HEAD is read from the file and local changes cannot be checked.
+    git_dir = '--git-dir=' + str(path / '.git')
+    folder = tempfile.gettempdir()
+    try:
+        head = subprocess_output.output_external([*prefix, git_dir, 'rev-parse', revision], cwd=folder, env=env, **hidden).strip()
+        dirty = subprocess_output.output_external([*prefix, git_dir, 'diff-index', '--cached', '--name-only', 'HEAD'], cwd=folder, env=env, **hidden)
+    except (subprocess.CalledProcessError, OSError):
+        # Git also refuses a --git-dir this long, so read what is asked from the files.
+        head = stored_head(path)
+        if revision == 'HEAD':
+            return head, ''
+        tree = stored_tree(path, head) if revision == 'HEAD^{tree}' else None
+        if tree is None:
+            raise
+        return tree, ''
+    return head, dirty
+
+
+def stored_head(path):
+    return io_path(path / '.git' / 'HEAD').read_text(encoding='utf-8').strip()
+
+
+def stored_tree(path, commit):
+    """The tree named by a loose commit object, or None (packed or missing objects)."""
+    import zlib
+    try:
+        if commit.startswith('ref:'):
+            name = commit.split(None, 1)[1]
+            commit = io_path(path / '.git' / name).read_text(encoding='utf-8').strip()
+        raw = zlib.decompress(io_path(path / '.git' / 'objects' / commit[:2] / commit[2:]).read_bytes())
+    except (OSError, ValueError, IndexError, zlib.error):
+        return None
+    header, _, body = raw.partition(b'\0')
+    first = body.split(b'\n', 1)[0]
+    if header.startswith(b'commit ') and first.startswith(b'tree ') and len(first) == 45:
+        return first[5:].decode('ascii')
+    return None
+
+
 def clone(target, url, commit, *, sparse=None, run=None, network=None, env=None):
     """Fetch only the pinned commit, atomically; never reset existing checkouts."""
     target = Path(target)
@@ -1209,25 +1366,24 @@ def clone(target, url, commit, *, sparse=None, run=None, network=None, env=None)
     env = proxy_environment(env)
     env.update(GIT_TERMINAL_PROMPT='0', GIT_HTTP_LOW_SPEED_LIMIT='1024', GIT_HTTP_LOW_SPEED_TIME='20')
     git = env.get('FREEVIDEO_GIT') or 'git'
-    version = git_version(git, env) if not target.exists() else None
+    version = git_version(git, env) if not io_path(target).exists() else None
     if version and version < MIN_GIT:
         raise RuntimeError(old_git_message(version))
     def execute(label, command):
-        command = [git, *command[1:]]
+        command = [git, '-c', 'core.longpaths=true', *command[1:]]
         if run:
             return run(label, command, env=env)
         result = subprocess_output.run_external(list(map(str, command)), env=env, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, **hidden_console())
         if result.returncode:
             raise RuntimeError(result.stdout[-2500:])
-    if target.exists():
-        if not (target / '.git').exists():
+    if io_path(target).exists():
+        if not io_path(target / '.git').exists():
             raise ValueError('Existing dependency is not a Git checkout of the pinned source: ' + str(target))
-        head = subprocess_output.output_external([git, '-C', str(target), 'rev-parse', 'HEAD'], env=env, **hidden_console()).strip()
-        dirty = subprocess_output.output_external([git, '-C', str(target), 'status', '--porcelain', '--untracked-files=no'], env=env, **hidden_console())
+        head, dirty = repository_state(target, git, env)
         if head != commit or dirty:
             raise ValueError('Existing dependency differs from the pinned clean source: ' + str(target))
         return target
-    target.parent.mkdir(parents=True, exist_ok=True)
+    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
     for name, endpoint, route in [(name, endpoint, route) for name, endpoint in candidates
@@ -1235,19 +1391,21 @@ def execute(label, command):
         env = route_environment(network, 'git', name, configured_env, route)
         temporary = target.with_name(target.name + '.partial-' + str(time.time_ns()))
         event(network, category='git', source=name, route=route, file=target.name, action='attempt', commit=commit)
+        scratch = git_scratch_folder() if git_folder_too_deep(temporary) else None
+        work = scratch / 'repo' if scratch else temporary
         try:
-            execute(target.name + '-init-' + name, ['git', 'init', '--quiet', temporary])
+            execute(target.name + '-init-' + name, ['git', 'init', '--quiet', work])
             # Never depend on a user's global CRLF conversion or long-path
             # settings when constructing the exact official patched tree.
-            execute(target.name + '-line-endings', ['git', '-C', temporary, 'config', 'core.autocrlf', 'false'])
-            execute(target.name + '-long-paths', ['git', '-C', temporary, 'config', 'core.longpaths', 'true'])
+            execute(target.name + '-line-endings', ['git', '-C', work, 'config', 'core.autocrlf', 'false'])
+            execute(target.name + '-long-paths', ['git', '-C', work, 'config', 'core.longpaths', 'true'])
             if sparse:
                 # The patterns file with cone mode off is what `sparse-checkout
                 # set --no-cone` writes; every Git reads it at checkout.
-                execute(target.name + '-sparse', ['git', '-C', temporary, 'config', 'core.sparseCheckout', 'true'])
-                execute(target.name + '-sparse-cone', ['git', '-C', temporary, 'config', 'core.sparseCheckoutCone', 'false'])
-                (temporary / '.git' / 'info').mkdir(exist_ok=True)
-                (temporary / '.git' / 'info' / 'sparse-checkout').write_text(''.join(p + '\n' for p in sparse), encoding='utf-8')
+                execute(target.name + '-sparse', ['git', '-C', work, 'config', 'core.sparseCheckout', 'true'])
+                execute(target.name + '-sparse-cone', ['git', '-C', work, 'config', 'core.sparseCheckoutCone', 'false'])
+                io_path(work / '.git' / 'info').mkdir(exist_ok=True)
+                io_path(work / '.git' / 'info' / 'sparse-checkout').write_text(''.join(p + '\n' for p in sparse), encoding='utf-8')
             # The URL reaches Git the way `git -c` hands settings to its own
             # children, in every Git version: never in process arguments, logs
             # or the repository's config, even if the fetch is killed.
@@ -1255,14 +1413,16 @@ def execute(label, command):
             env = dict(env, GIT_CONFIG_PARAMETERS=' '.join(filter(None, (
                 env.get('GIT_CONFIG_PARAMETERS', '').strip(), git_parameter('remote.origin.url=' + endpoint)))))
             try:
-                execute(target.name + '-fetch-' + name, ['git', '-C', temporary, '-c', 'fetch.fsckObjects=true', 'fetch', '--depth=1', '--no-tags', 'origin', commit])
+                execute(target.name + '-fetch-' + name, ['git', '-C', work, '-c', 'fetch.fsckObjects=true', 'fetch', '--depth=1', '--no-tags', 'origin', commit])
             finally:
                 env = original
-            execute(target.name + '-checkout', ['git', '-C', temporary, 'checkout', '--detach', commit])
-            actual = subprocess_output.output_external([git, '-C', str(temporary), 'rev-parse', 'HEAD'], env=env, **hidden_console()).strip()
+            execute(target.name + '-checkout', ['git', '-C', work, 'checkout', '--detach', commit])
+            actual = subprocess_output.output_external([git, '-c', 'core.longpaths=true', '-C', str(work), 'rev-parse', 'HEAD'], env=env, **hidden_console()).strip()
             if actual != commit:
                 raise ValueError('Downloaded source did not match its pinned commit')
-            temporary.rename(target)
+            if scratch:
+                relocate(work, temporary)
+            io_path(temporary).rename(io_path(target))
             source_health(network, 'git', name, True)
             route_health(network, 'git', name, route)
             event(network, category='git', source=name, route=route, file=target.name, action='verified', commit=commit)
@@ -1272,4 +1432,7 @@ def execute(label, command):
                 raise
             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))
diff --git a/freevideo_engine/policy.py b/freevideo_engine/policy.py
index 44965c3c..f52ee983 100644
--- a/freevideo_engine/policy.py
+++ b/freevideo_engine/policy.py
@@ -482,6 +482,9 @@ def int8_gpu_output_need(tokens, windows=False, full_tail=False):
         peak = at_low + (at_high - at_low) / (high - low) * (tokens - low)
     return int(peak * GiB) + margin
 MIN_GPU_RESERVE_GIB = .2
+# The smallest default RAM growth headroom a Windows foreground request may be
+# admitted with: the floor Linux already uses.
+MIN_RAM_RESERVE_GIB = 1.
 
 
 FF_STASH_WIDTH = 14336  # BF16 SwiGLU output channels kept for the per-tensor FP8 scale
@@ -888,6 +891,25 @@ def _choose(hardware: Hardware, *, vram_gib=None, ram_gib=None, attention='auto'
     reserve_gpu, reserve_ram = budget['gpu_system_reserve_bytes']/GiB, budget['ram_system_reserve_bytes']/GiB
     gpu_budget, ram_budget = budget['gpu_budget_bytes'], budget['ram_budget_bytes']
     desktop = hardware.system == 'Windows'
+    # The encoding phase already has a measured host floor. Do not discard a
+    # preloaded encoder because the later video stage needs a larger minimum;
+    # automatic_profile replans generation after conditioning has been saved.
+    ram_minimum = ENCODER_HOST_BYTES if stage == 'encoding' else 4*GiB
+    # Let a Windows foreground request spend RAM growth headroom before
+    # rejecting it, as the GPU's is spent below: the default 2 GiB is taken
+    # from live available memory, so a machine whose free RAM sits a few
+    # hundred MB under the minimum plus that reserve was refused outright.
+    # Only a request the default reserve alone refuses changes; admitted
+    # placements and an explicit reserve stay as they were. Neither the
+    # minimum nor the working-set estimate is relaxed. The reserve drops to
+    # the Linux floor, not to whatever leaves exactly the minimum: a
+    # 1344x768, 243-frame request on the smallest plan reached a 4.11 GiB
+    # working set against a 4.00 GiB budget and was stopped by the RAM guard
+    # when the VAE loaded, so the budget keeps every byte the floor leaves.
+    if (desktop and ram_reserve_gib is None and ram_budget < ram_minimum
+            and free_ram - int(MIN_RAM_RESERVE_GIB * GiB) >= ram_minimum):
+        reserve_ram = MIN_RAM_RESERVE_GIB
+        ram_budget = free_ram - int(reserve_ram * GiB)
     request_gpu_budget = gpu_budget
     gpu_budget -= placement_workspace
     # What this machine has actually held, for the weight cache only.
@@ -971,10 +993,6 @@ def _choose(hardware: Hardware, *, vram_gib=None, ram_gib=None, attention='auto'
         reserve_gpu = MIN_GPU_RESERVE_GIB
         request_gpu_budget = free_gpu - int(reserve_gpu * GiB)
         gpu_budget = request_gpu_budget - placement_workspace
-    # The encoding phase already has a measured host floor. Do not discard a
-    # preloaded encoder because the later video stage needs a larger minimum;
-    # automatic_profile replans generation after conditioning has been saved.
-    ram_minimum = ENCODER_HOST_BYTES if stage == 'encoding' else 4*GiB
     if gpu_budget < gpu_minimum or ram_budget < ram_minimum:
         raise ResourceBudgetError(hardware, gpu_total, ram_total, free_gpu, free_ram, reserve_gpu, reserve_ram,
                                   gpu_minimum_bytes=gpu_minimum + placement_workspace, ram_minimum_bytes=ram_minimum,
diff --git a/freevideo_engine/prepared_model.py b/freevideo_engine/prepared_model.py
index aaa4b260..1d4d719e 100644
--- a/freevideo_engine/prepared_model.py
+++ b/freevideo_engine/prepared_model.py
@@ -9,12 +9,16 @@
 from pathlib import Path, PurePosixPath
 import re
 import subprocess
+import time
 
 from . import execution_paths
 from .long_paths import io_path, resolved, plain, plain_errors
 from . import network
 
 CATALOG = Path(__file__).with_name('prepared_models.json')
+# (connect, total) seconds per pass; slow-network retries stop once the budget is spent.
+PREFLIGHT_PASSES = ((5, 12), (10, 30))
+PREFLIGHT_SECONDS = 150
 
 
 @plain_errors
@@ -129,17 +133,31 @@ def access_error(selection, networking, *, environ=None):
                 'Use --model-source source only if you want the larger original-model conversion.')
     manifest = next(r for r in files(selection) if r['file'].endswith('/manifest.json'))
     family = network.model_family(manifest)
+    attempts = []
     for source, url in network.model_urls(networking, manifest, env):
         headers = ['Authorization: Bearer ' + secret] if secret and source == 'official' else []
-        for route in network.route_order(networking, family, source, env):
+        attempts += [(source, route, url, headers) for route in network.route_order(networking, family, source, env)]
+    failures = {}
+    started = time.monotonic()
+    # A slow network gets one longer second chance, as the real download does.
+    for connect, total in PREFLIGHT_PASSES:
+        retry = []
+        for source, route, url, headers in attempts:
+            if time.monotonic() - started > PREFLIGHT_SECONDS:
+                break
             try:
                 result = execution_paths.run([
                     'curl', '--disable', '--silent', '--show-error', '--location', '--head',
-                    '--output', os.devnull, '--write-out', '%{http_code}', '--connect-timeout', '5',
-                    '--max-time', '12', '--config', '-'],
-                    input=network.curl_config(url, headers), capture_output=True, timeout=15,
+                    '--output', os.devnull, '--write-out', '%{http_code}', '--connect-timeout', str(connect),
+                    '--max-time', str(total), '--config', '-'],
+                    input=network.curl_config(url, headers), capture_output=True, timeout=total + 3,
                     env=network.route_environment(networking, family, source, env, route))
-            except (OSError, subprocess.TimeoutExpired):
+            except subprocess.TimeoutExpired:
+                failures[(source, route)] = 'timeout'
+                retry.append((source, route, url, headers))
+                continue
+            except OSError:
+                failures[(source, route)] = 'download tool unavailable'
                 continue
             status = result.stdout.decode('ascii', errors='ignore').strip()
             if result.returncode == 0 and status == '200':
@@ -148,10 +166,26 @@ def access_error(selection, networking, *, environ=None):
             if selection.get('private') and status in ('401', '403', '404'):
                 return ('Cannot access the pinned prepared model in ' + selection['repo'] +
                         '. Check your Hugging Face token and repository access; no large downloads started.')
-    return ('Could not reach the pinned prepared model sources. Check the proxy/network and retry; '
+            failures[(source, route)] = preflight_reason(result.returncode, status)
+            if result.returncode == 28:
+                retry.append((source, route, url, headers))
+        attempts = retry
+        if not attempts:
+            break
+    seen = ['%s (%s): %s' % (source, route, reason) for (source, route), reason in failures.items()]
+    return ('Could not reach the pinned prepared model sources' + (' [' + '; '.join(seen) + ']' if seen else '') +
+            '. Check the proxy/network and retry; '
             'existing files are retained. No automatic switch to the much larger source model.')
 
 
+def preflight_reason(returncode, status):
+    if returncode == 0 and re.fullmatch(r'[1-5]\d\d', status):
+        return 'HTTP ' + status
+    return {6: 'DNS lookup failed', 7: 'connection refused', 28: 'timeout', 35: 'TLS handshake failed',
+            47: 'too many redirects', 56: 'connection reset', 60: 'certificate not trusted'}.get(
+                returncode, 'download tool exit %d' % returncode)
+
+
 def verify_manifest(selection):
     from .adaln_assets import file_hash, asset_path
     row = next(r for r in files(selection) if r['file'].endswith('/manifest.json'))
diff --git a/freevideo_engine/prompt_vlm/assets.py b/freevideo_engine/prompt_vlm/assets.py
index a2f3177e..185e36d8 100644
--- a/freevideo_engine/prompt_vlm/assets.py
+++ b/freevideo_engine/prompt_vlm/assets.py
@@ -2,6 +2,7 @@
 import json
 from pathlib import Path
 import shutil
+import sys
 import threading
 from .. import disk_space
 
@@ -9,7 +10,31 @@
 
 
 def catalog():
-    return json.loads(Path(__file__).with_name('model.json').read_text(encoding='utf-8'))
+    """The model list shipped beside this file.
+
+    A frozen launcher also carries the engine source; its copy answers when the
+    unpacked one is gone (a cleaned temporary folder, antivirus, an interrupted
+    unpack). Nothing is remembered, so a restored file is used at once."""
+    name = 'model.json'
+    paths = [Path(__file__).with_name(name)]
+    bundle = getattr(sys, '_MEIPASS', None)
+    if bundle:
+        paths.append(Path(bundle) / 'engine-source' / 'freevideo_engine' / 'prompt_vlm' / name)
+    for path in paths:
+        try:
+            return json.loads(path.read_text(encoding='utf-8'))
+        except (OSError, ValueError) as error:
+            missing = error
+    raise missing
+
+
+def available():
+    """False while the model list cannot be read; nothing can be described or offered then."""
+    try:
+        catalog()
+    except (OSError, ValueError):
+        return False
+    return True
 
 
 def directory(root):
@@ -38,8 +63,8 @@ def settle(root):
 
 
 def _state(root, wait):
-    spec, folder = catalog(), directory(root)
     try:
+        spec, folder = catalog(), directory(root)
         receipt = json.loads((folder / 'verified.json').read_text(encoding='utf-8'))
         if receipt.get('revision') != spec['revision']:
             return False
@@ -73,8 +98,8 @@ def checking(root):
 
     A cheap look at the same file identity ready() uses, for a page that must not
     call the model missing, or offer its download, in the meantime."""
-    spec, folder = catalog(), directory(root)
     try:
+        spec, folder = catalog(), directory(root)
         receipt = json.loads((folder / 'verified.json').read_text(encoding='utf-8'))
         if receipt.get('revision') != spec['revision']:
             return False
diff --git a/freevideo_engine/prompt_vlm/rules.py b/freevideo_engine/prompt_vlm/rules.py
index f0dfb48f..8d4e8a60 100644
--- a/freevideo_engine/prompt_vlm/rules.py
+++ b/freevideo_engine/prompt_vlm/rules.py
@@ -23,8 +23,13 @@ def request(value):
         raise ValueError('too_many_images')
     roles = []
     for row in media:
-        if (not isinstance(row, dict) or set(row) != {'file', 'role'} or
-                not isinstance(row['file'], str) or row['role'] not in ('first', 'last', 'reference')):
+        # kind and number say which <Picture N> / <Video N> label the scene uses for the row; connected
+        # nodes and audio are left out of the list, so the numbers can skip.
+        if (not isinstance(row, dict) or not {'file', 'role'} <= set(row) <= {'file', 'role', 'kind', 'number'} or
+                not isinstance(row['file'], str) or row['role'] not in ('first', 'last', 'reference') or
+                row.get('kind', 'image') not in ('image', 'video') or
+                (row.get('kind') == 'video' and row['role'] != 'reference') or
+                (type(row.get('number', 1)) is not int or not 1 <= row.get('number', 1) <= 99)):
             raise ValueError('unsupported_media')
         roles.append(row['role'])
     if 'reference' in roles and any(r in roles for r in ('first', 'last')):
@@ -36,9 +41,40 @@ def request(value):
     # Anchor numbering must match the encoder even when the UI added last first.
     if mode != 'ref2va':
         media = sorted(media, key=lambda r: r['role'] != 'first')
+    if len({'number' in row for row in media}) > 1:
+        raise ValueError('unsupported_media')
+    counters, rows = {}, []
+    for row in media:
+        kind = row.get('kind', 'image')
+        counters[kind] = counters.get(kind, 0) + 1
+        rows.append(dict(file=row['file'], role=row['role'], kind=kind, number=row.get('number', counters[kind])))
+    media = rows
+    if len({label(row) for row in media}) != len(media):
+        raise ValueError('unsupported_media')
     return dict(text=prompt, seconds=float(seconds), media=media, mode=mode)
 
 
+LABEL = re.compile(r'<(Picture|Video|Audio) (\d+)>')
+
+
+def label(row):
+    return '<%s %d>' % ('Video' if row['kind'] == 'video' else 'Picture', row['number'])
+
+
+def unseen_labels(value):
+    """Labels the scene uses for media the model cannot look at (audio, connected nodes)."""
+    shown = {label(row) for row in value['media']}
+    found = []
+    for match in LABEL.finditer(value['text']):
+        if match.group(0) not in shown and match.group(0) not in found:
+            found.append(match.group(0))
+    return found
+
+
+def required_labels(value):
+    return [label(row) for row in value['media']] + unseen_labels(value)
+
+
 def instruction(value):
     rule = """Rewrite the user scene as an English MiniMax H3 video prompt. Follow the requested
 actions and camera movement exactly. Image references describe appearance, not whether an
@@ -57,7 +93,8 @@ def instruction(value):
 """
     if value['mode'] == 'ref2va':
         rule += """subject_definitions: Name each visible subject literally <Subject 1>, <Subject 2>, etc.,
-with its appearance and source <Picture N>. Preserve the supplied picture numbering.
+with its appearance and source <Picture N> or <Video N>. Preserve the supplied numbering.
+A video is shown as one still frame.
 summary: [reference generation] State what happens in the requested video.
 retention_analysis: Mark each subject fully_preserved, partially_preserved, attribute_transfer,
 or weak_reference. This describes retained visual attributes, not motion restrictions.
@@ -65,14 +102,18 @@ def instruction(value):
 """
     else:
         rule += 'integrated_multimodal_description: [Shot 1] Describe the target video.\n'
-        for i, row in enumerate(value['media'], 1):
+        for row in value['media']:
             timestamp = '0.00' if row['role'] == 'first' else f'{value["seconds"]:.2f}'
-            rule += (f'Anchor <Picture {i}> to the {row["role"]} frame at {timestamp} seconds; '
+            rule += (f'Anchor {label(row)} to the {row["role"]} frame at {timestamp} seconds; '
                      'keep its pictured layout at that anchor.\n')
     rule += """overall_soundscape: Environmental sounds, or N/A.
 non_diegetic_music: Music only if requested, otherwise N/A.
 Use only supplied media labels; do not invent audio or video references.
 """
+    unseen = unseen_labels(value)
+    if unseen:
+        rule += ('The scene also uses ' + ', '.join(unseen) + ', which you cannot see. '
+                 'Keep each of those labels exactly as written, where the scene uses it.\n')
     fields = REF_FIELDS if value['mode'] == 'ref2va' else BASE_FIELDS
     return (rule + f'Cover {value["seconds"]:.3f} seconds. All sections are required, in this order: '
             + ', '.join(fields) + f'. Begin with "{fields[0]}:".')
@@ -92,9 +133,23 @@ def sections(text, fields):
     return re.sub(r'\n+(?=(?:%s):)' % names, '\n\n', text).strip()
 
 
+class InvalidOutput(ValueError):
+    """Stable rejection metadata; never retain the prompt or rejected draft."""
+    def __init__(self, condition, text, value):
+        super().__init__('invalid_output')
+        draft = text if isinstance(text, str) else ''
+        required = required_labels(value)
+        pictures = [int(each[len('<Picture '):-1]) for each in required if each.startswith('<Picture ')]
+        indices = {int(i) for i in re.findall(r'<Picture (\d+)>', draft)}
+        self.details = dict(condition=condition, output_characters=len(draft),
+                            missing_pictures=sorted(i for i in pictures if f'<Picture {i}>' not in draft),
+                            out_of_range_pictures=sorted(i for i in indices if i not in pictures),
+                            retried=False)
+
+
 def validate_output(text, value):
     if not isinstance(text, str) or not 100 <= len(text.strip()) <= 20000:
-        raise ValueError('invalid_output')
+        raise InvalidOutput('type_or_length', text, value)
     text = text.strip()
     if text.startswith('```') and text.endswith('```'):
         text = '\n'.join(text.splitlines()[1:-1]).strip()
@@ -102,27 +157,28 @@ def validate_output(text, value):
     text = sections(text, fields)
     found = re.findall(r'(?m)^([a-z_]+):', text)
     if found != list(fields) or not text.startswith(fields[0] + ':'):
-        raise ValueError('invalid_output')
+        raise InvalidOutput('headings', text, value)
     if not re.search(r'\[Shot \d+\]', text):
         name = 'detailed_description' if value['mode'] == 'ref2va' else fields[0]
         text = text.replace(name + ':', name + ': [Shot 1]', 1)
     if '[Shot 1]' not in text:
-        raise ValueError('invalid_output')
+        raise InvalidOutput('missing_shot_1', text, value)
     for marker in fields:
         section = text.split(marker + ':', 1)[1].split('\n\n', 1)[0].strip()
         if not section:
-            raise ValueError('invalid_output')
+            raise InvalidOutput('empty_section', text, value)
     for minutes, seconds in re.findall(r'\bAt (\d{2}):(\d{2}\.\d{3})', text):
         if int(minutes) * 60 + float(seconds) >= value['seconds']:
-            raise ValueError('invalid_output')
-    for index in re.findall(r'<Picture (\d+)>', text):
-        if not 1 <= int(index) <= len(value['media']):
-            raise ValueError('invalid_output')
-    if re.search(r'<(?:Audio|Video) \d+>', text):
-        raise ValueError('invalid_output')
-    for i in range(1, len(value['media']) + 1):
-        if f'<Picture {i}>' not in text:
-            raise ValueError('invalid_output')
+            raise InvalidOutput('timestamp_not_before_duration', text, value)
+    # Only the labels the scene itself uses: the media given, plus audio and connected media it cannot see.
+    required = required_labels(value)
+    for found in LABEL.finditer(text):
+        if found.group(0) not in required:
+            raise InvalidOutput('picture_index_out_of_range' if found.group(1) == 'Picture'
+                                else 'unsupported_audio_video_reference', text, value)
+    for each in required:
+        if each not in text:
+            raise InvalidOutput('missing_picture_reference', text, value)
     return dialogue(text, value['text'])
 
 
diff --git a/freevideo_engine/prompt_vlm/service.py b/freevideo_engine/prompt_vlm/service.py
index 1a8acbff..4e8fc642 100644
--- a/freevideo_engine/prompt_vlm/service.py
+++ b/freevideo_engine/prompt_vlm/service.py
@@ -6,6 +6,7 @@
 import queue
 import secrets
 import subprocess
+import tempfile
 import threading
 import time
 import uuid
@@ -220,7 +221,10 @@ def spawn(root, machine, env, descriptor=None):
     def info(self, wait=True):
         """wait=False answers from the cheap check, for callers on an event loop or UI thread."""
         root, machine = self.installation()
-        spec = assets.catalog()
+        try:
+            spec = assets.catalog()
+        except (OSError, ValueError):
+            return dict(ready=False, model='', license='', bytes=0, busy=self.busy(), ram_bytes=self.ram_bytes(machine))
         return dict(ready=assets.settle(root) if wait else assets.ready(root), model=spec['repo'], license=spec['license'],
                     bytes=sum(row['bytes'] for row in spec['files']), busy=self.busy(),
                     ram_bytes=self.ram_bytes(machine))
@@ -287,12 +291,13 @@ def start(self, action, value):
                 raise ValueError('model_missing')
             value = rules.request(value)
             from ..comfy_assets import resolve_assets
-            resolved = resolve_assets(json.dumps(value['media']), self.input_directory())
+            resolved = resolve_assets(json.dumps([dict(file=row['file'], role=row['role']) for row in value['media']]),
+                                      self.input_directory())
             for i, row in enumerate(value['media']):
                 path = (resolved['references'][i]['path'] if row['role'] == 'reference'
                         else resolved[row['role']])
                 from ..comfy_assets import media_kind
-                if media_kind(path) != 'image':
+                if media_kind(path) != row['kind']:
                     raise ValueError('unsupported_media')
                 row['path'] = path
         with self.mutex:
@@ -370,6 +375,18 @@ def release_idle_image_worker():
         except Exception:
             pass
 
