feat: 缓存命中率有来源了（assistant/chunk 的 usage 事件）+ DSH 启动自动拉起浮窗

挖到真源：事件流里 assistant/chunk 带 chunk.type='usage'
  {inputTokens,outputTokens,cacheReadTokens,reasoningTokens}
从此算出真实指标（不再依赖 tokenMeter 的估算）：
  缓存命中率 = cacheRead/(input+cacheRead)     实测 0.9976（与 App 状态栏 99% 一致）
  总量       = output + input + cacheRead      实测 88.99M
  tok/s      = outputTokens/该步耗时            实测 37.3
  上下文占用 = 最后一步的 input+cacheRead / 窗口（不是历史累加）  实测 56.3%
插件与镜像桥两侧都接上了；镜像桥改为跟随插件认定的当前会话。

另：autoOpen（默认开）——DSH 启动后 1.5s 若浮窗没在跑就自动拉起（用户选定）。

lib/host.js

@@ -155,7 +155,7 @@ const SAMPLE_TYPES = new Set(['user/message', 'assistant/chunk', 'assistant/mess
 export function apply(ctx, config) {
   // 浮窗设置：DSH 设置页写进 profile 条目 config 的值 + 默认值
   const floatCfg = Object.assign(
-    { topmost: true, posLock: false, sizeLock: false, exclude: true, opacity: 0.92, scale: 1, backdrop: 'opaque', tool: true, button: true },
+    { topmost: true, posLock: false, sizeLock: false, exclude: true, opacity: 0.92, scale: 1, backdrop: 'opaque', tool: true, button: true, autoOpen: true },
     (config && typeof config === 'object') ? config : {},
   )
 
@@ -237,7 +237,7 @@ export function apply(ctx, config) {
 
   function counter(id) {
     let c = counts.get(id)
-    if (!c) { c = { turns: 0, steps: 0, outTokens: 0, decodeMs: 0, stepStart: 0, lastTps: null }; counts.set(id, c) }
+    if (!c) { c = { turns: 0, steps: 0, outTokens: 0, inTokens: 0, cacheRead: 0, decodeMs: 0, stepStart: 0, lastTps: null }; counts.set(id, c) }
     return c
   }
 
