v1.62.0 (AI Workbench, AI Builds Workflows on the Canvas, Repaint an Area, Materials Side Panel)

• 音频库搬进了“角色语音”页：页面顶上的标签现在是 角色 ｜ 配音 ｜ 音频库，左边栏少了一项。加了 AI 工作台以后左边栏有 10 项，数字键只有 1–9，音频库就没键可按了；现在 9 打开角色语音，并回到你上次停的那个标签。

• The audio gallery moved into Voices: the tabs at the top read Characters | Voice-over | Audio, and the rail is one entry shorter. With the AI workbench the rail had ten entries for nine number keys, leaving the gallery without one; 9 now opens Voices on the view used last.

• AI 能拿刚出的图接着改了：搭“在这张图上重画”的工作流时，加载图像节点也认输出文件夹里的图（xxx.png [output]，ComfyUI 本来就支持）；以前 input 文件夹里没有这张图就被拦下。左侧栏的“模型”“图库”不再显示成英文 models / gallery。

• The AI can redraw a picture it just made: a Load Image it builds now takes an output picture (name.png [output], as ComfyUI itself does) instead of being refused when the input folder lacks it. The rail shows Models / Gallery again instead of the raw keys.

• 修复完成后的任务接续：AI 带着原任务 ID 查看结果、补面板，也会留在刚完成的任务里，不会另开空任务或把面板放进另一个任务。错误标记只在对应重试完成并成功后消除；等待审批、仍在运行、已取消，以及另一个节点或设置的成功都不算修好。

• Fixed follow-ups carrying an explicit task ID: reads and panels immediately after completion stay with that task instead of creating an empty task or placing a panel in another one. Error recovery requires a completed successful retry of the recorded target; pending approval, running or cancelled attempts and success on another node or setting do not clear it.

• 参数名自动纠正现在也用于审批和任务记录：你批准后，AI 改用正确参数名重试不再重复问，换模型的结果卡片也不会漏掉。展台和底部提问条同步同一道题的草稿，刷新时保留原来的光标、选区和编辑位置。

• Corrected argument names now carry through approval and task recording: retrying with the proper name reuses an unused approval, and model changes keep their result cards. The stage and bottom question strip share each answer draft; refresh preserves the caret, selection and editing surface.

• 修复任务接续与等待回答：按你最近一次未读回复接续任务，后台进度更新不再打乱顺序；选择题回答也能接回原任务，仍有工具正在执行的任务不会被新会话自动接走。读取答案不指定面板时也会正确处理超时；到时间自动选择或恰好收到你的回答，都会在当次调用里交给 AI。等选择题时发的留言和图片不再被读走却漏传。

• Fixed task resumption and answer waiting: resume by the latest unread user reply, including panel answers, rather than background updates; a task with a tool still running is excluded from automatic takeover. Reading answers without a panel now handles expired questions too. Timeout choices and user answers arriving at that boundary reach the AI in the same call; messages and pictures sent while it waits for a question are no longer consumed without delivery.

• 任务列表更准：AI 改正过来的错（比如参数写错、下一次就对了）不再让整个任务显示成“出错”，只有 AI 没跨过去的才算，红标改叫“有一步没做成”；AI 说做完以后马上再看一眼结果，这一眼记在原来的任务里，不再多出一个只有一步的空任务；接着给你看前后对比、问你满不满意，也回到原来那个任务里（任务重新打开），不再另起一个；你点“新任务”写好的任务一直等到 AI 来接（或者你再开一个新的），不会过了 12 小时悄悄失效、而页面还写着“等 AI 接手”。

• AI 少走弯路、少猜：

  • AI 把必填参数写成了常见的另一个名字（比如打开工作流时把 workflow 写成 name、看图时把 image 写成 file）时，AMB 直接照它的意思做，并告诉 AI 正确的名字，省一次来回。只认考试里见过的这几种叫法；别的名字照旧报错、列出该写什么，免得猜错意思。
  • 时间线里 AI“放大看图”那一步显示成代码名，改成了中文名。
  • 去掉了“给 AI 的工具”里的 日常：所有工具都直接给。Antigravity、Claude Code、Cursor、VS Code 这些软件会自己按需读工具说明，AMB 再做一层“现查现用”只会让 AI 多绕路（一次实测 29 次调用里 18 次花在查工具和转一道上）。
  • AMB 更新以后，AI 软件还拿着旧的工具列表来调用时，任务标题下面会出现一行黄字，告诉你怎么让它重新连接；AI 那边也收到同样的提醒。AI 软件重新取了列表，黄字自己消失。
  • AI 加 LoRA 或换模型时，AMB 对比 LoRA 和主模型的底模：一个是 SD 1.5、一个是 SDXL 这种用不了的，会提醒 AI 并告诉它按底模去找；Pony 的 LoRA 用在 Illustrious 上这种能用但可能效果差的，也会轻轻提一句。只提醒，不拦着。
  • AI 看模型详情时，能看到作者示例图用的采样器、步数、CFG、尺寸，LoRA 还有作者用的权重（来自扫描时存下的 Civitai 资料），还有版本名。
  • 跑图被 ComfyUI 拒绝时，AI 收到的不再是原样的英文报错，而是“哪个节点、缺什么、该用哪个工具修”。
  • 用 Gemini 3.8 Flash 实际考了 7 道常见任务（出图、换模型加 LoRA、问你再改、缺模型、找图看参数、搭放大分支、修断线），失败多半出在第一次用某个工具时参数写错，于是：参数写错时，AI 收到的报错会列出这个工具该写哪些参数，下一次就对；同一个东西在各工具里叫同一个名字（写提示词和跑图都叫 prompt，打开工作流和跑图都叫 workflow）；等待时间写超了（比如 120 秒）按最长的 55 秒等，不再报错；按类型找模型时写单数（checkpoint、lora）也能找到。
  • 很久以前（两个多小时前）留下、AI 一直没收到的留言或标注，不再把一个新对话拉进那个旧任务。
  • 关卡窗口新增 等不到回答时：停下等我（默认）或 按推荐的继续。选后者时，到了等待时间你还没答，AI 标了推荐项的选择题就按推荐的选，展台和“决策”里都写着“到时间，按推荐选了”；没有推荐项的问题、你已经动手写了回答的问题、关卡里“先问我”的事，永远不会替你答。

• The task list says what is true: an error the AI got past (a wrong argument right on the next call) no longer marks the whole task as failed, only one it did not get past does, and the red mark reads "A step did not go through"; a look the AI takes right after saying it is done goes into that task instead of making an empty one-step task, and a before and after or a "happy with it?" it shows next opens that task again instead of starting another; a task you start with New task waits until an AI takes it (or you start another), instead of quietly lapsing after 12 hours while its page still says it waits.

• The AI takes fewer detours and guesses less:

  • When a call writes its one needed argument under a known other name (name for open_workflow's workflow, file for image), Anomalous takes it as meant and tells the AI the right name, saving a round trip. Only the names seen in the exams count; any other is answered with what the tool takes, so no meaning is guessed.
  • The timeline names the AI's "look closer at a picture" step instead of showing its code name.
  • The tool set's Daily choice is gone: AI apps get every tool. Antigravity, Claude Code, Cursor and VS Code load tool descriptions as they need them, so looking tools up through Anomalous only added detours (in one run 18 of 29 calls).
  • When an AI app still calls by the tool list of an older Anomalous, a yellow line under the task header says how to reconnect it, and the AI is told too; it goes once the app fetches the list again.
  • Adding a LoRA or swapping a model, Anomalous compares the base models of the LoRA and the main model: one that cannot work (SD 1.5 on SDXL) is pointed out with how to find a fitting one; one that loads but may look off (a Pony LoRA on Illustrious) gets a softer note. A note only, never a block.
  • Model details give the AI the sampler, steps, CFG and size the author's example pictures used, and for a LoRA the weight (from the Civitai record a scan keeps), and the version name.
  • When ComfyUI refuses a run, the AI gets which node, what is missing and the tool that fixes it, instead of ComfyUI's raw report.
  • Gemini 3.8 Flash took seven everyday tasks as an exam (a run, a model and LoRA swap, asking you first, a missing model, a picture's settings, an upscale branch, a broken link); most failed calls were a tool's first use with wrong arguments, so: an argument error now lists what the tool takes, so the next call is right; one thing has one argument name across tools (prompt for set_prompt as for run_workflow, workflow for open_workflow as for run_workflow); a wait longer than 55 s waits 55 s instead of failing; a model type in the singular (checkpoint, lora) finds them too.
  • A message or marks the AI never got, left more than two hours ago, no longer draws a new chat into that old task.
  • The Gates window has When I don't answer: Wait for me (the default) or Go with the recommendation, where a choice the AI marked a recommended option for takes it once your waiting time runs out, shown on the stage and under Decisions; a question without a recommendation or one you began to answer, and anything a gate asks first, is never answered for you.

• AI 问你的问题更好答：选项是一按就发出去的按钮，右边有箭头，问题下面写着“点一个就发给 AI”，点了马上变成选中的样子；让 AI 定 挪到问题右上角，不再跟选项挤在一起被当成“确认”；“都不是”那一行默认收起成 都不是，我自己说…，关卡卡片上的原因也收起成 说说想怎么改…，点了才出输入框。展台、留言框上方的提问条、关卡卡片都一样。关卡窗口里几条说明改对了：改画布也管加节点、连线、新建工作流；跑完以后那几句话在留言框上方；“先问我”和按钮上的字对上了。写了一半的回答不再被清掉：AI 又问了新问题、读走了另一个回答时，字和光标都还在，那个问题一直排在提问条最前面；窄窗口里你正在打字时，新问题只点亮展台标签，不把你切过去；用输入法按回车选字时，不再把半句回答发出去。

• Questions from the AI are easier to answer: options are buttons that send at a press, with an arrow and "One press sends it to the AI" above them, and the pressed one shows at once; Let the AI decide sits at the question's top right instead of under the options where it read as a confirm button; a line of your own stays folded as None of these, I'll say it…, and the Gates card's reason as Say how to change it…. The same on the stage, the question strip over the message box and the Gates card. Gates window texts fixed: changing the canvas includes adding nodes, wiring and new workflows; the lines after a run are over the message box; the intro names "Ask me" as the button does. A half-typed answer is no longer wiped when the AI asks something new or reads another answer: the text and the cursor stay, and that question stays first in the strip; in a narrow window a new question lights the stage tab instead of taking you there while you type; Enter that picks a word in an input method no longer sends half an answer.

