@leviyuan/lodestar 0.1.11 → 0.1.12

package/README.md

@@ -16,10 +16,12 @@ AI 不是帮手,是倍率。它放大的不是体力,是你——你的直觉、
 
 ## 你会得到什么
 
-- 🌊 **真·流式卡片** — 飞书 Card Kit v1 streaming,Claude 一个 token 一个 token 地打在同一张卡片里,不是发一堆零碎消息刷屏。
+- 🌊 **真·流式卡片** — 飞书 Card Kit v1 streaming,Claude 一个 token 一个 token 地打在同一张卡片里,不是发一堆零碎消息刷屏。每张 turn 卡片 footer 自带 `✅ ⏱ 12.3s · 📊 47% · 💰 $0.45`,本轮上下文占用 / 实付成本一眼可见。
 - 🧠 **思考过程透明** — `thinking` 流式渲染,turn 结束后自动收起为可展开面板。每次工具调用也是一格折叠面板:折起是概述,展开看完整 input/output。
 - 🔐 **权限审批就地完成** — 需要授权的工具调用,**原地**升级为 🔐 等审批状态,三颗按钮 `允许 / 始终允许 / 拒绝` 直接嵌在面板里。不弹独立卡片,不破坏时序。点完按钮,后续 output 接在同一条线上继续往下走。
 - ❓ **结构化追问** — Claude 的 `AskUserQuestion` 在群里呈现为可点击选项行;不满意?直接在群里**打字回答**,daemon 会把自由文本当作 custom answer 发回去。多题串行,有进度计数和"已答 N 题"折叠历史。
+- ⌨️ **Type-ahead 不打断** — Claude 跑着你继续连珠炮,daemon 全部接住排队,排队消息打 `⏳` 反应,消化后清空(`stop` 取消则换 `❌`)。daemon 还会给每条合并消息前面注 `[#N]\n` 序号,模型一眼分得清"这是 5 条独立消息"而不是一个长字符串。turn 中途有新消息进来 + 下一个 tool_use 边界 → 旧卡 `📨 转交新卡` 收尾(既不是 done 也不是打断),新卡续写,边界跟语义对齐。
+- ⏰ **定时唤醒可见化** — Claude 用 `CronCreate` / `ScheduleWakeup` 自己安排周期任务,到点子进程在 idle 间隙 fire,daemon 检测"非首次 init"自动开一张 `⏰ 定时唤醒` 卡片承接;这种自发 turn 不响加急(凌晨 3 点自检不该震你)。
 - 📦 **状态面板一键唤出** — 发 `hi` 弹一张控制台:model、上下文占用 %、累计 tokens/cost、上一轮 delta、session id、订阅额度(5h / 7d 真实 utilization,直读 Anthropic 官方 OAuth Usage API,凭据走 `~/.claude/.credentials.json`,token 过期自动 refresh)、本机所有活跃项目并列展示。
 - 📎 **图片 / 文件双向互传** — 用户在群里发图/文件,Claude 通过消息里的 `[file: /abs/path]` 提示就能读;Claude 在回复里写 `[[send: /abs/path]]`,标记被剥离,文件以独立消息发回群里。出站路径走 realpath + 白名单校验,只允许工作目录、`/tmp/lodestar-*`、inbox 三块,`/etc`、`~/.ssh`、`~/.config` 即使被符号链接绕也拒绝。
 - 📲 **加急锁屏推送** — 需要你回答问题、需要你批准操作、一轮跑完了——三种关键时刻自动触发飞书"应用内加急",直接打穿勿扰、亮屏推送。卡片摘要会同步改写成具体待办("🔐 等审批: Bash · rm -rf …"、"❓ 待回答 3 题: …"),锁屏一瞥就知道发生了什么。
@@ -42,11 +44,12 @@ AI 不是帮手,是倍率。它放大的不是体力,是你——你的直觉、
 | 指令 | 行为 |
 | --- | --- |
 | `hi` | 未运行时启动;运行中弹一张**状态卡片** |
+| `stop` | 软打断当前 turn + 清空 type-ahead 排队;子进程保活,刚排队中的消息会被打 `CrossMark` 反应表示取消 |
 | `kill` | 优雅关闭 Claude 进程;记住 `sessionId`,下次 `restart` 还能 resume |
 | `restart` | 用上一次的 `sessionId` 重启会话(保留上下文) |
 | `clear` | 杀掉进程并启动一个全新 session(等价于 Claude Code 的 `/clear`) |
 