+    def video_frames(self, value, folder):
+        """The model reads pictures: a reference video is shown as its middle frame."""
+        for row in value.get('media', ()):
+            if row.get('kind') != 'video':
+                continue
+            self.check()
+            try:
+                from ..agent.media import sample_frames
+                row['path'] = sample_frames(row['path'], Path(folder) / str(row['number']), count=1)[0]['path']
+            except Exception:
+                raise ValueError('unreadable_video') from None
+
     def work(self, root, machine, action, value):
         from ..locking import runtime_lock
         from ..monitoring import save
@@ -394,18 +411,22 @@ def work(self, root, machine, action, value):
                     self.check()
                     self.progress(dict(phase='complete'))
                 else:
-                    self.infer(root, machine, descriptor, value)
+                    with tempfile.TemporaryDirectory(prefix='freevideo-prompt-', ignore_cleanup_errors=True) as frames:
+                        self.video_frames(value, frames)
+                        self.infer(root, machine, descriptor, value)
         except InterruptedError:
             self.progress(dict(phase='cancelled'))
         except BlockingIOError:
             self.progress(dict(phase='failed', error='busy'))
         except Exception as error:
-            known = {'disk_space', 'ram_space', 'rewrite_failed', 'split_failed', 'worker_exit', 'timeout', 'worker_cleanup_failed'}
-            self.progress(dict(phase='failed', error=str(error) if str(error) in known
+            known = {'disk_space', 'ram_space', 'unreadable_video', 'rewrite_failed', 'split_failed', 'worker_exit', 'timeout', 'worker_cleanup_failed'}
+            self.progress(dict(phase='failed', error='invalid_output' if isinstance(error, rules.InvalidOutput)
+                               else str(error) if str(error) in known
                                else 'download_failed' if action == 'install'
                                else 'ask_failed' if action == 'ask' else 'split_failed' if action == 'split'
                                else 'rewrite_failed',
-                               exception=type(error).__name__))
+                               exception=type(error).__name__,
+                               **({'output_rejection': error.details} if isinstance(error, rules.InvalidOutput) else {})))
         finally:
             if action != 'install':
                 self.stop_prestart(self.prewarm_state.get('reason', 'none'))
@@ -413,10 +434,14 @@ def work(self, root, machine, action, value):
                 self.progress(dict(phase='cancelled'))
             self.result_expires = time.monotonic() + 600
             # Allowlisted metadata only, never original/rewritten text or input paths.
+            # Output rejections contain condition IDs, character counts, picture numbers and retry flags only.
             report = {key: item for key, item in self.state.items()
                       if key in ('phase', 'error', 'exception', 'error_details', 'timings', 'diagnostics',
-                                 'resources', 'cleanup', 'worker_returncode')}
-            spec = assets.catalog()
+                                 'resources', 'cleanup', 'worker_returncode', 'output_rejection')}
+            try:
+                spec = assets.catalog()
+            except (OSError, ValueError):
+                spec = dict(repo='', revision='')
             report.update(schema='freevideo.prompt-vlm', schema_version=2, action=action,
                           created_epoch=time.time(), report_id=uuid.uuid4().hex,
                           elapsed_seconds=time.monotonic() - started, model=spec['repo'],
@@ -580,7 +605,13 @@ def read_rows():
                 elif value.get('mode') == 'split':
                     rules.validate_parts(result.get('parts'), value)
                 else:
-                    rules.validate_output(result.get('text'), value)
+                    try:
+                        rules.validate_output(result.get('text'), value)
+                    except rules.InvalidOutput as error:
+                        stats = result.get('diagnostics') or {}
+                        error.details['retried'] = bool(stats.get('picture_retry') or stats.get('dialogue_retry'))
+                        self.progress(dict(diagnostics=stats))
+                        raise
             self.progress(dict(result, worker_returncode=child.returncode))
         finally:
             releasing = result_received_at[0] if result_received_at else time.monotonic()
diff --git a/freevideo_engine/prompt_vlm/worker.py b/freevideo_engine/prompt_vlm/worker.py
index 3a2b8162..0b6f97ec 100644
--- a/freevideo_engine/prompt_vlm/worker.py
+++ b/freevideo_engine/prompt_vlm/worker.py
@@ -156,7 +156,11 @@ def failure(error, mode='rewrite'):
     code = message if message in allowed else 'out_of_memory' if 'out of memory' in message.lower() else fallback
     frames = [dict(module=Path(f.filename).name, function=f.name, line=f.lineno)
               for f in traceback.extract_tb(error.__traceback__)]
-    return dict(error=code, exception=type(error).__name__, error_details=frames)
+    result = dict(error=code, exception=type(error).__name__, error_details=frames)
+    from .rules import InvalidOutput
+    if isinstance(error, InvalidOutput):
+        result['output_rejection'] = error.details
+    return result
 
 
 def begin(stats):
@@ -217,7 +221,7 @@ def tokenize(processor, messages, max_pixels=512 * 32 * 32):
 
 def kv_positions(input_tokens):
     """Cache positions reserved for a request: the input and its rewrite, then a
-    dialogue retry that resends both with a short correction and writes another.
+    correction retry that resends both with a short instruction and writes another.
     A retry that would need more is skipped when the model is held on the GPU."""
     return input_tokens + 2 * OUTPUT_TOKENS + 256
 
@@ -237,7 +241,7 @@ def kv_bytes(stored, input_tokens):
 
 
 def retry_allowed(tokens, stats):
-    """A dialogue retry must fit the context limit and, on the GPU, the reserved cache."""
+    """A correction retry must fit the context limit and, on the GPU, the reserved cache."""
     if tokens + OUTPUT_TOKENS > stats.get('kv_positions', tokens + OUTPUT_TOKENS):
         stats['dialogue_retry_skipped'] = 'gpu_space'
         return False
@@ -363,6 +367,16 @@ def run(folder, value, *, stats=None):
     return rewrite(folder, value, stats)
 
 
+def validate_rewrite(text, value, stats, retried=False):
+    from .rules import InvalidOutput, validate_output
+    try:
+        return validate_output(text, value)
+    except InvalidOutput as error:
+        error.details['retried'] = retried
+        stats.setdefault('output_rejections', []).append(error.details)
+        raise
+
+
 def rewrite(folder, value, stats):
     stats.update(stage='imports', pid=os.getpid(), image_count=len(value['media']),
                  mode=value['mode'], dtype='bfloat16', attention='sdpa', kv_cache='dynamic',
@@ -370,14 +384,14 @@ def rewrite(folder, value, stats):
                  persistent_model=False)
     torch, checkpoint = begin(stats)
     from transformers import AutoProcessor
-    from .rules import DIALOGUE_RETRY, instruction, language, missing_lines, user_text, validate_output
+    from .rules import DIALOGUE_RETRY, InvalidOutput, instruction, label, language, missing_lines, user_text
     checkpoint('preprocess')
     preparing = time.monotonic()
     processor = AutoProcessor.from_pretrained(folder, local_files_only=True, trust_remote_code=False)
     messages = [{'role': 'system', 'content': instruction(value)}]
     content = []
-    for index, row in enumerate(value['media'], 1):
-        content.extend([{'type': 'text', 'text': f'<Picture {index}> ({row["role"]})'},
+    for row in value['media']:
+        content.extend([{'type': 'text', 'text': f'{label(row)} ({row["role"]})'},
                         {'type': 'image', 'image': open_image(row['path'])}])
     content.append({'type': 'text', 'text': user_text(value)})
     messages.append({'role': 'user', 'content': content})
@@ -405,23 +419,51 @@ def on_token(timing):
         raise ValueError('context_length')
     text = processor.decode(ids, skip_special_tokens=True)
     missing = missing_lines(text, value['text'])
-    if missing:
-        # One correction turn when the user's dialogue was dropped; keep the better draft.
-        retry = messages + [{'role': 'assistant', 'content': text}, {'role': 'user', 'content': DIALOGUE_RETRY + '\n'.join(
-            '- [%s] %s' % (language(line), line) for line in missing)}]
+    rewritten, rejected = None, None
+    pictures = []
+    try:
+        rewritten = validate_rewrite(text, value, stats)
+    except InvalidOutput as error:
+        rejected = error
+        if error.details['condition'] == 'missing_picture_reference':
+            pictures = error.details['missing_pictures']
+    if missing or pictures:
+        # Both corrections share one turn and the existing context/cache allowance.
+        correction = ''
+        if pictures:
+            labels = ', '.join(f'<Picture {i}>' for i in pictures)
+            correction = (f'Your rewrite omitted these required image references: {labels}. '
+                          'Preserve all existing content and add each missing reference where it belongs. '
+                          'Return the complete prompt in the same required section format.\n')
+        if missing:
+            correction += DIALOGUE_RETRY + '\n'.join('- [%s] %s' % (language(line), line) for line in missing)
+        retry = messages + [{'role': 'assistant', 'content': text}, {'role': 'user', 'content': correction}]
         retry_inputs = tokenize(processor, retry)
         if retry_allowed(retry_inputs.input_ids.shape[-1], stats):
             again, _ = generate(torch, model, retry_inputs.to(device), 1800, time.monotonic(), lambda timing: None)
             second = processor.decode(again, skip_special_tokens=True)
-            stats.update(dialogue_retry=True, retry_output_tokens=len(again),
-                         rewrite_seconds=time.monotonic() - loaded)
+            stats.update(retry_output_tokens=len(again), rewrite_seconds=time.monotonic() - loaded)
+            if missing:
+                stats['dialogue_retry'] = True
+            if pictures:
+                stats['picture_retry'] = True
             try:
-                validate_output(second, value)
-                if len(again) < 1800 and len(missing_lines(second, value['text'])) < len(missing):
-                    text = second
-            except ValueError:
-                pass
-    rewritten = validate_output(text, value)
+                candidate = validate_rewrite(second, value, stats, retried=True)
+                if pictures and len(again) >= 1800:
+                    error = InvalidOutput('output_token_limit', second, value)
+                    error.details['retried'] = True
+                    stats['output_rejections'].append(error.details)
+                    raise error
+                if len(again) < 1800 and (pictures or len(missing_lines(second, value['text'])) < len(missing)):
+                    rewritten = candidate
+            except InvalidOutput:
+                if pictures:
+                    raise
+        elif pictures:
+            stats['picture_retry_skipped'] = stats.get('dialogue_retry_skipped', 'context_length')
+    if rewritten is None:
+        rejected.details = dict(rejected.details, retried=bool(stats.get('picture_retry') or stats.get('dialogue_retry')))
+        raise rejected
     stats['stage'] = 'complete'
     emit(phase='complete', text=rewritten, diagnostics=dict(stats))
 
diff --git a/freevideo_engine/release_notes.json b/freevideo_engine/release_notes.json
index 069a80a2..9ff162fa 100644
--- a/freevideo_engine/release_notes.json
+++ b/freevideo_engine/release_notes.json
@@ -1,31 +1,68 @@
 {
   "schema": 1,
-  "product_version": "0.4.17",
+  "product_version": "0.4.18",
   "en": {
-    "summary": "Image generation and editing come to Apple silicon Macs, a Mac update that fails to open is rolled back, and phones on the same network connect more reliably.",
+    "summary": "FreeVideo can be installed into deeply nested folders on Windows; many failure messages state the cause and the next step; Windows computers with little free memory can complete generation; consecutive requests run faster on computers with ample memory.",
     "changes": [
-      "Mac: Macs with Apple silicon and at least 16 GB of memory can now generate and edit images.",
-      "Mac: in later updates, if the new version stops right after opening, the previous version is put back. The previous version is kept in a hidden folder so it no longer shows up in search or Launchpad. When an update fails, the reason and both buttons now appear right below the description.",
-      "Phone access: the phone page connects when phone access is switched on in the launcher, and also works on older phone browsers.",
-      "Queue: a single failed read of the queue no longer shows the connection lost card. The card appears after three failures in a row and goes away when the connection returns, and a running loop is kept through a short outage.",
-      "Launcher: when the chosen folder is not a ComfyUI folder, the launcher names a ComfyUI folder found nearby. Windows: the engine now starts when its usual port falls inside a range that Windows has reserved.",
-      "Installation: an installation with no generation on record now also compiles the standard size in the background.",
-      "Failures: when the model file changed during generation, the failure text names what changed."
+      "Install: FreeVideo can be installed into deeply nested folders on Windows; when an update cannot be installed and the previous version is restored, the notice follows the interface language.",
+      "Launcher: errors state the step that failed and what to do next, and Report a problem includes a line about the environment; when the saved installation folder cannot be opened, Start opens the setup page with the saved paths and an explanation; a graphics driver older than 580 gets a notice with the installed version and the update steps; a damaged launcher settings file is kept as a copy before it is replaced; the launcher keeps working when the prompt model file list is missing, and that row appears once the list is available.",
+      "Model downloads: slow sources are retried with a longer time limit, and the last failure reason of each source and route is listed; when a source has problems, the launcher shows a dedicated card with a button to switch sources; the image model has its own download list with links and checksums, exported to a separate folder.",
+      "Failure messages: a failed video generation names the stage where it stopped (loading, sampling or decoding); failure feedback starts with the last error; automatic subtitles, subtitle downloads, media import and subtitle import in the editor state the reason when they fail, and GBK-encoded SRT subtitles can be imported; image generation, adding workflows and share export state when the disk is full; an unreadable queued reference image prompts for a different image; a failed part of a long video shows a short status, with details in the error card and the name of any missing file.",
+      "Generation: fixed an error when submitting a generation; when first/last frames and references cannot be combined, the failure card says which media to turn off; Generate with AI explains when the media list is full at 32 items; the estimate before generation marks first-time preparation, and a slow estimate gives a reason and a remedy; a queued request shows its wait in minutes or the number of requests ahead; a queued export shows Queued on its button; LoRA files in more naming styles are recognized, and an acceleration LoRA made for the original model gets its own explanation.",
+      "Prompt enhancement: scenes with audio, connected nodes or videos can be enhanced, with audio and connected nodes skipped and one frame read from each video; with reference images, a rewrite that leaves out an image reference is sent back to the model once to add it.",
+      "Reference images: uploading and reading show Uploading reference image and Reading reference image N%; the + button in the editor shows a spinner while loading; library import shows uploading, queued and processing in turn, and finishes once progress reaches 100%.",
+      "Autopilot: it can be stopped while waiting for a reply and shows its current action; after three minutes without a progress update it says generation is still running; a background tab title starts with Confirmation needed, Failed or Done; Edit prompt puts the summary back in the input box; the confirmation card shows the size and duration; when the entry check fails, the entry stays and the check can be retried; an unreadable saved conversation or film leaves the rest available; failure reasons are readable in the English interface.",
+      "Multi-shot films: for long videos and multi-shot films made by Autopilot, music is added on request: when requested, one continuous score is laid over the whole film after the shots are joined, and a music file from the user can be used directly.",
+      "Create and Your creations: saving or loading a prompt can be retried after a failure; key presses stay inside the creation panel; in the image editor, the History entry stays visible after a picture is placed from the history; the end of the Your creations list shows how many records cannot be read.",
+      "Memory and speed: when a Windows computer is refused only because of the default memory reserve, part of the reserve is used, keeping at least 1 GiB, so generation can start with about 5 GiB free; on computers with ample memory, consecutive requests skip reloading the models and run faster.",
+      "Mac: encoding with many reference images is much faster; accelerated computation falls back to the standard method when its results look wrong; after a one-click update, a new version that starts slowly is kept."
     ]
   },
   "zh": {
-    "summary": "Apple 芯片的 Mac 可以生成和编辑图片；Mac 更新后打不开时会还原旧版；同一网络下的手机连接更稳定。",
+    "summary": "Windows 上可以安装到层级较深的文件夹；多处失败提示写明原因和下一步操作；可用内存较少的 Windows 电脑可以继续完成生成；内存宽裕的电脑连续生成时速度更快。",
     "changes": [
-      "Mac：Apple 芯片、内存 16 GB 及以上的 Mac 可以生成和编辑图片。",
-      "Mac：以后的更新中，新版刚打开就停止时会自动换回旧版。旧版保存在隐藏文件夹里，不再出现在搜索和启动台中。更新失败时，失败原因和两个按钮紧跟在说明下方显示。",
-      "手机访问：在启动器中打开手机访问后，手机页面即可连接，较旧的手机浏览器也能使用。",
-      "队列：读取队列只失败一次时不再弹出“连接已断开”卡片；连续失败三次才出现，连接恢复后自动消失；短时间断线时，正在运行的循环会保留。",
-      "启动器：所选文件夹不是 ComfyUI 文件夹时，会指出附近找到的 ComfyUI 文件夹。Windows：引擎常用端口落在 Windows 保留的端口段里时，也能正常启动。",
-      "安装：没有任何生成记录的安装，后台也会预编译标准尺寸。",
-      "失败提示：生成过程中模型文件发生变化时，失败文字会写明变化的内容。"
+      "安装：Windows 上可以安装到层级较深的文件夹；更新未能装上并换回旧版时，提示按界面语言显示。",
+      "启动器：出错时写明出错的步骤和下一步操作，“报告问题”附带一行环境信息；已保存的安装文件夹无法打开时，点击“启动”会进入安装页，显示已保存的路径和说明；显卡驱动版本低于 580 时，显示当前驱动版本和更新步骤；启动设置文件损坏时，先保留一份副本再保存；提示词模型的文件清单缺失时，启动器照常工作，相应一行在清单可用后显示。",
+      "模型下载：下载源较慢时以更长的时限重试，并分别列出每个下载源和线路最近一次失败的原因；下载源出现问题时，启动器显示专门的卡片和切换下载源的按钮；图像模型有单独的下载清单，列出链接和校验值，并导出到单独的文件夹。",
+      "失败提示：视频生成失败时写明出错的阶段（载入、采样或解码）；生成失败后的反馈以最后一条错误开头；剪辑页的自动字幕、字幕下载、素材导入和字幕导入失败时写明原因，GBK 编码的 SRT 字幕可以导入；图像生成、补写工作流和分享导出遇到磁盘已满时写明原因；排队中的参考图无法读取时，提示更换图片；长视频的某一段失败时显示简短状态，详细内容在错误卡片中，并写出缺失文件的名称。",
+      "生成：修复了部分情况下提交生成时报错的问题；首尾帧与参考素材不能同时使用时，失败卡片说明原因并提示关闭哪一类素材；素材列表已满 32 项时，“AI 生成”说明原因；生成前的时间估计注明“含首次准备”，预计较慢时给出一个原因和一个处理办法；排队时显示“约 N 分钟后开始”或“前面还有 N 个”；导出排队时按钮显示“排队中”；更多命名方式的 LoRA 文件可以正确识别，为原版模型制作的加速 LoRA 会显示专门说明。",
+      "提示词增强：场景中有音频、连接节点或视频时也可以增强，音频和连接节点会被跳过，视频读取一帧画面；带参考图时，改写结果遗漏图片引用，会请模型补上一次。",
+      "参考图：上传和读取时显示“正在上传参考图”和“正在读取参考图 N%”；剪辑页的“+”按钮在加载时显示转圈；素材库导入依次显示上传中、排队中和处理中，进度到 100% 后直接完成。",
+      "AI 全自动：等待回复时可以停止，并显示当前动作；三分钟没有进度更新时，提示仍在生成中；后台标签页的标题前显示“待确认”“失败”或“已完成”；“改描述”会把摘要放回输入框；确认卡片显示尺寸和时长；入口检查失败时保留入口，可以重试；某个已保存的对话或影片无法读取时，其余内容照常可用；英文界面下的失败原因清楚可读。",
+      "多镜头作品：AI 全自动制作长视频和多镜头作品时，配乐按需添加：需要配乐时，在各镜头拼接完成后为整片铺一段连贯的配乐，用户提供的音乐文件也可以直接使用。",
+      "创作面板与我的作品：提示词保存或读取失败时可以重试；按键操作只作用于创作面板；图像编辑中从历史记录放入画布后，“历史”入口保持可见；“我的作品”列表末尾显示无法读取的记录数量。",
+      "内存与速度：Windows 电脑只因默认预留空间不足而无法开始生成时，会使用部分预留空间（至少保留 1 GiB），可用内存约 5 GiB 时也可以开始生成；连续生成多个请求时，内存宽裕的电脑无需重复加载模型，速度更快。",
+      "Mac：使用大量参考图时，编码速度明显提升；加速计算的结果异常时，自动改用标准计算；一键更新后，启动较慢的新版本同样会保留。"
     ]
   },
   "history": [
+    {
+      "product_version": "0.4.17",
+      "en": {
+        "summary": "Image generation and editing come to Apple silicon Macs, a Mac update that fails to open is rolled back, and phones on the same network connect more reliably.",
+        "changes": [
+          "Mac: Macs with Apple silicon and at least 16 GB of memory can now generate and edit images.",
+          "Mac: in later updates, if the new version stops right after opening, the previous version is put back. The previous version is kept in a hidden folder so it no longer shows up in search or Launchpad. When an update fails, the reason and both buttons now appear right below the description.",
+          "Phone access: the phone page connects when phone access is switched on in the launcher, and also works on older phone browsers.",
+          "Queue: a single failed read of the queue no longer shows the connection lost card. The card appears after three failures in a row and goes away when the connection returns, and a running loop is kept through a short outage.",
+          "Launcher: when the chosen folder is not a ComfyUI folder, the launcher names a ComfyUI folder found nearby. Windows: the engine now starts when its usual port falls inside a range that Windows has reserved.",
+          "Installation: an installation with no generation on record now also compiles the standard size in the background.",
+          "Failures: when the model file changed during generation, the failure text names what changed."
+        ]
+      },
+      "zh": {
+        "summary": "Apple 芯片的 Mac 可以生成和编辑图片；Mac 更新后打不开时会还原旧版；同一网络下的手机连接更稳定。",
+        "changes": [
+          "Mac：Apple 芯片、内存 16 GB 及以上的 Mac 可以生成和编辑图片。",
+          "Mac：以后的更新中，新版刚打开就停止时会自动换回旧版。旧版保存在隐藏文件夹里，不再出现在搜索和启动台中。更新失败时，失败原因和两个按钮紧跟在说明下方显示。",
+          "手机访问：在启动器中打开手机访问后，手机页面即可连接，较旧的手机浏览器也能使用。",
+          "队列：读取队列只失败一次时不再弹出“连接已断开”卡片；连续失败三次才出现，连接恢复后自动消失；短时间断线时，正在运行的循环会保留。",
+          "启动器：所选文件夹不是 ComfyUI 文件夹时，会指出附近找到的 ComfyUI 文件夹。Windows：引擎常用端口落在 Windows 保留的端口段里时，也能正常启动。",
+          "安装：没有任何生成记录的安装，后台也会预编译标准尺寸。",
+          "失败提示：生成过程中模型文件发生变化时，失败文字会写明变化的内容。"
+        ]
+      }
+    },
     {
       "product_version": "0.4.16",
       "en": {
@@ -77,9 +114,9 @@
         "changes": [
           "生图：本机生成过的图片都保存在历史记录里，按时间从新到旧排列，可放到画布、用作参考图、删除或打开所在文件夹。空白画布可选“标准 / 高清”尺寸。",
           "生图：6–8 GB 显存的显卡解码时占用的显存更少，速度更快。显存放得下整个模型的显卡，权重直接加载到显卡，不再在内存里留一份副本，加载后内存占用从约 15 GB 降到约 2 GB。",
-          "提交问题：生成、剪辑与导出、环境与模型、生图、智能体连接、作品库、提示词优化、分享、更新和启动器里，出现错误或失败提示的地方旁边都有“提交问题”按钮，点一下带上情况打开问题反馈。",
+          "提交问题：生成、剪辑与导出、环境与模型、生图、智能体连接、作品库、提示词优化、分享、更新和启动器里，出现错误或失败提示的地方旁边都有“提交问题”按钮，点击后带上情况打开问题反馈。",
           "AI 全自动：对话删不掉时会说明原因，并提供“停止并删除”；其他失败也会提示，并带“提交问题”按钮。",
-          "白模：现成场景只保留“橘猫跳上书架”，点一下直接开始。白模没做成时会说明原因和下一步，渲染中页面显示进度。问白模进度时会回答进度，不再当成新要求。",
+          "白模：现成场景只保留“橘猫跳上书架”，点击直接开始。白模没做成时会说明原因和下一步，渲染中页面显示进度。问白模进度时会回答进度，不再当成新要求。",
           "人物设定图：背面图里头和身体一起转向背后。",
           "Mac：支持分段生成长视频。",
           "启动器与更新：Windows 版在中国大陆更新更快。GitHub 慢时，启动器可以从国内镜像下载更新，镜像上的每个文件都先按发布签名校验。“确认安装”页适配窄窗口：侧栏收成图标，按钮始终可见。",
@@ -236,43 +273,6 @@
           "修复：用旧版本保存的工作流或恢复的标签页生成时提示步间缓存参数无效。"
         ]
       }
-    },
-    {
-      "product_version": "0.4.9",
-      "en": {
-        "summary": "Quality levels are reworked: Medium is one pass of 8 steps, and High and Max use step caches to finish much sooner; decoding on Windows reads ahead earlier on slow drives, and the editor handles ranges, splits and text tracks more reliably.",
-        "changes": [
-          "Quality levels: Medium is one pass of 8 steps; High and Max run 20 steps with step caches and finish much sooner than plain 20 steps.",
-          "Decoding on Windows: read-ahead is sized by physical memory and starts earlier on slow drives, so the end of a run waits less for the disk.",
-          "When Windows runs out of virtual memory, in any system language, FreeVideo now retries the generation automatically in a fresh process and reports a memory shortage if it still fails.",
-          "The progress estimate counts reference images, so requests with reference images show a closer total time from the start.",
-          "Editor: dragging a range's left edge to the start of the timeline no longer pushes its right edge later; splitting a sped-up or reversed part keeps the fade-in only on the first half and the fade-out only on the second.",
-          "Editor: drag-selecting parts is smooth again on timelines with many parts.",
-          "Editor text: clicking the side Content box right after pressing T keeps the new text track, and dropping text onto a locked text track says the track is locked.",
-          "Long videos: each part's description can also be edited in a larger window.",
-          "Chat: when a conversation is full, the notice appears once under the message and its button starts a new chat instead of retrying.",
-          "Export: the sizes offered in the export dialog are rounded the same way as the saved work.",
-          "Diagnostic reports record how much of the graphics card other programs were using during each sampling step.",
-          "Launcher: when the browser cannot be opened automatically, the message says to click Copy address and open it in a browser, in Chinese as well. Canvas margin labels stay readable and aligned. Nodes: width, height, duration and seed have Chinese labels."
-        ]
-      },
-      "zh": {
-        "summary": "质量档位调整：标准为单遍 8 步，精细和极致用步间缓存，快得多；Windows 解码在慢硬盘上更早预读；编辑器的区间、切分和文字轨更可靠。",
-        "changes": [
-          "质量档位：标准为单遍 8 步；精细和极致为 20 步加步间缓存，比纯 20 步快得多。",
-          "Windows 解码：预读按物理内存决定大小，并在慢硬盘上更早开始，生成收尾时等硬盘的时间更短。",
-          "Windows 虚拟内存耗尽时，不论系统语言，FreeVideo 现在会在新进程中自动重试生成，仍失败时提示内存不足。",
-          "进度估计把参考图计入，带参考图的请求一开始就显示更接近实际的总时间。",
-          "编辑器：把区间左边拖到时间线开头，右边不再被往后推；切分加速或倒放的片段时，淡入只留在前半段，淡出只留在后半段。",
-          "编辑器：片段很多时，框选重新变得流畅。",
-          "编辑器文字：按 T 后马上点侧边“内容”框，新文字轨会保留；把文字拖到已锁定的文字轨上会提示该轨已锁定。",
-          "长视频：每段描述也能在放大窗口里编辑。",
-          "对话：对话已满时，提示只在消息下出现一次，按钮换成“新对话”。",
-          "导出：导出窗口里给出的尺寸和保存的作品用同样的方式取整。",
-          "诊断报告记录每一步采样时其他程序占用显卡的比例。",
-          "启动器：浏览器没能自动打开时，提示点“复制地址”到浏览器里打开，中文界面也显示中文。画布边距标签更清楚、对齐。节点：宽度、高度、时长和种子有中文标签。"
-        ]
-      }
     }
   ]
 }
diff --git a/freevideo_engine/request_assets.py b/freevideo_engine/request_assets.py
index 85136b13..dcddb4f6 100644
--- a/freevideo_engine/request_assets.py
+++ b/freevideo_engine/request_assets.py
@@ -61,7 +61,11 @@ def _prompt_vlm(root):
     from .prompt_vlm import assets
     if assets.settle(root):
         return []
-    spec = assets.catalog()
+    try:
+        spec = assets.catalog()
+    except (OSError, ValueError):
+        # Nothing to download from, so the request is not ready for the rewrite step.
+        return [dict(id='prompt-vlm', kind='prompt-vlm', files=0, bytes=0, source='', action='unavailable')]
     return [dict(id='prompt-vlm', kind='prompt-vlm', files=len(spec['files']),
                  bytes=sum(row['bytes'] for row in spec['files']), source='', action='consent')]
 