• 局部重绘前先对一下范围：你圈得不一定准，AI 看了标好的图可以挪动、改形状（方框 / 椭圆）、往外扩、把边缘做软，然后在展台把 你圈的 和 AI 调整后 并排给你看，你点 用 AI 调整的 / 用我圈的 / 我重新圈，或者写一句想怎么改。选好以后 AMB 把带这个遮罩的图放进 input 文件夹，AI 收到文件名和一句常用重绘节点的提示；画布上怎么加节点、怎么连线仍然由 AI 自己搭，不套固定模板。

• Agree on the area before repainting: your marks are rough, so the AI can move, reshape (box or ellipse), grow and soften the area and show Your marks beside The AI's area on the stage; you pick Use the AI's / Use mine / I'll mark it again or write what to change. Anomalous then puts the picture with that mask into the input folder and the AI gets its name with a short note on the usual repaint nodes; how to build the repaint on the canvas stays the AI's, no fixed template.

• AI 问你问题后会等你：关卡窗口新增 等我回答（1 分钟 / 3 分钟 / 10 分钟，默认 3 分钟）。一次调用最多等 55 秒，你没答它就接着等，直到这个时间；你在回答框或留言框里打字时，停手一分钟内都算还在答。AI 写完说明或放出展台面板后会被提醒：收工前先问你下一步做什么、要不要开始（跑图、下载、扫描时除外），说明里写了问句的，提醒它改用提问。问题上的“这题跳过”改叫 让 AI 定，AI 收到后自己决定、记下决定，不再把同一个问题再问一遍。

• The AI waits for your answer: the Gates window has Waiting for your answer (1, 3 or 10 minutes, 3 by default). One call waits up to 55 s and the AI then waits on until that time, and while you type in an answer or the message box until a minute after your last key. After a note or a panel it is reminded to ask what next or whether to go ahead before it ends its turn (not while a run, download or scan goes on), and to put a question in its note into ask_user. "Skip this one" is now Let the AI decide: the AI decides, records it and does not ask again.

• 记录里不再出现“你在 AMB 里改了：保存图像 $$canvas-image-preview”：那是 ComfyUI 显示预览图时自己改的内部字段，不是你的改动，以前会被当成你的改动发给 AI。

• The record no longer says you changed a Save Image node's $$canvas-image-preview: ComfyUI sets it to show a node's picture, and it used to reach the AI as your change.

• AI 已经把任务做完了，你再留言或在图上标注：任务重新打开，留言框上方出现那句一键复制的话（“我在 Anomalous 里回复了，看看再接着做”），AI 回来接着在这个任务里做、回复也挂在你那条下面，不再另开一个新任务。给更早的任务留言也一样：AI 手上的任务做完或停下后，会回到你留言的那个任务。

• When the AI had already finished a task and you write to it or mark a picture, the task opens again: the line to copy for the AI shows over the message box, and the AI goes on in this task, answering under what you wrote, instead of starting a new one. The same holds when you write to an older task while the AI had moved on: once its own task is done or paused, it comes back to the one you wrote to.

• 同一个东西只用一个名字：界面上的“配方”统一叫 工作流（工作流页的卡片、保存和导出、操作记录、备份说明），“配方包”保留，它是发给别人的 .zip；AI 工作台里统一用“素材”“留言”“节点结构”“连接”“副本”“撤销”。

• One name for each thing: what the screen called a recipe is a workflow everywhere (the Workflows page, saving and exporting, the activity log, backups); a recipe package stays the .zip you give to others. The AI workbench says materials, message, node map and copy throughout.

• 标注窗口的“同时做成重绘遮罩”勾选框换成 怎么改 三选一：交给 AI 判断、只改标出的地方（其余保持原样，AMB 做好遮罩，AI 只重绘这些地方）、整张重新出图。你选的会作为你的要求写在 AI 收到的第一句，时间线里那条标注也写着“改法：…”；用“涂重绘区”笔刷时自动选中“只改标出的地方”。

• The marker's "make a repaint mask" checkbox is now How to change it, one of three: leave it to the AI, only the marked areas (the rest stays; Anomalous makes the mask and the AI repaints just those), or the whole picture anew. The choice is the first thing the AI reads, as yours, and the timeline row says it; painting with the repaint brush picks "only the marked areas".

• 任务列表更跟手：收起时列表也提前排好藏在左边，鼠标碰到左上角按钮（或从页面往左移到最左边缘）下一帧就滑出来；每几秒的自动刷新只重画变了的任务。点任务马上显示：看过的任务直接出来，鼠标停在任务上时先读好，再悄悄补上最新内容；展台里的图先用缩略图，显示得比缩略图大时才换原图。

• The task list keeps up: folded, it stays laid out out of sight and slides in from the next frame when the pointer reaches its button (or moves left onto the page's left edge); the regular refresh redraws only the tasks that changed. A picked task shows at once, from the copy read before or read ahead while the pointer rested on it, then fills in what changed; the stage's pictures load as thumbnails first and as the file itself only when shown larger.

• 任务标题栏的“专用标签页”改叫 AI 改副本 ▾：点开是带说明的二选一（改副本 / 改当前画布），不再点一下就悄悄切换；旁边的按钮叫 切到副本。快捷指令和关卡窗口用同样的说法。

• The header's "Own tab" is now AI edits a copy ▾, opening the two ways with what each means instead of switching at a click; beside it, Show the copy. The / commands and the Gates window say the same.

• 素材的“来源”列表：鼠标移到哪一行，哪一行就亮起来，当前选中的那项有底色和勾（之前被一条样式盖掉了）；名字没被截断时不再弹白色小提示；搜索框里能用 ↑ ↓ 选、回车确定。

• The materials' Source list lights the row under the pointer and tints and ticks the one shown (a later style had hidden both); a name gets a tooltip only when it is cut short; ↑ ↓ and Enter work from the search.

• 展台入口显示面板数量；收起或查看其他栏时出现新增、更新，会亮主题色小圆点，打开查看后消失。

• The Stage button shows its panel count and a theme-colored dot for new or updated content while hidden; viewing the stage clears the dot.

• “回到最新一步”改为约 240 毫秒的短滚动，工作台刷新不会打断它；滚轮或触摸可随时接管，减少动态效果时直接定位，并补上按钮按下反馈。

• Jump to latest now uses a short 240 ms scroll that survives workbench refreshes, yields to user navigation, respects reduced motion, and gives immediate pressed feedback.

• AI 工作台新增右侧“成果”页：按步骤汇总图片，集中查看工作流、搭配、最新图片的提示词、模型下载请求和音频；复用图库的保存、参数对比与提示词复制入口。时间线上方改为简洁的成果统计，右侧按钮提示尚未查看的新成果。

• The AI workbench now collects pictures, workflows, combos, prompts, model download requests and audio in a Results side view, reusing gallery keep/comparison actions. A compact timeline summary and unseen-result badge replace the old results strip.

• AI 搭工作流支持一次添加并自动接线：insert_after 接回选中的下游，wire_from / wire_to 按唯一类型批量连接。正负条件歧义、多分支、已有输入冲突和循环在写入前报明；连接失败恢复原接线。成功回执包含实际连线、缺失输入与输出可达性，支持原生 Ctrl+Z 撤销。

• 现状简报新增有长度限制的流程链；节点搜索只返回少量摘要，具体参数按需读取，减少逐线调用和重复上下文。

• Canvas tools can add and wire a node in one call, safely splice a selected branch, and connect unique concrete types by node IDs. Ambiguity/conflicts/cycles are rejected before mutation; failed connections restore previous wiring. Compact receipts, bounded flow chains and brief node search reduce context.

• AI 加节点时能插在已有的线前面：wire_to 指向的输入原来接着线的，改接新节点，原来的来源接进新节点，比如在采样器前加 ControlNet（正向接正向、反向接反向），或在采样器和两个提示词前加 LoRA，一次调用接好。同类型的几个插口名字相同时自动配对；insert_after 会把 LoRA 的模型线和 CLIP 线一起接进去。

• add_node splices in front of a node: an input wire_to feeds that was already wired now goes through the new node, its old source feeding the new node (a ControlNet before a KSampler, positive to positive and negative to negative; a LoRA before the sampler and both prompts). Same-type sockets of the same name pair up, and insert_after splices every type a node passes through (a LoRA's MODEL and CLIP).

• 改参数不再靠固定的填空：run_workflow 和 set_values 都能按参数名改任何节点的参数，第三方节点也一样（比如 {"steps": 30}）。只有一个节点有这个参数就直接改；几个节点都有时（高清修复的两个采样器）什么都不改，列出是哪几个，让 AI 写上节点号。几个正向提示词框接进同一个采样器（分区提示词）时，prompt 也不再把同一句话填进所有框，而是停下来让 AI 分别写。

• Settings go by name instead of a growing list of blanks: run_workflow's and set_values' values set any node's setting, a third-party node's too ({"steps": 30}). The one node that has it is changed; a name several nodes have changes nothing and lists them, for the AI to name the node ({"12": {"denoise": 0.5}}). prompt no longer gives one text to several positive prompt boxes combined into one sampler (regional prompts).

• 一次调用加一组节点：add_nodes 按 ComfyUI 的 API 格式写一组节点，组里互相连，也能连画布上已有的节点（"#12"），AMB 先检查类型，排成一块放在空位上，接好线，整组算一步撤销。原来要十几次调用的局部重绘分支，现在一次。

• add_nodes adds a group of nodes in API form to the open canvas in one call: links inside the group, to canvas nodes as "#12" (outputs by slot or name) and wire_to into canvas nodes, checked first, laid out as one block and undone in one step.

• 图跑完后，第一个告诉 AI "跑完了" 的回复（不管是查进度还是别的工具）直接带着这次出的图，最多 4 张缩略图，每次运行只带一次。AI 不用再去取图，看完就能留言。

• The answer that first tells the AI its run ended, run_status or any other, carries up to four thumbnails of the run's pictures, once per run, so it need not fetch them with get_image_info.

