dsh-recent-usage-metrics 0.2.2 → 0.3.0

package/CHANGELOG.md

@@ -1,5 +1,30 @@
 # 更新记录
 
+## 0.3.0
+
+状态从三种扩到四种，加入上下文压缩的显示，并给抽屉补上悬浮提示。
+
+- **压缩中**：状态栏显示图标、开始时刻、「压缩上下文中」与实时秒数；抽屉第 0 行同步计时。
+  图标闪动，速度列写「压缩上下文」，首字列固定为破折号。
+- **压缩完成**：图标停止闪动，文案改为「压缩完成」，保持到下一条请求发出。该行进入
+  抽屉 #1，补上输出 Token 与折算速度（output ÷ wall），但**不计入底部平均值**。
+- 压缩没有首字与生成速度，这两项数据源里本就不存在；总用时取 wall，是日志事实，
+  刷新后不变。
+- 压缩用哪个模型：优先显式配置，否则跟随最后一次请求（复用 KV 缓存）。压缩结束后
+  compaction/summary 会记录真实模型并显示；压缩期间尚无此记录，显示对话模型名。
+- **悬浮提示**：时间列显示完整时间戳，速度列显示模型，输出 Token 列显示输入 token
+  与缓存命中率；七个表头各自说明该列含义，# 列的说明直接画出四种状态的图标。
+- 提示用自建浮层而非 title，因为 title 只能显示纯文本，而 # 列的说明需要图标。
+- 压缩阶段的状态栏预留空闲态宽度，切换时整体尺寸与图标位置都不变。
+
+## 0.2.3
+
+- 中断的行现在显示它跑了多久：被手动停止的请求没有 model，总用时回退到 wall
+  （endTime - startTime），刷新后仍是同一个数。
+- 「中断」标记移到速度列。首字若已返回就照常显示实测值，总用时也照常显示——
+  缺的只是生成速度，由速度列说明原因。
+- 中断行不计入底部平均值，但标题仍数满十行。
+
 ## 0.2.2
 
 - 修复：切换模型后抽屉卡在旧的一轮。数据源的 target 曾被缓存，而会话重建

package/README.en.md

@@ -95,75 +95,67 @@ Two further details make the edges actually coincide:
 Click the pill to open the drawer; click it again, or press Escape, or click outside, to
 close it. There is no hover-open and no tooltip.
 
-## Three states
+## Four states
 
-This plugin has two kinds of busy and one kind of idle, and all three look different:
+The plugin sorts "what is happening right now" into four phases, each with its own glyph and
+wording:
 
-| State | Decided by | Pill icon | Drawer header | Live row |
-| --- | --- | --- | --- | --- |
-| **Model writing** | `model` is `null` | ⚡ pulsing bolt | `Streaming` | yes (11 rows) |
-| **Tools running** | `model` set, `endTime` `null` | **pulsing T** | **`Calling tools`** | yes (11 rows) |
-| **Idle** | `endTime` set, or no running step | still | `last N` | no (10 rows) |
+| State | Decided by | Pill icon | Drawer header |
+| --- | --- | --- | --- |
+| **Model writing** | `model` is `null` | ⚡ pulsing | `Streaming` |
+| **Tools running** | `model` set, `endTime` `null` | T pulsing | `Calling tools` |
+| **Compacting** | newest compaction `running` | ≡ pulsing | `Compacting context` |
+| **Compacted** | compaction finished, no request since | ≡ still | `Context compacted` |
+| **Idle** | otherwise | ⚡ still | `Last N` |
 
-### Why the tool phase is its own state
+When a compaction finishes the pill says "Context compacted" and stops pulsing, and it stays
+that way until the next request starts — at which point the writing phase takes over again.
 
-This plugin measures **model output**, and `step/end` only arrives after the step's tool calls
-have all returned. Keying off `running` (which `step/end` clears) therefore kept the plugin in
-"streaming" for the entire tool phase after the reply had already finished: the bolt kept
-pulsing and the live row kept occupying an eleventh row.
+### Compaction is its own activity
 
-The field that separates the first two is `model`, which is `messageTime − startTime`:
+It belongs to no step: it is opened by `compaction/start` and closed by `compaction/end`, it
+lives on the turn's `compactions[]`, it runs between two steps, and it can take tens of
+seconds. No step is running during that time, so the pill used to keep showing the previous
+step's figures and the pause looked like a hang.
 