@@ -75,7 +79,8 @@ def inspect(root, machine, request, *, lora_names=None):
     lora_names: the LoRA files that exist; None skips the LoRA check.
     Every item: id, kind, files, bytes, source (id of the source tried first,
     '' when not applicable) and action: 'auto' is fetched when Generate starts,
-    'consent' asks first, 'manual' means the user adds the file.
+    'consent' asks first, 'manual' means the user adds the file, 'unavailable' means
+    the item's own list of files cannot be read, so it cannot be fetched at all.
     """
     root = Path(root)
     sampled, items = _sampling(root, machine, request)
diff --git a/freevideo_engine/resident_models.py b/freevideo_engine/resident_models.py
index c6fb1858..ac675357 100644
--- a/freevideo_engine/resident_models.py
+++ b/freevideo_engine/resident_models.py
@@ -262,6 +262,25 @@ def ram_fits(self, need=2*GiB):
         return (used is not None and inference_headroom(sample) >= need + 2*GiB and
                 (self.ram_budget is None or used + need <= self.ram_budget))
 
+    def settle(self, pinned=0):
+        """End of a request: keep the engine and encoder only if the next cold encoder still fits.
+
+        The RAM guard counts the whole session, so retained engine weights
+        plus the next encoder's start (about 8 GiB) must stay inside the
+        budget, and so must the host weights a fresh plan would pin: measured,
+        a retained copy left no room to pin and sampling took 95 s instead of
+        38 s. Otherwise the engine goes as it did with one worker per request.
+        """
+        for role in ('engine', 'encoder'):
+            if role in self.entries and not self.ram_fits(8*GiB + int(pinned)):
+                self.drop(role, 'Retained weights leave no host room for the next request')
+        gc.collect()
+        try:
+            import ctypes
+            ctypes.CDLL('libc.so.6').malloc_trim(0)
+        except (OSError, AttributeError):
+            pass
+
     def input_vae_room(self, model, canvas, *, audio=False):
         # Input encoding has its own live workspace; it cannot assume an output
         # decoder's memory footprint. Charge actual retained CPU model tensors
@@ -401,7 +420,7 @@ def engine(self, request, factory):
         peak = estimate if observed is None else observed
         # Sampling is the long stage: an idle model stays through it only where
         # the device still holds both beside every other application.
-        self.make_room('engine', peak, model=found, device=True)
+        self.make_room('engine', peak, model=found, device=True, ram_need=8*GiB)
         self.decisions[-1]['evidence'] = dict(source=source if observed is None else 'complete_local_stage_peak',
             cold_estimate_bytes=estimate, observed_peak_bytes=observed, video_and_reference_tokens=rows,
             head_chunk=options.get('head_chunk', 8), window_batch=options.get('window_batch', 1),
@@ -446,7 +465,11 @@ def encoder_room(self, clip, tokens, fallback_peak):
             peak = max(weights, gpu_bytes(clip)) + max(2*GiB, workspace)
         else:
             source = 'encoder_checkpoint_estimate'
-        self.make_room('encoder', min(self.gpu_budget, peak), model=clip, ram_need=2*GiB)
+        # Measured: a reused encoder grew the whole session's guarded working
+        # set by 4 GiB beside 21.5 GiB of retained engine weights, passing a
+        # 25 GiB budget that a 2 GiB admission had accepted. Charge 4 GiB so
+        # the retained host copy is parked first on tight machines.
+        self.make_room('encoder', min(self.gpu_budget, peak), model=clip, ram_need=4*GiB)
         self.decisions[-1]['evidence'] = dict(source=source, cold_estimate_bytes=int(fallback_peak),
                                             requested_peak_bytes=int(peak))
 
diff --git a/freevideo_engine/resident_process.py b/freevideo_engine/resident_process.py
index e95fd231..cb211719 100644
--- a/freevideo_engine/resident_process.py
+++ b/freevideo_engine/resident_process.py
@@ -31,6 +31,9 @@
 AGENT_HOLD_SECONDS = 30.
 OTHER_GPU_GROWTH = 512 * 2**20
 HOST_FLOOR = 4 * GiB
+# Below this much host memory (free plus the idle worker's own), a kept worker's
+# base leaves the next part less room than a fresh one; measured on 16 GiB.
+KEEP_MIN_HOST = 24 * GiB
 RELEASE_LOG = 20
 
 
@@ -382,7 +385,7 @@ def finished(self, *, next_part=None, agent=False, gpu=None):
             hold = min(HOLD_SECONDS, AGENT_HOLD_SECONDS) if agent else HOLD_SECONDS
             if not next_part or hold <= 0:
                 return self.release('finished' if not next_part else 'no_hold')
-            reason = pressure(gpu, None)
+            reason = pressure(gpu, None) or self._tight()
             if reason:
                 return self.release(reason)
             self.generation += 1
@@ -392,6 +395,21 @@ def finished(self, *, next_part=None, agent=False, gpu=None):
                              name='freevideo-resident-hold', daemon=True).start()
             return self.record(action='kept', reason=next_part, seconds=hold, pid=self.process.pid)
 
+    def _tight(self):
+        from .system import windows
+        if windows():
+            # Windows keeps its earlier behaviour: the available-memory reading
+            # here has not been checked against its commit accounting.
+            return None
+        try:
+            from .background_room import available_memory
+            own = worker_memory(self.process.pid).get('worker_rss_bytes', 0)
+            if available_memory() + own < KEEP_MIN_HOST:
+                return 'host_memory_tight'
+        except (OSError, RuntimeError, AttributeError, ValueError, ImportError):
+            pass
+        return None
+
     def _watch(self, generation, since, hold, gpu, baseline):
         while True:
             time.sleep(.5)
diff --git a/freevideo_engine/runtime_alias.py b/freevideo_engine/runtime_alias.py
index 37c4eb3f..599e598d 100644
--- a/freevideo_engine/runtime_alias.py
+++ b/freevideo_engine/runtime_alias.py
@@ -20,6 +20,7 @@
 from .package_paths import ALIAS_ROOT_RESERVE, KNOWN_RUNTIME_PATH_CHARS, WINDOWS_PATH_LIMIT
 
 RESERVE = 40
+ONLINE_LOADED_PATH_CHARS = 100  # below the engine folder: envs\\unified\\Lib\\site-packages\\nvidia\\cudnn\\bin\\<dll>
 PROGRAMS = ('.exe',)
 LOADED = PROGRAMS + ('.dll', '.pyd')  # What Windows starts or loads, never from beyond the limit.
 # The licence texts a package carries are a runtime's deepest files, and no program opens them.
@@ -544,6 +545,17 @@ def needs_view(root, longest=None, spelling_only=False) -> bool:
     return path_chars(root) + 1 + longest + RESERVE > WINDOWS_PATH_LIMIT
 
 
+def online_loaded_too_long(engine) -> bool:
+    """An online installation whose deepest program or library is beyond what Windows loads.
+
+    The loader ignores the system's long-path setting for a DLL such as torch\\lib\\c10.dll.
+    """
+    if not _backend.windows() or os.environ.get('FREEVIDEO_RUNTIME_ALIAS') == '0':
+        return False
+    engine = _ordinary(resolved(checked_real_path(engine)))
+    return path_chars(engine) + 1 + ONLINE_LOADED_PATH_CHARS > WINDOWS_PATH_LIMIT
+
+
 def is_junction(path) -> bool:
     """Any Windows directory junction or directory symlink, wherever it lives."""
     if not _backend.windows():
diff --git a/freevideo_engine/tuning.py b/freevideo_engine/tuning.py
index fd6fd976..dc206593 100644
--- a/freevideo_engine/tuning.py
+++ b/freevideo_engine/tuning.py
@@ -103,8 +103,8 @@ def identity(machine, hardware):
     for checkout, ref, expected in ((machine['vdn_root'], 'HEAD', specs['vdn']['commit']),
             (machine['comfy_root'], 'HEAD', specs['encoder']['comfy_commit']),
             (str(Path(machine['vdn_root']) / 'diffusers'), 'HEAD^{tree}', specs['diffusers']['patched_tree'])):
-        observed_ref = subprocess_output.output_external(['git', '-C', checkout, 'rev-parse', ref], timeout=10).strip()
-        dirty = subprocess_output.output_external(['git', '-C', checkout, 'status', '--porcelain', '--untracked-files=no'], timeout=10)
+        from . import network
+        observed_ref, dirty = network.repository_state(checkout, 'git', None, ref, timeout=10)
         if observed_ref != expected or dirty:
             raise ValueError('Pinned model code changed; saved tuning cannot be reused: ' + checkout)
     return {'schema': SCHEMA, 'system': hardware.system,
diff --git a/freevideo_engine/worker.py b/freevideo_engine/worker.py
index b7e12697..9b6950fe 100644
--- a/freevideo_engine/worker.py
+++ b/freevideo_engine/worker.py
@@ -733,6 +733,12 @@ def keep_audio_lead_in(value):
             resident.device_free, resident.device_reserve = None, 0
         if live_budget is not None:
             live_budget.close()
+        if resident is not None and metrics['success'] and engine is not None:
+            try:
+                resident.settle(float(request['engine_options'].get('pin_host_gb') or 0)*1e9)
+            except Exception:
+                # Trimming is best effort; it must not turn a finished request into a failure.
+                pass
         if resident is not None:
             metrics['resident_admission'] = list(resident.decisions)
         from .adaptive import classify_failure
diff --git a/scripts/test_freevideo_entry.mjs b/scripts/test_freevideo_entry.mjs
index 8a289573..8e25a9fe 100644
--- a/scripts/test_freevideo_entry.mjs
+++ b/scripts/test_freevideo_entry.mjs
@@ -141,7 +141,7 @@ test('main browser entry registers both views and preserves node hooks', async t
             .replace("import { imageGenerationSupported } from './image_generation.js';", 'const imageGenerationSupported = () => true;')
             .replace("import { uploadMedia } from './media_upload.js';", 'const {uploadMedia} = globalThis.__freevideoEntryTest;')
             .replace("import { createCastLibrary } from './cast_library.js';", 'const {createCastLibrary} = globalThis.__freevideoEntryTest;')
-            .replace("import { agentInterfaceEnabled } from './agent_connect.js';", 'const agentInterfaceEnabled = async () => false;')
+            .replace("import { agentInterfaceState } from './agent_connect.js';", "const agentInterfaceState = async () => 'off';")
             .replace("import { installNativeErrorGuard } from './native_errors.js';", 'const installNativeErrorGuard = () => {};')
             .replace("import { installTranslateGuard } from './translate_guard.js';", 'const installTranslateGuard = () => {};');
         await import('data:text/javascript;base64,' + Buffer.from(source).toString('base64'));
diff --git a/web/agent_activity.js b/web/agent_activity.js
index 37c75079..4cd4dd3f 100644
--- a/web/agent_activity.js
+++ b/web/agent_activity.js
@@ -77,7 +77,7 @@ function row(job, t, api, refresh, open, cancellationErrors) {
     if (job.status === 'running' && p.remaining_seconds > 0) meta.push(left(p.remaining_seconds, t));
     const reason = job.status === 'failed' && REASONS[job.error?.code];
     if (reason) meta.push(t(...reason));
-    else if (job.status === 'failed') meta.push(t(job.error?.message || 'Did not finish', job.error?.error_zh || '没有完成'));
+    else if (job.status === 'failed') meta.push(t(job.error?.error_en || 'Did not finish', job.error?.error_zh || '没有完成'));
     if (job.project) meta.push(job.project);
     const foot = el('div', 'fv-agent-foot');
     const details = el('p', 'fv-agent-meta', meta.join(' · '));
diff --git a/web/agent_connect.js b/web/agent_connect.js
index 1bee48ba..605438f1 100644
--- a/web/agent_connect.js
+++ b/web/agent_connect.js
@@ -12,14 +12,19 @@ export function connectionSentence(address, t) {
     return t(`Connect to FreeVideo: open ${url} and follow the instructions.`, `请连接 FreeVideo：打开 ${url} 并按说明操作。`);
 }
 
-// Whether this FreeVideo has the Agent interface on. Any failure counts as off: the surface shows nothing.
-export async function agentInterfaceEnabled(api) {
+// Whether this FreeVideo has the Agent interface on: 'on', 'off' or 'unknown'. Only an explicit `enabled: false`
+// is off; a check that failed (no answer, an error status, an unreadable reply) is not an answer.
+export async function agentInterfaceState(api) {
     try {
         const response = await api.fetchApi('/freevideo/agent', {cache: 'no-store'});
-        return response.ok && (await response.json()).enabled === true;
-    } catch { return false; }
+        if (!response.ok) return 'unknown';
+        const enabled = (await response.json())?.enabled;
+        return enabled === true ? 'on' : enabled === false ? 'off' : 'unknown';
+    } catch { return 'unknown'; }
 }
 
+export async function agentInterfaceEnabled(api) { return await agentInterfaceState(api) === 'on'; }
+
 const ICONS = {plug: ['M9 7V3M15 7V3M7 7h10v4a5 5 0 0 1-10 0V7zM12 16v5', 1.8], copy: ['M11 9h7a2 2 0 0 1 2 2v7a2 2 0 0 1-2 2h-7a2 2 0 0 1-2-2v-7a2 2 0 0 1 2-2zM5 15V5a2 2 0 0 1 2-2h8', 2]};
 function icon(name) {
     const svg = document.createElementNS('http://www.w3.org/2000/svg', 'svg');
@@ -32,7 +37,9 @@ function icon(name) {
 // The card alone: who can connect and one button that copies the sentence to paste to the agent. The sentence
 // stays out of sight; it shows, selected, only when the browser refuses the clipboard. shown() says whether the
 // user can see the card now; the surface calls sync() whenever that may have changed.
-export function agentConnectCard(t, api, shown = () => true) {
+// A failed check leaves the card up, so its own retry is the way back; an address that loads then asks again and
+// reports the answer to changed().
+export function agentConnectCard(t, api, shown = () => true, changed = () => {}) {
     const css = document.createElement('link'); css.rel = 'stylesheet'; css.href = new URL('./agent_connect.css', import.meta.url).href; document.head.append(css);
     const body = el('section', 'fv-agent-connect-body');
     const mark = el('span', 'fv-agent-connect-icon'); mark.append(icon('plug'));
@@ -94,6 +101,7 @@ export function agentConnectCard(t, api, shown = () => true) {
             if (!connect?.address) throw new Error('Missing address');
             text = connectionSentence(connect.address, t);
             failures = 0; render();
+            if (await state === 'unknown') { state = agentInterfaceState(api); changed(await state); }
         } catch { if (current === generation) { failures++; render(); } }
         finally {
             if (current === generation) {
@@ -106,7 +114,8 @@ export function agentConnectCard(t, api, shown = () => true) {
         if (!active()) { stop(); return; }
         if (!text && failures < 3 && !controller && !timer) load();
     }
-    // Resolves to whether this FreeVideo has the Agent interface on; a surface shows nothing when it is off.
-    const enabled = agentInterfaceEnabled(api);
-    return {element: body, enabled, sync, stop};
+    // `state` is the first answer; `enabled` resolves to whether it is on.
+    let state = agentInterfaceState(api);
+    const first = state, enabled = first.then(value => value === 'on');
+    return {element: body, state: first, enabled, sync, stop};
 }
diff --git a/web/autopilot.js b/web/autopilot.js
index ee9302f6..7fe11558 100644
--- a/web/autopilot.js
+++ b/web/autopilot.js
@@ -10,7 +10,7 @@ import { productVersion } from './updates.js';
 import { openSetup } from './setup.js';
 import { feedbackButton } from './feedback.js';
 import { openExtendEditor, openStudio, showNodes } from './studio.js';
-import { viewChanged, viewSwitch } from './view_navigation.js';
+import { agentTabReady, viewChanged, viewSwitch } from './view_navigation.js';
 import { wordmark } from './wordmark.js';
 
 const languageOverride = typeof location !== 'undefined'
@@ -46,7 +46,11 @@ export function openAutopilotView(node, {studio = null, initialPrompt = '', init
     header.append(brand, navigation, tools, line);
     const inSight = () => dialog.open && dialog.dataset.closing !== 'true';
     const body = el('div', 'fv-autopilot-body');
-    const card = agentConnectCard(t, api, () => inSight() && !card.element.hidden);
+    const card = agentConnectCard(t, api, () => inSight() && !card.element.hidden, value => {
+        on = value; show();
+        if (value !== 'unknown') agentTabReady(value);
+        if (value === 'on') chat.sync();
+    });
     // Leaving for another view waits until this page has closed, so the view underneath is never covered twice.
     const closed = () => dialog.open ? new Promise(done => { dialog.addEventListener('close', done, {once: true}); closeDialog(dialog); }) : null;
     // The next video at the studio's current settings; the agent may choose others, so this is only what is missing now.
@@ -70,10 +74,11 @@ export function openAutopilotView(node, {studio = null, initialPrompt = '', init
     function show() {
         const connected = last.agent.connected === true, label = connectionLabel(last.agent, t);
         if (connected) wantCard = false;
-        const ready = on === true && last.known, talking = ready && (connected || (last.any && !wantCard));
+        const ready = on === 'on' && last.known, talking = ready && (connected || (last.any && !wantCard));
         // The card gives way to the conversation: typing can start at once.
         const arrived = talking && chat.element.hidden && (card.element.contains(document.activeElement) || document.activeElement === lamp);
-        card.element.hidden = !ready || talking; chat.element.hidden = !talking;
+        // A check that failed keeps the card up: its retry is the way back.
+        card.element.hidden = !(ready || on === 'unknown') || talking; chat.element.hidden = !talking;
         body.dataset.view = talking ? 'chat' : 'connect';
         if (talking) chat.sync();
         if (arrived) chat.input.focus();
@@ -98,7 +103,7 @@ export function openAutopilotView(node, {studio = null, initialPrompt = '', init
     go.onclick = () => chat.scrollToJob(currentId);
     halt.onclick = () => chat.stopAll();
     body.append(card.element, chat.element);
-    card.enabled.then(value => { on = value; show(); if (value) chat.sync(); });
+    card.state.then(value => { on = value; show(); if (value === 'on') chat.sync(); });
     show();
     dialog.append(header, body);
     dialog.addEventListener('cancel', event => { event.preventDefault(); closeDialog(dialog); });
diff --git a/web/autopilot_chat.css b/web/autopilot_chat.css
index 8e5edad3..014d7d44 100644
--- a/web/autopilot_chat.css
+++ b/web/autopilot_chat.css
@@ -120,6 +120,7 @@
 /* The agent has the sentence and has not answered yet. */
 .fv-chat-dots{display:inline-flex;gap:5px;align-items:center;height:22px}
 .fv-chat-dots>i,.fv-chat-dots:empty::before{content:"";width:6px;height:6px;border-radius:50%;background:var(--fv-muted);animation:fv-chat-breathe 1.4s var(--fv-ease) infinite}
+.fv-chat-thinking{display:flex;align-items:center;gap:10px;min-width:0}.fv-chat-activity{min-width:0;overflow:hidden;white-space:nowrap;text-overflow:ellipsis;color:var(--fv-muted);font-size:12.5px}.fv-chat-activity[hidden]{display:none}
 .fv-chat-dots>i:nth-child(2){animation-delay:.2s}.fv-chat-dots>i:nth-child(3){animation-delay:.4s}
 @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}
@@ -145,6 +146,7 @@
 .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}
+.fv-studio .fv-chat-send.fv-chat-halt svg{fill:currentColor}
 .fv-studio .fv-chat-send:focus-visible{outline-offset:2px}
 @media(max-width:720px){
  .fv-studio .fv-chat-conversations .fv-chat-conversation-row:has(> .fv-chat-more:not([hidden])) .fv-chat-conversation{padding-right:calc(var(--fv-chats-pad) + 36px)}
diff --git a/web/autopilot_chat.js b/web/autopilot_chat.js
index e9e2168d..3526d822 100644
--- a/web/autopilot_chat.js
+++ b/web/autopilot_chat.js
@@ -32,6 +32,12 @@ export const WHITEBOX_PREVIEW = Object.freeze({
 });
 let chatNumber = 0;
 
+// What the agent is doing before it answers, by the key it reports. Words the page does not know are not shown.
+export const STAGE_TEXT = Object.freeze({
+    reading: ['Reading the request', '正在理解要求'], scripting: ['Writing the script', '正在写剧本'],
+    storyboarding: ['Planning the shots', '正在拆分镜头'], prompting: ['Writing the shot prompts', '正在写画面描述'],
+    checking: ['Checking the results', '正在检查结果'], assembling: ['Joining the shots', '正在拼接镜头'],
+});
 const WORKING = ['running', 'queued', 'waiting', 'cancelling'], VIDEOS = ['generate', 'assemble'];
 // What a job with no picture of its own is doing, in the reader's language.
 const QUIET = {vlm: ['Checking', '检查中'], enhance: ['Enhancing prompt', '增强提示词']};
@@ -45,10 +51,12 @@ const WHITEBOX = {
 const GENERIC_FAILURE_ZH = '没有完成';
 // Sends and failure cards use the same public reason, including the engine's Chinese mapping.
 function failureReason(value, t) {
-    const error = value.error, reason = error?.error_zh || value.error_zh;
+    const error = value.error, reason = error?.error_zh || value.error_zh, reasonEn = error?.error_en || value.error_en;
     if (error?.code === 'session_full' && (!reason || reason === GENERIC_FAILURE_ZH))
         return t('This chat is full. Click "New chat".', '这个对话已满，请点“新对话”。');
-    return t(error?.message || 'Did not finish', reason || GENERIC_FAILURE_ZH);
+    // `message` is written for programs; people read the engine's own English reason. A refusal the engine
+    // did not classify keeps its sentence, as before.
+    return t(reasonEn || error?.message || 'Did not finish', reason || GENERIC_FAILURE_ZH);
 }
 
