dsh-smooth-stream 0.3.1 → 0.3.3

package/README.en.md

@@ -14,15 +14,17 @@ Left: default Web UI. Right: dsh-smooth-stream.
 
 ![Left: without the plugin. Right: with dsh-smooth-stream.](docs/compare.gif)
 
-## What it does
-
-- **Reveal tracks the model.** Assistant text appears at a cadence that follows the arrival rate. Fast bursts do not dump a whole paragraph; a slow stream does not sit still and then jump.
-- **Markdown stays markdown.** Code, emphasis, and the rest render while the reply is still coming. There is no plain-text tail that later swaps into formatted markdown.
-- **Wraps glide in.** A new line or a growing tool card eases into view instead of snapping the transcript up by a line.
-- **You keep the scroll.** Scroll up to read earlier text and the overlay lets go. Follow resumes only when you return to the bottom — the to-bottom button counts.
-- **Think stays the built-in row.** Reasoning uses the usual disclosure. It opens while thinking is the live tail and closes when thinking ends; the chevron still toggles by hand.
-- **The rest of the turn moves with it.** Running tool cards, model retries, and workflow runs share the same follow, so the whole turn slides instead of only the assistant text.
-- **It backs off when it should.** `prefers-reduced-motion` shows the finished text at once and does not take follow. If the frame rate drops below 30 fps and the reply is off-screen, reveal pauses and catches up when the view is healthy again.
+## Current behavior
+
+- **The whole Agent turn uses one extensible pipeline.** Assistant text, Think, Context, Retry, Command, Bash, Glob, Read, tool calls, and newly registered renderers all enter progressive reveal and bottom-follow through the same boundary, without a tool-name allowlist.
+- **Reveal speed adapts to queue pressure.** Small updates keep a soft cadence while large or fast bursts catch up promptly. Once the producer completes, remaining source is committed immediately instead of continuing to type long after the Agent has stopped.
+- **Markdown remains mounted throughout streaming.** Code blocks, tables, emphasis, and other formatting do not begin as plain text and later swap trees. Historical messages also do not replay their reveal animation when remounted.
+- **Scroll room opens only when a wrap is actually likely.** The predictor combines buffered source with the current line's remaining width before opening its runway. Long replies still absorb line wraps smoothly, while a short same-line answer after Think does not pre-scroll and rebound.
+- **Conversation chrome stays fixed.** `Deep diving...`, the composer, and the to-bottom button never ride the message transform. Fast output and low-frame-rate catch-up cannot paint through the status row or disappear behind the composer.
+- **One continuous spring owns motion.** The engine carries velocity and displacement between frames instead of repeatedly starting native smooth-scroll calls. Line wraps, code, tables, and growing tool rows converge along the same trajectory; completion lands on the natural floor and retires temporary state without a flash or overshoot rebound.
+- **Reader input wins immediately.** A small upward wheel, touch, or keyboard gesture releases automatic follow. Ownership returns only after the reader actually reaches the bottom again.
+- **Think respects the user's preference.** With auto-expand enabled it keeps the Harness disclosure interaction and collapses when thinking ends. With it disabled, collapsed reasoning can keep updating without fake height motion, and a manual toggle is not wrestled back by stream state.
+- **Performance guards preserve the final position.** `prefers-reduced-motion` shows complete content without taking follow. Off-screen DOM commits pause under low FPS, then catch up under control and still finish exactly at the bottom.
 
 ## Install
 
@@ -60,7 +62,7 @@ The bundle installs with `preset: balanced`. Change it in the profile `cordis.pa
 | `balanced` | Default |
 | `silky` | More buffer, slower catch-up |
 
-`maxScrollSpeedPxPerSec` (default `1000`) is a ceiling so the first large lag does not teleport.
+Legacy `mode`, `revealCharsPerSec`, `scrollSpeedPxPerSec`, and `maxScrollSpeedPxPerSec` fields are still accepted so existing profiles keep loading; the current adaptive engine uses only `preset` to tune cadence.
 