• 跑完叫我：关卡窗口新增“等跑图”，默认“跑完叫我”。AI 提交后最多等 1 分钟（跑得快的当场就拿到结果），还没跑完就告诉你在跑，然后收工，不再一遍遍查进度（以前一张图要查 10 次左右）。跑完后，留言框上方有几句话，点一下复制，粘到 AI 的对话框里，它就接着看结果和图；AI 知道结果后这几句话自动消失。选“AI 自己等”就和以前一样。

• Tell AI when done: a new setting in the Gates window, on by default. The AI waits up to a minute after queueing (a quick run comes back with its pictures at once) and, if the run still goes, says so and ends its turn instead of polling run_status. When the run ends, lines over the message box are ready to copy into the AI's chat; they go once the AI knows. "AI waits" keeps the old way.

• 工具说明压短：常用工具的说明重写，去掉重复和绕的话，规则和“不要做”的都保留；开始任务时收到的规则从约 1900 字减到约 900 字。

• Shorter tool descriptions: the everyday tools' descriptions are rewritten without repeats, keeping their rules; begin_task's rules are about 900 characters instead of 1,900.

• 工具目录：AI 开始任务时，回复里附带一张按“想做的事”排好的工具目录（出图、画布、模型、提示词、图片、和你沟通……），AI 看一眼就直接去用对的工具，不用先到处试。目录从工具登记表自动生成，只列当前给 AI 的工具，新工具会自己出现。

• A directory of the tools by what the AI wants to do comes with begin_task's answer (and get_situation's while there is no task): each tool offered under its stage of the work with its title, built from the catalog so new tools show up by themselves.

• 给 AI 的话常驻在消息框上方：AI 停下来的时候（没在调用工具、没在跑图、没在等你回答），AI 工作台消息框上方总有一行话，点一下就复制，粘到 AI 的对话框里。话跟着情况变：图跑完了是“跑完了，看看结果”这几句；你在 AMB 里回复了、AI 还没收到，是“我在 Anomalous 里回复了，看看再接着做”；别的时候是“继续”。最后一句“用 Anomalous 接着做任务：…”给新开的 AI 对话用。原来挂在跑图那一步下面的几句话挪到了这里。“交给 AI”给你的那句话也不再叫 AI 用旧的 show_panel 问你。

• Lines for the AI's chat stay over the message box: whenever the AI is not in a call, a run or a question, the AI workbench offers a line to copy that fits what waits (about the run that ended; "I replied in Anomalous, take a look and go on" when it has not got your reply yet; else "Go on"), and one that picks the task up in a new chat. The lines under a finished run's step moved here. The Hand to AI sentence no longer tells the AI to ask with show_panel.

• AI 在 AMB 里问你时至少等 30 秒：就算它自己说不等（Gemini 常这样），你在 AMB 里点了选项，它在同一步里就收到，接着做，不用再回 AI 的对话框说“继续”。“加节点”工具的说明也提醒 AI：要加一组互相连着的节点（重绘、放大一段），用一次调用的“加一组节点”。

• ask_user waits at least 30 s even when the AI asks it not to, so an answer given in Anomalous right away reaches the same call and the AI goes on without a word in its chat. add_node's description points to add_nodes for several nodes wired to each other (an inpaint or upscale branch).

• 时间线更短：AI 连着做的事（查资料、加节点、连线、写提示词、改参数、记决策）三步以上、其中至少两步不是查资料时，叠成一行“做了 N 步”，下面一排小标签写做了什么（加了节点 ×5、连了线 ×10……），点开就是原来的每一步。行上有“撤销这一串”，一下撤掉这一串在画布上的改动。你的话、AI 留言、问你的、面板、出图、出错的、等你允许的、正在做的一直露在外面。AI 还在做时，这一行写着最后一步是什么。一个修手的任务从 43 行变成 18 行。

• A shorter timeline: three or more of the AI's working steps in a row (reads, canvas changes, prompts, settings, decisions; two of them not reads) fold into one row, "Did N steps", with chips of what they did and "Undo these" for their canvas changes; it opens to the steps as before. Your words, the AI's notes and questions, panels, runs, errors, steps waiting for you and running ones stay in sight. While the AI is at it, the row names its last step. A hand-fix task went from 43 rows to 18.

• 展台自动打开：AI 往展台放了新面板，右栏自己打开并跳到它，你关掉了或者在看节点结构、决策、素材也一样；AI 只是改已有的面板（比如进度）时不会。展台顶上有开关“有新东西就打开”，默认开着。窗口窄时不抢时间线，只在“展台”标签上亮个点；AI 问你问题时仍然切过去。

• The stage opens by itself: a new panel from the AI opens the side column at it, even when you closed it or were on another view; the AI changing a panel (a progress bar) does not. The switch "Open on new panels" in the stage's head turns it off. In a narrow window it lights the stage's tab instead of taking the timeline's place; a question still brings the stage forward.

• 打开右栏的按钮更醒目：时间线里每个面板那一行变成卡片，左边小图标表示里面是什么（问题、图、对比、节点结构），标题下一行写内容（“1 张图”“问你 · 2 个选项”），右边是实心按钮“在右侧打开”；右栏正在显示的那个面板，按钮换成绿点“正在右侧显示”。跑图那一步的“在右侧看大图”也换成同样的按钮。“在画布上找到”“撤销”“在图库里看”这些蓝色小字换成带图标的小按钮。

• Ways to the right are easier to see: each panel's line in the timeline is a card with an icon of what it holds, a line of what is in it ("1 picture", "Asks you · 2 options") and a solid "Show on the right" button, which turns into a green "On the right now" while that panel is in view there. A run's "See it large on the right" is the same button. "Find on canvas", "Undo" and "See it in the gallery" are small buttons with icons instead of blue text.

• 修复任务画布标签关闭或切换失败后误用当前工作流的问题：操作会明确报错，重新打开或明确指定新画布后才能继续。下载缺失模型现在也跟随任务画布；留言里的 @ 搭配和工作流按完整名称查找，不再因模糊搜索前几项占满而误报缺失。

• Task canvas actions now fail when their recorded tab is closed or cannot be activated instead of using another workflow. Missing-model downloads follow the task tab, and message references resolve exact combo/workflow names before search limits.

• AI 工作台任务工具栏新增小巧的“运行状态”按钮与浮窗（移除遮挡时间线的常驻大横条）：点开即可查看真实运行状态、排队、节点进度、用时和结果，运行中直接在按钮上显示进度百分比；没有收到运行信息时明确提示，不再空转显示“AI 正在工作”或把沉默当成完成。页面可见时自动补查状态。运行工具默认等待 15 秒并返回实际结果；外部提交可用 track_run 接入准确的运行 ID，不会重新出图。

• The AI workbench now places confirmed execution state, queue/node progress, elapsed time and results inside a compact toolbar button and popover (instead of a tall sticky banner), with visible-page refresh. Untracked work and silence are explicit. Run calls wait briefly for real results; track_run attaches external submission IDs without resubmitting.