 export {about};
@@ -125,6 +133,18 @@ export function cardLine(said, job, t) {
     if (job.status === 'running' && (stepped(p) || SAVING.includes(p.phase))) said.last = job;
     return jobLine(said.last || job, t);
 }
+// The generation page's rule: three minutes without a new step or stage and the card says it is still going.
+export const STALL_MS = 180000;
+export const progressKey = job => { const p = job.progress || {}; return JSON.stringify([job.status, p.phase, p.step, p.block, p.fraction]); };
+// The one thing a hidden tab announces in its title, most pressing first: a question, a failure, a finished job.
+// `seen` holds the finished jobs the user already had in front of them when the tab was hidden.
+export function tabNotice(jobs, asked, seen) {
+    if (asked) return 'ask';
+    const fresh = Object.values(jobs).filter(job => !QUIET[job.kind] && !WORKING.includes(job.status) && !seen.has(job.id));
+    if (fresh.some(job => jobMode(job) === 'failed')) return 'failed';
+    return fresh.some(job => jobMode(job) === 'done') ? 'done' : null;
+}
+const TAB_PREFIX = {ask: ['Needs your OK · ', '要确认 · '], failed: ['Failed · ', '出错了 · '], done: ['Done · ', '已完成 · ']};
 export function jobMeta(job, t) {
     const parts = [];
     if (job.width > 0 && job.height > 0) parts.push(`${job.width} × ${job.height}`);
@@ -224,6 +244,25 @@ export function agentChat(t, api, {open, edit, shown = () => true, changed = ()
     const form = el('form', 'fv-chat-input'), input = el('textarea'), sendButton = el('button', 'fv-chat-send');
     input.rows = 1; input.maxLength = 8000; input.setAttribute('aria-label', t('Message the agent', '对 Agent 说'));
     sendButton.type = 'submit'; sendButton.setAttribute('aria-label', t('Send', '发送')); sendButton.append(glyph('M12 19V5M5 12l7-7 7 7'));
+    // While the agent is thinking and has started nothing, the send key is the stop key: same place, a solid square.
+    let stopMode = false;
+    const SEND_GLYPH = 'M12 19V5M5 12l7-7 7 7', STOP_GLYPH = 'M7 7h10v10H7z';
+    function refreshSend() {
+        const label = stopMode ? t('Stop', '停止') : t('Send', '发送');
+        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();
+    }
+    async function abandon() {
+        const session = state.session;
+        if (!session || state.abandoned) return;
+        state = {...state, abandoned: true, stage: null}; render();
+        input.focus();
+        let done = false;
+        try { done = (await dataLayer.abandon(session)).ok; } catch {}
+        if (!done && state.session === session) { state = {...state, abandoned: false}; render(); say(t('Could not stop. Click again.', '没能停止，请再点一次。')); }
+    }
     let whiteboxPreview = false;
     try { whiteboxPreview = localStorage.getItem(WHITEBOX_PREVIEW.storageKey) === 'true'; } catch {}
     const whiteboxButton = press('fv-chat-whitebox', null, () => setWhiteboxPreview(!whiteboxPreview));
@@ -632,6 +671,8 @@ export function agentChat(t, api, {open, edit, shown = () => true, changed = ()
         return row ? row.seconds : null;
     }
     const active = () => !stopped && element.isConnected && !document.hidden && shown();
+    // A hidden page keeps reading, slowly, so that its title can tell what happened.
+    const reading = () => !stopped && element.isConnected && shown();
     const inSelectedChat = () => !creating && (!selectedSession || selectedSession === state.session);
     const composeAllowed = () => inSelectedChat() && state.agent.connected === true;
     const sendingAllowed = () => inSelectedChat() && canSend(state);
@@ -705,6 +746,13 @@ export function agentChat(t, api, {open, edit, shown = () => true, changed = ()
         const stop = press('fv-chat-stop', null, () => void cancelJobs([job.id]));
         foot.append(left, stop); info.append(head, middle, foot); item.replaceChildren(box, info);
         const said = {};
+        // Three minutes without a new step or stage: the card says so, once, in the line that names the stage.
+        let signature = null, changed = performance.now(), quiet = false;
+        const stalled = next => {
+            const key = progressKey(next);
+            if (key !== signature) { signature = key; changed = performance.now(); }
+            return next.status === 'running' && performance.now() - changed >= STALL_MS;
+        };
         // The number rides on the water, and goes under the line once there is no room left above it.
         const level = next => {
             const fraction = fractionOf(next);
@@ -715,8 +763,9 @@ export function agentChat(t, api, {open, edit, shown = () => true, changed = ()
             const seconds = leftOf(next);
             left.textContent = seconds === null ? '' : about(seconds, t);
         };
-        item.tick = () => level(state.jobs[job.id] || job);
-        return next => {
+        let current = job, draw = null;
+        item.tick = () => { current = state.jobs[job.id] || job; level(current); if (stalled(current) !== quiet) draw?.(current); };
+        return draw = next => {
             const cells = jobParts(next);
             level(next);
             name.textContent = jobName(next, t); meta.textContent = jobMeta(next, t); meta.hidden = !meta.textContent;
@@ -726,8 +775,10 @@ export function agentChat(t, api, {open, edit, shown = () => true, changed = ()
             parts.hidden = !cells.length;
             const block = next.progress?.block;
             const where = cardLine(said, next, t);
-            part.textContent = cells.length ? t(`Part ${block.index} / ${block.total}`, `第 ${block.index} / ${block.total} 段`) : where;
-            step.textContent = cells.length ? where : '';
+            quiet = stalled(next);
+            const waiting = t('Progress has not updated recently; generation is still running.', '进度暂未更新，仍在生成中。');
+            part.textContent = cells.length ? t(`Part ${block.index} / ${block.total}`, `第 ${block.index} / ${block.total} 段`) : quiet ? waiting : where;
+            step.textContent = cells.length ? quiet ? waiting : where : '';
             const going = stopping.has(next.id) || next.status === 'cancelling';
             stop.disabled = going; stop.textContent = going ? t('Stopping…', '正在停止…') : t('Stop', '停止');
             stop.setAttribute('aria-busy', String(going));
@@ -908,7 +959,13 @@ export function agentChat(t, api, {open, edit, shown = () => true, changed = ()
         };
         buttons.append(press('fv-chat-yes fv-primary', t(...WHITEBOX.generate), answer(true)),
             press('fv-whitebox-edit', t(...WHITEBOX.edit), () => {
-                if (!disabled && whiteboxState(entry) === 'asked') input.focus();
+                if (disabled || whiteboxState(entry) !== 'asked') return;
+                // The summary goes into an empty box with the cursor after it; what the user has already typed stays.
+                const text = entry.whitebox.summary || entry.whitebox.scene_summary || '';
+                if (text && !input.value.trim()) { input.value = text; input.dispatchEvent(new Event('input')); }
+                input.focus();
+                input.setSelectionRange(input.value.length, input.value.length);
+                input.scrollTop = input.scrollHeight;
             }), press('fv-chat-no fv-quiet', t(...WHITEBOX.decline), answer(false)));
         item.replaceChildren(preview, summary, seconds, buttons, status, error);
         return {update(next) {
@@ -967,7 +1024,9 @@ export function agentChat(t, api, {open, edit, shown = () => true, changed = ()
             buttons.querySelectorAll('button').forEach(button => { button.disabled = disabled; });
             const question = el('p', null, entry.question);
             const estimate = el('p', 'fv-chat-meta', entry.estimate_seconds > 0 ? about(entry.estimate_seconds, t) : ''); estimate.hidden = !estimate.textContent;
-            item.replaceChildren(question, estimate, buttons);
+            // What is about to be made, written as on the job card; a question that carries no size or length shows only the time.
+            const shape = el('p', 'fv-chat-meta', jobMeta(entry, t)); shape.hidden = !shape.textContent;
+            item.replaceChildren(question, shape, estimate, buttons);
             if (error) {
                 const notice = el('p', 'fv-chat-error', error); notice.setAttribute('role', 'alert');
                 item.append(notice, issueButton(t, () => ({where: CHAT_ISSUE, message: error})));
@@ -990,8 +1049,9 @@ export function agentChat(t, api, {open, edit, shown = () => true, changed = ()
     }
     function dots() {
         const item = el('li', 'fv-chat-thinking'), mark = el('span', 'fv-chat-dots'); mark.append(el('i'), el('i'), el('i'));
-        item.setAttribute('role', 'status'); item.setAttribute('aria-label', t('Thinking', '正在思考')); item.append(mark);
-        return {element: item, update() {}};
+        const activity = el('span', 'fv-chat-activity');
+        item.setAttribute('role', 'status'); item.setAttribute('aria-label', t('Thinking', '正在思考')); item.append(mark, activity);
+        return {element: item, update(text) { activity.textContent = text || ''; activity.hidden = !text; }};
     }
     const MAKE = {hello, examples, dots, user: () => message('user'), agent: () => message('agent'), job: jobCard, confirm: confirmCard};
 
@@ -1029,8 +1089,10 @@ export function agentChat(t, api, {open, edit, shown = () => true, changed = ()
         if (!said) wanted.push(['examples', 'examples', null]);
         // The agent has the sentence, or is waiting for one, and has not answered yet.
         const tail = localPending.at(-1) || entries.at(-1);
-        if (tail?.type !== 'job' && tail?.type !== 'confirm' && (localPending.some(entry => !entry.error) || !tail?.error && tail?.role === 'user') && state.agent.connected === true
-            && (state.agent.listening !== false || tail.taken === true)) wanted.push(['dots', 'dots', null]);
+        const thinking = tail?.type !== 'job' && tail?.type !== 'confirm' && (localPending.some(entry => !entry.error) || !tail?.error && tail?.role === 'user') && state.agent.connected === true
+            && (state.agent.listening !== false || tail.taken === true) && (localPending.length > 0 || !state.abandoned);
+        stopMode = thinking && localPending.length === 0;
+        if (thinking) wanted.push(['dots', 'dots', Object.hasOwn(STAGE_TEXT, state.stage) ? t(...STAGE_TEXT[state.stage]) : null]);
 
         const stick = follow || pinned;
         const before = log.children.length, keep = new Set(wanted.map(([key]) => key));
@@ -1044,7 +1106,7 @@ export function agentChat(t, api, {open, edit, shown = () => true, changed = ()
         const sendable = sendingAllowed();
         input.disabled = state.agent.connected !== true || creating || !!selectedSession && selectedSession !== state.session;
         input.placeholder = state.agent.connected === true && !sendable ? t('Ask your agent to continue', '请在 Agent 里让它继续') : t('Message the agent…', '对 Agent 说…');
-        sendButton.disabled = !composeAllowed() || draftSending() || !input.value.trim();
+        refreshSend();
         if (focusSession === state.session && focusSession && !input.disabled && active()) { focusSession = null; input.focus(); }
         if (stick) { pinned = true; scroller.scrollTop = scroller.scrollHeight; down.hidden = true; }
         else if (log.children.length > before) down.hidden = false;
@@ -1143,7 +1205,7 @@ export function agentChat(t, api, {open, edit, shown = () => true, changed = ()
         // At most five lines; the box scrolls after that.
         const line = (typeof getComputedStyle === 'function' && parseFloat(getComputedStyle(input).lineHeight)) || 22;
         input.style.height = `${Math.min(input.scrollHeight, Math.round(line * 5) + 14)}px`;
-        sendButton.disabled = !composeAllowed() || draftSending() || !input.value.trim();
+        refreshSend();
     }
     input.addEventListener('input', () => {
         if (entryDraft !== null) entryDraft = input.value;
@@ -1160,10 +1222,11 @@ export function agentChat(t, api, {open, edit, shown = () => true, changed = ()
     // Enter sends, Shift+Enter starts a new line; Enter that confirms an input-method candidate does neither.
     input.addEventListener('keydown', event => {
         if (event.key !== 'Enter' || event.shiftKey || event.isComposing || event.keyCode === 229) return;
-        event.preventDefault(); form.requestSubmit();
+        event.preventDefault(); if (!stopMode) form.requestSubmit();
     });
     form.addEventListener('submit', event => {
         event.preventDefault();
+        if (stopMode) { void abandon(); return; }
         const text = input.value;
         const draft = saveDraft();
         if (!text.trim() || !composeAllowed() || draftSending()) return;
@@ -1181,9 +1244,9 @@ export function agentChat(t, api, {open, edit, shown = () => true, changed = ()
         clearTimeout(timer); timer = null;
         if (stopped) return;
         if (busy) { reread = true; return; }
-        if (!active()) return;
+        if (!reading()) return;
         busy = true; reread = false;
-        void readConversations();
+        if (!document.hidden) void readConversations();
         const readSelection = selection, requested = selectedSession;
         let again = false;
         try {
@@ -1231,8 +1294,8 @@ export function agentChat(t, api, {open, edit, shown = () => true, changed = ()
         known = true;
         // The next read is booked before the page hears of this one, so the page's sync() does not start another.
         const idle = state.agent.connected !== true && !state.active;
-        timer = setTimeout(poll, chatDelay({idle, interval, failures}));
-        render(); report();
+        timer = setTimeout(poll, chatDelay({hidden: document.hidden, idle, interval, failures}));
+        render(); report(); tab();
         if (sendingAllowed()) for (const entry of [...pending]) {
             if (entry.session !== state.session || !entry.automatic || entry.sending || entry.acknowledged) continue;
             await send(entry.text, entry, entry.draft, true);
@@ -1245,7 +1308,28 @@ export function agentChat(t, api, {open, edit, shown = () => true, changed = ()
         for (const item of log.children) item.tick?.();
         report();
     }, 250);
-    const visible = () => { if (!document.hidden) sync(); };
+    // The title of a hidden tab carries one fixed prefix: no flashing, no counts. It is the page's own title again when the user returns.
+    let titleBase = null, titleShown = null, seen = null;
+    const finished = () => new Set(Object.values(state.jobs).filter(job => !WORKING.includes(job.status)).map(job => job.id));
+    const asking = () => [...state.items.values()].some(entry => entry.type === 'confirm'
+        && (entry.whitebox ? whiteboxState(entry) : answers.get(cardKey(state.session, entry.id)) || state.confirms[entry.id] || entry.state || 'asked') === 'asked');
+    function tab() {
+        if (titleShown !== null && document.title !== titleShown) titleBase = titleShown = null;
+        if (stopped || !document.hidden || !known) {
+            if (titleShown !== null && document.title === titleShown) document.title = titleBase;
+            titleBase = titleShown = null; seen = null; return;
+        }
+        seen ||= finished();
+        const notice = tabNotice(state.jobs, asking(), seen);
+        if (!notice) {
+            if (titleShown !== null) document.title = titleBase;
+            titleBase = titleShown = null; return;
+        }
+        if (titleBase === null) titleBase = document.title;
+        // The browser trims and collapses spaces in a title; what it shows is what the next look compares with.
+        document.title = t(...TAB_PREFIX[notice]) + titleBase; titleShown = document.title;
+    }
+    const visible = () => { if (document.hidden) { seen = finished(); tab(); } else { tab(); if (reading()) void poll(); } };
     document.addEventListener('visibilitychange', visible);
     function sync() {
         if (pinned) scroller.scrollTop = scroller.scrollHeight;
@@ -1265,7 +1349,7 @@ export function agentChat(t, api, {open, edit, shown = () => true, changed = ()
             try { await cancelJobs(ids); }
             finally { bulkStopping = false; if (!stopped) { report(); void poll(); } }
         },
-        stop() { closeMenu(); stopped = true; dismissImage?.(); closed.abort(); clearInterval(moving); clearTimeout(timer); timer = null;
+        stop() { closeMenu(); stopped = true; tab(); dismissImage?.(); closed.abort(); clearInterval(moving); clearTimeout(timer); timer = null;
             document.removeEventListener('visibilitychange', visible); document.removeEventListener('pointerdown', outside, true);
             document.removeEventListener('keydown', keys, true); window.removeEventListener('resize', resize); list.removeEventListener('scroll', listScroll);
             log.querySelectorAll('video').forEach(video => video.pause()); }};
diff --git a/web/capability_view.js b/web/capability_view.js
index d0c01f11..4e5dd3d1 100644
--- a/web/capability_view.js
+++ b/web/capability_view.js
@@ -383,7 +383,25 @@ export function createCapabilityView({node, api, t, getState, onPick, onEstimate
         const i = grid.sizes.findIndex(s => s.width === Number(value(node, 'width')) && s.height === Number(value(node, 'height')));
         const j = grid.lengths.findIndex(l => l.seconds === Number(value(node, 'seconds')));
         const seconds = tier?.estimate_seconds?.[i]?.[j];
-        return positive(seconds) ? aboutTime(seconds, t) : '';
+        if (!positive(seconds)) return '';
+        // The suffix stays in one piece so a narrow line moves it whole to the next line.
+        return aboutTime(seconds, t) + (grid.machine?.compile_owed === true ? t(' · \u2060includes\u00a0first-time\u00a0setup', ' · \u2060含\u2060首\u2060次\u2060准\u2060备') : '');
+    }, warning(steps, twoPass, refine, stepCache = null, upscale = 2) {
+        if (!grid || !surface) return '';
+        const tier = gridTier(grid, steps, twoPass, refine, stepCache, upscale);
+        const i = grid.sizes.findIndex(s => s.width === Number(value(node, 'width')) && s.height === Number(value(node, 'height')));
+        const j = grid.lengths.findIndex(l => l.seconds === Number(value(node, 'seconds')));
+        const code = tier?.plan?.[i]?.[j];
+        if (!(code > 0) || !positive(tier.estimate_seconds?.[i]?.[j])) return '';
+        const why = tier.slow_by?.[i]?.[j];
+        const head = why === 'gpu_memory' ? t('Not enough graphics memory, so this will be slow.', '显存不足，这次会很慢。')
+            : why === 'memory' ? t('Not enough RAM, so this will be slow.', '内存不足，这次会很慢。') : t('This will be slow.', '这次会很慢。');
+        // One remedy only, and only when it really removes the slowdown.
+        const levels = effortsFrom(grid), at = levels.findIndex(l => grid.tiers?.[l.name.toLowerCase()] === tier);
+        const lower = at > 0 && levels.slice(0, at).some(l => { const row = grid.tiers?.[l.name.toLowerCase()]; return row?.plan?.[i]?.[j] === 0 && positive(row.estimate_seconds?.[i]?.[j]); });
+        if (lower) return head + t(' ', '') + t('Choose a lower Quality level.', '请把质量降低一档。');
+        const smaller = tier.plan.slice(0, i).some((row, k) => row[j] === 0 && positive(tier.estimate_seconds?.[k]?.[j]) && grid.sizes[k].pixels < grid.sizes[i].pixels);
+        return smaller ? head + t(' ', '') + t('Choose a smaller size.', '请缩小尺寸。') : head;
     }, dispose() {
         disposed = true; controller?.abort(); clearTimeout(retry); observer.disconnect();
         api.removeEventListener('freevideo_capabilities', changed); window.removeEventListener('pointerdown', outside);
diff --git a/web/chain_support.js b/web/chain_support.js
index 81014159..ee82d2c4 100644
--- a/web/chain_support.js
+++ b/web/chain_support.js
@@ -4,6 +4,17 @@ import { effortFor, effortsFrom } from './sampling_effort.js';
 const clockDefault = {now: () => Date.now(), setTimeout: (fn, ms) => setTimeout(fn, ms), clearTimeout: id => clearTimeout(id)};
 export const chainRoute = (id, action) => '/freevideo/chain' + (id ? '/' + encodeURIComponent(id) : '') + (action ? '/' + action : '');
 