-- the model is still writing → `null`
-- `assistant/message` has landed → a fixed value, **even while tools run**
+It has no first token, no generation rate and no streamed token count — all three belong to a
+settled reply. So the drawer's columns are filled like this:
 
-`endTime` then separates the tool phase from idle. Both fields come from the Host's
-`sessionStats` fold — the same data the Statistics page renders.
-
-### The live row stays through the tool phase
-
-The request is genuinely still in flight, so it keeps its row; only the marker changes from the
-bolt to a T, and the drawer header switches from "Streaming" to "Calling tools". That makes all
-three states distinguishable at a glance.
+| Column | While compacting | Once compacted |
+| --- | --- | --- |
+| First token | `—` | `—` |
+| Speed | `Compaction` | `Compaction` |
+| Tokens | `—` | the real figure (e.g. `4.9k tok`) |
+| Time | counts up live | frozen at `wall` |
+| Amortised | `—` | `output ÷ wall` |
 
-Both glyphs share **the same 14×14 svg box** (`.dsh-rum-trigger svg`) and the same `pulse`
-animation. Switching changes the shape and colour inside the box, never the box, so the status
-row does not shift. The T is drawn narrower on purpose so it does not read heavier than the bolt.
+The total is `wall` (`endTime − startTime`): a log fact, so a reload shows the same span.
+A compaction row is excluded from the footer average — its total is wall-clock time, and
+folding it into the mean would pull it toward a number describing neither case.
 
-### Fallback
+### Which model compacts
 
-A composition without the statistics fold has no `model` to read. There, `partial` and a
-Trajectory request can only mean the model is writing, so the state is always "model writing" —
-it never degrades to idle just because the fold is missing.
+The summarizer picks, in order: an explicitly configured `summarizationProvider/Model` → **the
+model the last request used** → the Agent's own model. The default is the second, because
+replaying the whole conversation is only cheap when it hits the same provider's KV cache.
 
-**Drawer** — slides up from the status row, right-aligned to that pill, `position: fixed`
-on `document.body` so no scroll container can clip it. Its material is the shipped
-statistics dialog's own, copied token for token: the `--dsw-specific-menu` fill
-(`#303136f0` in dark mode, i.e. nearly opaque, the same surface behind the native
-numbers), the `--dsw-menu-backdrop-filter` 40px/150% blur, `--dsw-elevation-prominent`
-stroke and shadow, `--dsw-radius-lg`, and 16px padding. It lists the last ten requests,
-newest first, with no explanatory footer:
+`compaction/summary` records the model actually used, so **after** a compaction the real name
+is shown; **while** one runs the log has no such record yet, so the chat model is named, and
+the generic wording is the last resort.
 
-| # | At | TTFT | Speed | Tokens | Time | Amortised |
-| --- | --- | --- | --- | --- | --- | --- |
-| 1 | 18:42:07 | 2.3 s | 445 tok/s | 1024 tok | 4.6 s | 223 tok/s |
+### Hover hints
 
-The `At` column is when the request was **issued** (`step/start`), matching the Statistics
-view. It is not the end time: a row stamped 18:42:07 with a 4.6 s total actually finished
-at 18:42:12, so an end stamp would read as "it finished then" and would also order the list
-by completion rather than by issue.
+Every drawer column and every header carries a hint:
 
-A new request slides in at the top; the row that falls off the ten-row cap is removed in
-the **same frame** rather than fading out on its own. An exit phase would keep an
-eleventh row laid out for the animation's duration, growing the drawer by a row and
-shrinking it back — a visible height jump on every insertion. Opening the drawer also
-establishes a baseline instead of animating, so the whole list never slides in at once.
+| Where | Content |
+| --- | --- |
+| Clock column | the full timestamp `11-06 10:47:41.828` |
+| Speed column | the model that served that row |
+| Tokens column | input tokens and cache hit rate |
+| Headers | what the column measures (the `#` header draws the four state glyphs) |
 
-While a step is streaming, the drawer shows it as a live row above the history, and the
-status reading switches to a live style. Under a rule below the history sits a summary
-row — the average of the visible requests, in the same grid so each mean lands under its
-column. A column with no samples (say no TTFT on an interrupted request) prints a dash
-rather than a misleading zero.
+The hints are a real overlay rather than `title`: `title` can only show plain text, and the
+`#` header has to draw glyphs.
 
 ## How the numbers are measured
 
@@ -412,6 +404,25 @@ carries `ttft`, `model`, `decode`, `tps`, `output` and `wall`.
 | **Time** | **`model`** | **model time, not the Statistics view's total** |
 | Amortised | `output ÷ model` | this plugin's own figure |
 