右边一栏一次一样，按钮固定在右上角 / One view at a time on the right, switched from the top right

• 工作台右边那一栏一次只显示一样东西，用窗口右上角的三个按钮切换：展台、节点结构（AI 跑过的节点结构图，大图，可以选看哪一次运行）、素材（原来停在屏幕右边的素材小窗，现在就在这一栏里）。再点一下亮着的按钮收起这一栏；切走再切回来，内容都还在。时间线里每次运行都有 在右侧看大图。
• 任务标题和那一排按钮在时间线往下翻时一直停在顶上。
• 右边这一栏改成一张单独的卡片，不再用一条线硬隔开；默认窄一些，给中间的时间线留出地方。它左边的缝里有三个小点，鼠标移上去变成一条竖线，拖动调宽窄（记得住，双击恢复）。这条缝和时间线的滚动条分开，不会再一抓就抓错。
• 左边的任务列表用窗口左上角的按钮收起 / 展开；收起后鼠标移到按钮上，列表会临时浮出来，点任务或移开就收回。
• 素材在右边这一栏里时，鼠标移过卡片不再自动跳进详细说明（之前很难操作），点卡片上的 ⓘ 才看。
• 左边任务列表收起后，鼠标从按钮往下移去点任务时列表不会再中途消失（之前移得慢一点就会收回）：鼠标在按钮和列表围成的范围里都算还在用，离开后才收回。
• The folded task list no longer vanishes on the way down from its button: it stays while the pointer is anywhere in the box around the button and the list, however slowly it moves.
• 模型补丁不再被当成 ControlNet：Anima LLLite、Qwen-Image 分块 ControlNet、Z-Image Fun ControlNet 这类文件名字里常带 ControlNet，但只有「加载模型补丁」节点能读，它只看 model_patches 文件夹。现在拖进模型页会自动放进 model_patches；已经放在 controlnet 里的，模型页的“整理”会提示挪过去。
• Model patches (Anima LLLite, the Qwen-Image block-wise and Z-Image Fun ControlNets) are told from ControlNets by their header: imported into model_patches, and Tidy offers to move those already in controlnet.
• AI 找模型时直接知道该用哪个节点：search_models / get_model 给每个模型写上能加载它的节点；放错文件夹的会说该放哪、放在这里加载不了，同名文件已经在对的地方时会直接指给 AI。新工具 move_model 把放错的模型挪过去（新关卡“挪动模型文件”，默认先问你；从不覆盖文件）。AI 不必再去命令行翻 ComfyUI 的源码。AI 加节点时如果下拉列表是空的，会被告知“ComfyUI 没找到文件，可能放错了文件夹”。
• search_models and get_model say which node types load each model, and where a misfiled one belongs; move_model moves it there behind a new "files" gate (asks by default, never overwrites). An empty drop-down in add_node now says ComfyUI found no file for it.
• AI 第一次在任务的副本标签页上动手时，切换标签页不再被记成“你在画布上的改动”告诉 AI。
• The AI switching to its task tab is no longer reported to it as a canvas change of the user's.
• 工作台里直接用图库的图：留言框左边新增图片按钮，打开图库（最新的在前，可按提示词、模型、seed 搜索），点选几张后 加进留言；打 @ 时也能看到图库最新的图，打字会顺带在图库里搜。AI 收到的是图片本身和它的生成参数。
• Gallery pictures in the AI workbench: the message box's picture button picks them from the gallery (newest first, searchable), and "@" offers the gallery's newest pictures and searches it as you type; the AI gets each picture with what made it.
• 等 AI 接手时先备好素材：“等 AI 接手”的任务是一页：最上面是 把这句话发给你的 AI 和 复制这句话，往下滚时它停在顶上，挑完图不用翻回去；下面的 写一句要做什么、从图库挑图给 AI（大图库可搜索，点一张就放到最下面的留言框，再点一下拿掉）都是可选的，点了才展开，行上写着已经放了几张；备好的都列在下面。AI 接手时一起收到。那句话改成让 AI 照任务里的交代做，没有交代才问你，不会一接手就停下来问。
• A task waiting for an AI is one page: Send this to your AI with Copy leads and stays at the top while the page scrolls; Say what to do and Pictures from the gallery (a big searchable grid feeding the message box) are optional folds that open when pressed and say how many pictures are in; what is ready is listed under them. The sentence now has the AI do what the task says, asking only when it says nothing.
• 不全屏时中间更宽敞：AI 工作台的窗口不够高时（没有最大化），任务标题、状态、时间步数、“接着”和右边的工具并成一行，上下留白也收紧；输入框下面的提示只在出错时显示（平时鼠标停在输入框上能看到）；等 AI 接手的页面图片变小、步骤靠紧。同样的窗口里，中间能多显示约两成内容；最大化时样子不变。
• In a low window (not maximised) the AI workbench keeps the task's head to one line, tightens the room around it, shows the message box's hint only for a failure (its words are the input's tooltip) and shrinks the waiting page's picture: about a fifth more room for the timeline.
• 换类型：模型详情页多了 换类型 按钮，给“导入时类型选错了 / 代码认不出、放错了文件夹”的模型用。窗口里写明现在放在哪一类、文件头看起来像什么（看得出就替你选好并标“推荐”，不确定标“可能是”，看不出就不猜），每一类都写着用哪个节点加载；可以改挪到的根目录和子文件夹，下面显示完整路径。挪动和“整理”用的是同一套：封面、信息、备注一起走，从不覆盖文件，挪完 ComfyUI 的模型列表马上更新。导入单个模型后，提示里也有 类型选错了？，点一下直接改。
• "Change type" on a model's page, and "Wrong type?" in the toast after importing one model: the window says where it is and what its header suggests (picked and marked when it can tell, nothing guessed when it cannot), offers every type with its loader and where it goes, and moves it as Tidy does.
• 跨盘挪动、复制模型时，除了比大小，还会比对文件开头、中间、结尾各 1 MB 是否一模一样，对不上就不删原文件（只多花几毫秒，不需要事先有指纹）。
• A model copied to another drive is checked at its start, middle and end (a megabyte each) besides its size before the original goes; no fingerprint needed, no felt cost.
• 挪动、导入模型时，就算目标位置在最后一刻冒出一个同名文件，也绝不会被盖掉：这次挪动会失败，两边的文件都原样保留（之前在 Windows 上理论上有极小的机会覆盖）。
• Moving or importing a model never replaces a file that took the name at the last moment: the move fails and both files stay (on Windows a rename could replace it before).
• AI 能放大看图、看预览，不用再自己写代码：新工具 zoom_image 把一张图的某一块放大给 AI 看，还能把两张图的同一块并排比较（改之前和改之后）；输出、输入、临时预览的图都能看，只看不改。出图结果里会列出这次运行的预览图（比如 OpenPose 骨骼、遮罩的预览），AI 用它们就能检查中间结果，不必自己去跑预处理器。AI 开任务、收到遮罩时都会被明确告知：要改图就跑 ComfyUI 的工作流，绝不用自己的代码去画、去改、去覆盖图片文件。
• zoom_image shows part of a picture larger (output, input or temp folder), with the same part of a second picture to compare, reading only; run_status lists a run's PreviewImage pictures; begin_task and the repaint note tell the AI to change pictures only by running workflows, never by editing picture files with its own code.
• AI 能直接看输入文件夹里的图：get_image_info 加了 folder=input，能看 ComfyUI 输入文件夹里的图（你标注时 AMB 存的重绘遮罩图就在那里）：尺寸、要重画的区域在哪、占多大，也能附上图本身。交给 AI 遮罩图时也明说“它已经在输入文件夹里了，不用跑命令检查”，Gemini 不会再为这个弹窗问你。
• get_image_info with folder=input reads a picture in ComfyUI's input folder (the inpaint picture Anomalous writes for a repaint): its size and the area LoadImage gives as MASK, optionally attached; the marks message says the inpaint picture is in place, so the AI need not check it with a command.
• AI 少去翻文件：AMB 工具返回给 AI 的内容去掉了多余的空格、画布里 ComfyUI 自己的内部字段不再列出、操作记录默认只给最近 10 条，这样 Antigravity 不必把结果另存成文件再让 Gemini 去读。AI 开始任务时也会被告知：模型、图片、节点、画布和运行都有 AMB 的工具，不要自己去读 ComfyUI 的文件夹、源码或接口（那样每一步都要你点允许）。
• Tool results reach the AI as compact JSON, without the canvas's internal "$$" widgets, and recent_activity gives the last 10 by default, so apps such as Antigravity keep more of them in the chat instead of a file the AI must open; begin_task tells the AI that AMB's tools cover models, pictures, nodes, canvas and runs, so it need not read ComfyUI's folders, source or routes.
• 修复输入框上方的出图进度停在“节点 0/29 0%”不动的问题：现在它和时间线里那一步、标题栏的进度按钮一起更新。
• The run line over the message box now follows the run's progress with the timeline's bar and the header's button, instead of staying at the start.
• 最下面就能看到 AI 在做什么、要你做什么：输入框上方多了一条“现在”栏：AI 出图时显示进度（节点、步数、百分比、进度条）和停止按钮；AI 某一步还在进行时说在做什么；AI 想做要你点头的事（比如挪模型、下载）时，直接在这里 允许 / 不要；AI 问你问题时，问题和选项就在这里，点一下就回答（有推荐的带 ☆，也能自己写一句或跳过），多选和挑图会带你去展台。
• A "now" strip over the message box: a run's progress with stop, the AI's step still going, a gate to allow or refuse right there, and the AI's open question with its options as buttons (one's own words or skip too; several picks and pictures to pick lead to the stage).
• 拖进来、粘贴进来的图片也能发给 AI：把电脑里的图片拖到输入框上，或者直接 Ctrl+V 粘贴截图，它会存进 ComfyUI 输出文件夹里的 ai_inbox（图库里也看得到，从不覆盖文件，只收 PNG、JPEG、WebP、GIF，最大 25 MB），并加到这条消息上。从 ComfyUI 页面里拖来的输出图片直接加上，不另存。
• Pictures dropped on the message box or pasted are kept in the output folder's ai_inbox (seen in the gallery; never over a file; PNG, JPEG, WebP or GIF up to 25 MB) and attached; an output picture dragged from ComfyUI's page is attached as it is.
• 很长的 AI 留言先折叠：超过八行左右的留言只显示开头，点 展开全文 看全部。
• Long AI notes are folded to their first lines with "Show all".
• 输入框固定在最下面，像聊天软件一样：不管时间线多长、翻到哪里，留言框都停在任务这一栏的最下方。它上面是这条消息要带的东西：图片显示成小图，每张有 ✎ 标注（画框、圈、写要改哪里，标好后单独发给 AI）和 ×；@ 提到的模型、搭配显示成小标签。从图库挑的图不再往文字里塞“@文件名”。左边是 图库 和 素材 两个按钮，给了东西就亮起并显示数量。输入框会随内容变高：Enter 发送，Shift+Enter 换行；只发图片不写字也可以。往上翻了一屏以上，或者往上翻着看时 AI 有了新步骤，输入框右上方会出现圆形 ↓ 按钮（有新步骤时角上带数字），点一下回到最新；自己发出的消息会自动滚到眼前。时间线里你发的图片和 AI 出的图一样，可以查看、对比、收藏、标注。
• The message box is docked at the foot of the task column like a chat app's: over the input what the message carries (pictures with ✎ to mark up and ×, chips for what "@" names; gallery pictures are attached, not typed in), beside it the gallery and materials buttons with their counts, an input that grows (Enter sends, Shift+Enter breaks the line; pictures alone are fine), a round ↓ button back to the newest step (with the count of new ones) when you read more than a screen up or the AI adds steps meanwhile, and your own message scrolled into view. Pictures you sent carry the same tools as the AI's. The side column's grip sits in its own gap, three dots, beside the timeline's scrollbar instead of over it.
• 打开任务就停在最新一步：像聊天软件一样，打开任务直接看到最新的进展。你停在最下面时，AI 的新步骤来了就跟着往下走，不管任务现在是什么状态（之前只在 AI 正在干活时才跟，打开任务时还会停在上一个任务翻到的位置）；往上翻着看就不动你。
• A task opens at its newest step; read at the foot, the timeline follows new steps whatever the task's state (it used to follow only while the AI worked, and opened where the previous task was scrolled to); read further up, it stays put.
• 面板上的进度条不再冒充实时进度：AI 画在面板上的进度条是它自己写的数，不会自己动。任务往前走了（那次出图结束了、AI 又画了新面板或留了言）而它没去更新，这条进度条会变成灰色，并写明“这是 AI 在几点写下的数，之后没有再更新”。出图的真实进度一直在时间线那一步、输入框上方和标题栏的运行状态里。AI 也会被告知：出图不用另画进度条，自己画的要自己更新。
• A panel's progress bar is the AI's own numbers: once the task goes past it without an update (a run ends, a later panel or note), the stage greys it and says when it was written. The AI is told a run's real progress shows already and to keep its own bars current.