+// The line for a chain that waits behind others: its own time when one video is ahead and its time left is known.
+export function waitingText(waiting, t) {
+    const ahead = waiting?.ahead, left = waiting?.remaining_seconds;
+    if (!Number.isInteger(ahead) || ahead < 1) return t('Queued', '排队中');
+    if (ahead > 1) return t(`Queued · ${ahead} ahead`, `排队中 · 前面还有 ${ahead} 个`);
+    if (!Number.isFinite(left) || left < 0) return t('Queued · waiting for the previous one', '排队中 · 等前一个完成');
+    const time = left < 60 ? t(`about ${Math.max(1, Math.ceil(left))} s`, `约 ${Math.max(1, Math.ceil(left))} 秒`)
+        : t(`about ${Math.ceil(left / 60)} min`, `约 ${Math.ceil(left / 60)} 分钟`);
+    return t(`Queued · starts in ${time}`, `排队中 · ${time}后开始`);
+}
+
 // No overlapping polls and no delivery after disposal, including a slow response.
 export function createPoller(read, deliver, {clock = clockDefault, interval = 2000, onError = () => {}} = {}) {
     let active = true, timer = null, running = null;
diff --git a/web/chat_data.js b/web/chat_data.js
index fe7fec75..592c74ac 100644
--- a/web/chat_data.js
+++ b/web/chat_data.js
@@ -8,7 +8,7 @@ export async function sendError(response) {
     try { return (await (response.clone?.() || response).json())?.error || null; }
     catch { return null; }
 }
-export const emptyChat = () => ({session: undefined, items: new Map(), jobs: {}, failures: {}, confirms: {}, last: 0, read: 0, agent: {}, current: null, active: 0});
+export const emptyChat = () => ({session: undefined, items: new Map(), jobs: {}, failures: {}, confirms: {}, last: 0, read: 0, agent: {}, current: null, active: 0, abandoned: false, stage: null});
 // One answer from the server folded into what the page already has: entries add up by their number, jobs are
 // replaced by id (a card keeps its last state after the server stops listing its job), and so are the answers to its questions.
 export function chatMerge(state, data) {
@@ -17,7 +17,8 @@ export function chatMerge(state, data) {
     const read = Math.max(state.read, Number(data.read_seq) || 0, ...(data.items || []).filter(item => item.read === true).map(item => item.seq));
     return {session: data.session ?? null, items, jobs: {...state.jobs, ...(data.jobs || {})}, failures: data.failures || {}, confirms: {...state.confirms, ...(data.confirms || {})}, read,
         last: Math.max(state.last, Number(data.last_seq) || 0, ...items.keys()), agent: data.agent || {},
-        current: data.current || null, active: Number(data.active) || 0};
+        current: data.current || null, active: Number(data.active) || 0,
+        abandoned: data.abandoned === true, stage: typeof data.stage === 'string' ? data.stage : null};
 }
 // The entries in order. A user's sentence counts as taken once the agent has read it or has said or done anything after it.
 export function chatEntries(state) {
@@ -91,6 +92,7 @@ export function chatData(api) {
         create: () => agentPost(api, '/freevideo/agent/chat/new', {}),
         remove: (session, stop = false) => agentPost(api, '/freevideo/agent/chat/delete', stop ? {session, stop: true} : {session}),
         confirm: body => agentPost(api, '/freevideo/agent/chat/confirm', body),
+        abandon: session => agentPost(api, '/freevideo/agent/chat/abandon', session ? {session} : {}),
         cancel: id => retryAgentPost(api, '/freevideo/agent/activity/cancel', {id}),
         activity: (cursor, options) => read('/freevideo/agent/activity' + (cursor ? '?cursor=' + encodeURIComponent(cursor) : ''), options),
         sendMessage: (entry, options) => sendChatMessage(api, entry, options),
@@ -99,7 +101,9 @@ export function chatData(api) {
     };
 }
 export function chatDelay({hidden = false, phone = false, idle = false, interval = 1500, failures = 0} = {}) {
-    return phone ? (hidden ? 10000 : 2000) : (idle ? Math.max(interval, 3000) : interval) * Math.min(4, 1 + failures);
+    // A page nobody is looking at asks less often; it only has to notice what the tab title announces.
+    if (phone) return hidden ? 10000 : 2000;
+    return Math.max(hidden ? 10000 : 0, idle ? Math.max(interval, 3000) : interval) * Math.min(4, 1 + failures);
 }
 // Serial polling, visibility rescheduling and disposal work without a browser.
 export function chatPoller(read, {delay = () => 2000, error = () => {}, clock = globalThis} = {}) {
diff --git a/web/editor.css b/web/editor.css
index fa1681ae..fbeee9c6 100644
--- a/web/editor.css
+++ b/web/editor.css
@@ -115,6 +115,10 @@
 .fv-ed-ref-chip{border-color:transparent;background:var(--fv-bg)}
 .fv-ed-ref-add:not(:disabled):not([aria-disabled=true]):hover,.fv-ed-ref-add[aria-expanded=true],.fv-ed-ref-chip:not(:disabled):not([aria-disabled=true]):hover{border-color:var(--fv-accent);color:var(--fv-ink)}
 .fv-ed-ref-add:disabled,.fv-ed-ref-chip:disabled,.fv-ed-ref-add[aria-disabled=true],.fv-ed-ref-chip[aria-disabled=true]{opacity:.4;cursor:default}
+.fv-ed-ref-add[data-uploading=true]{opacity:1;color:transparent}
+.fv-ed-ref-add[data-uploading=true]::after{content:'';position:absolute;inset:0;margin:auto;width:12px;height:12px;border:2px solid var(--fv-accent);border-right-color:transparent;border-radius:50%;animation:fv-ref-spin .9s linear infinite}
+@keyframes fv-ref-spin{to{transform:rotate(360deg)}}
+@media(prefers-reduced-motion:reduce){.fv-ed-ref-add[data-uploading=true]::after{animation:none}}
 .fv-ed-ref-thumb{position:absolute;inset:0;display:flex;align-items:center;justify-content:center;font-size:11px;color:var(--fv-muted)}
 .fv-ed-ref-thumb img{display:block;width:100%;height:100%;object-fit:cover}
 .fv-ed-ref-menu{position:absolute;bottom:calc(100% - 4px);min-width:180px;max-width:260px;padding:6px;background:var(--fv-card);border:1px solid var(--fv-border);border-radius:8px;z-index:12;display:flex;flex-direction:column;font:13px var(--fv-font)}
diff --git a/web/editor.js b/web/editor.js
index e89323c9..5055fab9 100644
--- a/web/editor.js
+++ b/web/editor.js
@@ -32,7 +32,7 @@ import {createFirstRunGuide} from './first_run_guide.js';
 import {SHOWCASE_SCENE} from './entry_showcase.js';
 import {createEditorFailures} from './editor_failure.js';
 import {fitEditorHeader} from './editor_header_fit.js';
-import {optionalText} from './optional_inputs.js';
+import {optionalText, clonePrompt} from './optional_inputs.js';
 import {outputNotWritable} from './error_panel.js';
 import {repairAction} from './repair_hint.js';
 import {expandCast} from './cast_mention.js';
@@ -497,14 +497,15 @@ export function openEditor({api, app, from = null, panel = {}, components: given
         title.title = state.record.name;
         exportButton.classList.toggle('fv-primary', !!model.exportFrames(state.record) && !state.selection && !jobActive() && !exportStatus);
         const running = exportPreparing || ['running', 'queued'].includes(exportStatus?.state);
-        if (['running', 'queued'].includes(exportStatus?.state)) exportButton.textContent = t(`Export ${Math.round((exportStatus.fraction || 0) * 100)}%`, `导出 ${Math.round((exportStatus.fraction || 0) * 100)}%`);
+        if (['running', 'queued'].includes(exportStatus?.state) && exportStatus.queued) exportButton.textContent = t('Queued', '排队中');
+        else if (['running', 'queued'].includes(exportStatus?.state)) exportButton.textContent = t(`Export ${Math.round((exportStatus.fraction || 0) * 100)}%`, `导出 ${Math.round((exportStatus.fraction || 0) * 100)}%`);
         else if (exportPreparing) exportButton.textContent = queue.dirty ? t('Saving changes…', '正在保存改动…') : t('Exporting', '正在导出');
         else exportButton.textContent = exportStatus?.state === 'done' && exportMatches() ? t('View', '查看') : t('Export', '导出');
         // Never disabled: a press that cannot export says why next to it.
         exportButton.disabled = false; exportButton.title = '';
     }
     async function snapshot() {
-        let value = panel.snapshot ? structuredClone(await panel.snapshot()) : structuredClone(await app.graphToPrompt());
+        let value = panel.snapshot ? clonePrompt(await panel.snapshot()) : clonePrompt(await app.graphToPrompt());
         if (value.output) {
             const node = String(panel.node?.id || panel.node || Object.keys(value.output).find(id => value.output[id].class_type === 'FreeVideoGenerate') || '');
             value = {node, prompt: value.output, extra_data: {extra_pnginfo: {workflow: value.workflow || {}}}};
@@ -1042,7 +1043,7 @@ export function openEditor({api, app, from = null, panel = {}, components: given
             download: report => new Promise((resolve, reject) => {
                 const step = value => {
                     const install = value?.install;
-                    if (install?.state !== 'running') { resolve(install?.state === 'done'); return false; }
+                    if (install?.state !== 'running') { resolve(install?.state === 'done' || install?.reason || false); return false; }
                     report(install.total > 0 ? install.done / install.total : 0); return true;
                 };
                 store.installEars().then(value => { if (step(value)) store.watchEarsPack(step); }, reject);
diff --git a/web/editor_entry.js b/web/editor_entry.js
index 0a576e25..e976249c 100644
--- a/web/editor_entry.js
+++ b/web/editor_entry.js
@@ -3,7 +3,7 @@ import {openEditor} from './editor.js';
 import {api} from '../../scripts/api.js';
 import {app} from '../../scripts/app.js';
 import {openStudio, showNodes} from './studio.js';
-import {cleanOptionalPrompt} from './optional_inputs.js';
+import {cleanOptionalPrompt, clonePrompt} from './optional_inputs.js';
 import {openAutopilot, viewChanged} from './view_navigation.js';
 import {mediaReferences} from './editor_media_references.js';
 import {turnNodeOn} from './studio_queue.js';
@@ -18,7 +18,7 @@ function editorPanel(node) {
             // A muted or bypassed node would be left out of the request; if no request comes out, it stays as it was.
             const held = turnNodeOn(node);
             let value;
-            try { value = cleanOptionalPrompt(structuredClone(await app.graphToPrompt())); } catch (error) { held?.restore(); throw error; }
+            try { value = cleanOptionalPrompt(clonePrompt(await app.graphToPrompt())); } catch (error) { held?.restore(); throw error; }
             held?.done();
             return {node: String(node.id), prompt: value.output, extra_data: {extra_pnginfo: {workflow: value.workflow || {}}}};
         }};
diff --git a/web/editor_genbar.js b/web/editor_genbar.js
index 59cf5e7e..19d515dc 100644
--- a/web/editor_genbar.js
+++ b/web/editor_genbar.js
@@ -4,6 +4,7 @@ import {createEditorReferences} from './editor_references.js';
 import {createExpandButton} from './editor_expand.js';
 import {createRequestAssets} from './request_assets.js';
 import {issueButton} from './issue_button.js';
+import {waitingText} from './chain_support.js';
 const el = (tag, cls, text) => { const e = document.createElement(tag); if (cls) e.className = cls; if (text != null) e.textContent = text; return e; };
 export const timeEstimate = (seconds, t) => seconds == null ? '' : seconds < 60
     ? t(`about ${Math.max(1, Math.ceil(seconds))} s`, `约 ${Math.max(1, Math.ceil(seconds))} 秒`)
@@ -137,7 +138,7 @@ export function createEditorGenbar(host, {state, t, actions, text = ''}) {
         cancel.hidden = !busy || opening || cancelling || !held && finalization?.cancelable === false; resume.hidden = !stopped;
         const waiting = status?.state === 'queued' && !(status.blocks || []).some(block => ['queued', 'running'].includes(block.state) && block.prompt_id);
         if (opening) hint.textContent = t('Starting…', '正在开始…');
-        else if (held || waiting) hint.textContent = t('Queued', '排队中');
+        else if (held || waiting) hint.textContent = waiting && !held ? waitingText(status.waiting, t) : t('Queued', '排队中');
         else if (cancelling) hint.textContent = t('Stopping…', '正在停止…');
         else if (running) {
             const blocks = status?.blocks || [], done = blocks.filter(block => block.state === 'done').length;
diff --git a/web/editor_library.js b/web/editor_library.js
index cac1cf80..261d2ac0 100644
--- a/web/editor_library.js
+++ b/web/editor_library.js
@@ -1,6 +1,7 @@
 import {editorIcon} from './editor_icons.js';
 import {sourceKey} from './editor_store.js';
 import {issueButton} from './issue_button.js';
+import {connectionLostText} from './browser_extension.js';
 const LIBRARY_ISSUE = ['Editor media library', '剪辑器素材库'];
 const el = (tag, cls, text) => { const e = document.createElement(tag); if (cls) e.className = cls; if (text != null) e.textContent = text; return e; };
 const stamp = frames => { const seconds = Math.floor((frames || 0) / 24); return `${Math.floor(seconds / 60)}:${String(seconds % 60).padStart(2, '0')}`; };
@@ -24,6 +25,10 @@ export function createEditorLibrary(host, {store, t, actions, state}) {
     const uploading = new Map(), recentImports = new Map(), pendingAdds = new Map();
     const preparing = () => t('Preparing this clip. It goes on the timeline when ready.', '素材准备好后会自动放到时间线上。');
     const cannotOpen = () => t('This file cannot be opened. Supported formats: ', '这个文件打不开。支持的格式：') + FORMATS;
+    // Why an import did not work, and where to report it: only a file that could not be read lists the formats and asks for a report.
+    const importFailure = (reason, status) => status === 413 ? [t('This file is over 8 GiB. Trim or compress it, then try again.', '文件超过 8 GiB，请先剪短或压缩后再导入。')]
+        : reason === 'disk_full' ? [t('The disk is full. Free some space, then import again.', '磁盘空间不足，请清理出空间后再导入。')]
+        : reason === 'import_failed' ? [connectionLostText(t)] : [cannotOpen(), LIBRARY_ISSUE];
     function metadata(row, kind) {
         const g = row.geometry || row, cached = store.sources.get(`${kind}:${row.id}`);
         return {frames: cached?.frames || Math.round(g.frames || (g.seconds || 0) * (g.fps || 24)), hasVideo: cached?.has_video ?? (kind === 'work' || row.kind === 'video'),
@@ -45,7 +50,7 @@ export function createEditorLibrary(host, {store, t, actions, state}) {
     function requestAdd(row, kind, plus = false) {
         if (disposed) return;
         if (kind === 'work' || row.state === 'ready') return add(row, kind, plus).catch(actions.error);
-        if (row.state === 'failed') { actions.notify?.(cannotOpen(), LIBRARY_ISSUE); return; }
+        if (row.state === 'failed') { actions.notify?.(...importFailure(row.reason)); return; }
         const key = `${kind}:${row.id}`;
         if (!pendingAdds.has(key)) pendingAdds.set(key, {plus, project: actions.projectToken?.()});
         actions.notify?.(preparing());
@@ -58,7 +63,7 @@ export function createEditorLibrary(host, {store, t, actions, state}) {
             if (row.state === 'ready') {
                 pendingAdds.delete(key);
                 add(row, 'import', pending.plus, {...pending, queued: true}).catch(actions.error);
-            } else if (row.state === 'failed') { pendingAdds.delete(key); actions.notify?.(cannotOpen(), LIBRARY_ISSUE); }
+            } else if (row.state === 'failed') { pendingAdds.delete(key); actions.notify?.(...importFailure(row.reason)); }
         }
     }
     function card(row, kind) {
@@ -74,9 +79,13 @@ export function createEditorLibrary(host, {store, t, actions, state}) {
         else if (row.kind === 'audio') picture.append(el('span', 'fv-ed-audio-icon', '♫'));
         if (ready) picture.append(el('span', 'fv-ed-source-time', stamp(metadata(row, kind).frames)));
         else if (row.state === 'failed') {
-            picture.append(el('span', 'fv-ed-source-preparing', t('Cannot open', '打不开')));
-        } else picture.append(el('span', 'fv-ed-source-preparing', t('Preparing', '准备中')
-            + (Number.isFinite(row.fraction) ? ` ${Math.round(Math.max(0, Math.min(1, row.fraction)) * 100)}%` : '')));
+            picture.append(el('span', 'fv-ed-source-preparing', row.reason === 'disk_full' ? t('Disk full', '磁盘已满') : t('Cannot open', '打不开')));
+        } else {
+            // Sending, then waiting behind another import, then being made; only the first and the last have a percentage.
+            const percent = Number.isFinite(row.fraction) ? ` ${Math.round(Math.max(0, Math.min(1, row.fraction)) * 100)}%` : '';
+            picture.append(el('span', 'fv-ed-source-preparing', row.sending ? t('Uploading', '上传中') + percent
+                : row.started ? t('Processing', '处理中') + percent : t('Queued', '排队中')));
+        }
         button.append(picture);
         const used = el('span', 'fv-ed-source-used'); used.append(editorIcon('check')); used.setAttribute('aria-label', t('Used', '已用')); used.hidden = true; picture.append(used);
         button.ondblclick = () => requestAdd(row, kind);
@@ -97,7 +106,7 @@ export function createEditorLibrary(host, {store, t, actions, state}) {
         plus.ondblclick = event => { event.preventDefault(); event.stopPropagation(); };
         plus.onkeydown = event => { if (event.key === 'Enter' || event.key === ' ') insert(event); };
         container.append(button, plus);
-        if (kind !== 'work' && row.state === 'failed') container.append(issueButton(t, () => ({where: LIBRARY_ISSUE, message: t('Cannot open', '打不开')})));
+        if (kind !== 'work' && row.state === 'failed' && row.reason !== 'disk_full') container.append(issueButton(t, () => ({where: LIBRARY_ISSUE, message: t('Cannot open', '打不开')})));
         return container;
     }
     function updateUsed() {
@@ -122,8 +131,11 @@ export function createEditorLibrary(host, {store, t, actions, state}) {
         for (const row of shown) grid.append(card(row, tab));
         updateUsed();
         message.textContent = ''; // Empty states stay silent.
-        for (const row of shown) if (row.state === 'failed') message.textContent = t('This file cannot be opened. Supported formats: ', '这个文件打不开。支持的格式：') + FORMATS;
-        if (message.textContent) message.append(issueButton(t, () => ({where: LIBRARY_ISSUE, message: cannotOpen()})));
+        const failed = shown.findLast(row => row.state === 'failed');
+        if (failed) {
+            const [text, where] = importFailure(failed.reason); message.textContent = text;
+            if (where) message.append(issueButton(t, () => ({where, message: text})));
+        }
         if (scroll > 0) {
             const replacement = anchor && [...grid.querySelectorAll('.fv-ed-source')].find(node => node.dataset.key === anchor.dataset.key);
             host.scrollTop = replacement ? scroll + replacement.getBoundingClientRect().top - anchorTop : scroll;
@@ -212,7 +224,7 @@ export function createEditorLibrary(host, {store, t, actions, state}) {
         const place = Number.isSafeInteger(at) && at >= 0 ? {at} : null;
         for (const value of files) {
             if (disposed) break;
-            const id = 'upload-' + Math.random(), row = {id, name: value.name, kind: value.type.startsWith('audio') ? 'audio' : 'video', state: 'preparing', fraction: 0};
+            const id = 'upload-' + Math.random(), row = {id, name: value.name, kind: value.type.startsWith('audio') ? 'audio' : 'video', state: 'preparing', fraction: 0, sending: true};
             uploading.set(id, row); render();
             try {
                 const item = await store.upload(value, fraction => { row.fraction = fraction; render(); });
@@ -225,7 +237,7 @@ export function createEditorLibrary(host, {store, t, actions, state}) {
                     if (place) { placing.push({id: item.id, place}); placeReady(imported); }
                     else startPending([item]);
                 }
-            } catch { actions.notify(cannotOpen(), LIBRARY_ISSUE); }
+            } catch (error) { actions.notify(...importFailure(error?.message === 'import_failed' ? 'import_failed' : error?.reason, error?.status)); }
             finally { pendingAdds.delete(`import:${id}`); uploading.delete(id); render(); }
         }
     }
diff --git a/web/editor_references.js b/web/editor_references.js
index 35c00876..1dfc1d6d 100644
--- a/web/editor_references.js
+++ b/web/editor_references.js
@@ -24,7 +24,7 @@ export function createEditorReferences({t, source, size = () => null, error = ()
     const menu = el('div', 'fv-ed-ref-menu'); menu.hidden = true; menu.setAttribute('role', 'menu');
     const picker = el('input'); picker.type = 'file'; picker.accept = 'image/*'; picker.multiple = true; picker.hidden = true;
     root.append(add, chips, menu, picker);
-    let shown = false, continuing = false, busy = false, working = false, opener = null, disposed = false, entries = [], waited = false;
+    let shown = false, continuing = false, busy = false, working = false, uploading = false, opener = null, disposed = false, entries = [], waited = false;
     const active = () => source.rows().map((row, index) => ({row, index})).filter(({row}) => row.enabled !== false);
     const isImage = file => IMAGE.test(file || '');
     // The pictures this part uses: all of them, or the first reference images of a part that continues
@@ -80,7 +80,7 @@ export function createEditorReferences({t, source, size = () => null, error = ()
     };
     async function addFiles(files) {
         if (working) return;
-        working = true; update();
+        working = uploading = true; update();
         let skipped = false;
         try {
             for (const file of files) {
@@ -91,7 +91,7 @@ export function createEditorReferences({t, source, size = () => null, error = ()
             }
             if (skipped && !disposed) notify(t('Could not read this picture. Please choose another one.', '无法读取这张图片，请换一张。'), ['Editor reference images', '剪辑器参考图']);
         } catch (value) { error(value); }
-        finally { working = false; picker.value = ''; update(); }
+        finally { working = uploading = false; picker.value = ''; update(); }
     }
     async function generate() {
         if (working) return;
@@ -151,8 +151,9 @@ export function createEditorReferences({t, source, size = () => null, error = ()
         const isFull = full();
         // While a part is being made the controls stay clickable and say why they do nothing.
         add.disabled = working;
+        add.dataset.uploading = String(uploading);
         add.setAttribute('aria-disabled', String(busy || isFull));
-        add.title = busy ? waitText() : isFull ? fullText() : t('Reference image', '参考图');
+        add.title = uploading ? t('Uploading reference image…', '正在上传参考图…') : busy ? waitText() : isFull ? fullText() : t('Reference image', '参考图');
         chips.replaceChildren(...used().map(entry => {
             const chip = el('button', 'fv-ed-ref-chip'); chip.type = 'button'; chip.disabled = working; chip.setAttribute('aria-disabled', String(busy));
             chip.dataset.role = entry.row.role || 'reference';
diff --git a/web/editor_store.js b/web/editor_store.js
index 896b11f8..caa8bc12 100644
--- a/web/editor_store.js
+++ b/web/editor_store.js
@@ -391,8 +391,9 @@ export function createEditorStore(api, {t = en => en, notify = () => {}, onGener
                 xhr.upload.onprogress = event => { if (event.lengthComputable) progress(event.loaded / event.total); };
                 xhr.onload = () => {
                     release();
-                    if (xhr.status === 404) { notice('import', 'Import is unavailable until you restart after updating.', '更新后重启即可导入素材。'); resolve(null); return; }
-                    if (xhr.status < 200 || xhr.status >= 300) { reject(errorFor(xhr)); return; }
+                    if (xhr.status === 404) { notice('import', 'Import is unavailable until you restart after updating.', '更新后重启可导入素材。'); resolve(null); return; }
+                    // The server's reason (a full disk) comes along, as with a failed save.
+                    if (xhr.status < 200 || xhr.status >= 300) { let body = null; try { body = JSON.parse(xhr.responseText); } catch {} reject(Object.assign(errorFor(xhr), {reason: body?.reason})); return; }
                     try { resolve(JSON.parse(xhr.responseText).import); } catch (error) { reject(error); }
                 };
                 xhr.onerror = () => { release(); reject(new Error('import_failed')); };
diff --git a/web/editor_text.js b/web/editor_text.js
index 71e8403a..df42b3fb 100644
--- a/web/editor_text.js
+++ b/web/editor_text.js
@@ -3,7 +3,7 @@
 import * as model from './editor_model.js';
 import {FONT_KEYS} from './fonts.js';
 import {row, button, slider, number, segmented, toggle} from './editor_adjust.js';
-import {isText, isTextTrack, addText, changeText, changeSubtitleStyle, styleOf, activeText, setTextTrack, parseSRT, importSRT, addSubtitles, defaults, cleanStyle, SUBTITLE_DEFAULTS} from './editor_text_model.js';
+import {isText, isTextTrack, addText, changeText, changeSubtitleStyle, styleOf, activeText, setTextTrack, parseSRT, decodeSRT, importSRT, addSubtitles, defaults, cleanStyle, SUBTITLE_DEFAULTS} from './editor_text_model.js';
 import {loadCoverage, loadTextFonts, textFontStack, missingGlyphs, missingHint, overTwoHint, textLayout, drawText} from './editor_text_layout.js';
 import {createSelectionActions, isTyping} from './editor_select.js';
 import {plan} from './editor_geometry.js';
@@ -69,10 +69,14 @@ export function createEditorText({element,state,t,actions,adjust,player,toolbar,
     function speechLabel(fraction){speech.shown=fraction;const p=Math.round(Math.min(1,Math.max(0,Number(fraction)||0))*100),size=Math.max(1,Math.round(speech.need/1e6)),text=fraction==null?(speech.need?t(`Auto subtitles · Download ${size} MB`,`自动字幕 · 下载 ${size} MB`):t('Auto subtitles','自动字幕')):speech.loading?t(`Cancel · Downloading ${p}%`,`取消 · 下载 ${p}%`):t(`Cancel · ${p}%`,`取消 · ${p}%`);for(const b of speech.buttons){b.textContent=text;b.hidden=!(speech.available||speech.need);}}
     function speechEnd(text,where=null){speech.stop?.();speech.stop=null;speech.job=null;speech.busy=false;speech.loading=false;speechLabel(null);if(text)actions.notify?.(text,where);}
     const speechFailed=()=>t('Could not create subtitles','没能生成字幕'),speechSilent=()=>t('No speech found','没有听到说话声');
-    const speechReason=reason=>reason==='no_sound'?speechSilent():reason==='language'?t('Language not supported yet','暂不支持这种语言'):speechFailed();
+    const speechDisk=()=>t('The disk is full. Free some space, then try again.','磁盘空间不足，请清理出空间后重试。');
+    // Causes the person can act on are named and carry no issue button; anything else is unexplained.
+    const speechReasons={no_sound:speechSilent,language:()=>t('Language not supported yet','暂不支持这种语言'),too_long:()=>t('Auto subtitles work on up to 2 hours of sound.','自动字幕最多处理 2 小时的声音。'),
+        source_missing:()=>t('A source file is missing. Import it again from the library.','有素材的原文件找不到了，请在素材库重新导入。'),unreadable:()=>t('The sound in a clip cannot be read.','有素材的声音无法读取。'),disk_full:speechDisk};
+    const speechReason=reason=>(Object.hasOwn(speechReasons,reason)?speechReasons[reason]:speechFailed)();
     function speechResult(job){
         if(job.state==='running'){speechLabel(job.fraction);return true;}
-        if(job.state!=='done'){speechEnd(job.state==='cancelled'?null:speechReason(job.reason),job.state==='cancelled'||['no_sound','language'].includes(job.reason)?null:SPEECH_ISSUE);return false;}
+        if(job.state!=='done'){speechEnd(job.state==='cancelled'?null:speechReason(job.reason),job.state==='cancelled'||Object.hasOwn(speechReasons,job.reason)?null:SPEECH_ISSUE);return false;}
         const cues=(Array.isArray(job.result?.lines)?job.result.lines:[]).filter(l=>typeof l?.text==='string'&&l.text.trim()&&l.end>l.start).map(l=>({start:l.start,frames:l.end-l.start,text:l.text.trim()}));
         speechEnd();if(!cues.length){actions.notify?.(speechSilent());return false;}
         finishInline();commitDraft();
@@ -88,10 +92,11 @@ export function createEditorText({element,state,t,actions,adjust,player,toolbar,
                 speech.loading=true;speechLabel(0);
                 const done=await actions.speech.download(fraction=>{if(run===speech.run)speechLabel(fraction);});
                 if(run!==speech.run)return;
-                const ability=done?await actions.speech.ability():null;
+                // `done` is true, or the reason the download stopped.
+                const ability=done===true?await actions.speech.ability():null;
                 if(run!==speech.run)return;
                 if(ability){speech.available=ability.available;speech.need=ability.need;}
-                if(!done||speech.need){speechEnd(t('Could not download. Try again.','没能下载，请再试一次。'),SPEECH_ISSUE);return;}
+                if(done!==true||speech.need){speechEnd(done==='disk_space'?speechDisk():t('Could not download. Try again.','没能下载，请再试一次。'),done==='disk_space'?null:SPEECH_ISSUE);return;}
                 if(!speech.available){speechEnd(speechFailed(),SPEECH_ISSUE);return;}
                 speech.loading=false;
             }
@@ -164,9 +169,16 @@ export function createEditorText({element,state,t,actions,adjust,player,toolbar,
         }else{
             const file=el('input');file.type='file';file.accept='.srt';file.hidden=true;
             const importButton=button(t('Import subtitle file','导入字幕文件'),{onClick:()=>file.click()});importButton.classList.add('fv-ed-import-srt');
-            // Only a file that cannot be read says so; a refused edit (a locked track) flashes red like any other.
-            file.onchange=async()=>{const f=file.files?.[0];if(!f)return;let text;const result={};try{if(f.size>4*1024*1024)throw new Error();text=await f.text();parseSRT(text,result);}catch{actions.notify?.(t('Could not read this subtitle file','读不出这个字幕文件'));return;}finally{file.value='';}
-                if(apply(t('Import subtitle file','导入字幕文件'),r=>importSRT(r,text))){const missing=projectMissing(state.record);if(missing.length)actions.notify?.(missingHint([...new Set(missing.flatMap(r=>r.chars))],t));if(result.truncated)actions.notify?.(t('Only the first 500 subtitles were imported.','只导入了前 500 条字幕。'));}};
+            // A file that cannot be used says why; a refused edit (a locked track) flashes red like any other.
+            file.onchange=async()=>{const f=file.files?.[0];if(!f)return;let text;const result={};
+                try{
+                    if(f.size>4*1024*1024){actions.notify?.(t('Subtitle files can be up to 4 MB.','字幕文件不能超过 4 MB。'));return;}
+                    try{text=decodeSRT(await f.arrayBuffer(),t('windows-1252','gb18030'));}catch{actions.notify?.(t('Could not read this subtitle file','读不出这个字幕文件'));return;}
+                    try{parseSRT(text,result);}catch{actions.notify?.(t('No subtitles found. Only SRT files can be imported.','未找到字幕。仅支持导入 SRT 字幕文件。'));return;}
+                }finally{file.value='';}
+                if(apply(t('Import subtitle file','导入字幕文件'),r=>importSRT(r,text))){const missing=projectMissing(state.record);if(missing.length)actions.notify?.(missingHint([...new Set(missing.flatMap(r=>r.chars))],t));
+                    if(result.truncated)actions.notify?.(t('Only the first 500 subtitles were imported.','只导入了前 500 条字幕。'));
+                    else if(result.skipped)actions.notify?.(t(`${result.skipped} subtitles could not be read and were skipped.`,`${result.skipped} 条字幕无法读取，已跳过。`));}};
             // Shown only where this computer can recognise speech.
             const hear=button(t('Auto subtitles','自动字幕'),{onClick:toggleSpeech});hear.classList.add('fv-ed-auto-subtitles');for(const b of speech.buttons)if(!b.isConnected)speech.buttons.delete(b);speech.buttons.add(hear);speechLabel(speech.busy?speech.shown:null);
             host.append(importButton,hear,file);
diff --git a/web/editor_text_model.js b/web/editor_text_model.js
index b7d32711..944f1072 100644
--- a/web/editor_text_model.js
+++ b/web/editor_text_model.js
@@ -101,6 +101,14 @@ export function activeText(record, frame) {
     return record.tracks.filter(t => isTextTrack(t) && !t.hidden).sort((a,b) => (a.kind === 'text' ? -1 : 1) - (b.kind === 'text' ? -1 : 1))
         .flatMap(t => t.clips.filter(c => c.start <= frame && c.start+c.frames > frame).map(c => ({...c, style: styleOf(record,c), opacity: c.text_kind === 'text' && styleOf(record,c).fade ? Math.min(1,(frame-c.start+.5)/7.2,(c.start+c.frames-frame-.5)/7.2) : 1})));
 }
+// A subtitle file's bytes as text: UTF-8, UTF-16 with its byte-order mark, or else the older local encoding
+// (`legacy`, e.g. GBK for Chinese), so such a file never turns into replacement characters.
+export function decodeSRT(bytes, legacy = 'windows-1252') {
+    const view = new Uint8Array(bytes);
+    if (view[0] === 0xff && view[1] === 0xfe) return new TextDecoder('utf-16le').decode(view);
+    if (view[0] === 0xfe && view[1] === 0xff) return new TextDecoder('utf-16be').decode(view);
+    try { return new TextDecoder('utf-8', {fatal: true}).decode(view); } catch { return new TextDecoder(legacy).decode(view); }
+}
 export function parseSRT(text, result = {}) {
     check(typeof text === 'string' && text.length <= 4*1024*1024, 'srt');
     const blocks = text.replace(/^\ufeff/, '').replace(/\r\n?/g,'\n').trim().split(/\n\s*\n/);
@@ -126,7 +134,7 @@ export function parseSRT(text, result = {}) {
     if (skipped) console.warn(`FreeVideo SRT import skipped ${skipped} unreadable blocks.`);
     if (trimmed) console.warn(`FreeVideo SRT import trimmed ${trimmed} overlapping subtitles.`);
     check(entries.length > 0, 'srt');
-    result.truncated = entries.length > 500;
+    result.skipped = skipped; result.truncated = entries.length > 500;
     if (result.truncated) console.warn('FreeVideo SRT import kept only the first 500 subtitles.');
     return entries.slice(0, 500);
 }
diff --git a/web/error_panel.js b/web/error_panel.js
index c268f973..fbdeee1b 100644
--- a/web/error_panel.js
+++ b/web/error_panel.js
@@ -12,6 +12,10 @@ const record = value => value && typeof value === 'object' && !Array.isArray(val
 
 // The engine's own messages for errors people can fix themselves.
 const PLAIN_ERRORS = [
+    {key: 'engine-busy', message: 'Setup, testing or generation is already using this installation/GPU lock. Retry after it finishes.',
+        title: ['FreeVideo is busy with another task', 'FreeVideo 正在处理另一个任务'],
+        summary: ['Try again when it finishes. If nothing else is running, close duplicate FreeVideo instances, then try again.',
+            '请在它完成后重试。若没有其他任务在进行，请关闭重复运行的 FreeVideo 后重试。']},
     {key: 'continuation-missing', message: 'The previous video has no usable continuation data. Generate it again with this version of FreeVideo.',
         title: ['This video cannot be continued', '这个视频暂时不能接续'],
         summary: ['Generate the previous video again, then continue it.', '请重新生成上一段视频，再接续。']},
@@ -39,6 +43,9 @@ const PLAIN_ERRORS = [
     {key: 'continuation-media', message: 'Continue from video cannot be used with these settings. Connect up to 3 pictures as references, and remove video, audio and pre-encoded inputs.',
         title: ['A continued part takes picture references only', '接续段只能用图片参考'],
         summary: ['Connect up to 3 pictures as references, and remove video, audio and pre-encoded inputs.', '最多接 3 张图片作为参考图，并移除视频、音频参考和预编码输入。']},
+    {key: 'media-mixed-layout', message: 'Choose keyframes or experimental references; mixed layouts are not supported',
+        title: ['First and last frames cannot be used with references', '首尾帧不能和参考素材同时使用'],
+        summary: ['Turn off the media cards or inputs you are not using, then generate again.', '请关闭本次不用的素材卡片或输入，再重新生成。']},
     {key: 'continuation-frames', message: 'Frame count must be 5 + 17n and at least 39 (39, 56, 73, 90, ...).',
         aliases: ['Frames must be 0 to use Duration, or 5 + 17n and at least 39 (for example, 39, 56, 141 or 243). The total includes 22 overlap frames when continuing.'],
         title: ['Choose a supported frame count', '请选择支持的帧数'],
@@ -58,6 +65,10 @@ const PLAIN_ERRORS = [
             'Unsupported LoRA tensor (no patches are ignored): ', 'Adapter contains no LoRA factors', 'Expected matching 2-D LoRA factors', 'Incomplete LoRA pair'],
         title: ['This LoRA cannot be used', '此 LoRA 无法使用'],
         summary: ['This LoRA is not made for the FreeVideo model. Turn it off or remove it, then generate again.', '此 LoRA 不适用于 FreeVideo 的模型。请关闭或移除后重新生成。']},
+    {key: 'lora-acceleration', message: 'This acceleration LoRA was made for the original model and cannot be used. The FreeVideo model is already accelerated and does not need one.',
+        title: ['This LoRA cannot be used', '此 LoRA 无法使用'],
+        summary: ['The FreeVideo model is already accelerated and does not need an acceleration LoRA. Turn this LoRA off or remove it, then generate again.',
+            'FreeVideo 的模型已经加速，无需加速 LoRA。请关闭或移除此 LoRA 后重新生成。']},
     {key: 'lora-int8', message: 'This LoRA changes parts of the model that the int8 version cannot merge yet, so it cannot be used for now. LoRAs that change only linear layers work.',
         title: ['This LoRA cannot be used', '此 LoRA 无法使用'],
         summary: ['The int8 version cannot merge the layers this LoRA changes yet. Turn it off or remove it, then generate again.',
@@ -96,6 +107,24 @@ function knownError(text) {
         || error.prefixes?.some(prefix => line.startsWith(prefix)))?.key ?? null;
 }
 