 ## User settings
 

package/README.md

@@ -14,15 +14,17 @@
 
 ![左：未使用插件。右：使用 dsh-smooth-stream。](docs/compare.gif)
 
-## 它做什么
-
-- **揭示跟着模型走。** 助手文本按到达速率出现。突发不会整段倒出来，慢流也不会停住再猛地补上。
-- **一直是 Markdown。** 代码、强调等在流式过程中就按格式渲染，没有先出纯文本再换成排版的交接。
-- **换行是滑进来的。** 新的一行或正在变高的工具卡片会滑入视野，而不是把整段记录往上顶一格。
-- **滚动条归你。** 往上翻看前文时 overlay 会松手。只有回到底部才会继续跟随——点「回到底部」也算。
-- **思考仍是内置那一行。** 推理用原来的 disclosure。它是当前流式尾部时展开，思考一结束就收起；箭头仍可手动开关。
-- **整轮一起动。** 运行中的工具卡片、模型重试、workflow run 共用同一套跟随，所以滑的是整轮回复，不只是助手正文。
-- **该停的时候会停。** `prefers-reduced-motion` 会直接给出全文，也不接管跟随。帧率低于 30 fps 且回复在屏外时，揭示会暂停，画面恢复后再补上。
+## 当前实现
+
+- **整轮 Agent 输出统一接管。** 助手正文、Think、Context、Retry、Command、Bash、Glob、Read、工具调用以及后续注册的新渲染器，都通过同一个可扩展入口进入渐进揭示和底部跟随，不依赖工具名称白名单。
+- **吐字速度会跟随积压自动变速。** 小批内容保持柔和节拍，大批或高速到达时及时追赶；完成信号到达后立即提交剩余正文，不让 Agent 已结束而文字还长时间继续输出。
+- **流式过程始终使用 Markdown 渲染。** 代码块、表格、强调等不会先显示成纯文本再整体换树，历史消息也不会在重新挂载时重播动画。
+- **换行前才准备滚动空间。** 引擎结合待揭示内容和当前行剩余宽度判断是否可能增高，只在真实换行风险出现时打开预测 runway。长回复仍能平滑吸收换行，Think 后的短同一行正文不会多滚再回弹。
+- **状态区域保持稳定。** `Deep diving...`、输入框和「滚动到底部」按钮不参与消息 transform；高速输出或低帧率下，正文也不会越过状态区域或藏到输入框后面。
+- **滚动使用持续弹簧而不是反复启动原生平滑滚动。** 每帧保留速度和位移状态，换行、代码块、表格及工具卡片增高都沿同一轨迹收敛；结束时落在自然底部并安静撤销临时状态，不闪烁、不越位回弹。
+- **用户输入拥有最高优先级。** 向上滚轮、触控拖动或键盘滚动会在轻微手势时立即解除自动跟随；只有用户真正回到底部后才重新接管。
+- **Think 尊重用户设置。** 自动展开开启时沿用 Harness 的 disclosure 交互并在思考结束后收起；关闭时折叠内容持续更新但不引发虚假高度和上下闪动，手动展开也不会被流状态抢回。
+- **性能保护不会破坏最终状态。** `prefers-reduced-motion` 直接显示完整内容且不接管跟随；低帧率且回复在屏外时暂停 DOM 提交，恢复后受控追赶，最终仍准确停在底部。
 
 ## 安装
 
@@ -60,7 +62,7 @@ Host 日志里应出现 `[dsh-smooth-stream] plugin loaded!`。
 | `balanced` | 默认 |
 | `silky` | 缓冲更大，追上更慢 |
 
-`maxScrollSpeedPxPerSec`（默认 `1000`）是速度上限，避免第一次滞后过大时瞬移。
+旧版的 `mode`、`revealCharsPerSec`、`scrollSpeedPxPerSec` 和 `maxScrollSpeedPxPerSec` 字段仍可被加载，以兼容已有 profile；当前自适应引擎仅使用 `preset` 调整节拍。
 
 ## 用户设置