• 还没出图就问进度不再算“出错”：AI 在出图前调用 run_status，现在得到“这个任务还没有出图”的回答，任务不会因此标上“有一步出错”。
• run_status before any run answers state "none" instead of failing, so the task no longer shows a failed step for it.
• 宽窗口里时间线收在中间：右边一栏关着、窗口又很宽时，时间线、标题和输入框不再横跨整个窗口，而是收成中间约 880 像素宽的一列（像聊天软件那样），滚动条仍在最右边。时间线里的图片按自己的大小从左排开，一张图不再被撑成横跨整行的大灰框。
• AI 每一步都知道下一步做什么：开始任务、出图和看进度、加节点连线、问你问题得到回答，这些回复里都多了一句 next：比如“还在跑，等 55 秒再看”“跑完了，先放大看图再告诉你结果”“还缺这几根线，照清单连”“缺模型，先查缺哪个”。小模型不擅长自己规划，但很会照着做。
• Answers of begin_task, run_workflow and run_status, the canvas changes and an answered ask_user carry next: the one thing to do now.
• 照着跑：日常出图一步完成：AI 出图时可以直接说“用哪个存好的工作流、改什么”：run_workflow 带上工作流名，再带上要改的提示词、反向提示词、种子、尺寸或一张图（比如你标注过的重绘图），AMB 就在任务的副本标签页里打开它（第一次打开，之后接着用同一个标签页），填好这几处，其余照存好的样子不动，然后出图、等结果。尺寸只改起始的空白潜空间，高清修复里的 Latent 放大保持原来的放大尺寸。原来要 6、7 步（找工作流、打开、看画布、改提示词、改种子……），现在一步，小模型也不容易绕晕。填的内容和出图一起算一步，改画布那部分同样能 Ctrl+Z；如果你给“改画布”设的关卡比“用显卡”严，这一步按改画布的关卡来问你。画布报错是因为缺模型时，“现状一览”会让 AI 先查缺哪个模型，而不是乱换选项。
• run_workflow takes blanks for a daily run in one call: workflow (a saved one, opened in the task's tab once and reused), prompt, negative, seed, width, height and image (image_node for several LoadImage nodes); the rest stays as saved. The size goes to the latent a run starts from, never to a hires fix's latent upscale. With blanks a run asks as the stricter of the GPU's and the canvas's gate.
• 现状一览，AI 一次就知道到哪一步了：新工具 get_situation，AI 调一次就拿到：任务（你最后说了什么、有几个问题在等你、定了哪些决策、出过哪些图、“AMB 之外”几次）、画布简报（每个节点一行：谁接到谁；还缺哪几根线、各要什么类型、画布上哪些现有输出能接上；ComfyUI 报错的节点）、最近一次出图，以及下一步该做什么。回复控制在约 3600 字以内，Antigravity 不会把它存成文件再让 AI 去读。新开对话或拿不准时先调它；连接说明里的那句话和规则也写上了。小模型（比如 Gemini Flash）不用再把整张图记在脑子里。
• get_situation: the task, the canvas in brief (a line per node, the inputs still unwired with the outputs that fit, ComfyUI's errors), the latest run and the next step in one answer under about 3,600 characters, for a new chat or an AI that lost track.
• AI 每连一根线都知道还缺什么：add_node、connect_nodes、remove_nodes 的回复里附上画布上还没接的输入，以及哪些现有输出能接上，像一张清单，照着补完就能跑。
• add_node, connect_nodes and remove_nodes answer with the inputs still unwired and what fits them.
• AMB 之外做的事也看得见：AMB 只看得到 AI 发给它的调用，AI 软件自己跑的命令和脚本它看不到。现在 AI 每调用一次 AMB，AMB 就看一眼输出和输入文件夹：上次调用以来有没有不是出图产生的新图片、有没有图片被原地改写（比如之前那张被覆盖了 6 次的图）。有的话，时间线上多一行橙色的 AMB 之外，列出是哪些文件，输出文件夹里的图直接显示；AI 同时收到提醒：可能是你自己做的，如果是它做的，改图只能跑工作流。只在 AI 调用时看，不在后台扫描；ComfyUI 正在出图时，新图等出图结束再判断，不会把正常出的图当成外面来的。AI 也多了一个工具 log_outside：用了别的工具（命令、脚本、读写文件）后说清做了什么、为什么、动了哪些文件，同样显示为一行 AMB 之外；任务的统计里会写“AMB 之外 N 次”。
• Pictures changed outside Anomalous show in the timeline: when the AI calls, Anomalous looks at the top of the output and input folders for pictures no run saved and pictures rewritten in place since its previous call, and tells the AI with its answer; new pictures wait while ComfyUI runs a prompt. log_outside is the AI's own word on a command or script it ran. Both are "Outside AMB" steps, counted on the task.
• 连接说明里多了一段给 AI 的规则：第 ④ 步可以复制一段规则（先用 AMB 的工具、只通过跑工作流改图、在 AMB 里问和说明、用了别的工具就说一声），放进 AI 软件的“规则”里（Antigravity：Customizations → Rules → + Global），它每次对话都会读到。
• The connect guide has rules to paste into the AI app's own rules: use AMB's tools, change pictures only by running workflows, ask and explain in AMB, and say what was done outside it.
• In a wide window the task (its head, timeline and message box) keeps to a centred column of about 880 px, as chat apps do; the scrollbar stays at the edge. A step's pictures are cards of about their own size from the left, no longer one stretched across the row.
• 点得到、看得见：选中的图片有粗框、发光和大对勾，图片罩上一层颜色，和鼠标停留一眼分得开；留言框左边的图片按钮在图库面板打开时亮起，留言里有图时显示张数。素材列表、AI“看了资料”那一行（带 › 可展开）、关卡选项、优先用 / 只能用、来源下拉、发送和复制按钮、模型页的标签、首页的“一键准备”、记录里可展开的一行，鼠标移上去都有明显的变化。点 展台 时即使还没有内容，也会打开并说明会放什么，不再没反应。“交给 AI · 0”在什么都没选时不能点。挪模型那一步的标题写文件名和挪到哪，不再是一串编号；只改了一个节点时不再出现两个“在画布上找到”；“接着哪个任务…”的字号和旁边一致。
• Clearer feedback: picked pictures get a thick glowing frame, a tint and a big tick; the picture button lights while its panel is open and counts the message's pictures; materials rows, the folded reads row (now with a ›), gate options, the mode switch, the source menu, send and copy buttons, the models page's tabs, the home page's set-up button and expandable activity rows all react clearly to the pointer. The Stage button opens an empty stage that says what will appear there; "Hand to AI · 0" is disabled with nothing picked; move_model's step names the file and where it goes; a one-node canvas change shows one "find on canvas"; the "Goes on from…" chip matches the size around it.
• 右边那一栏的宽度把手更好抓了：卡片左边整条空隙（从上到下）都能拖，把手平时就淡淡地显示，鼠标靠近时整条边亮起；点一下后也能用 ← → 调宽窄。
• The side column's grip catches the pointer across the whole gap left of the card, top to bottom, shows faintly at rest and lights the edge when pointed at; focused, ← → resize it.
• 节点结构的简单看法：除了完整的方块图，现在默认是一张一眼能看懂的说明：每一段采样用哪个模型加哪些 LoRA（带权重）、正负提示词、从空白画布还是从上一段的结果开始（用哪个放大模型、缩放多少），最后保存在哪；和上一段一样的就写“和第 1 段一样”。点任意一项定位到画布上的节点；“简单 / 完整”随时切换，AMB 记得你的选择。
• The run map has a Simple view (now the default): per sampler, the model and LoRAs feeding it, its prompts, where it starts (an empty canvas, a picture, or the part before, upscaled) and where it is saved, "same as part N" when nothing changed; Simple / Full is kept on this browser.
• 删掉任务里的图以后：时间线、本次成果、展台上的那张图显示成“图片已删除”的占位（鼠标停上去说明可在回收站找回），不再是破图，也点不开；@ 列表里不再出现它。AI 那边会收到“你删除了这张图，别再用它”；它要是还去读，会被告知“可能已被用户删除”。
• A task's pictures deleted since show as "Picture deleted" instead of a broken picture, and "@" no longer offers them; the AI hears plainly that the user deleted one, and a tool reading a gone picture says it may have been deleted.
• 关联任务：从一个任务新开的任务会记住它“接着”哪个任务（AI 开的新任务、你看着某个任务时点的“新任务”都会）。标题下面显示“接着：…”“后续：…”，点一下跳过去；任务列表里后续任务带 ↳。可以解开，也可以自己选“接着哪个任务…”。AI 接手后续任务时会先收到上一个任务的决策。
• Tasks remember the task they go on from (an AI's new task while on another, or one the user starts from a task): links under the title lead both ways, the list marks follow-ups, the user can link or unlink, and an AI beginning a follow-up hears what was decided before.
• 决策记录：聊天在哪里都行，结论留在 AMB。AI 用新工具 log_decision 把商量好的结论记成时间线上的 📌 卡片（定了什么、为什么、没选哪些、谁定的、跟哪几步有关）；你在 AMB 里的选择（面板回答、允许或不允许、交出去的素材、AI 改画布的地方、停止）自动算作决策。右上角新增 决策，按时间列出全部决策，点一下回到那一步；AI 很久没记录时会轻轻提醒。AI 重新接手任务时会先收到这些决策。
• Key decisions stay in Anomalous: the AI records what was settled with log_decision (a card on the timeline: what, why, what was left out, who decided, the steps it is about); the user's own choices in Anomalous count as decisions too; a Decisions view in the side column lists them all with a link to each step, notes work done since the last one, and begin_task repeats them to an AI picking the task up.
• 留言框里的 @ 和 /：打 @ 弹出能搜索的列表（这个任务的图、搭配、模型、工作流、提示词、节点包、节点），选中的写成“@名字”并在下面显示成小标签，AI 收到的是确切的那个东西（图片连同生成参数一起）。开头打 / 是快捷指令：停止、撤销、展台、节点结构、素材、副本、当前画布、新任务马上生效；重跑、解释、继续帮你写好一句话。
• "@" in the AI workbench's message box names the task's pictures, combos, models, workflows, prompts, node packs and node types; the AI gets each in full with the message (pictures attached). "/" lists commands that act at once (stop, undo, side views, new task, canvas mode) or fill in words to send.
• 整个自定义节点包交给 AI：素材里点“节点”，再在来源那一排点某个自定义节点包，上面会出现“整包交给 AI”。不知道该选哪个节点时，把整个包给它，AI 自己从包里挑；选“只能用”时，这个包里的节点都允许，别的第三方和付费节点照样拦下。AI 用 get_toolbox 能看到包里有哪些节点，find_nodes 搜索时也把它们排在前面。
• A whole custom node pack can be picked (from its source in the nodes' second column): the AI gets its node types in get_toolbox, finds them first, and with "only" every node of the pack is allowed.
• 小窗里的“来源”不再是一条要横着拖的长条，改成一个“来源：全部 ▾”按钮：点开是一张列表（来源多时上面有搜索框），点一个就选上并收起；Esc 或点别处也收起。
• In the side panel the sources (or types) are one "Source: All ▾" button opening a list to pick from, with a search when it is long, instead of a strip to scroll sideways.
• 素材在右边这一栏（小窗）里时改成一行一个的列表：左边是图标或小封面，中间是名字和来源，右边是勾选；搜索框和“全部 / 已选 / 常用”并成一行。同样的高度能看到的条目多了好几倍；展开成大窗口还是原来的卡片。
• In the side panel the materials picker lists one line per item (badge or small cover, name, where it is from, the tick) and puts the search and the filters on one line; the large window keeps the cards.
• AI 工作台和素材窗口里的滚动条变细，平时看不见，鼠标移到能滚动的那一块上才显出来。
• Scrollbars on the AI page and in the materials picker are thin and show only while the pointer is over what scrolls.
• The side column is a card of its own, narrower by default and resizable from its left edge (kept; double-click resets). The task list folds from a title-bar button and peeks over the page while pointed at. In the docked materials picker, hovering a card no longer opens its detail; its ⓘ does.
• The AI workbench's side column shows one view at a time, switched from the window's title bar: the stage, a run's node map at full size (pick the run), or the materials picker docked there. Pressing the shown one hides the column; each view keeps its state. Each run's step has "See it large". The task's header stays on top while the timeline scrolls.