-> 这四个词被全局保留:在群里发 "hi" 当问候也会触发控制台卡片,不会到 Claude 那边。换来的是手机上单手打字的便利。
+> 这五个词被全局保留:在群里发 "hi" 当问候也会触发控制台卡片,不会到 Claude 那边。换来的是手机上单手打字的便利。
 
 ## 安装
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@leviyuan/lodestar",
-  "version": "0.1.11",
+  "version": "0.1.12",
   "publishConfig": {
     "access": "public"
   },

package/src/claude-process.ts

@@ -93,6 +93,12 @@ export class ClaudeProcess extends EventEmitter {
   lastResult: ClaudeResultMeta = {
     cost_usd: null, duration_ms: null, num_turns: null, usage: null,
   }
+  /** Context-window capacity of the model that ran the latest turn —
+   * lifted from `result.modelUsage[model].contextWindow` so we don't
+   * have to hardcode `[1m]` vs stock variants. 200K is the safe
+   * default if no result has landed yet (e.g. between spawn and the
+   * first turn close). */
+  lastContextWindow: number = 200_000
 
   constructor(opts: SpawnOpts) {
     super()
@@ -242,6 +248,19 @@ export class ClaudeProcess extends EventEmitter {
         num_turns: typeof msg.num_turns === 'number' ? msg.num_turns : null,
         usage: msg.usage ?? null,
       }
+      // modelUsage maps "<model id>" → { contextWindow, maxOutputTokens, … }.
+      // For mixed-model runs the SDK reports one entry per model used in
+      // the turn; we take the one matching `lastModel` (the assistant's
+      // latest model id) and fall back to any single entry if it's the
+      // only one — covers the common single-model case.
+      const mu = msg.modelUsage
+      if (mu && typeof mu === 'object') {
+        const entry = (this.lastModel && mu[this.lastModel])
+          || (Object.keys(mu).length === 1 ? mu[Object.keys(mu)[0]!] : null)
+        if (entry && typeof entry.contextWindow === 'number' && entry.contextWindow > 0) {
+          this.lastContextWindow = entry.contextWindow
+        }
+      }
       this.emit('result', msg)
       return
     }

package/src/session.ts

@@ -511,7 +511,15 @@ export class Session {
     const wasBusy = this.currentTurn !== null || this.openingTurn
     this.pendingUserMessageCount++
     this.lastUserOpenId = userOpenId
-    this.proc!.sendUserText(text, files)
+    // When this msg will be merged with siblings into a multi-content
+    // user turn (i.e. the SDK queued it because the daemon was busy),
+    // prepend a `[#N]\n` ordinal so the model can tell the merged
+    // blocks apart. Without it the harness renders multi-content text
+    // back-to-back ("1"+"2"+"5"+"56"+"89" → "1255689") and the model
+    // can't see the original boundaries — surfaced 2026-05-16 when a
+    // 5-msg accumulator test got mis-summed as one big number.
+    const wireText = wasBusy ? `[#${this.pendingUserMessageCount}]\n${text}` : text
+    this.proc!.sendUserText(wireText, files)
     if (wasBusy && msgId) {
       // Hold the slot in the map even if the API call hasn't returned
       // yet — empty string is a sentinel meaning "we tried to react;
@@ -883,6 +891,16 @@ export class Session {
     return this.lastTurnDelta?.inputTokens ?? 0
   }
 