+**Interrupted rows:**
+
+A request that was stopped never produced an `assistant/message`, so `model`, `decode` and
+`tps` are all null. That row's **total falls back to `wall`** (`endTime − startTime`) — how
+long the request actually ran before it was stopped. It is a log fact, so it survives a reload
+unchanged, which is exactly what that column is for here.
+
+**The first-token column falls back to `wall` too**, in two shapes: if the upstream never returned a
+token, the first-token figure is the wait itself and equals the total; if a token had arrived,
+`ttft` is already fixed and the row keeps that measurement, so the first-token figure is smaller
+than the total — which is how a reader tells "it started answering and was cut off" from "it never
+got to answer".
+
+The word **"interrupted" sits in the speed column**, not the first-token column: both the
+first-token figure and the total are real readings, and the only missing figure is the generation
+rate, so that is the column that says why. Interrupted rows are
+excluded from the footer average — their total is wall-clock time, and folding it into the mean
+would pull it toward a number describing neither case — but the caption still counts all ten.
+
 **Why the total is `model` and not `wall`:**
 
 The Statistics view's total column is `wall = endTime − startTime`, which **includes the step's

package/README.md

@@ -85,68 +85,60 @@ DSH 的极客级请求指标。它在 **设置 → 通用 → 性能与用量**
 
 点击气泡打开抽屉；再次点击、按 Esc、或点击外部即可关闭。没有悬停打开，也没有悬停提示。
 
-## 三种状态
+## 四种状态
 
-插件只该有两种「忙」和一种「闲」，三者的判定与呈现都不同：
+插件把「现在在干什么」分成四段，每段有自己的图标与文案：
 
-| 状态 | 判定 | 状态栏图标 | 抽屉右上角 | 实时行 |
-| --- | --- | --- | --- | --- |
-| **模型输出中** | `model` 为 `null` | ⚡ 闪电闪动 | `生成中` | 有（共 11 行） |
-| **工具调用中** | `model` 有值、`endTime` 为 `null` | **T 闪动** | **`调用工具中`** | 有（共 11 行） |
-| **空闲** | `endTime` 有值，或没有运行中的步骤 | 静止 | `最近 N 条` | 无（共 10 行） |
+| 状态 | 判定 | 状态栏图标 | 抽屉右上角 |
+| --- | --- | --- | --- |
+| **模型输出中** | `model` 为 `null` | ⚡ 脉冲 | `生成中` |
+| **工具调用中** | `model` 有值、`endTime` 为 `null` | T 脉冲 | `调用工具中` |
+| **上下文压缩中** | 最近一次压缩 `running` | ≡ 脉冲 | `上下文压缩中` |
+| **上下文压缩完成** | 压缩已结束、之后没有新请求 | ≡ 静止 | `上下文压缩完成` |
+| **空闲** | 其余 | ⚡ 静止 | `最近 N 条` |
 
-### 为什么要分「工具调用中」
+压缩完成后状态栏显示「压缩完成」并停止闪动，直到下一条请求发出——那时自动回到模型输出状态。
 
-本插件统计的是**模型输出性能**，而 `step/end` 要等这一步的工具调用全部跑完才到。
-若只看 `running`（它由 `step/end` 清零），模型回复早已结束、工具还在跑的整段时间里，
-插件都会停在「生成中」：闪电一直闪、实时行一直占着第 11 行。
+### 压缩是一种独立的活动
 
-分隔前两者的字段是 `model`，也就是 `messageTime − startTime`：
+压缩不属于任何一步：它由 `compaction/start` 与 `compaction/end` 界定，挂在回合的
+`compactions[]` 上，跑在两步之间，可以持续几十秒。在这段时间里没有步骤在跑，
+所以之前只能沿用上一步的数字，看起来像是卡住了。
 
-- 模型还在写 → `null`
-- `assistant/message` 落地 → 定值，**即使工具还在跑**
+它没有首字、没有生成速度，也没有流式 token —— 那三项都依附于一条已结算的回复。
+因此抽屉里那几列，压缩行是这样处理的：
 
-再用 `endTime` 把工具阶段与空闲分开。这两个字段都来自宿主的 `sessionStats` 折叠，
-与统计页显示的是同一份数据。
-
-### 工具阶段仍保留实时行
-
-这一阶段请求确实还在飞行，所以它继续占一行；只是标记从闪电换成 T，
-抽屉标题也从「生成中」改成「调用工具中」。三个状态因此一眼可分。
+| 列 | 压缩期间 | 压缩完成后 |
+| --- | --- | --- |
+| 首字 | `—` | `—` |
+| 速度 | `压缩上下文` | `压缩上下文` |
+| 输出 Token | `—` | 实际值（如 `4.9k tok`） |
+| 总用时 | 实时递增 | 冻结为 `wall` |
+| 折算速度 | `—` | `output ÷ wall` |
 
-两种图标共用**同一个 14×14 的 svg 盒**（`.dsh-rum-trigger svg`）与同一段 `pulse` 动画。
-切换时改变的是盒内的字形与颜色，盒子本身不动，所以状态栏不会位移。
-T 的字形特意收窄，避免比闪电看起来更重。
+总用时取 `wall`（`endTime − startTime`）：这是日志事实，刷新后仍是同一个数。
+压缩行不计入底部平均值——它的总用时是墙钟时间，混进均值会拉偏。
 