+// The media node's own check (comfy_assets.resolve_assets) names a file that is no longer in ComfyUI's
+// input folder. Only ComfyUI's custom-validation entry counts, so typed text quoting the sentence never does.
+// Returns the file's name, or null. A part of a long video carries the entry after its message (comfy_chain).
+function missingMedia(detail) {
+    if (reports.has(detail)) return null;   // A saved report holds no chosen input.
+    let nodes = record(record(detail?.response).node_errors ?? detail?.node_errors);
+    const part = /^Prompt outputs failed validation; (\{.*\})$/s.exec(typeof detail === 'string' ? detail : scalar(detail?.message));
+    if (!Object.keys(nodes).length && part) {
+        try { nodes = {part: JSON.parse(part[1])}; } catch { /* not the part's entry */ }
+    }
+    for (const node of Object.values(nodes)) for (const reason of list(node?.errors)) {
+        const found = reason?.type === 'custom_validation_failed'
+            && /^\w+ - Upload or select the missing media file: (.+)$/s.exec(scalar(reason.details));
+        if (found) return found[1].split(/[\\/]/).filter(Boolean).at(-1) || '';
+    }
+    return null;
+}
+
 // When a diagnostic report exists the card hides Export report; the same advice
 // then names Download report in the preview or node panel.
 const DOWNLOAD_ADVICE = ['If it still fails, use Download report in the preview and submit it in a GitHub issue.',
@@ -246,7 +275,8 @@ export function createErrorReport(detail, context = {}) {
     for (const key of ['done', 'total']) if (Number.isFinite(context.progress?.[key])) progress[key] = context.progress[key];
     const message = scalar(response.error?.message) || scalar(response.error) || scalar(response.message)
         || scalar(detail?.exception_message) || scalar(detail?.message) || scalar(detail) || 'FreeVideo generation failed.';
-    const known = knownError(message);
+    // The missing file's name is not kept: the report never holds typed or chosen input.
+    const known = knownError(message) ?? (missingMedia(detail) === null ? null : 'media-missing');
     const kind = failureAdvice(errorText(detail), (en) => en, known).kind;
     const found = detectBrowserExtension(detail, {kind, known, scrub: clean.inputsOnly});
     // A script name the report would redact is not shown or saved either.
@@ -322,6 +352,9 @@ export function failureAdvice(value, t, known = knownError(value)) {
         if (key === known)
             return {...row(key, title, summary, ['', '']), plain: true};
     }
+    if (known === 'media-missing')
+        return {...row('media-missing', ['A media file is missing', '找不到素材'],
+            ['Upload or select the missing media file again, then generate.', '请重新上传或选择缺少的素材后再生成。'], ['', '']), plain: true};
     if (/Prompt outputs failed validation|Required input is missing|Value not in list|Return type mismatch|Failed to validate prompt|value (?:\S+ )?(?:smaller|bigger) than/i.test(text))
         return row('validation', ['Check generation settings', '检查生成设置'],
             ['The generation settings did not pass validation; see the error details below.', '生成设置未通过检查，可查看下方错误详情。'],
@@ -451,9 +484,12 @@ export function adviceFor(report, t, value) {
         advice.title = t('The queue operation did not complete', '队列操作未完成');
         advice.summary = t('Check that ComfyUI is still running and connected, then retry.', '请确认 ComfyUI 仍在运行、连接正常后重试。');
     }
+    // The name comes from the error as it arrived; a card shown again from its report says it without one.
+    const media = advice.kind === 'media-missing' ? missingMedia(value) : null;
+    if (media) advice.summary = t(`"${media}" could not be found. Upload or select it again, then generate.`, `素材“${media}”已不存在，请重新上传或选择后再生成。`);
     // Every failure except a rejected input (which the user fixes in place) has one layout: written title,
     // Export report / Submit to GitHub, the raw error only inside the technical details.
-    const generated = advice.kind !== 'validation';
+    const generated = !['validation', 'media-missing'].includes(advice.kind);
     // A generation that stopped without a known cause is named by the stage it stopped in.
     if (generated && advice.kind === 'unknown' && report.stage === 'execution') {
         const stage = failureStage(report);
@@ -577,7 +613,8 @@ export function createErrorPanel(t) {
             `${node.type || t('Node', '节点')} #${node.id}${error.input ? ' · ' + error.input : ''}: ${error.message}${error.details ? ': ' + error.details : ''}`));
         const causeText = shownCause(reasons.length ? reasons.join('\n') : [report.message, report.error_details].filter(Boolean).join('\n'));
         const internal = advice.kind === 'unknown' && !['submission', 'queue'].includes(report.stage);
-        const causeInDetails = generated;
+        // A missing file is said in the explanation; where it happened and ComfyUI's wording stay in the technical details.
+        const causeInDetails = generated || advice.kind === 'media-missing';
         location.hidden = causeInDetails; detailLocation.hidden = !causeInDetails;
         detailLocation.textContent = causeInDetails ? location.textContent : '';
         if (causeInDetails) location.textContent = '';
@@ -604,7 +641,7 @@ export function createErrorPanel(t) {
         copy.hidden = generated; copyInside.hidden = !generated; copyInside.textContent = copy.textContent;
         // For a generation failure the note on removed text sits in the technical details, beside the copy button.
         privacy.hidden = generated; privacyInside.hidden = !generated;
-        const lora = advice.kind === 'lora-unusable' || advice.kind === 'lora-int8';
+        const lora = ['lora-unusable', 'lora-acceleration', 'lora-int8'].includes(advice.kind);
         // An int8 limit is not a judgement, so there is no mistake to report.
         github.hidden = !generated || advice.kind === 'lora-int8'; githubLora = lora;
         github.textContent = lora ? t('Report a mistake', '报告误判') : t('Report a problem', '提交问题');
diff --git a/web/failure_issue.js b/web/failure_issue.js
index 069d5a49..26a7a0dd 100644
--- a/web/failure_issue.js
+++ b/web/failure_issue.js
@@ -13,8 +13,10 @@ const PHASES = [
 export function failureStage(report) {
     if (report?.stage === 'submission') return ['submitting the task', '提交任务'];
     if (report?.stage === 'queue') return ['queue operation', '队列操作'];
+    // A resource-planning snapshot label says where the resources were last read, not where the run stopped.
     const named = [report?.progress?.phase, report?.progress?.stage,
-        /^Stage: (\S+)/m.exec(String(report?.details || ''))?.[1]].filter(Boolean);
+        /^Stage: (\S+)/m.exec(String(report?.details || ''))?.[1]]
+        .filter(name => name && !/^(?:before|after)-encoding$|^retry-\d+$/.test(name));
     for (const name of named)
         for (const [pattern, words] of PHASES) if (pattern.test(name)) return words;
     return null;
diff --git a/web/first_run_guide.js b/web/first_run_guide.js
index 1f8ce232..a18fb6aa 100644
--- a/web/first_run_guide.js
+++ b/web/first_run_guide.js
@@ -30,7 +30,7 @@ export const CHAPTERS = [
         en: 'For a longer video, enter a longer duration and it is made in parts automatically. You can also add parts one by one in "Edit".',
         zh: '填写更长的时长会自动分段生成，也可在“剪辑视图”中逐段接续。'},
     {id: 'agent', anchor: '[data-fv-guide="agent"]', added: true,
-        en: 'To have an AI agent such as Claude Code or Codex make videos for you, choose "Autopilot".', zh: '要让 Claude Code、Codex 等 Agent 替你生\u2060成\u2060视\u2060频，点“AI 全自动”。'},
+        en: 'To have an AI agent such as Claude Code or Codex make videos for you, choose "Autopilot".', zh: '要让 Claude Code、Codex 等 Agent 生\u2060成\u2060视\u2060频，点“AI 全自动”。'},
     {id: 'editor-timeline', group: 'editor', anchor: '[data-fv-guide="editor-timeline"]', added: true,
         en: 'Drag a video here.', zh: '把视频拖到这里。'},
     {id: 'editor-trim', group: 'editor', anchor: '[data-fv-guide="editor-trim"]', added: true,
diff --git a/web/freevideo.js b/web/freevideo.js
index d8ca0461..081d66a8 100644
--- a/web/freevideo.js
+++ b/web/freevideo.js
@@ -15,7 +15,7 @@ import { startDoneNotify, askDesktopPermission, DESKTOP_SETTING } from './done_n
 import { installNavigation, refreshNavigation, preferredView, collapseWorkspaces } from './view_navigation.js';
 import { isLaunchVisit } from './launch_visit.js';
 import { openFreeVideoWorkflow } from './launcher.js';
-import { agentInterfaceEnabled } from './agent_connect.js';
+import { agentInterfaceState } from './agent_connect.js';
 import { attachReferencePicker, referenceItems, syncReferencePrompt, referenceInputs, mediaRows } from './prompt_references.js';
 import { expandCast } from './cast_mention.js';
 import { outputDownloadURL } from './output_download.js';
@@ -229,7 +229,8 @@ function mediaPanel(node, mount = null) {
         finally { node.isUploading = false; add.disabled = false; addAudio.disabled = false; }
     }
     async function newImage() {
-        if (node.isUploading || read().length >= 32) return;
+        if (node.isUploading) return;
+        if (read().length >= 32) { message(text('The media list is full (32 items, including disabled ones). Remove some to add more.', '素材已满 32 个（含未启用的），请先移除不用的素材。'), true); return; }
         const library = read().filter(row => mediaKind(row.file) === 'image')
             .map(row => ({file: row.file, label: (row.source || row.file).split('/').pop()}));
         let outcome;
@@ -237,7 +238,8 @@ function mediaPanel(node, mount = null) {
             outcome = await openImageEditor({target: videoSize(), role: role.value, t: text,
                 resolve: viewURL, upload: uploadOne, library});
         } catch { message(text('This image cannot be edited right now; choose an image again.', '图片暂时无法编辑，可重新选择图片。'), true); return; }
-        if (!outcome?.blob || read().length >= 32) return;
+        if (!outcome?.blob) return;
+        if (read().length >= 32) { message(text('The media list is full (32 items, including disabled ones). Remove some to add more.', '素材已满 32 个（含未启用的），请先移除不用的素材。'), true); return; }
         node.isUploading = true; add.disabled = addAudio.disabled = generateImage.disabled = true;
         try {
             const result = await uploadOne(new File([outcome.blob], outcome.name, {type: outcome.blob.type}));
@@ -700,7 +702,7 @@ app.registerExtension({
     beforeConfigureGraph() { configuringGraph = true; configuration++; },
     async setup() {
         // One failing part must not take the others down; health.js shows what failed.
-        for (const [where, start] of [['progress', () => progressConnection.start()], ['navigation', () => installNavigation(showStudio, agentInterfaceEnabled(api), {openWorkflow: openFreeVideoWorkflow})],
+        for (const [where, start] of [['progress', () => progressConnection.start()], ['navigation', () => installNavigation(showStudio, agentInterfaceState(api), {openWorkflow: openFreeVideoWorkflow, recheck: () => agentInterfaceState(api)})],
             ['updates', startUpdateChecks], ['done-notify', () => startDoneNotify({api, text, desktop: desktopNoticeEnabled, isFreeVideoNode})],
             ['whitebox', () => startWhiteboxWorker({api})]]) {
             try { start(); } catch (error) { health({type: 'error', where, error}); }
diff --git a/web/image_editor.js b/web/image_editor.js
index 4cd1efc8..af8bc958 100644
--- a/web/image_editor.js
+++ b/web/image_editor.js
@@ -231,7 +231,7 @@ export async function openImageEditor({source = null, name = source || '', targe
     const genClient = imageGenClient || createImageGenClient(api.fetchApi.bind(api));
     const generation = createImageGeneration({client: genClient,
         renderer, t, getDoc: () => doc, getSelection: () => selection, commit: commitInserted, el, button, segmented,
-        library, resolve, icon, onRatioChange: setBlankCanvasRatio, onPlace: (file, label) => addImageLayer(file, label), productVersion, fetchApi: (path, options) => api.fetchApi(path, options),
+        library, resolve, icon, onRatioChange: setBlankCanvasRatio, onPlace: (file, label) => addImageLayer(file, label, {keepTool: true}), productVersion, fetchApi: (path, options) => api.fetchApi(path, options),
         resumeKey: `${role}|${source || ''}|${name}`});
 
     // ----- tools ----------------------------------------------------------------
@@ -538,7 +538,7 @@ export async function openImageEditor({source = null, name = source || '', targe
         let text = '';
         if (['eraser', 'heal', 'clone', 'blur'].includes(tool) && !rasterKinds.has(layer.kind)) text = t('Pick an image or paint layer to work on its pixels.', '请选择图片图层或绘画图层再处理像素。');
         else if (tool === 'clone' && !opts.clone.source) text = t('Alt-click to set the source, then paint.', '先按住 Alt 点击设置取样点，再涂抹。');
-        else if (tool === 'move' && layer.locked && doc.layers.length > 1) text = t('The original image is locked. Click another layer to move it.', '原图图层已锁定；点击其他图层即可移动。');
+        else if (tool === 'move' && layer.locked && doc.layers.length > 1) text = t('The original image is locked. Click another layer to move it.', '原图图层已锁定；点击其他图层可移动。');
         else if (tool === 'lasso' && opts.lasso.type === 'polygon') text = t('Click to add corners; double-click or press Enter to close.', '单击添加顶点；双击或按 Enter 闭合。');
         hint.textContent = text; hint.hidden = !text;
     }
@@ -916,12 +916,12 @@ export async function openImageEditor({source = null, name = source || '', targe
         const n = doc.layers.filter(l => l.kind === 'paint').length + 1;
         commitInserted(commands.addPaintLayer(doc, t(`Paint ${n}`, `绘画 ${n}`)));
     }
-    async function addImageLayer(file, label) {
+    async function addImageLayer(file, label, {keepTool = false} = {}) {
         try {
             const img = await renderer.image(file); renderer.images.get(file).value = img;
             commitInserted(commands.addImageLayer(doc, {src: file, nw: img.naturalWidth, nh: img.naturalHeight,
                 name: label || file.split('/').pop()}));
-            select('move');
+            if (!keepTool) select('move');
         } catch { message(t('The image could not be added. Choose another one.', '图片没能添加，请换一张。'), ['Image editor', '图片编辑器']); }
     }
     filePicker.onchange = async () => {
diff --git a/web/image_generation.js b/web/image_generation.js
index bc720902..c2dd08bb 100644
--- a/web/image_generation.js
+++ b/web/image_generation.js
@@ -164,6 +164,7 @@ export function imageGenError(code, info, t) {
         device_unavailable: ['The graphics card is not available right now. Try again later.', '显卡暂时用不了，稍后再试。'],
         worker_exit: ['Image generation stopped unexpectedly. Try again.', '图像生成意外停止了，请再试一次。'],
         timeout: ['This image took too long and was stopped. Try again.', '这一张用时太长，已停止。请再试一次。'],
+        disk_full: ['Saving the image failed: the disk is full. Free some space, then generate again.', '保存图片时出错：磁盘空间不够，空出空间后再生成。'],
         cancelled: ['Image generation was cancelled.', '已取消图像生成。'],
         invalid_request: ['The image settings are invalid. Adjust them and try again.', '图片设置有误，请调整后重试。'],
         invalid_media: ['The reference image cannot be read. Choose another image.', '参考图读不出来，请换一张图。'],
@@ -462,11 +463,13 @@ export function createImageGenSession({client, prepare, apply, onChange = () =>
     async function start(request) {
         if (state.running || disposed || !request.prompt.trim() || !state.info?.installed) return;
         const current = ++epoch;
-        emit({running: true, job: null, fraction: 0, stage: '', error: '', report_id: null, waiting: '', fitted: null});
+        // A picture that is edited sends its reference images first; there is no job to report on yet.
+        emit({running: true, job: null, fraction: 0, stage: request.mode === 'edit' ? 'upload' : '', error: '', report_id: null, waiting: '', fitted: null});
         const alive = () => current === epoch && !disposed;
         try {
             const {context, images = [], mask} = await prepare(request, alive);
             if (!alive()) return;
+            if (state.stage === 'upload') emit({stage: ''});
             const withImages = images.length > 0;
             const row = await client.generate({prompt: request.prompt, width: context.width, height: context.height, images, mask});
             const {job} = row;
@@ -735,7 +738,9 @@ export function createImageGeneration({client, renderer, t, getDoc, getSelection
                 : state.running && state.waiting ? t(...(QUEUE_LABELS[state.waiting] || QUEUE_LABELS.image))
                 : state.running && state.fitted ? t(...memoryWords([`Not enough GPU memory. Retrying at ${state.fitted.width}\u00a0×\u00a0${state.fitted.height}`,
                     `显存不足，改用 ${state.fitted.width}\u00a0×\u00a0${state.fitted.height} 重试`], state.info?.unified_memory)) + ' · ' + percent
-                : state.running ? (state.stage === 'load' ? t('Loading the model', '加载模型') + ' ' : '') + percent
+                : state.running && state.stage === 'upload' ? t('Uploading reference images', '正在上传参考图')
+                : state.running ? (state.stage === 'load' ? t('Loading the model', '加载模型') + ' '
+                    : state.stage === 'encode' ? t('Reading reference images', '读取参考图') + ' ' : '') + percent
                 : state.fitted ? fittedNote(state.fitted, t, state.info?.unified_memory) : '';
         }
         note.hidden = !note.textContent;
diff --git a/web/launcher.js b/web/launcher.js
index 5fa58417..ed1eafac 100644
--- a/web/launcher.js
+++ b/web/launcher.js
@@ -3,6 +3,7 @@ import { api } from '../../scripts/api.js';
 import { restorePromptDraft } from './prompt_draft.js';
 import { takePortMarker } from './port_notice.js';
 import { launcherRestoreFinished } from './workspace_restore.js';
+import { clonePrompt } from './optional_inputs.js';
 
 // health.js names what failed; files loaded before it queue their reports.
 const health = item => (window.__freevideoHealth ||= []).push(item);
@@ -93,7 +94,7 @@ async function existingStudioWorkflow(target, bundled) {
     // Only the shared tab is ours to replace; a workflow the user saved under another name stays.
     const current = target.path === 'workflows/FreeVideo.json' ? currentStudioWorkflow(saved, bundled) : null;
     if (current) return restorePromptDraft(api, current, target.path);
-    const graph = structuredClone(saved);
+    const graph = clonePrompt(saved);
     if (!graph.nodes?.some(node => node.type === 'FreeVideoGenerate')) {
         const node = structuredClone(bundled.nodes.find(node => node.type === 'FreeVideoGenerate'));
         node.id = Math.max(graph.last_node_id || 0, ...graph.nodes.map(node => Number(node.id) || 0)) + 1;
@@ -154,7 +155,7 @@ async function keepInFront(workflow, opened) {
         if (studioWorkflow(app.graph)) {
             const current = app.graph.serialize(), target = store?.activeWorkflow;
             const replaced = target?.path === 'workflows/FreeVideo.json' ? currentStudioWorkflow(current, workflow) : null;
-            const restored = await restorePromptDraft(api, replaced || structuredClone(current), target?.path || '__existing__');
+            const restored = await restorePromptDraft(api, replaced || clonePrompt(current), target?.path || '__existing__');
             if (JSON.stringify(restored) !== JSON.stringify(current))
                 await app.loadGraphData(restored, true, true, target || 'FreeVideo');
             return;
diff --git a/web/library.css b/web/library.css
index c3474247..8523b7f5 100644
--- a/web/library.css
+++ b/web/library.css
@@ -71,5 +71,6 @@
 .fv-library .fv-library-undo:not([hidden])~.fv-library-selection:not([hidden]){margin-top:0}
 .fv-library .fv-library-undo button{min-height:28px;padding:2px 10px;font-size:13px;color:var(--fv-accent)}
 .fv-library .fv-library-skipped{flex-basis:100%;font-size:12px;color:var(--fv-muted)}
+.fv-library .fv-library-unreadable{margin:8px;text-align:center}.fv-library .fv-library-unreadable:empty{display:none}
 @media(max-width:680px){.fv-library-tools{gap:0}.fv-library .fv-library-header{grid-template-columns:auto minmax(0,1fr) auto}.fv-library-header h2{text-align:center}.fv-library-header button{white-space:nowrap}.fv-studio.fv-library .fv-library-header button{padding-left:6px;padding-right:6px}}
 .fv-library-message .fv-issue-button{margin:4px 0 4px 8px;vertical-align:middle}
diff --git a/web/library.js b/web/library.js
index 957bc7d0..495455b4 100644
--- a/web/library.js
+++ b/web/library.js
@@ -93,7 +93,7 @@ export async function libraryPage(t, {limit = 24, before = null, signal} = {}) {
     const query = new URLSearchParams({limit: String(limit)});
     if (before) query.set('before', before);
     const response = await api.fetchApi('/freevideo/library?' + query, {signal});
-    if (response.status === 404) throw Object.assign(new Error(t('Restart ComfyUI to open your creations after updating.', '更新后重启 ComfyUI，即可打开作品。')), {restart: true});
+    if (response.status === 404) throw Object.assign(new Error(t('Restart ComfyUI to open your creations after updating.', '更新后重启 ComfyUI，可打开作品。')), {restart: true});
     if (!response.ok) throw new Error(t('Could not load your creations. Try Refresh.', '作品暂时无法读取，请点击刷新。'));
     return response.json();
 }
@@ -144,15 +144,18 @@ async function workflowBackfill(collection, t) {
         const result = value?.result || {};
         heading.textContent = result.embedded ? t(`Workflows added to ${result.embedded} videos`, `已为 ${result.embedded} 个视频补上工作流`) : t('No video was changed', '没有视频被改动');
         text.textContent = result.embedded
-            ? t('Drag them onto the canvas to restore their workflow. The originals are kept next to them as video.before-workflow.mp4.', '把它们拖到画布上即可还原工作流。原视频保留在各自目录，文件名为 video.before-workflow.mp4。')
+            ? t('Drag them onto the canvas to restore their workflow. The originals are kept next to them as video.before-workflow.mp4.', '把它们拖到画布上可还原工作流。原视频保留在各自目录，文件名为 video.before-workflow.mp4。')
             : t('Videos that could not be checked safely were left as they were.', '无法安全确认的视频都保持原样。');
         if (result.failed) text.textContent += ' ' + t(`${result.failed} videos were left unchanged and will be offered again.`, `另有 ${result.failed} 个视频未改动，之后会再次提示。`);
-        actions.replaceChildren(button(t('Done', '好的'), () => box.remove(), 'fv-quiet'));
+        const full = Math.min(Number.isSafeInteger(result.no_space) ? result.no_space : 0, result.failed || 0);
+        if (full > 0) text.textContent += ' ' + t(full === result.failed ? 'The disk is full. Free some space first.' : `${full} of them could not be changed because the disk is full. Free some space first.`,
+            full === result.failed ? '磁盘空间不够，请先空出空间。' : `其中 ${full} 个因磁盘空间不够，请先空出空间。`);
+        actions.replaceChildren(button(t('Got it', '知道了'), () => box.remove(), 'fv-quiet'));
     }
     if (value.state === 'running') { follow(); return; }
     heading.textContent = t('Add workflows to earlier videos', '为之前的视频补上工作流');
     text.textContent = t(`${value.count} earlier videos can get the workflow they were made with, so dropping them on the canvas restores it. Nothing changes until you confirm, and each original is kept as a backup (about ${gb(value.bytes)} GB).`,
-        `有 ${value.count} 个之前生成的视频可以补上生成时的工作流，补写后拖到画布即可还原。确认前不会改动任何文件，每个原视频都会保留一份备份（约 ${gb(value.bytes)} GB）。`);
+        `有 ${value.count} 个之前生成的视频可以补上生成时的工作流，补写后拖到画布可还原。确认前不会改动任何文件，每个原视频都会保留一份备份（约 ${gb(value.bytes)} GB）。`);
     const run = button(t('Add workflows', '补写工作流'), async () => {
         run.disabled = later.disabled = true;
         try { if (!(await post('run')).ok) throw new Error(); follow(); }
@@ -193,6 +196,8 @@ export function openLibrary(t, {upscale = null, selectedWork = null, extend = nu
     const grid = el('div', null, 'fv-library-grid');
     const message = el('p', '', 'fv-library-message'); message.setAttribute('role', 'status');
     const more = button(t('Show more', '加载更多'), () => load(true), 'fv-quiet'); more.hidden = true;
+    // Creations whose saved record could not be read stay on disk; say how many the loaded pages skipped.
+    const unreadable = el('p', '', 'fv-library-skipped fv-library-unreadable'); let unread = 0;
     // Selecting several creations: a strip above the cards, the delete bar below them.
     const picking = el('div', null, 'fv-library-picking'); picking.hidden = true;
     const pickAll = button(t('Select all', '全选'), () => selectAll(), 'fv-quiet');
@@ -207,7 +212,7 @@ export function openLibrary(t, {upscale = null, selectedWork = null, extend = nu
     bar.append(barLabel, barText, barActions);
     // The undo note and the delete bar share one sticky dock so they stack instead of covering each other.
     const dock = el('div', null, 'fv-library-dock'); dock.append(undoBar, bar);
-    collection.append(picking, grid, message, more, dock);
+    collection.append(picking, grid, message, more, unreadable, dock);
     workflowBackfill(collection, t);
     const detail = el('section', null, 'fv-library-detail');
     const all = button(t('All creations', '全部作品'), () => {
@@ -661,7 +666,9 @@ export function openLibrary(t, {upscale = null, selectedWork = null, extend = nu
         message.textContent = rows.size ? '' : t('Loading your creations…', '正在读取作品…');
         try {
             const data = await libraryPage(t, {before: appendPage ? next : null, signal: abort.signal}); if (disposed) return;
-            if (!appendPage) { thumbnails.disconnect(); grid.replaceChildren(); cards.clear(); rows.clear(); }
+            if (!appendPage) { thumbnails.disconnect(); grid.replaceChildren(); cards.clear(); rows.clear(); unread = 0; }
+            unread += Number.isSafeInteger(data.unreadable) && data.unreadable > 0 ? data.unreadable : 0;
+            unreadable.textContent = unread ? t(`${unread} ${unread === 1 ? 'creation' : 'creations'} could not be read`, `另有 ${unread}\u00a0个作品无法读取`) : '';
             for (const row of data.items || []) append(row);
             for (const row of arrivals) append(row, true);
             next = data.next; more.hidden = !next;
diff --git a/web/lora_check.js b/web/lora_check.js
index 39f0c871..1889d7a4 100644
--- a/web/lora_check.js
+++ b/web/lora_check.js
@@ -13,6 +13,8 @@ export function describeVerdict(verdict, t) {
     else if (verdict.reason === 'int8-unmergeable') reason = t('This LoRA cannot be used', '此 LoRA 无法使用');
     else if (verdict.reason === 'format') reason = t('not a LoRA file FreeVideo can read', '不是 FreeVideo 能读取的 LoRA 文件');
     else if (verdict.reason === 'shape') reason = t('does not match the current model', '和当前模型的结构不匹配');
+    else if (verdict.reason === 'acceleration') reason = t('an acceleration LoRA made for the original model. The FreeVideo model is already accelerated and does not need one',
+        '这是为原版模型制作的加速 LoRA。FreeVideo 的模型已经加速，无需此类 LoRA');
     else reason = family ? t(`made for ${family}`, `这是给 ${family} 训练的 LoRA`) : t('not made for the FreeVideo model', '不适用于 FreeVideo 的模型');
     const summary = verdict.summary || {};
     const lines = [];
diff --git a/web/optional_inputs.js b/web/optional_inputs.js
index 7f3404a1..0cbd4f1a 100644
--- a/web/optional_inputs.js
+++ b/web/optional_inputs.js
@@ -52,3 +52,14 @@ export function cleanOptionalPrompt(prompt) {
     }
     return prompt;
 }
+
+// A graph taken from the host can carry a value structuredClone refuses, such as a reactive proxy another
+// extension left in the workflow's extra data. The request is JSON on the wire, so when the native copy
+// fails the JSON copy stands in for it; a graph that clones natively is copied exactly as before.
+export function clonePrompt(value) {
+    try { return structuredClone(value); }
+    catch (error) {
+        if (error?.name !== 'DataCloneError') throw error;
+        return JSON.parse(JSON.stringify(value));
+    }
+}
diff --git a/web/prompt_draft.js b/web/prompt_draft.js
index f28e5839..7490ef7b 100644
--- a/web/prompt_draft.js
+++ b/web/prompt_draft.js
@@ -16,6 +16,15 @@ const equal = (a, b) => JSON.stringify(a) === JSON.stringify(b);
 const watched = new WeakMap(), dirty = new Set();
 const historyObservers = new WeakSet(), restoringHistory = new WeakSet();
 let timer = null, pending = null, inFlight = null, sending = null;
+// What the page could not do with the draft file, so the prompt box can say so and retry.
+const trouble = {save: false, restore: false}, troubleWatchers = new Set();
+function setTrouble(key, value) {
+    if (trouble[key] === value) return;
+    trouble[key] = value;
+    for (const watcher of troubleWatchers) watcher({...trouble});
+}
+export function promptDraftTrouble() { return {...trouble}; }
+export function watchPromptDraftTrouble(watcher) { troubleWatchers.add(watcher); return () => troubleWatchers.delete(watcher); }
 try { pending = JSON.parse(globalThis.sessionStorage?.getItem(OUTBOX) || 'null'); } catch {}
 