+  /** Context-window capacity for the model the subprocess is currently
+   * running — sourced authoritatively from `result.modelUsage[model]
+   * .contextWindow` captured by ClaudeProcess on each turn close, so
+   * the daemon doesn't have to enumerate model ids itself (was the
+   * source of a "560K/200K" display bug — model id didn't include
+   * `[1m]` so the hardcoded fallback won). */
+  private contextWindowMax(): number {
+    return this.proc?.lastContextWindow ?? 200_000
+  }
+
   private async openTurnCard(userText: string, userOpenId: string, trigger: 'user_message' | 'scheduled'): Promise<void> {
     const turn = ++this.turnCounter
     const card = cards.mainConversationCard({
@@ -950,12 +968,13 @@ export class Session {
       if (this.openingTurn) this.rotationBuffer.push({ kind: 'assistant', delta })
       return
     }
-    if (this.wantsRotation) {
-      this.wantsRotation = false
-      this.rotationBuffer.push({ kind: 'assistant', delta })
-      void this.rotateCard()
-      return
-    }
+    // Note: assistant text DOES NOT trigger rotation, even if a mid-turn
+    // user message landed and set `wantsRotation`. Rotating mid-segment
+    // would chop the model's in-progress reply (often a response to the
+    // ORIGINAL prompt that started this card) onto a fresh card,
+    // visually associating it with the queued msg — which is the bug
+    // the user surfaced 2026-05-16. The rotation defers to the next
+    // tool_use, which is a clean section boundary.
     if (!this.currentTurn.currentAssistantSegmentId) {
       const i = this.currentTurn.assistantSegmentCount++
       const segId = cards.ELEMENTS.assistant(i)
@@ -985,12 +1004,8 @@ export class Session {
       if (this.openingTurn) this.rotationBuffer.push({ kind: 'thinking', delta })
       return
     }
-    if (this.wantsRotation) {
-      this.wantsRotation = false
-      this.rotationBuffer.push({ kind: 'thinking', delta })
-      void this.rotateCard()
-      return
-    }
+    // Thinking, like assistant text, doesn't trigger rotation — it's
+    // preamble to the same response, not a section break.
     this.currentTurn.thinkingText += delta
     cardkit.streamTextThrottled(
       this.currentTurn.cardId,
@@ -1303,11 +1318,28 @@ export class Session {
       await cardkit.replaceElement(cardId, cards.ELEMENTS.thinking, cards.thinkingCollapsedPanel(thinkingText))
     }
     const sendNote = sendPaths.length ? ` · 📎 ${sendPaths.length}` : ''
-    // Suffix REPLACES the trailing `✅ done` — it represents a terminal
-    // state distinct from natural completion (e.g. `📨 转交新卡` for a
-    // mid-turn rotation). When absent, the turn ended cleanly.
-    const stateMark = suffix ? ` · ${suffix}` : ' · ✅ done'
-    const footer = `⏱ ${elapsed}s${sendNote}${stateMark}`
+    // State marker leads the footer (✅ for natural completion, or the
+    // suffix verbatim for non-natural states like `📨 转交新卡`). The
+    // trailing "done" word is gone — the ✅ already carries that
+    // meaning. User-confirmed footer order 2026-05-16.
+    const stateMark = suffix ? suffix : '✅'
+    // Per-turn metrics: context-window occupancy (as a real percentage,
+    // not a token count) and dollar cost. Only meaningful on a clean
+    // close — suffix-tagged turns (rotation / interrupt) didn't fire
+    // the `result` event that populates `lastTurnDelta`, so these
+    // numbers would be stale and misleading.
+    let metrics = ''
+    if (!suffix) {
+      const ctxTokens = this.currentContextTokens()
+      const ctxMax = this.contextWindowMax()
+      if (ctxTokens > 0 && ctxMax > 0) {
+        const pct = Math.round((ctxTokens / ctxMax) * 100)
+        metrics += ` · 📊 ${pct}%`
+      }
+      const cost = this.lastTurnDelta?.costUsd ?? 0
+      if (cost > 0) metrics += ` · 💰 $${cost.toFixed(3)}`
+    }
+    const footer = `${stateMark} ⏱ ${elapsed}s${metrics}${sendNote}`
     await cardkit.streamText(cardId, cards.ELEMENTS.footer, footer)
     // Final chat-list preview: clean finish shows "⏱ Xs · NK tokens";
     // interrupted shows the suffix instead (no usage event landed).