回答发出去了没有，一眼看清 / See at a glance whether the AI has your answer

• 面板上每个回答、你的留言、批注、发给 AI 的图片下面都有醒目的状态标签：琥珀色 已发送 · 等 AI 来取，AI 拿到后变成绿色 AI 已收到 · 时间（取代原来那行灰色小字）。鼠标停在“已发送”上会提示 AI 停下时该怎么办。
• 面板上的问题可以点 这题跳过：可选的问题不再一直显示“等你回答”，AI 会收到“你跳过了”，自己拿主意。
• Each answer, message, set of marks and picture handed to the AI shows a pill: "Sent · waiting for the AI", then "The AI has it" with the time. Panel questions can be skipped; the AI reads skipped: true.

看得见 AI 跑了什么、AI 改副本、图片直接给 AI / What the AI ran, a copy for the AI, pictures straight to the AI

• 每次运行下面都有 节点结构：一行写明底模、几个 LoRA、尺寸、步数，点开是方块图（LoRA 带权重、正负提示词、采样器设置……，点方块定位到画布上的节点），还用黄色小块标出 比上次改了什么，比如“步数 20 → 30”“LoRA 权重 0.65 → 0.8”“新增「加载LoRA」”。图是 AMB 按 ComfyUI 实际收到的工作流画的，不靠 AI 描述；AI 讲解时用新工具 show_run_map 一次把它放上展台。
• AI 改副本（默认开）：AI 第一次改画布前，AMB 把你正开着的工作流复制到以任务命名的新标签页（副本），之后它只在副本里改和运行；你切去别处，它下次动手前会切回来；你原来的工作流一笔不动。任务标题栏的“AI 改副本 ▾”能换成“改当前画布”，“切到副本”跳过去；关卡窗口里设定新任务的默认做法。
• 图片直接发给 AI：图库右键“发给 AI”或看图窗口的 ✦，可以发给 AI 正在做的任务（也能开新任务），可附一句话。AI 收到图片本身和从图里读出的模型、LoRA、提示词、尺寸、采样设置。AI 搜图库时也能让结果直接带上前 6 张图。
• Each run shows the shape ComfyUI actually ran (boxes, links, settings; boxes locate their nodes) and what changed since the run before; show_run_map puts it on the stage. By default the AI works in a tab of the task's own: the open workflow is copied before its first change and brought back before later ones, the user's tab untouched (per task, and a default in the Gates window). Output pictures can be handed to the task an AI is on (gallery menu, inspector): the AI gets them attached with what made them; search_images(with_pictures) attaches the first six.

运行时看得更清楚，也能一键停 / See more of a run, and stop it in one click

• AI 运行工作流时，进度条旁边多了红色的 ■ 停止（任务工具栏的运行状态按钮旁边也有）：一点就停下这一次运行，排队中的移出队列、正在跑的立刻中断，只动这次运行。时间线上记着“你停止了运行”，AI 也会收到，不会自己再跑一遍；AI 自己也能用新工具 stop_run 停掉自己的运行。
• 进度条下面多了当前节点约剩多久、每步几秒、显存还剩多少，以及 ComfyUI 控制台最新的一行（比如放大模型加载、Model Initializing ...），点开看最近几行。AI 收到的进度里也有这些，不用再去翻控制台。
• 出图一排上队，AMB 就读这次工作流，算出每个采样器要在多大的图上采样。放大模型放大后又整张重新采样、尺寸超出这张显卡的舒适范围、要解码特别大的图时，会出现黄色提醒，说清可能很慢或爆显存，以及怎么改（只想放大就单用放大模型，或先缩小、减步数、换分块解码）。AI 收到同样的提醒，并被要求先问你要不要停。
• 运行状态浮窗开着时按 Esc 只关浮窗，不再把整个 AMB 窗口一起关掉。
• A stop button on a running run (and beside the run-status button): only this run's prompts leave the queue or are interrupted; the step ends "cancelled", the AI hears that the user stopped it, and stop_run lets the AI stop its own. Runs show time left and per step, VRAM, and ComfyUI's console (the bar's note, models loading), for the user and the AI alike. Once queued, each sampler's working size is estimated from the prompt, with warnings for an upscaled picture sampled again, a size past the card or a huge decode. Esc closes the run-status card only.

素材面板的侧边栏小窗 / The materials picker as a side panel

• “素材”默认在屏幕右边打开成侧边栏小窗，不挡工作台和画布：种类是一排标签，卡片两列，点卡片在小窗里看说明。右上角 ⤢ 展开成原来的大窗口，大窗口里也能收回侧栏；AMB 记得你上次用的哪种。
• The materials picker opens as a side panel by default (the page stays usable), and expands to the large window and back; the last choice is remembered.

AI 在画布上搭工作流 / The AI builds workflows on the canvas

• AI 可以在你眼前的画布上搭工作流了：查这台 ComfyUI 有哪些节点（包括自定义节点）、每个节点有哪些输入输出和可选项；新建一个工作流标签页（空白的，或者一次搭好整套并自动排好位置，不动你正在开的那个）；加节点、连线、改参数、删节点。每一步都在画布上当场发生，是一次 Ctrl+Z，记在操作记录里，并在 AI 工作台那一步下面显示改了什么，连线也会写成“某节点的输入 model：（空）→ 加载模型 #4 · MODEL”。连不上的类型、不存在的选项、写错的参数会先被拦下并告诉 AI 正确的写法，画布不会留下半截。新建的标签页名字带时间，不会打开或覆盖你存过的工作流。连线会绕成死循环时也会先拦下。运行被 ComfyUI 拒绝时，AI 收到的是 ComfyUI 给的具体原因（哪个节点、什么问题），不再只有一句“运行失败”。
• The AI can build workflows on the canvas the user sees: look up this ComfyUI's node types (custom nodes too) with their inputs, outputs and choices; open a new tab, blank or built and laid out from a workflow in API form (the open one is left alone); add, wire, set and remove nodes. Each is one Ctrl+Z step, logged as the AI's and shown under its AI workbench step, rewiring included. Wrong types, choices and settings are refused with what fits, before anything changes. Wiring that would loop is refused too, and a run ComfyUI rejects reports ComfyUI's own reason.

局部重绘：涂一涂就交给 AI / Repaint an area: paint it, hand it to the AI

• 标注窗口多了“▧ 涂重绘区”笔刷（粗细可调）：涂上要重画的地方（框、圈、画笔圈的地方也算），选“只改标出的地方”发给 AI，AMB 就在 ComfyUI 的 input 文件夹里做好一张带遮罩的原图（anomalous_inpaint_….png），AI 用 LoadImage 读进来就能局部重绘；AI 同时收到怎么接、重绘强度怎么选的说明。时间线上你的标注下面写着“已做成重绘遮罩”。AI 搭工作流时如果刚生成的文件还不在下拉列表里，AMB 会自动刷新一次列表。
• The marker's repaint brush and "only the marked areas": the picture with the painted area as its alpha goes to ComfyUI's input folder for LoadImage, and the AI is told how to inpaint with it. Canvas tools refresh ComfyUI's lists once when a choice is missing.

AI 的问题在 AMB 里问 / The AI asks in Anomalous

• 你在图上标注或给 AI 留言后，AI 的分析和修改方案也回在 AMB 里：它的留言会写明“回复第几步”，你那条标注下面出现“AI 已回复 · 第几步”，两边点一下就互相跳过去。AI 的留言现在有排版：小标题、列表、加粗、代码 都会正常显示（只认这几种写法，不会运行网页代码）。AMB 告诉 AI：收到标注和留言就在 AMB 里回复，要你选修法就用选择题问。
• The AI replies in Anomalous: report takes reply_to (the user's message or marks), both steps link to each other; its notes are laid out (headings, lists, bold, code) with text only.
• AI 接了 AMB 的任务以后，要你做选择（比如“先测哪种题材？”）不再弹在 AI 软件的对话框里，而是出现在 AI 工作台（展台上有提示，你不在工作台时右下角也会提醒）：几个选项按钮，AI 推荐的那个标着“推荐”，每个选项可以带一句说明；下面还有一行“都不是？写下你的想法”。你点了或写了，AI 在同一步里马上拿到答案，接着干活。AMB 告诉 AI 的说明里写明了：干 AMB 的活时，问题都这样问。
• New ask_user tool: a question with options (one may be marked recommended) and a line for an answer of the user's own, shown in the AI workbench; the AI waits for the answer in the same call. The server instructions tell AIs to ask this way, not in their own chat.