 function retain() {
@@ -177,14 +186,16 @@ export async function savePromptDraft(api, keepalive = false) {
         try {
             const reply = await api.fetchApi(FILE, {cache: 'no-store', keepalive});
             if (!reply.ok && reply.status !== 404) throw new Error('draft_read');
-            const current = reply.ok ? await reply.json() : null;
+            // A damaged file can never be merged into; the next save replaces it.
+            const current = reply.ok ? await reply.json().catch(() => null) : null;
             const saved = await api.fetchApi(FILE + '?overwrite=true', {method: 'POST',
                 headers: {'Content-Type': 'application/json'}, body: JSON.stringify(combine(current, patch)), keepalive});
             if (!saved.ok) throw new Error('draft_write');
-            sending = null; retain(); return true;
+            sending = null; retain(); setTrouble('save', false); return true;
         } catch {
             const newer = pending; pending = patch; sending = null;
             if (newer) addPatch(newer); else retain();
+            setTrouble('save', true);
             return false;
         }
     };
@@ -195,11 +206,13 @@ export async function savePromptDraft(api, keepalive = false) {
 export async function restorePromptDraft(api, workflow, project = null) {
     // A refreshed page retries its own unsaved patch before launcher restoration.
     if (pending || inFlight || dirty.size) await savePromptDraft(api);
-    let draft;
+    let draft, unread = false;
     try {
         const reply = await api.fetchApi(FILE, {cache: 'no-store'});
-        if (reply.ok) draft = await reply.json();
-    } catch { /* Keep the workflow and any local edits if the server is offline. */ }
+        // Only a missing file means there is no earlier prompt.
+        if (reply.ok) draft = await reply.json(); else unread = reply.status !== 404;
+    } catch { unread = true; /* Keep the workflow and any local edits if the server is offline. */ }
+    setTrouble('restore', unread);
     // A failed retry remains authoritative over the older remote copy. Retain
     // the outbox until a later explicit save succeeds, including across reloads.
     const local = pending || sending ? mergePatches(sending, pending) : null;
diff --git a/web/prompt_enhance.js b/web/prompt_enhance.js
index 3193315c..a12259a4 100644
--- a/web/prompt_enhance.js
+++ b/web/prompt_enhance.js
@@ -125,11 +125,14 @@ export function createPromptEnhancer({node, input, editor, api, context, setText
         busy: t('Finish the current queue before enhancing this prompt.', '请等待当前队列完成后再改写。'),
         prompt_required: t('Type a prompt first; enhancement rewrites it before generating.', '请先输入提示词，生成前会自动改写。'),
         setup_required: t('Complete FreeVideo installation first.', '请先完成 FreeVideo 安装。'),
-        model_missing: t('Download the optional model to enable enhancement.', '下载可选模型后即可启用。'),
-        download_failed: t('Download paused. Retry to resume.', '下载已暂停，重试即可继续。'),
+        model_missing: t('Download the optional model to enable enhancement.', '下载可选模型后可启用。'),
+        download_failed: t('Download paused. Retry to resume.', '下载已暂停，重试可继续。'),
         disk_space: t('Not enough disk space for the optional model.', '可选模型的磁盘空间不足，请清理后重试。'),
         ram_space: t(`This model needs about ${gib(error?.ramBytes || ramBytes)} GiB of available RAM to load. You can keep using the original.`, `此模型加载需约 ${gib(error?.ramBytes || ramBytes)} GiB 可用内存，可先使用原文生成。`),
-        unsupported_media: t('Use images added in Media for enhancement; connected nodes, video and audio are not supported yet.', '增强暂支持在素材中添加的图片；连接节点、视频和音频暂不支持。'),
+        unsupported_media: t('Enhancement can read only images and videos added in Media. You can keep using the original.', '增强只能读取素材中添加的图片和视频，可先使用原文生成。'),
+        connected_media: t('Enhancement cannot read connected nodes. Add images or videos in Media, or use the original.', '增强读不到连接节点的素材，请在素材中添加图片或视频，或使用原文。'),
+        audio_media: t('Enhancement cannot read audio. Add images or videos in Media, or use the original.', '增强读不到音频，请在素材中添加图片或视频，或使用原文。'),
+        unreadable_video: t('Could not read a frame from this video. Replace it with another video or an image, or use the original.', '读不出这个视频的画面，请换一个视频或图片，或使用原文。'),
         too_many_images: t('Use up to 8 reference images for enhancement.', '增强暂支持最多 8 张参考图。'),
         invalid_prompt: t('Enter a prompt of up to 12,000 characters.', '请输入不超过 12,000 字符的提示词。'),
         invalid_output: t('The rewrite was incomplete. Retry or use the original.', '改写结果不完整，请重试或使用原文。'),
diff --git a/web/prompt_references.css b/web/prompt_references.css
index 377bba73..faddfc1c 100644
--- a/web/prompt_references.css
+++ b/web/prompt_references.css
@@ -1,6 +1,8 @@
 .fv-prompt-editor{position:relative;display:flex;flex-direction:column;height:170px;min-height:148px;max-height:340px;resize:vertical;overflow:hidden;background:var(--fv-card)}
 .fv-studio .fv-prompt-editor textarea{flex:1;min-height:0;height:0;max-height:none;resize:none;padding-bottom:8px;outline:none}
 .fv-prompt-tools{display:flex;flex:none;align-items:center;justify-content:flex-end;height:36px;padding-right:9px}
+.fv-studio .fv-prompt-draft-note{min-height:27px;height:27px;margin:0 auto 0 9px;padding:0 9px;border:0;background:transparent;color:var(--fv-warning);font-size:12px;overflow:hidden;text-overflow:ellipsis;white-space:nowrap}
+.fv-studio .fv-prompt-draft-note[hidden]{display:none}
 .fv-studio .fv-reference-trigger{min-height:27px;height:27px;width:30px;padding:0;font-size:17px;border:0;background:var(--fv-raised);color:var(--fv-muted)}
 @media(min-height:960px) and (min-width:681px){.fv-prompt-editor{height:clamp(170px,24dvh,340px)}}
 .fv-reference-picker{box-sizing:border-box;position:fixed;inset:auto;margin:0;padding:6px;border:1px solid var(--fv-border,#3a414c);border-radius:14px;background:var(--fv-card,#1a202a);color:var(--fv-ink,#eee);box-shadow:0 14px 42px #0006;font:13px/1.4 var(--fv-font,system-ui);z-index:10000;max-height:280px;overflow:hidden;animation:fv-reference-appear .12s ease-out}
diff --git a/web/prompt_references.js b/web/prompt_references.js
index 48cd4cd3..483e381e 100644
--- a/web/prompt_references.js
+++ b/web/prompt_references.js
@@ -1,5 +1,6 @@
 // H3 numbers each reference kind separately, in the order of comfy_media.export.
 import {caretRect} from './caret_rect.js';
+import {listenDialogKey} from './dialog_keys.js';
 const titles = {image: 'Picture', video: 'Video', audio: 'Audio'};
 const widget = (node, name) => node?.widgets?.find(w => w.name === name);
 const linkFor = (node, name) => node?.graph?.links?.[node.inputs?.find(i => i.name === name)?.link];
@@ -73,6 +74,17 @@ export function referenceItems(node) {
         token: `<${titles[item.kind]} ${counters[item.kind]}>`}));
 }
 
+// What the prompt model can read: pictures and, as one still frame, videos added in Media. Audio and connected
+// nodes are left out, and keep their labels in the scene; the error names why only when nothing is left.
+export function enhancementMedia(items, conditioningLinked = false) {
+    const media = items.filter(row => row.file && ['image', 'video'].includes(row.kind))
+        .map(row => ({file: row.file, role: row.role || 'reference', kind: row.kind, number: row.number}));
+    if (media.length) return media;
+    const connected = conditioningLinked || items.some(row => !row.file);
+    if (connected || items.length) throw new Error(connected ? 'connected_media' : 'audio_media');
+    return media;
+}
+
 export function remapReferences(text, previous, current, removed = '[Removed reference]') {
     const byKey = new Map(current.map(item => [item.key, item.token]));
     const replacement = new Map(previous.map(item => [item.token, byKey.get(item.key) || removed]));
@@ -135,7 +147,10 @@ export function attachReferencePicker(input, {items, t = en => en, view, host, a
     input.setAttribute('aria-controls', list.id); input.setAttribute('aria-expanded', 'false');
     let matches = [], active = 0, range = null, composing = false, disposed = false;
     const listeners = [];
-    const listen = (target, name, fn, options) => { target.addEventListener(name, fn, options); listeners.push(() => target.removeEventListener(name, fn, options)); };
+    const listen = (target, name, fn, options) => {
+        if (name.startsWith('key')) { listeners.push(listenDialogKey(target, name, fn, options)); return; }
+        target.addEventListener(name, fn, options); listeners.push(() => target.removeEventListener(name, fn, options));
+    };
     function close() {
         if (popup.hidePopover && popup.matches(':popover-open')) popup.hidePopover();
         popup.hidden = true; input.setAttribute('aria-expanded', 'false'); input.removeAttribute('aria-activedescendant');
@@ -203,7 +218,7 @@ export function attachReferencePicker(input, {items, t = en => en, view, host, a
         if (!matches.length) {
             const empty = document.createElement('div'); empty.className = 'fv-reference-empty';
             empty.textContent = invalid ? t('Check your media selection.', '请检查素材用途是否冲突。')
-                : available.length || everyone.length ? t('No matching characters or media', '没有匹配的角色或素材') : t('Add reference media to use @', '添加参考素材后即可使用 @');
+                : available.length || everyone.length ? t('No matching characters or media', '没有匹配的角色或素材') : t('Add reference media to use @', '添加参考素材后可使用 @');
             list.append(empty);
             if (!available.length && !everyone.length && !invalid && addMedia) {
                 const add = document.createElement('button'); add.type = 'button'; add.textContent = t('Add media', '添加素材');
diff --git a/web/prompt_terms.js b/web/prompt_terms.js
index d3f95895..8ce5a0aa 100644
--- a/web/prompt_terms.js
+++ b/web/prompt_terms.js
@@ -1,6 +1,7 @@
 // Camera and motion words offered while a prompt is typed. The list is local; nothing is sent anywhere,
 // a word is written only when it is chosen, and the typed language decides which form is inserted.
 import {caretRect} from './caret_rect.js';
+import {listenDialogKey} from './dialog_keys.js';
 export const TERMS = [
     ['dolly in', '推镜头'], ['dolly out', '拉镜头'], ['pan left', '左摇镜头'], ['pan right', '右摇镜头'],
     ['tilt up', '上摇镜头'], ['tilt down', '下摇镜头'], ['tracking shot', '跟拍镜头'], ['orbit shot', '环绕镜头'],
@@ -56,7 +57,10 @@ export function attachTermSuggest(input, {t = en => en, host, suppressed = () =>
     let found = null, active = -1, composing = false, disposed = false, dismissed = null;
     let saved = null;
     const listeners = [];
-    const listen = (target, name, fn, options) => { target.addEventListener(name, fn, options); listeners.push(() => target.removeEventListener(name, fn, options)); };
+    const listen = (target, name, fn, options) => {
+        if (name.startsWith('key')) { listeners.push(listenDialogKey(target, name, fn, options)); return; }
+        target.addEventListener(name, fn, options); listeners.push(() => target.removeEventListener(name, fn, options));
+    };
     function close() {
         if (popup.hidePopover && popup.matches(':popover-open')) popup.hidePopover();
         popup.hidden = true; found = null;
diff --git a/web/report_issue.js b/web/report_issue.js
index a57fb8de..e93b3143 100644
--- a/web/report_issue.js
+++ b/web/report_issue.js
@@ -177,12 +177,12 @@ export function showReportTip(bar, anchor, t, {mode = 'invite', submit, onHide =
         title.textContent = t('Last step', '最后一步');
         lead.textContent = savedName?.endsWith('.zip')
             ? t(`The report is large, so it was saved as ${savedName}. Please drag it into "Diagnostic report", then select Submit new issue.`,
-                `报告较大，已保存为 ${savedName}。请将它拖入「诊断报告」，再点击 Submit new issue 即可。`)
+                `报告较大，已保存为 ${savedName}。请将它拖入「诊断报告」，再点击 Submit new issue。`)
             : savedName
             ? t(`Please drag ${savedName} into "Diagnostic report", then select Submit new issue.`,
-                `请将 ${savedName} 拖入「诊断报告」，再点击 Submit new issue 即可。`)
+                `请将 ${savedName} 拖入「诊断报告」，再点击 Submit new issue。`)
             : t('Please drag the downloaded report into "Diagnostic report", then select Submit new issue.',
-                '请将下载的诊断报告拖入「诊断报告」，再点击 Submit new issue 即可。');
+                '请将下载的诊断报告拖入「诊断报告」，再点击 Submit new issue。');
         const link = el('a', null, t("Page didn't open? Go to GitHub", '页面没有打开？点此前往 GitHub'));
         link.href = address; link.target = '_blank'; link.rel = 'noopener';
         note.replaceChildren(link);
diff --git a/web/request_assets.js b/web/request_assets.js
index 8aa02df6..8bae5c03 100644
--- a/web/request_assets.js
+++ b/web/request_assets.js
@@ -56,6 +56,7 @@ export function createRequestAssets({t, api, request, onError = () => {}}) {
         const wait = timeAfter(Object.values(bytes).reduce((sum, n) => sum + n, 0), rate, t);
         if (parts.length) rows.push(t(`Downloads before generating: ${parts.join(', ')}${wait}`, `生成前会先下载：${parts.join('、')}${wait}`));
         for (const item of items) {
+            if (item.action === 'unavailable') continue;
             if (item.kind === 'prompt-vlm') {
                 const wait = timeAfter(item.bytes, rate, t);
                 rows.push(t(`First prompt rewrite downloads ${size(item.bytes)}${wait}`, `首次改写提示词会先下载 ${size(item.bytes)}${wait}`));
diff --git a/web/result_actions.js b/web/result_actions.js
index 2a22572c..cb15dd77 100644
--- a/web/result_actions.js
+++ b/web/result_actions.js
@@ -1,7 +1,7 @@
 import { api } from '../../scripts/api.js';
 import { outputDownloadURL } from './output_download.js';
 import { shareButton } from './share.js';
-import { canExtend, extendText, extendSourceExists } from './studio_extend.js';
+import { canExtend, extendText, extendSourceProblem, extendSourceState } from './studio_extend.js';
 import { issueButton } from './issue_button.js';
 
 const css = document.createElement('link');
@@ -74,7 +74,7 @@ export function secondPassAction(t, run) {
         try { await run(); label.textContent = button.title = t('Queued', '已加入队列'); }
         catch {
             button.disabled = false;
-            label.textContent = button.title = t('Second pass did not start · Retry', '二采没有开始 · 重试');
+            label.textContent = button.title = t('Second pass did not start · retry', '二采没有开始 · 重试');
             problem = issueButton(t, () => ({where: ['Second pass', '二采'], message: label.textContent}));
             button.after(problem);
         } finally { button.removeAttribute('aria-busy'); }
@@ -152,20 +152,21 @@ export function resultActions(record, t, {diagnostic = null, secondPass = null,
             problem = issueButton(t, () => ({where: ['Continue video', '延长视频'], message: label.textContent}));
             button.after(problem);
         };
-        const available = exists => {
-            label.textContent = exists ? extendText(t, 'action') : t('Video unavailable · retry', '视频不可用 · 重试');
+        // Only a video the server says is gone keeps the issue button; a check that could not finish just retries.
+        const available = state => {
+            label.textContent = state === 'ok' ? extendText(t, 'action') : `${extendSourceProblem(t, state)} · ${t('retry', '重试')}`;
             button.title = '';
-            if (exists) problem?.remove(); else fault();
+            if (state === 'missing') fault(); else problem?.remove();
+            return state === 'ok';
         };
-        extendSourceExists(api, record).then(available);
+        extendSourceState(api, record).then(available);
         button.onclick = async () => {
             if (button.disabled) return;
             problem?.remove();
             button.disabled = true;
             button.title = t('Checking video…', '正在检查视频…');
             try {
-                const exists = await extendSourceExists(api, record); available(exists);
-                if (!exists) return;
+                if (!available(await extendSourceState(api, record))) return;
                 if (await extend(record) === null) throw new Error(t('Could not open the editor. Try again.', '无法打开剪辑，请重试。'));
             } catch (error) {
                 label.textContent = error.message || t('Could not open the editor. Try again.', '无法打开剪辑，请重试。');
diff --git a/web/sampling_effort.css b/web/sampling_effort.css
index 1d68b873..e9391fa6 100644
--- a/web/sampling_effort.css
+++ b/web/sampling_effort.css
@@ -7,6 +7,7 @@
 .fv-effort{--fv-position:0;--fv-move:cubic-bezier(.25,1.15,.45,1);--fv-tone:var(--fv-accent,#6db8fa)}
 .fv-effort-heading{display:flex;align-items:baseline;gap:8px;margin:0 0 10px;color:var(--fv-muted,#969a9f);font:12px/1.5 var(--fv-font,system-ui)}
 .fv-effort-estimate{margin-left:auto;font-variant-numeric:tabular-nums;text-align:right;font-size:11px}
+.fv-effort-warning{margin:-4px 0 10px;color:var(--fv-warning);font:11px/1.5 var(--fv-font,system-ui);text-align:left}
 .fv-effort-rail{position:relative;cursor:pointer;touch-action:pan-y;user-select:none;-webkit-user-select:none;outline:none;border-radius:20px}
 /* A thick track in the interface's own accent; the knob travels inside it. */
 .fv-effort-track{position:relative;display:block;height:36px;border-radius:18px;background:color-mix(in srgb,var(--fv-ink,#eee) 7%,transparent);box-shadow:inset 0 1px 2px #0005}
diff --git a/web/sampling_effort.js b/web/sampling_effort.js
index 4cb0fabf..9656f6b2 100644
--- a/web/sampling_effort.js
+++ b/web/sampling_effort.js
@@ -99,6 +99,7 @@ export function createSamplingEffort(t, {onChange, onPreview = () => {}, levels
     const header = el('div', 'fv-effort-heading');
     const label = el('span', '', t('Quality', '质量'));
     const estimate = el('span', 'fv-effort-estimate');
+    const warning = el('div', 'fv-effort-warning'); warning.hidden = true;
     // The names under the track show the level; the heading does not repeat it.
     header.append(label, estimate);
     const rail = el('div', 'fv-effort-rail'); rail.tabIndex = 0;
@@ -119,7 +120,7 @@ export function createSamplingEffort(t, {onChange, onPreview = () => {}, levels
         names.append(b); return b;
     });
     const custom = el('span', 'fv-effort-custom', t('Custom steps', '自定义步数')); custom.hidden = true;
-    element.append(header, rail, names, custom);
+    element.append(header, warning, rail, names, custom);
     let selected = 0, disabled = false, pointer = null, draft = 0, currentSteps = 8, currentRefine = 3, twoPass = true, stepCache = null, upscale = 2, reason = '';
     function paint(position) {
         element.style.setProperty('--fv-position', position / 3);
@@ -200,5 +201,5 @@ export function createSamplingEffort(t, {onChange, onPreview = () => {}, levels
         efforts = next;
         update({baseSteps: currentSteps, refineSteps: currentRefine, twoPass, stepCache, upscale, disabled, reason});
     }
-    return {element, update, setLevels, levels: () => efforts, setEstimate(text, title = '') { estimate.textContent = text; if (title) estimate.title = title; else estimate.removeAttribute('title'); }, dispose: cancel};
+    return {element, update, setLevels, levels: () => efforts, setWarning(text) { warning.textContent = text; warning.hidden = !text; }, setEstimate(text, title = '') { estimate.textContent = text; if (title) estimate.title = title; else estimate.removeAttribute('title'); }, dispose: cancel};
 }
diff --git a/web/setup.js b/web/setup.js
index e775cfa4..d9c3dbb3 100644
--- a/web/setup.js
+++ b/web/setup.js
@@ -287,7 +287,7 @@ export async function openSetup() {
             : planErrors ? t("Fix the problems marked in red above, then click Detect & review again.", "上面有标红的问题，处理后请重新点“检测并预览”。")
             : ready && !planId ? t(INSTALLED_HINT.en, INSTALLED_HINT.zh)
             : !planId ? ""
-            : !accept.checked ? t("Tick the agreement above to enable Install / repair.", "勾选上方的同意选项后，即可点击“安装／修复”。") : "";
+            : !accept.checked ? t("Tick the agreement above to enable Install / repair.", "勾选上方的同意选项后，可点击“安装／修复”。") : "";
         blocked.textContent = why; blocked.hidden = !why;
     };
     // A changed setting makes a shown plan outdated: clear it rather than leave it beside a disabled button.
diff --git a/web/share.js b/web/share.js
index 7b66491a..100a2df4 100644
--- a/web/share.js
+++ b/web/share.js
@@ -159,6 +159,7 @@ export async function openShare(record,t){
                 const response=await api.fetchApi('/freevideo/share/video',{method:'POST',headers:{'Content-Type':'application/json'},signal:exportRequest.signal,
                     body:JSON.stringify({id,template,rect:shape.rect})});
                 if(response.status===409)throw Object.assign(new Error(t('Another export is running. Try again shortly.','已有分享视频正在导出，请稍后再试。')),{reportable:false});
+                if(response.status===507)throw Object.assign(new Error(t('Exporting the share video failed: the disk is full. Free some space, then export again.','导出分享视频时出错：磁盘空间不够，空出空间后再导出。')),{reportable:false});
                 if(!response.ok)throw new Error(t('Export could not complete. Try again.','分享视频导出未完成，请重试。'));
                 exported=await response.json();
             }
diff --git a/web/studio.css b/web/studio.css
index 23089455..22482d4e 100644
--- a/web/studio.css
+++ b/web/studio.css
@@ -97,6 +97,7 @@
 .fv-studio .fv-stage[data-short=true] .fv-generation-heading{flex-direction:row;gap:16px}
 .fv-studio .fv-stage[data-short=true] .fv-generation-count{font-size:20px}
 .fv-studio .fv-stage[data-short=true] .fv-generation-times{display:none}
+.fv-studio .fv-stage .fv-generation-progress.fv-generation-waiting .fv-generation-count,.fv-studio .fv-stage .fv-generation-progress.fv-generation-waiting .fv-generation-times{display:none}
 .fv-studio .fv-stage-media video{animation:fv-preview-appear .55s ease}
 @keyframes fv-preview-appear{from{opacity:0}to{opacity:1}}
 /* One metrics card; the 1 px gaps draw the dividers in any column count. */
diff --git a/web/studio.js b/web/studio.js
index 0e686187..2f33f86d 100644
--- a/web/studio.js
+++ b/web/studio.js
@@ -1,8 +1,7 @@
 import {createLanAccess} from './lan_access.js';
 import { createErrorPanel, errorText, createErrorReport } from './error_panel.js';
-import {topDialog} from './dialog_keys.js';
 import {createExpandButton} from './editor_expand.js';
-import {cleanOptionalPrompt} from './optional_inputs.js';
+import {cleanOptionalPrompt, clonePrompt} from './optional_inputs.js';
 import { app } from '../../scripts/app.js';
 import { api } from '../../scripts/api.js';
 import { openSetup } from './setup.js';
@@ -10,7 +9,7 @@ import { feedbackButton, feedbackIssueURL } from './feedback.js';
 import { wordmark } from './wordmark.js';
 import { createGenerationProgress, referenceTrimText, thermalText, sharedGpuText, encoderFileText, encoderFilePath } from './generation_progress.js';
 import { viewSwitch, viewChanged, agentTabReady, openAutopilot } from './view_navigation.js';
-import { agentInterfaceEnabled } from './agent_connect.js';
+import { agentInterfaceState } from './agent_connect.js';
 import { createUpdateNotice, createVersionInfo, productVersion } from './updates.js';
 import { createPortNotice } from './port_notice.js';
 import { reportIssue, reportTipClosed, showReportTip } from './report_issue.js';
@@ -22,7 +21,7 @@ import { createRecentVideos } from './studio_recent.js';
 import { createAgentActivity } from './agent_activity.js';
 import { createStudioQueue, randomSeed, turnNodeOn } from './studio_queue.js';
 import { isNetworkTypeError, connectionLostText } from './browser_extension.js';
-import { attachReferencePicker, referenceItems, syncReferencePrompt, referenceInputs, mentionAt, handleNames, mediaRows } from './prompt_references.js';
+import { attachReferencePicker, referenceItems, enhancementMedia, syncReferencePrompt, referenceInputs, mentionAt, handleNames, mediaRows } from './prompt_references.js';
 import { attachTermSuggest } from './prompt_terms.js';
 import { castInRows, expandCast, mentionHandle, mentionState, mentionedIds, withCharacter, MAX_PICTURES_EACH } from './cast_mention.js';
 import { attachMentionOverlay } from './cast_overlay.js';
@@ -34,7 +33,7 @@ import { continuationUpscale, createSamplingEffort, effortFor, effortsFrom, pass
 import { resultActions, previewBadge, icon, actionsUnavailable } from './result_actions.js';
 import { createPromptEnhancer } from './prompt_enhance.js';
 import { createRequestAssets, enabledLoras, referenceFiles } from './request_assets.js';
-import { rememberPromptDraft, trackPromptDraft } from './prompt_draft.js';
+import { promptDraftTrouble, rememberPromptDraft, restorePromptDraft, savePromptDraft, trackPromptDraft, watchPromptDraftTrouble } from './prompt_draft.js';
 import { canvasDimensions, imageSizeSelection } from './studio_size.js';
 import { createCapabilityView } from './capability_view.js';
 import { openEditorView } from './editor_loader.js';
@@ -44,7 +43,9 @@ import { SHOWCASE_SCENE } from './entry_showcase.js';
 import { createStudioLength, savedSeconds } from './studio_length.js?v=20261011-length';
 import { firstFrameRow } from './first_frame_row.js';
 import { createStudioChain, chainPosition, chainParts, chainSplit, composePartText, readChainResult, sumPartEstimates, singleRunSnapshot } from './studio_chain.js?v=20261019-words';
-import { createExtendState, extendText, extendSourceExists, readExtendRecord, extendSnapshot, mountExtendThumbnail, videoUnavailableText } from './studio_extend.js';
+import { waitingText } from './chain_support.js';
+import { createExtendState, extendText, extendSourceProblem, extendSourceState, readExtendRecord, extendSnapshot, mountExtendThumbnail, videoUnavailableText } from './studio_extend.js';
+import { topDialog, isolateDialogKeys, listenDialogKey, routeDialogKey } from './dialog_keys.js';
 