@@ -275,11 +275,25 @@ export function apply(ctx, config) {
       }
       else if (type === 'step/start') c.stepStart = Date.now()
       else if (type === 'step/end') c.steps++
+      else if (type === 'assistant/chunk') {
+        // 真实用量在这里（实测）：data.chunk = {type:'usage', usage:{inputTokens,outputTokens,cacheReadTokens,reasoningTokens}}
+        const ch = d2(event)?.chunk
+        if (ch?.type === 'usage' && ch.usage) {
+          const u = ch.usage
+          c.inTokens += u.inputTokens || 0
+          c.cacheRead += u.cacheReadTokens || 0
+          c.outTokens += u.outputTokens || 0
+          c.lastPrompt = (u.inputTokens || 0) + (u.cacheReadTokens || 0) // 当前上下文占用（最后一步输入）
+          const ms = c.stepStart ? Date.now() - c.stepStart : 0
+          if (u.outputTokens && ms > 0) { c.decodeMs += ms; c.lastTps = u.outputTokens / (ms / 1000) }
+        }
+      }
       else if (type === 'assistant/message') {
-        const usage = event?.usage || event?.message?.usage || {}
+        // 兜底：有些版本把 usage 挂在 message 上
+        const usage = d2(event)?.usage || d2(event)?.message?.usage || {}
         const out = usage.outputTokens ?? usage.output_tokens ?? usage.completionTokens
         const ms = c.stepStart ? Date.now() - c.stepStart : 0
-        if (out && ms > 0) { c.outTokens += out; c.decodeMs += ms; c.lastTps = out / (ms / 1000) }
+        if (out && ms > 0 && !c.outTokens) { c.outTokens += out; c.decodeMs += ms; c.lastTps = out / (ms / 1000) }
       }
       const row = toRow(event)
       lastRow = row ? { type, row } : { type, row: null }
@@ -297,13 +311,16 @@ export function apply(ctx, config) {
   let meterKeysLogged = false
   function metricsFor(session) {
     const id = session?.id || lastSessionId || 'unknown'
-    const c = counts.get(id) || { turns: 0, steps: 0, outTokens: 0, decodeMs: 0, lastTps: null }
+    const c = counts.get(id) || { turns: 0, steps: 0, outTokens: 0, inTokens: 0, cacheRead: 0, decodeMs: 0, lastTps: null }
+    const promptTokens = (c.inTokens || 0) + (c.cacheRead || 0)
     const out = {
       turns: c.turns || null,
       steps: c.steps || null,
       tps: c.lastTps,
-      tokens: null,
+      // 总量 = 输出 + (未缓存输入 + 命中缓存输入)
-      cacheHit: null,
+      tokens: (c.outTokens || 0) + promptTokens || null,
+      // 命中率 = 命中缓存输入 / 全部输入（与 App 状态栏口径一致）
+      cacheHit: promptTokens > 0 ? (c.cacheRead || 0) / promptTokens : null,
       contextUsed: null,
     }
     // 有 session-projection 就用它（更权威），否则用事件计数兜底
@@ -322,7 +339,8 @@ export function apply(ctx, config) {
       if (m) {
         // 第一次把真实字段名记进状态文件，便于在用户机器上核对映射
         if (!meterKeysLogged) { meterKeysLogged = true; log('tokenMeter keys: ' + Object.keys(m).join(',')) }
-        out.tokens = m.totalTokens ?? m.surfaceTokens ?? null
+        // 事件流里已经拿到真实用量时不要被 tokenMeter 的估算覆盖
+        if (!out.tokens) out.tokens = m.totalTokens ?? m.surfaceTokens ?? null
         const u = m.usage || m.providerUsage || m.usageBuckets
         if (u) {
           const read = u.cacheRead ?? u.cache_read ?? u.cacheReadTokens ?? 0
@@ -331,10 +349,16 @@ export function apply(ctx, config) {
           if (denom > 0) out.cacheHit = read / denom
         }
         let capacity = contextWindows.get(id) || modelContext() || 1000000 // 兜底：DSH 常用 1M 上下文
-        const used = m.surfaceTokens ?? m.totalTokens
+        // 上下文占比：优先用"最后一步输入"（事件流里拿得到，就是当前 surface），其次 tokenMeter
+        const used = c.lastPrompt || m.surfaceTokens || m.totalTokens
         if (capacity && used) out.contextUsed = Math.min(1, used / capacity)
       }
     } catch (e) { if (!meterKeysLogged) { meterKeysLogged = true; log('tokenMeter measure failed: ' + String(e?.message || e)) } }
+    // tokenMeter 拿不到时，用事件流里的"最后一步输入"兜底算上下文占比
+    if (out.contextUsed == null && c.lastPrompt) {
+      const cap = contextWindows.get(id) || modelContext() || 1000000
+      if (cap) out.contextUsed = Math.min(1, c.lastPrompt / cap)
+    }
     return out
   }
 
@@ -785,5 +809,18 @@ border:1px solid rgba(120,140,180,.35);background:rgba(22,26,34,.92);color:#e6ea
   } catch (e) { log('settings registration failed:', String(e?.message || e)) }
   writeStatus()
 
+  // ---------------------------------------------------------------- DSH 启动时自动拉起浮窗
+  // 用户选的就是这条：DSH 每次重启都会把浮窗（进程树后代）带走，所以启动后自动补一个。
+  // 用 setTimeout 延后，让宿主先起完；回调里不碰 ctx（那时候它可能已经失效）。
+  if (floatCfg.autoOpen !== false) {
+    setTimeout(() => {
+      try {
+        if (shellAlive()) { log('autoOpen: float already running'); return }
+        log('autoOpen: spawning float window')
+        spawnShell().then((r) => log('autoOpen -> ' + short(r, 120))).catch((e) => log('autoOpen failed: ' + String(e?.message || e)))
+      } catch (e) { log('autoOpen error: ' + String(e?.message || e)) }
+    }, 1500)
+  }
+
   log('ready; endpoints:', `${ROUTE}/events`, `${ROUTE}/state`, `${ROUTE}/health`)
 }