AI 工作台换新外观 / A new look for the AI workbench

• 左边任务栏：“新任务”成了醒目的主按钮，“连接 AI”和“关卡”变成带图标的按钮；下面写着最近连接的是哪个 AI、什么时候；有任务在进行、等你回答或出错时，上面出现对应的筛选按钮和数量；每个任务前的状态换成清楚的图标，选中的任务左边有一条色条。
• 任务标题栏：状态标签带图标；“素材”“全部撤销”“显示展台”是统一样式的按钮，“改名”和“从列表移除”换成铅笔和垃圾桶图标（鼠标移上去有说明，删除悬停时变红）。
• 等 AI 接手的任务：左边一张插图，右边标题；下面是三个带编号、连成一条线的步骤（想让 AI 做什么 → 把这句话发给 AI，大号“复制这句话”按钮 → AI 的每一步会出现在这里，附淡淡的示意）；最下面“第一次用？先把 AI 连上 AMB”是一张带插图的卡片，点“查看连接方法”就在原地展开。没有任何任务时的欢迎页也换成插图加同样的卡片。插图是用 ChatGPT 画的。
• The task list leads with a New task button, icon buttons for Connect an AI and the gates, which AI came last, quick filters by state and state icons; a task's header uses icon buttons; a waiting task shows a picture, three numbered steps and a connect card that opens in place; the empty page matches. Illustrations made with ChatGPT.

给 AI 的素材 / Materials for the AI

• AI 工作台每个任务上多了“素材”按钮：按来源分组列出这台 ComfyUI 的全部节点（自带的、你装的每个第三方节点包、只在网页里的小工具、会花积分的付费 API 节点单独放在最后），还有你的搭配、工作流、提示词和模型。勾选这次想让 AI 用的，点“交给 AI”；可以选“优先用这些”或“只能用这些节点”（自带节点照常可用，AI 想加别的第三方或付费节点会被拦下）。☆ 设为常用后，没挑素材的任务自动带上常用的。AI 下一步就会收到，时间线上也记着你给了什么、AI 收到没有。
• AI 能在新标签页里打开你存的工作流，在它的基础上改；工作流详情页也有“✦ 交给 AI”。AI 查节点时会标出你挑的节点和付费节点。
• 素材窗口重做：大窗口，顶部一个搜索框搜所有素材（按 / 直接搜），左边分类显示各自挑了几个、搜到几个；节点可以按来源筛选，每个第三方节点包有自己的彩色徽标；节点卡片画成一个迷你节点，输入输出的颜色和 ComfyUI 画布上一样，写着有几个参数；搭配、工作流、模型显示封面；提示词显示原文，并且能直接挑搭配和工作流里的提示词。鼠标移到卡片上，右边显示完整说明（节点的每个输入、参数、输出，提示词全文）。下面一排是挑好的素材，带缩略图，可以一个个去掉；“优先用 / 只能用”做成了开关。图标和节点包徽标是矢量画的，跟着深浅色主题变；空白时的工具箱插图，以及没有封面的搭配、工作流、模型用的占位封面，是用 ChatGPT 画的透明背景插图（每张几十 KB）。
• 打开不用等了：节点列表直接用网页里已有的节点信息，其他素材一次读完、提前读好、留一分钟，打开窗口时先显示留着的，再在后台刷新。
• 实测：你装的 10 个第三方节点包（Impact Pack、IPAdapter、Easy-Use、Comfyroll、rgthree、ControlNet 预处理器等）共 1642 种节点，AI 都能放到画布上；用 Comfyroll 节点搭的工作流真实跑出了图。
• A Materials button on each AI task: every node type of this ComfyUI grouped by source (built-in, each custom node pack, page-only helpers, paid API nodes last), plus combos, workflows, prompts and models; picks go to the AI as prefer or only-these-nodes (built-in nodes stay allowed), favourites stand in for tasks without picks. The AI can open a saved workflow in a new tab; a workflow's page can hand it to the AI.

AI 后端修复 / AI backend fixes

• 出图结果按本次提交返回的准确运行 ID 归属，不再通过全局队列差集猜测；其他 AI、其他页面或后续手动排队不会混入。批次只提交成功一部分时，保留这些运行的进度和图片，并明确报告未全部提交；不会自动重跑整批。旧页面需刷新后才能接手出图，不支持请求 ID 的旧版 ComfyUI 前端会在提交前提示更新。同一页面正在提交时另一 AI 收到忙碌提示，待提交结束即可重试。
• Workflow runs now follow exact submission IDs in batch order. Partial submissions keep accepted runs/results and report the missing portion; no automatic resubmission. Stale pages cannot claim runs until refreshed, unsupported host frontends fail before queueing, and a second AI waits while the same page is submitting. Native widget/seed behavior is preserved and later manual batches do not delay the AI receipt.
• “允许一次”只用于一次调用，不会被相同参数的后续调用重复使用；无法建立确认记录时不会直接执行。修复确认超时后用户刚好回答、工具返回又覆盖答案的问题；重启后的待确认步骤仍能把后续答案交给 AI，回答也会唤醒正在等待的 AI。
• 指定任务或会话读取留言时，不再取走同一 AI 客户端其他任务的留言、标记和回答。
• 没找到排队记录、无法读取运行记录、跟踪超时均明确标为未完成，不再误报完成或永久显示运行中；批次内出错及时记录，批次进度保留总数。跟踪停止不会取消 ComfyUI 实际运行。
• 标图的校验和保存移到工作线程；跨盘图片路径返回校验错误，错误类型的 MCP 工具名返回协议错误。
• One-shot approvals cannot be reused, missing approval records fail closed, and timeout/restart races preserve user decisions and wake waiting readers. Task/session input stays isolated. Missing or unreadable run history and monitoring timeouts no longer imply success or leave receipts running forever. Marked-image validation/storage runs off the event loop; malformed image paths and tool names return controlled errors.

AI 工作台与插件能力融合 / Integrated task results

• 工作台的出图结果和 AI 展台图片：鼠标移到图片上，右上角出现 ⓘ 图片与参数、⇄ 和当前画布对比、☆ 保存为（同一个保存菜单存成工作流、搭配和提示词），不再在每张图下面排三个大按钮。左上角可以勾选：勾两张（不同步骤、不同任务都行），页面底部出现“查看参数差异”，用插件现有的参数对比看真实差异；勾第三张会替换最早那张。“本次成果”直接显示缩略图，不用再点开。视频和临时文件没有这些参数操作，缩略图按需加载。
• AI 干活时你在 AMB 里做的改动（画布改动、改模型信息、存搭配和提示词等）会作为“你在 AMB 里改了”出现在同一条时间线上，画布改动附带修改前后；AI 下一次调用任何工具时会收到这条改动，并被告知保留它，不要改回去；时间线上写着 AI 是否已经知道。操作记录页里 AI 做的、以及 AI 干活期间你做的记录，旁边有“AI 任务 →”可以跳到对应任务。
• ComfyUI 在 AI 某一步进行到一半时重启，这一步不再永远显示“正在执行…”，而是标成没有结束。
• AI 软件里直接显示 AMB 卡片（MCP Apps，Claude、ChatGPT、VS Code 等支持的软件）：AI 开始任务、出图、画面板、留言时，对话里会出现一张 AMB 卡片，显示它在做哪几步、最新的几张图，和它问你的问题。在卡片里点选项就算回答，AI 马上收到并接着做。为了安全，“关卡”只能在 AMB 里点允许，卡片里只提示你去 AMB。Gemini 命令行和 Claude Code 终端版不显示卡片，照常用 AMB 的工作台。
• 在图上标给 AI 看：AI 工作台里的图片，鼠标移上去多了一个 ✎。点开后可以用框、圈、箭头、画笔（红黄青三种颜色）在图上标出要改的地方，每个标记写一句“这里要怎么改”，再写整体的想法，点“发给 AI”。AI 下一步就会收到这些说明和这张标好的图（能看图的 AI 能直接看到标在哪），时间线上也会显示你标的图和说明。标好的图存在 AMB 自己的记录里，不会混进图库。
• 运行进度 AI 也看得见：AI 让工作流开始运行后，时间线上那一步下面有进度条：排队时写“前面还有几个”，运行时写正在跑哪个节点、第几步（比如 采样器 14/28）和节点做完了几个。AI 多了一个 run_status 工具，可以随时问进度，也可以等到跑完（每次最多等 55 秒），跑完直接拿到出的图；运行的回复里也会提醒它用这个，别凭时间猜有没有跑完。进度条右边显示百分比（和 ComfyUI 浏览器标签上的一样：正在跑的节点走到第几步）。AI 不用自己去问：运行期间它调用任何工具，回复里都附带一句当前进度（比如“采样器 14/28，50%，节点 3/9，已经 42 秒，还没跑完”），跑完后的下一次回复会告诉它跑完了、出了哪几张图，只说一次。
• 提示词改动一眼看清：AI（或你自己）改了提示词时，记录里先列出“去掉 N 个”（红色划掉）和“加上 N 个”（绿色）的词，点“看词义”用 AMB 的翻译在每个词下面显示中文意思；“看完整的修改前 / 后”里，两段全文中变化的词也标了颜色，没变的变暗。其他文字改动会把不同的那一段标出来。操作记录里提示词改动最多保留 2000 字（以前 400 字），整段提示词都能对比，不会把被截断的半个词算成改动。
• 等 AI 接手的任务重新排版：中间一个输入框写想让 AI 做什么（就是任务名），下面是要发给 AI 的那句话和“复制这句话”，再加三个小步骤；第一次连接 AI 的说明收进“AI 还没连上 AMB？”里，不再整屏铺开。等 AI 的时候不显示留言框。
• 交给 AI：模型页、图片详情（右上角 ✦）和搭配卡片上多了“交给 AI”。写一句想让 AI 做什么（已经帮你填好，比如“用这个模型出一张图”），AMB 会建好一个带着这个模型、图片或搭配的任务，再给你一句话贴给 AI；AI 接手时会直接拿到它们，不用再自己去找。
• 斜杠命令（MCP prompts）：在 Claude 等软件里可以直接选“在工作台开一个任务”“出几张让我挑”“补齐缺的模型”“讲讲现在的画布”，AI 会按适合在工作台里看的方式去做。
• 流程图：任务标题下面一排“模型 › 搭配 › 提示词 › 运行 › 挑图”，按 AI 做事的先后排列，每段标着做完了、正在做、等你点头或出了问题；点一下跳到时间线上那一步。
• 关卡：任务列表上方的“关卡”里，改画布、用显卡出图配音、下载模型、联网扫描四类事，各自可以选“直接做 / 先问我 / 不许做”（默认只有下载要先问）。选了“先问我”，AI 的这一步会停在时间线上，列出它要做什么，等你点“允许”或“不要”（可以写上原因或想怎么改，AI 会收到）；勾上“以后直接做”就不再问。你不在工作台时右下角会提示“AI 在问你”。AI 等不及先走了也没关系：你回答后它下一步会收到，允许的话再做一次，不会问第二遍。
• 画布操作直接显示记录中的“修改前 / 修改后”、新增或删除的节点；与操作记录页共用展示，保留定位和适用时的撤销。旧记录缺失时明确提示，不用当前画布猜历史。任务页更新隔离迟到请求，并整理了结果卡片、时间线和窄布局。
• 修复参数对比窗口遮挡撤销提示的问题，撤销按钮现在可直接点击。
• Task outputs and AI panel pictures now use the existing image inspector, Keep menu and image/canvas parameter comparison, as tools over each picture on hover; tick any two pictures (any step or task) and compare their settings from the bar at the foot of the page. The results strip shows thumbnails straight away. Changes you make in Anomalous while an AI works join its timeline, and the AI is told of them (to keep, not undo) with its next call; activity entries link to their AI task. Steps a restart interrupted are marked as not finished instead of running forever. In apps that support MCP Apps (Claude, ChatGPT, VS Code…) an Anomalous card in the conversation shows the task's steps, latest pictures and the AI's questions, answered with a click (gates are only allowed in Anomalous). Four ready-made prompts appear as slash commands. "Hand to AI" on a model's page, an image's inspector and a combo card starts a task carrying it, with a sentence to send the AI. A flow over each task shows its stages (models, combo, prompt, run, pictures) in order with their state. Gates: changing the canvas, using the GPU, downloading and scanning online can each go ahead, ask you first or never happen (downloads ask by default); a step that asks waits on the timeline with what it would do, for Allow or No (with your reason for the AI). Canvas changes share recorded before/after receipts with the activity page. Unavailable history is explicit; stale task responses cannot replace a newer selection. Refined result cards and timeline styling, and fixed Undo toasts being blocked by the comparison overlay.