 const languageOverride = typeof location !== 'undefined'
     ? new URLSearchParams(location.search).get('freevideo_lang') : null;
@@ -155,7 +156,7 @@ function createChainPartList(t, save, {onShow = () => {}, onHide = () => {}, red
         editor.maxLength = 20000; editor.value = row.draft = row.part.text;
         editor.setAttribute('aria-label', t(`Part ${row.part.index + 1} description`, `第 ${row.part.index + 1} 段描述`));
         editor.addEventListener('input', () => { row.draft = editor.value; });
-        editor.addEventListener('keydown', event => {
+        listenDialogKey(editor, 'keydown', event => {
             // Escape leaves the edit, never the studio.
             if (event.key === 'Escape') { event.preventDefault(); event.stopPropagation(); finish(row, false); row.text.focus(); }
         });
@@ -613,7 +614,9 @@ export function showNodes(node, dialog = null) {
 
 const studioQueues = new WeakMap();
 export async function extendVideo(record, node = current?.node || app.graph?._nodes?.find(row => row.type === 'FreeVideoGenerate')) {
-    if (!node || !await extendSourceExists(api, record)) throw new Error(extendText(t, 'unavailable'));
+    if (!node) throw new Error(extendText(t, 'unavailable'));
+    const state = await extendSourceState(api, record);
+    if (state !== 'ok') throw new Error(extendSourceProblem(t, state));
     return openStudio(node, {extendFrom: await readExtendRecord(api, record)});
 }
 
@@ -654,7 +657,7 @@ export function openStudio(node, {extendFrom = null} = {}) {
     const versionInfo = createVersionInfo(cn); tools.append(versionInfo.element); cleanup.push(versionInfo.dispose);
     tools.append(feedbackButton(t), button(t('Settings', '设置'), openSetup, 'fv-quiet'));
     // The Agent's page is the fourth tab, shown where the Agent interface is on; the guide's last step circles it.
-    agentTabReady(agentInterfaceEnabled(api));
+    agentTabReady(agentInterfaceState(api), () => agentInterfaceState(api));
     const navigation = viewSwitch('studio', () => {}, () => showNodes(node, dialog),
         () => openEditorView(node, {studio: dialog}), () => openAutopilot(node, {studio: dialog}));
     // The guide starts after the tab settles: it fixes its steps from the anchors present at that moment.
@@ -724,7 +727,28 @@ export function openStudio(node, {extendFrom = null} = {}) {
     mention.setAttribute('aria-label', t('Reference media (@)', '引用素材（@）'));
     mention.title = mention.getAttribute('aria-label');
     mention.disabled = prompt.disabled; mention.onpointerdown = e => e.preventDefault();
-    const promptTools = el('div', null, 'fv-prompt-tools'); promptTools.append(mention); promptEditor.append(promptTools);
+    // The prompt file could not be written, or the last prompt could not be read at launch.
+    const draftNote = button('', async () => {
+        draftNote.disabled = true;
+        draftNote.textContent = promptDraftTrouble().save ? t('Saving…', '正在保存…') : t('Loading…', '正在读取…');
+        try {
+            if (promptDraftTrouble().save) await savePromptDraft(api);
+            else {
+                // Read again into the open project; edits not yet saved still win over the file.
+                const current = app.graph.serialize(), target = app.extensionManager?.workflow?.activeWorkflow;
+                const restored = await restorePromptDraft(api, structuredClone(current), target?.path || '__existing__');
+                if (JSON.stringify(restored) !== JSON.stringify(current)) await app.loadGraphData(restored, true, true, target || 'FreeVideo');
+            }
+        } catch { /* The note stays with its retry. */ }
+        draftNote.disabled = false; showDraftTrouble(promptDraftTrouble());
+    }, 'fv-prompt-draft-note');
+    const showDraftTrouble = ({save, restore}) => {
+        if (draftNote.disabled) return;
+        draftNote.hidden = !save && !restore;
+        draftNote.textContent = save ? t('Prompt not saved · retry', '描述没能保存 · 重试') : t('Last prompt not loaded · retry', '上次的描述没能读取 · 重试');
+    };
+    showDraftTrouble(promptDraftTrouble()); cleanup.push(watchPromptDraftTrouble(showDraftTrouble));
+    const promptTools = el('div', null, 'fv-prompt-tools'); promptTools.append(draftNote, mention); promptEditor.append(promptTools);
     cleanup.push(() => references.dispose());
     const terms = attachTermSuggest(prompt, {t, host: dialog, suppressed: () => !!mentionAt(prompt.value, prompt.selectionStart, prompt.selectionEnd, handleNames(castChoices()))});
     cleanup.push(() => terms.dispose());
@@ -749,9 +773,8 @@ export function openStudio(node, {extendFrom = null} = {}) {
         onReport: row => { node.freevideoPromptReportId = row.rewrite_report_id || row.report_id; },
         onChange: () => { rememberPromptDraft(api, node); requestAssets?.refresh(); },
         setText: text => { set(node, 'text', text || (untouched() ? starter : '')); placeholder(); }, context: () => {
-        const items = referenceItems(node);
-        if (linked(node, 'conditioning') || items.some(row => !row.file || row.kind !== 'image')) throw new Error('unsupported_media');
-        return {seconds: Number(value(node, 'seconds')), media: items.map(row => ({file: row.file, role: row.role || 'reference'}))};
+        const media = enhancementMedia(referenceItems(node), linked(node, 'conditioning'));
+        return {seconds: Number(value(node, 'seconds')), media};
     }});
     cleanup.push(() => enhancer.dispose());
     requestAssets = createRequestAssets({t, api, request: () => requestAssetsQuery?.()});
@@ -822,7 +845,7 @@ export function openStudio(node, {extendFrom = null} = {}) {
         '先出半分辨率预览，满意后在结果里点“继续二采”，成片与直接生成一致。');
     const previewPassHint = passOptions[2].title;
     passMode.append(...passOptions);
-    passMode.addEventListener('keydown', event => {
+    cleanup.push(listenDialogKey(passMode, 'keydown', event => {
         const step = {ArrowRight: 1, ArrowDown: 1, ArrowLeft: -1, ArrowUp: -1}[event.key];
         if (!step) return;
         event.preventDefault();
@@ -830,7 +853,7 @@ export function openStudio(node, {extendFrom = null} = {}) {
         const current = enabled.findIndex(option => option.getAttribute('aria-checked') === 'true');
         const next = enabled[(current + step + enabled.length) % enabled.length];
         if (next) { next.click(); next.focus(); }
-    });
+    }));
     twoPassLabel.append(el('span', t('Two-pass', '二次采样')), twoPassStatus, passMode);
     let effortEstimate = {}, effortEstimateKey = '', effortEstimateTimer = null, effortEstimateRequest = null, lengthControl = null, chainController = null;
     // A level's step cache; custom steps and workflows without the input sample every step.
@@ -902,7 +925,7 @@ export function openStudio(node, {extendFrom = null} = {}) {
     let releaseExtendThumbnail = () => {};
     cleanup.push(() => releaseExtendThumbnail());
     if (mediaNode?.freevideoCreateMediaEditor) { cleanup.push(mediaNode.freevideoCreateMediaEditor(mediaMount)); mediaDetails.open = mediaFilled(); }
-    else mediaMount.append(el('p', t('Connect a FreeVideo Media panel in node view to upload here.', '在节点视图连接 FreeVideo 素材面板，即可在这里上传。'), 'fv-legacy'));
+    else mediaMount.append(el('p', t('Connect a FreeVideo Media panel in node view to upload here.', '在节点视图连接 FreeVideo 素材面板，可在这里上传。'), 'fv-legacy'));
     paintMediaHint(); mediaDetails.addEventListener('toggle', paintMediaHint);
     controls.addEventListener('change', paintMediaHint, true); controls.addEventListener('input', paintMediaHint, true);
     const connectedLora = upstream(node, 'loras');
@@ -1030,6 +1053,7 @@ export function openStudio(node, {extendFrom = null} = {}) {
         return level?.name === 'Medium' ? t('A continuation uses the Medium level.', '接着生成使用标准档。') : t('A continuation uses the Light level.', '接着生成使用轻量档。');
     }
     function showEffortEstimate(steps = Number(baseSteps.value), twoPassPlan = twoPass.checked, cache = stepCacheNow(), factor = upscaleNow()) {
+        effort.setWarning('');
         const parts = lengthControl.plan?.count > 1 ? t(`${lengthControl.plan.count} parts`, `分 ${lengthControl.plan.count} 段`) : '';
         // A request that fits one run keeps its own estimate while the longest run is unknown.
         if ((lengthControl.capabilityState === 'pending' || lengthControl.capabilityState === 'unavailable')
@@ -1049,7 +1073,11 @@ export function openStudio(node, {extendFrom = null} = {}) {
             return;
         }
         const rangeEstimate = capabilityView?.estimate(steps, twoPassPlan, arguments.length ? 3 : Number(refineSteps.value), cache, factor);
-        if (rangeEstimate) { effort.setEstimate(rangeEstimate); return; }
+        if (rangeEstimate) {
+            effort.setEstimate(rangeEstimate);
+            effort.setWarning(capabilityView.warning(steps, twoPassPlan, arguments.length ? 3 : Number(refineSteps.value), cache, factor));
+            return;
+        }
         const mode = twoPassPlan ? 'two_pass' : 'single';
         // Keys: "<steps>[:<cache>][x<factor> when two-pass and not 2]".
         const estimate = effortEstimate[mode]?.[`${steps}${cache ? ':' + cache : ''}${twoPassPlan && factor !== 2 ? 'x' + factor : ''}`];
@@ -1384,9 +1412,10 @@ export function openStudio(node, {extendFrom = null} = {}) {
             const position = chainPosition(chain, message);
             const joining = readChainResult(chain).joining, waiting = chain.state === 'queued' && message.phase === 'prepare';
             setChainLine(joining ? t('Joining the parts.', '正在接合各段。')
-                : waiting ? t(`Waiting in queue · Part ${position.index + 1} of ${position.count}`, `正在排队 · 第 ${position.index + 1} 段，共 ${position.count} 段`)
+                : waiting ? (chain.waiting ? waitingText(chain.waiting, t) : t(`Waiting in queue · Part ${position.index + 1} of ${position.count}`, `正在排队 · 第 ${position.index + 1} 段，共 ${position.count} 段`))
                 : t(`Part ${position.index + 1} of ${position.count}`, `第 ${position.index + 1} 段，共 ${position.count} 段`));
             progress.element.querySelector('.fv-generation-label').hidden = joining || waiting;
+            progress.element.classList.toggle('fv-generation-waiting', waiting);
             progress.updateReport(message); showProgress(message);
         },
         onState: chain => {
@@ -1405,7 +1434,7 @@ export function openStudio(node, {extendFrom = null} = {}) {
                 : t(`Part ${position.index + 1} of ${position.count}`, `第 ${position.index + 1} 段，共 ${position.count} 段`));
             if (active && node.properties.freevideo_chain?.width) previewSize(node.properties.freevideo_chain.width, node.properties.freevideo_chain.height);
             if (!active) { cancelling = false; hideProgress(); }
-            if (!active) progress.element.querySelector('.fv-generation-label').hidden = false;
+            if (!active) { progress.element.querySelector('.fv-generation-label').hidden = false; progress.element.classList.remove('fv-generation-waiting'); }
             if (chain?.state === 'cancelled') status.textContent = t('Cancelled', '已取消');
             // After a stop with videos queued the next one starts at once; its line says so until something else is done.
             if (stopNote) { status.dataset.error = 'false'; status.textContent = stopNote; }
@@ -1413,9 +1442,9 @@ export function openStudio(node, {extendFrom = null} = {}) {
             updateRunButton();
         },
         onResult: (result, outcome) => { node.freevideoLastResult = result; showResult(result, {chosen: outcome.partial});
-            if (outcome.partial && !outcome.joined) status.textContent = t('Showing the last completed part.', '正在显示最后完成的一段。'); },
-        onError: error => showQueueError(new Error(error.joining ? t(`Could not join the parts: ${error.message}`, `未能接合各段：${error.message}`)
-            : t(`Part ${error.index + 1} failed: ${error.message}`, `第 ${error.index + 1} 段失败：${error.message}`)))});
+            // A part that failed keeps its sentence; the part list already marks which ones finished.
+            if (outcome.partial && !outcome.joined && status.dataset.error !== 'true') status.textContent = t('Showing the last completed part.', '正在显示最后完成的一段。'); },
+        onError: showPartError});
     cleanup.push(() => chainController.dispose());
     node.freevideoStudioChain = chainController;
     node.freevideoEditorLooping = () => queueController.state().looping;
@@ -1522,6 +1551,18 @@ export function openStudio(node, {extendFrom = null} = {}) {
         lastQueueError = null; failure.clear(); node.freevideoFailure = ''; node.freevideoFailureReport = null;
         status.dataset.error = 'false'; status.textContent = '';
     }
+    // A part that stopped is one sentence on the status line; what ComfyUI said goes in the card, as a generation failure.
+    function showPartError(error) {
+        if (disposed) return;
+        status.dataset.error = 'true';
+        status.textContent = error.joining ? t('Could not join the parts. See the details below.', '未能接合各段，原因见下方。')
+            : t(`Part ${error.index + 1} could not be generated. See the details below.`, `第 ${error.index + 1} 段未能生成，原因见下方。`);
+        if (lastQueueError === error.message) return;
+        lastQueueError = error.message;
+        const report = failure.show(error.message, false,
+            {stage: 'execution', node, graph: app.graph, ...(error.joining ? {progress: {phase: 'save'}} : {})});
+        node.freevideoFailureReport = report; node.freevideoFailure = errorText(report);
+    }
     function updateRunButton() {
         const submitting = capturing || queueState?.submitting;
         updateFirstFrame();
@@ -1536,7 +1577,7 @@ export function openStudio(node, {extendFrom = null} = {}) {
         const chainWait = Boolean(chainController?.active && (!chainController.canQueue || !multiple));
         const chainReason = !chainWait ? '' : readChainResult(chain)?.joining
             ? t('Joining parts…', '正在拼接…')
-            : chain?.state === 'queued' ? t('Queued…', '排队中…')
+            : chain?.state === 'queued' ? (chain.waiting ? waitingText(chain.waiting, t) : t('Queued…', '排队中…'))
             : t(`Generating part ${position.index + 1}…`, `正在生成第 ${position.index + 1} 段…`);
         generate.disabled = submitting || chainWait || lengthControl.blocked || (runMode === 'loop' && queueState?.looping);
         generate.textContent = submitting ? t('Adding…', '正在添加…')
@@ -1691,13 +1732,14 @@ export function openStudio(node, {extendFrom = null} = {}) {
             if (lengthPlan?.count > 1 && chosenMode !== 'single') { status.textContent = t('Videos made in parts cannot use batch or loop.', '分段视频不能批量或循环生成。'); return; }
             const lengthOptions = lengthControl.options();
             const source = extension.source;
-            if (source && !await extendSourceExists(api, source)) throw new Error(extendText(t, 'unavailable'));
+            const sourceState = source ? await extendSourceState(api, source) : 'ok';
+            if (sourceState !== 'ok') throw new Error(extendSourceProblem(t, sourceState));
             if (source && lengthPlan?.count > 1) lengthOptions.after = {work: source.video};
             // Only serialization briefly locks the editor. Network submission
             // and all GPU execution leave the next draft fully editable.
             for (const section of controls.querySelectorAll('.fv-section')) section.inert = true;
             held = turnNodeOn(node);
-            const snapshot = singleRunSnapshot(extendSnapshot(cleanOptionalPrompt(structuredClone(await app.graphToPrompt())), node.id, source,
+            const snapshot = singleRunSnapshot(extendSnapshot(cleanOptionalPrompt(clonePrompt(await app.graphToPrompt())), node.id, source,
                 {multiple: lengthPlan?.count > 1}), node.id, lengthPlan);
             if (!snapshot) { fault(t('Could not prepare this run. Try again.', '没能准备这次生成，请重试。')); return; }
             const seedIndex = node.widgets.findIndex(w => w.name === 'seed');
@@ -1850,7 +1892,10 @@ export function openStudio(node, {extendFrom = null} = {}) {
             return;
         }
         generate.click();
-    } }; dialog.addEventListener('keydown', escape);
+    } };
+    cleanup.push(listenDialogKey(dialog, 'keydown', escape));
+    cleanup.push(isolateDialogKeys(dialog, event => routeDialogKey(event, dialog)));
+
     dialog.addEventListener('cancel', e => { e.preventDefault(); closeDialog(dialog); });
     dialog.onclose = () => { disposed = true; for (const f of cleanup) f(); stage.querySelector('video')?.pause(); dialog.remove(); if (current?.dialog === dialog) { current = null; viewChanged('nodes', node); } };
     capabilityView = createCapabilityView({node, api, t, getState: () => ({ratio,
diff --git a/web/studio_extend.js b/web/studio_extend.js
index fa8971b1..08de9da3 100644
--- a/web/studio_extend.js
+++ b/web/studio_extend.js
@@ -6,6 +6,7 @@ export const EXTEND_TEXT = Object.freeze({
     remove: ['Remove', '移除'],
     prompt: ['Prompt', '描述'],
     unavailable: ['Video unavailable', '视频不可用'],
+    unchecked: ['Could not check the video', '没能检查视频'],
 });
 export const extendText = (t, key) => t(...EXTEND_TEXT[key]);
 
@@ -24,21 +25,32 @@ export function extendVideoURL(record) {
         subfolder: split < 0 ? '' : record.video.slice(0, split), type: 'output'});
 }
 
-export async function extendSourceExists(api, record) {
-    if (!canExtend(record)) return false;
-    const url = extendVideoURL(record);
+// 'ok'; 'missing' only when the server answered 404 or 410; 'unknown' when offline, timed out or any other answer.
+export async function extendSourceState(api, record) {
+    if (!canExtend(record)) return 'missing';
+    const url = extendVideoURL(record), gone = status => status === 404 || status === 410;
+    let missing = false;
     try {
-        if ((await api.fetchApi(url, {method: 'HEAD', signal: AbortSignal.timeout(10000)})).ok) return true;
+        const head = await api.fetchApi(url, {method: 'HEAD', signal: AbortSignal.timeout(10000)});
+        if (head.ok) return 'ok';
+        missing = gone(head.status);
     } catch { /* HEAD can fail even when the browser can still read the video. */ }
     try {
         const response = await api.fetchApi(url, {method: 'GET', headers: {Range: 'bytes=0-0'},
             signal: AbortSignal.timeout(10000)});
         // A server may ignore Range. Release its stream without downloading the video.
         await response.body?.cancel().catch(() => {});
-        return response.ok;
-    } catch { return false; }
+        return response.ok ? 'ok' : gone(response.status) ? 'missing' : 'unknown';
+    } catch { return missing ? 'missing' : 'unknown'; }
 }
 
+export async function extendSourceExists(api, record) {
+    return await extendSourceState(api, record) === 'ok';
+}
+
+// What a continuation says when its source video is not 'ok'.
+export const extendSourceProblem = (t, state) => extendText(t, state === 'missing' ? 'unavailable' : 'unchecked');
+
 export const videoUnavailableText = t => t('This video cannot be opened. If you moved the file, refresh the page.',
     '视频打不开，如果移动过文件，请刷新页面。');
 
diff --git a/web/studio_queue.js b/web/studio_queue.js
index 75802379..3d3ffad8 100644
--- a/web/studio_queue.js
+++ b/web/studio_queue.js
@@ -1,7 +1,8 @@
 // ComfyUI owns execution. This controller only submits frozen workflows and
 // replenishes a loop after its previous request has actually completed.
 import {fetchQueueRead, detectBrowserExtension, isNetworkTypeError} from './browser_extension.js';
-const clone = value => structuredClone(value);
+import {clonePrompt} from './optional_inputs.js';
+const clone = clonePrompt;
 
 // One dropped read is usually a connection reset or an engine restart. Only this many
 // polls failing in a row are shown; the next successful read clears the notice.
diff --git a/web/view_navigation.js b/web/view_navigation.js
index 2be364d0..a58d1c39 100644
--- a/web/view_navigation.js
+++ b/web/view_navigation.js
@@ -33,10 +33,29 @@ export function viewChanged(view, node) {
     refreshNavigation();
 }
 
-// Whether this FreeVideo has the Agent interface on. The surfaces that open first (the toolbar, the studio) ask
-// and tell it here; a view switch shows its fourth tab only once the answer is yes.
-export function agentTabReady(answer) {
-    agentTab = Promise.resolve(answer).then(on => on === true, () => false);
+// Whether this FreeVideo has the Agent interface on: 'on', 'off' or 'unknown' (true and false read as on and off).
+// The surfaces that open first (the toolbar, the studio) ask and tell it here. The fourth tab hides only when the
+// interface is off; a check that failed leaves it in place and asks again, waiting longer each time.
+let agentState = null, agentRun = null;
+const agentTabs = new Set();
+const asState = on => on === true || on === 'on' ? 'on' : on === false || on === 'off' ? 'off' : 'unknown';
+// On the Autopilot page itself the tab always shows.
+function showAgentTab(button, state = agentState) {
+    if (!button.dataset.here) button.hidden = state === 'off';
+    if (state === 'on') button.dataset.fvGuide = 'agent'; else delete button.dataset.fvGuide;
+}
+function settleAgentTab(state, recheck, run, tries) {
+    if (run !== agentRun) return;
+    agentState = state;
+    for (const ref of agentTabs) { const button = ref.deref(); if (button) showAgentTab(button); else agentTabs.delete(ref); }
+    if (state !== 'unknown' || !recheck || tries >= 6) return;
+    setTimeout(() => {
+        if (run === agentRun) Promise.resolve().then(recheck).then(asState, () => 'unknown').then(next => settleAgentTab(next, recheck, run, tries + 1));
+    }, 2000 * 2 ** tries);
+}
+export function agentTabReady(answer, recheck = null) {
+    const run = agentRun = {};
+    agentTab = Promise.resolve(answer).then(asState, () => 'unknown').then(state => { settleAgentTab(state, recheck, run, 0); return state; });
     return agentTab;
 }
 
@@ -60,8 +79,9 @@ export function viewSwitch(view, onStudio, onNodes, onEditor, onAutopilot) {
         button.setAttribute('aria-pressed', String(view === id)); button.onclick = action;
         if (id === 'autopilot') {
             button.hidden = view !== id;
+            if (view === id) button.dataset.here = 'true';
             // Settled once the tab is shown or known to stay away; the first-run guide waits for it.
-            group.ready = (agentTab || Promise.resolve(false)).then(on => { if (on) { button.hidden = false; button.dataset.fvGuide = 'agent'; } });
+            group.ready = (agentTab || Promise.resolve('off')).then(state => { agentTabs.add(new WeakRef(button)); showAgentTab(button, agentState ?? state); });
         }
         group.append(button);
     }
@@ -99,9 +119,9 @@ export function openCreativeWorkspace() {
     return true;
 }
 
-export function installNavigation(openStudio, agent = false, {openWorkflow = null} = {}) {
+export function installNavigation(openStudio, agent = false, {openWorkflow = null, recheck = null} = {}) {
     open = openStudio;
-    if (!toolbar) agentTabReady(agent);
+    if (!toolbar) agentTabReady(agent, recheck);
     if (toolbar) { refreshNavigation(); return; }
     toolbar = document.createElement('aside'); toolbar.className = 'fv-view-navigation';
     toolbar.setAttribute('aria-label', t('FreeVideo workspace', 'FreeVideo 工作区'));
@@ -139,7 +159,7 @@ export function installNavigation(openStudio, agent = false, {openWorkflow = nul
     settings.textContent = t('Settings', '设置'); settings.className = 'fv-view-settings'; settings.onclick = openSetup;
     row.append(grip, brand, viewSwitch('nodes', () => choose(node => { open?.(node); }, true), () => {}, () => choose(async (node, created) => {
         const handle = await openEditorView(node);
-        if (created) handle?.notify(t('Created an editing project for you.', '已为你新建一个剪辑工程'));
+        if (created) handle?.notify(t('Created an editing project for you.', '已新建一个剪辑工程'));
         return handle;
     }, true), () => choose(node => openAutopilot(node), true)), createLanAccess(t, api).element, createVersionInfo(cn).element, feedbackButton(t, 'fv-view-settings'), settings); toolbar.append(row);
     toolbar.append(createUpdateNotice(cn).element);