-### 降级
+### 压缩用哪个模型
 
-拿不到统计折叠时的组成（`uiConversation` 缺失）没有 `model` 可读，此时 `partial`
-与轨迹域的请求都只可能表示模型正在输出，因此一律按「模型输出中」处理——
-不会因为缺少折叠而误判为空闲。
+压缩的模型按顺序取：显式配置的 `summarizationProvider/Model` → **最后一次请求实际用的模型**
+→ Agent 自身的模型。默认走第二条，因为重放整段对话时复用同一个模型才能命中 KV 缓存。
 
-**抽屉**——从状态行向上滑出，右边缘与该气泡对齐，`position: fixed` 挂在
-`document.body` 上，因此不会被任何滚动容器裁剪。它的材质逐字复刻了原生统计弹窗：
-`--dsw-specific-menu` 底色（深色模式下是 `#303136f0`，即近乎不透明，正是原生数字背后的那层）、
-`--dsw-menu-backdrop-filter` 的 40px/150% 模糊、`--dsw-elevation-prominent` 描边与阴影、
-`--dsw-radius-lg`、以及 16px 内边距。它按最新在上的顺序列出最近十次请求，底部不加说明文字：
+`compaction/summary` 会记录实际用的模型，所以**压缩结束后**显示的是真实模型名；
+**压缩期间**日志里还没有这个信息，此时显示对话模型名，拿不到则写「使用对话模型」。
 
-| # | 时间 | 首字 | 速度 | 输出 Token | 总用时 | 折算速度 |
-| --- | --- | --- | --- | --- | --- | --- |
-| 1 | 18:42:07 | 2.3 s | 445 tok/s | 1024 tok | 4.6 s | 223 tok/s |
+### 悬浮提示
 
-「时间」列取请求**发起**的时刻（`step/start`），与统计页一致。它不取结束时刻：
-否则一行写着 18:42:07、总用时 4.6 秒，实际却是 18:42:12 才完成，读起来会误解成「那一刻就结束了」；
-而且列表会变成按完成时间排序，而不是按发起时间。
+抽屉的每一列与表头都有提示：
 
-新请求从顶部滑入；被十行上限挤掉的那一行在**同一帧**内移除，而不是自己淡出。
-离场动画会让第 11 行在整个动画期间参与布局，使抽屉先长高一行再缩回——
-每次插入都会看到一次高度跳动。打开抽屉时也只是建立基线，不播放动画，
-所以整个列表不会在打开的瞬间一起滑入。
+| 位置 | 内容 |
+| --- | --- |
+| 时间列 | 完整时间戳 `11-06 10:47:41.828` |
+| 速度列 | 该行使用的模型 |
+| 输出 Token 列 | 输入 token 与缓存命中率 |
+| 表头 | 该列的含义（`#` 一列用图标展示四种状态） |
 
-某个步骤正在流式生成时，抽屉把它显示为历史上方的一条 live 行，状态读数也切换为 live 样式。
-历史下方有一条分隔线，线下面是汇总行——对当前可见请求取平均值，复用同一套 grid，
-因此每个均值都落在它所汇总的那一列下面。完全没有样本的列（比如中断请求没有首字时间）
-打印破折号，而不是容易误解的 0。
+提示是自建的浮层而不是 `title`：`title` 只能显示纯文本，而 `#` 列的说明要画出图标。
 
 ## 数字是怎么算出来的
 
@@ -370,6 +362,16 @@ const target = uiConversation.binding(sessionId).target('stats')
 | **总用时** | **`model`** | **模型用时，不是统计页的「总用时」列** |
 | 折算速度 | `output ÷ model` | 本插件独有 |
 
+**中断的行：**
+
+被手动停止的请求没有 `assistant/message`，因此 `model`、`decode`、`tps` 全为 null。
+那一行的**总用时回退到 `wall`**（`endTime − startTime`），也就是这条请求从发出到被
+停止实际过了多久——它是日志里的事实，刷新多少次都不变，正是这个场合要的数字。
+
+**首字也回退到 `wall`**，分两种情况：上游一直没返回 token，那么首字就是它等待的时长，
+缺的只是生成速度，所以由速度列说明原因。中断行不计入底部的平均值——它的总用时是墙钟
+时间，混进均值会拉向一个既不代表完成、也不代表中断的数；但平均行的标题仍数满十行。
+
 **总用时为何取 `model` 而不是 `wall`：**
 
 统计页的「总用时」列是 `wall = endTime − startTime`，它**把工具调用也算了进去**。