Fixes / 修复

• AI 工作台与 MCP 调用实时生命周期及消息送达：每个工具调用开始时立即写入“正在执行…”步骤并接手等待中的任务，完成后原位更新同一步骤（不再等到耗时操作结束后才显示，也不会重复生成步骤）；用户在工作台留下的消息和展台面板回答，除了可通过 get_answers 读取外，还会在 AI 下一次成功调用任意工具时自动附带送达一次，并立即在界面上更新为“AI 已经看到”；多会话与显式 task ID 互不串任务。
• Live in-flight MCP step lifecycle and automatic delivery of user messages and panel answers: every tool call now starts a "running" step immediately and updates the same step in place on completion or failure; unread user messages and panel answers are piggybacked once onto the next successful MCP tool response (in addition to get_answers) and immediately notify the workbench UI; concurrent sessions and explicit task ids stay isolated.
• 展台内容按出现顺序保留：update_panel 新增的成图、说明等接在已有内容下面，即使 AI 把整张看板按“结果在前”重发，也不会倒置旧内容。省略旧内容不再将它删除；重复提交不重复追加。明确指定积木 ID 时仍可原位更新进度、问题或修正内容，已有回答保留。超出 30 块时提示新开面板，不截掉旧内容。旧记录保留其已保存的排版。
• Panel updates append new output after what the user already saw, preserve omitted blocks, and avoid duplicating unchanged resubmissions. Stable block ids allow deliberate in-place edits; invalid or over-limit updates leave the panel untouched. Historical records keep their saved layout.
• 展台右上角增加“关闭展台”，关闭后时间线展开；任务标题旁的“打开展台”或时间线里的面板入口可重新打开，保留面板和未提交的输入。当前任务更新不会强行重开。面板编号改为“来自第 N 步”，明确它的创建位置。The stage can be closed and reopened without losing panels or drafts; step labels now say “From step N”.
• AI 工作台：AI 先查询再调用 begin_task，现在会继续已接手的任务，保留用户标题和之前的步骤；重复调用不会拆任务。另一件事可用 new_task=true 明确新建，已结束或长时间闲置的任务仍分开记录。
• 展台打开历史任务时定位到底部；新增面板、更新最新面板及图片加载后继续跟随底部。往上翻看时保留位置，用“查看最新”回到底部，时间线的定位链接仍跳到对应面板开头。
• The workbench keeps preliminary reads and repeated begin_task calls in the active task; new_task=true explicitly starts separate work. The stage opens at the bottom, follows updates and loaded images while there, and preserves the reader's position when browsing earlier content.

✨ AI workbench: see everything the AI did, open it, undo it (AI 工作台)

• A new page on the rail, AI workbench: every step an AI app (Claude Code, Cursor, Cherry Studio… connected through Anomalous's MCP) takes is kept, grouped into tasks. The AI names a task when it starts (a new begin_task tool); otherwise one app's calls that follow each other make one task.
• A task is a timeline: what it looked up (folded into one line, a click shows each), what it changed on the canvas, the workflow runs, downloads, scans and spoken lines, and the AI's own notes (a new report tool, also for pictures it made another way). Each step says when, how long it took, and why it failed when it did. Steps are numbered in order (1, 2, 3… down the line), each dot in its kind's colour; while the AI works, the next number waits at the foot.
• What it made stays, and leads there: the pictures a run made are kept with the task (watched until the run ends) and shown on top as What it made; a click opens the picture, a model opens its page, a combo opens the combo, a spoken line plays, a canvas change is found on the canvas or undone (also all of a task's canvas changes at once, stopping at anything you changed since).
• Live: a task the AI is still working on grows by itself and says The AI is working…; the page follows the newest task unless you picked another. While you are elsewhere, the rail's AI workbench gets a dot. Home gets What the AI did lately.
• New task (新任务): start a task yourself (with what the AI should do, or empty): it waits for an AI (Waiting for an AI), and the next AI app that calls takes it over, so its work starts there instead of joining its last task.
• Connect an AI (连接 AI): always beside the task list, and on an empty page or a waiting task: Anomalous's address, a sentence to copy to your AI so it adds the server itself (and starts on the waiting task), and the way for Claude Code, Gemini CLI, Cursor, VS Code and Cherry Studio.
• Write to the AI from the workbench (给 AI 留言): a box under a task's timeline. The message becomes a step of the task (You) and reaches the AI the next time it asks (get_answers, which now also brings the user's messages, and with no panel named waits for one); it says whether the AI has seen it. Anomalous cannot wake the AI: when it has stopped, the box says to ask in the AI app.
• The task list is grouped by day (today, yesterday, a date), each task with an icon for where it stands: working ●, waiting for you ?, a step failed !, done ✓.
• The stage (展台): the AI's panels are shown in a column on the right, beside the timeline, so they stay in view however many steps come after them; the timeline keeps one line per panel (… on the stage ▸), which brings the stage to it. Like the timeline it reads from the top down: the task's panels in the order they came, each with its step number, the newest at the bottom; a new one scrolls the stage down to it while you are at the bottom, otherwise ↓ New panel says it came. The strip of what a task made also runs oldest to newest, left to right. In a window too narrow for three columns, the timeline and the stage are two tabs (the stage tab lights up when a question waits, and comes forward when one arrives). The task list can be folded away (‹) for more room.
• While the AI works, only the steps that changed are drawn again: a diagram or a strip of pictures scrolled sideways no longer jumps back, and text half-typed into a panel or the message box stays.
• Diagrams read in order: a box that only feeds one further on (a prompt into the sampler) stands right before it instead of at the far left with an arrow across everything, and boxes in a column are sorted by what feeds them, so arrows cross less.
• Tasks can be renamed or removed from the list (only the record goes: pictures and the canvas stay). The newest 300 are kept, in ComfyUI's user folder. Without any AI connected, the page shows the address to add in the AI app.
• Only what goes through Anomalous's MCP is seen; the AI can report the rest. The rail now has ten entries: the number keys 1–9 open the first nine.

🧩 The AI draws panels, and asks you on them (AI 画面板，在面板上问你)

• A new show_panel tool: the AI lays out a panel in its task, in Anomalous's own style, from fixed blocks: headings and text (also as a tip or a warning), pictures (a row, a grid, or two side by side to compare), your models as cards with their covers (a click opens them), tables, lists, settings, a progress bar.
• It can ask: buttons (one pick or several) and a line to type into. Your answer is handed over and waits for the AI, which reads it with the new get_answers (it can wait up to a minute for you); each question says whether the AI has seen your answer yet. update_panel changes a panel (a progress bar going up) or closes its questions.
• Safe by design: nothing the AI sends is run or read as markup. Every block is checked on the server (only these kinds, plain text of bounded length, pictures only from ComfyUI's output folder, models only from your folders) and written into the page as text.
• When the AI shows a panel while the browser is elsewhere, a small card in the corner says so (The AI asks you: …, Have a look); a task with a question waiting says Waiting for you in the list.

🎞️ The AI's stage: marks on pictures, before / after, picks, diagrams (AI 展台)

• Marks on a picture: the AI circles or boxes places on a picture it made (① the fingers are fixed, ② softer light); each mark has its number, the labels are listed under the picture, and a click on one zooms the picture there (again: back).
• Before / after slider: two pictures over each other, a line to drag across to compare the details.
• Picks: pictures side by side, each with the AI's reason and Take this one (the choice goes to the AI); the AI can mark the one it recommends. Each can also be put on the canvas (its saved workflow opens in a new tab, as dropping the picture on ComfyUI does) or dragged onto the canvas.
• Diagrams: boxes and arrows, laid out by Anomalous (in columns or rows), with a label on an arrow; a box naming a canvas node (◎) finds it on the canvas.
• The AI only sends data (places as fractions of the picture, what connects to what); every shape and line is drawn by Anomalous, so it looks the same whichever AI is connected, and no drawing code of the AI's is taken.