tools/mirror-bridge.mjs

@@ -87,6 +87,13 @@ function rowsFromLog(text) {
   let streamStart = 0
   let streamChars = 0
   let lastChunkAt = 0
+  let outTokens = 0
+  let inTokens = 0
+  let cacheRead = 0
+  let stepStart = 0
+  let decodeMs = 0
+  let lastTps = null
+  let lastPromptTokens = 0 // 最后一步的输入(=未缓存+命中缓存)：这才是"当前上下文占用"
   const est = (chars) => Math.round(chars / 3.2) // 中英混排的粗略 token 估算
 
   for (const line of text.split('\n')) {
@@ -97,7 +104,7 @@ function rowsFromLog(text) {
     const d = e?.data || e || {}
     const t = e?.time || 0
     if (type === 'turn/start') turns++
-    else if (type === 'step/start') steps++
+    else if (type === 'step/start') { steps++; stepStart = t || Date.now() }
     else if (type === 'request/context' && d.contextWindow) contextWindow = d.contextWindow
     else if (type === 'session/title' && d.title) titles.set('title', String(d.title))
     else if (type === 'user/message') {
@@ -110,6 +117,17 @@ function rowsFromLog(text) {
     } else if (type === 'assistant/chunk') {
       const c = d.chunk || {}
       const kind = String(c.type || '')
+      if (kind === 'usage' && c.usage) {
+        // 真实用量：{inputTokens,outputTokens,cacheReadTokens,reasoningTokens}
+        const u = c.usage
+        inTokens += u.inputTokens || 0
+        cacheRead += u.cacheReadTokens || 0
+        outTokens += u.outputTokens || 0
+        lastPromptTokens = (u.inputTokens || 0) + (u.cacheReadTokens || 0)
+        const ms = stepStart ? (t || Date.now()) - stepStart : 0
+        if (u.outputTokens && ms > 0) { decodeMs += ms; lastTps = u.outputTokens / (ms / 1000) }
+        continue
+      }
       const piece = c.text ?? c.delta ?? (kind.includes('text') ? c.content : '')
       if (piece) {
         const now = t || Date.now()
@@ -129,16 +147,20 @@ function rowsFromLog(text) {
   }
   for (const r of rows) delete r.streaming
   const streaming = lastChunkAt && Date.now() - lastChunkAt < 8000
-  const tps = streaming && streamStart ? est(streamChars) / Math.max(0.5, (lastChunkAt - streamStart) / 1000) : null
+  // tok/s：优先用真实 outputTokens/步耗时；否则退化成"流式字符估算"
+  const tps = lastTps ?? (streaming && streamStart ? est(streamChars) / Math.max(0.5, (lastChunkAt - streamStart) / 1000) : null)
+  const promptTokens = inTokens + cacheRead
+  const realUsage = promptTokens + outTokens > 0
   return {
     rows, titles, turns, steps, streaming,
     metrics: {
       turns,
       steps,
       tps,
-      tokens: outChars ? est(outChars + userChars) : null,
+      // 有真实用量就用真实值；否则退化成字符估算（界面会带 ≈）
-      cacheHit: null,
+      tokens: realUsage ? outTokens + promptTokens : (outChars ? est(outChars + userChars) : null),
-      contextUsed: contextWindow ? Math.min(1, est(userChars + outChars) / contextWindow) : null,
+      cacheHit: promptTokens > 0 ? cacheRead / promptTokens : null,
+      contextUsed: contextWindow ? Math.min(1, (realUsage ? lastPromptTokens : est(userChars + outChars)) / contextWindow) : null,
     },
   }
 }
@@ -164,9 +186,16 @@ function buildSnapshot() {
     })
   }
 
-  // 当前会话：watch.json 指定优先，否则"日志最新"（真正在跑的那个）
+  // 当前会话：watch.json 指定 > 插件自己认定的当前会话 > 日志最新
   let active = null
   if (watch?.sessionId) active = entries.find((e) => e.id === watch.sessionId || e.id.replace(/^session-/, '') === watch.sessionId) || null
+  if (!active) {
+    // 插件写的 snapshot.json 带 sessionId：跟它走，指标缺口才能对上同一个会话
+    const plug = readJson(SNAPSHOT)
+    if (plug?.sessionId && !plug.source) {
+      active = entries.find((e) => e.id === plug.sessionId || e.id.replace(/^session-/, '') === plug.sessionId) || null
+    }
+  }
   if (!active) active = entries.sort((a, b) => b.mtime - a.mtime)[0] || null
   if (!active) return null