dsh-personal-system-rules 0.6.0 → 0.7.0

package/CHANGELOG.md

@@ -1,16 +1,28 @@
 # 更新日志
 
+## 0.7.0 — 2026-11-05
+
+- 静默守卫从「步数」改为「时长」：新增 `maxSilentSeconds`（默认 60 秒，`0` 关闭），按最后一条可见文字消息到当前时刻的间隔判定，同一段沉默每再满一个阈值提醒一次。提醒文案的 `{count}` 现在是秒数。
+- 原因（会话 `session-2debd25a` 实测 258 个 step、43 次提醒）：43 次里有 41 次发生在距上次说明不足 60 秒（中位 18 秒、最小 7 秒），而 18 个「已沉默 ≥60 秒」的 step 中有 16 个完全没提醒，最长沉默 629 秒。步数口径既太快打断，也看不见单步长耗时造成的长时间沉默。
+- 新增纯函数 `silentMs(agent, now)`（读 surface 事件自带的时间戳，无可见文字时回落到本轮起点），`test/loop-guard.test.js` 覆盖测量与回落，`test/guards.test.js` 覆盖阈值、重置与分段提醒。
+- 设置页字段改为「多久没有可见输出就提醒（秒）」，中英文案、README、locale 描述同步。
+- 旧的 `maxSilentSteps` 键保留在 profile 里也无害（schemastery 不报未知键），但不再生效；迁移时换成期望的秒数。
+
+## 0.6.1 — 2026-11-05
+
+- 修复任务状态守卫从不触发的问题：`staleTodoSteps()` 原先在 `session.surface.nodes` 里查找 `todo/write` 事件，但 surface 只承载消息类事件（`system`/`developer`/`user`/`assistant` 消息与 `tool/result`），`todo/write` 是仅落日志的事件，永远不会出现在 surface 上，于是每步都被判成「本轮没写列表」，守卫一直静默。
+- 改为读取 assistant 消息里模型自己的 `todo_write` 工具调用（`arguments.todos`）作为列表来源，步数从最近一次该调用算起；日志里的 `todo/write` 事件不再参与判断。
+- 回放会话 `session-2debd25a` 的真实 surface：修复前注入 0 次，修复后在第 6/12/18/24/30/36 步各注入一次。
+- `test/loop-guard.test.js`、`test/guards.test.js` 的用例改为真实形状（`todo_write` 工具调用），并新增「仅落日志的 `todo/write` 事件不构成列表」「以最近一次调用为准」「非法 arguments 不误判」三个回归用例。
+
 ## 0.6.0 — 2026-11-04
 
 - 新增任务状态守卫：本轮的 todo 列表里还有 `in_progress` 项、且连续 `maxStaleTodoSteps` 步没有重写列表时，经 `agent/pre-step` 注入一条提醒；之后每再累积同样步数提醒一次。默认 6 步，`0` 关闭。
 - 本轮没有 todo 列表（短任务可以主动跳过）或列表已全部 `completed` 时都不触发，避免把有意不写列表的任务也当成「没更新」。
 - 设置页「守卫」卡片新增「任务状态」分组，阈值与文案都可在弹窗编辑，文案支持 `{count}`（未更新步数）与 `{inProgress}`（进行中项数）。
 - 检测逻辑新增纯函数 `staleTodoSteps()`，`test/loop-guard.test.js` 与 `test/guards.test.js` 各有覆盖。
-
-## 0.5.1 — 2026-11-04
-
 - 声明并验证对 DeepSeek Harness 0.2.2 预发布版的支持：`settings.section` 与 `plugins.bundle.config` 的注册契约、`describe/mutate` 设置读写、locale 的 `register(ns, {zh, en})` 形式、`loader/volatile-update` 热更新路径在 0.2.2 均未变。`engines.dsh` 已覆盖该版本。
-- 不使用 Session Projection、spill 策略或 `dsh-fs-observation-policy`，因此 0.2.2 对这几项的改动不命中本插件。无代码改动，已用 DSH 0.2.2-rc.1 内核验证加载与全部测试。
+- 不使用 Session Projection、spill 策略或 `dsh-fs-observation-policy`，因此 0.2.2 对这几项的改动不命中本插件。已用 DSH 0.2.2-rc.1 内核验证加载与全部测试。
 
 ## 0.5.0 — 2026-10-31
 

package/README.md

@@ -4,12 +4,12 @@
 
 把个性化规则文件注入 DeepSeek Harness（dsh）的**系统提示**，并提供**静默提醒**、**重复操作循环检测**与**任务状态守卫**；规则正文与守卫参数都能在设置弹窗「个性化规则」里编辑。
 
-A dsh plugin that injects a personal rules file as a system-prompt section and adds silent-step / repeated-action loop guards and a stale-todo guard. The rules body and guard options are editable from the dsh settings dialog.
+A dsh plugin that injects a personal rules file as a system-prompt section and adds a silent (elapsed-time) reminder, a repeated-action loop guard and a stale-todo guard. The rules body and guard options are editable from the dsh settings dialog.
 
 ## 功能
 
 1. **系统提示段**：把 `~/.dsh/PERSONAL-SYSTEM-RULES.md`（或 `config.file` 指定文件）注册成段 `user:agents-md`，order 默认 10150，原样注入；`watchFile` 1 秒轮询，改完下一次模型请求生效，不用重启。
-2. **静默提醒**：连续 `maxSilentSteps`（默认 2）步没有可见文本时，注入一条提醒要求模型说明进展；同一条链只提醒一次，`0` 关闭。
+2. **静默提醒**：距离上一次**可见文字**超过 `maxSilentSeconds`（默认 60 秒）时，注入一条提醒要求模型说明进展；同一段沉默每再满一个阈值提醒一次，`0` 关闭。判定用时长而不是步数：单个长耗时步骤不会再掩盖十分钟的沉默，两步之间的十几秒也不会提前打断。
 3. **循环检测**：连续 `repeatLimit` 步执行**完全相同**的工具调用（工具名 + 归一化参数）时触发，`repeatAction` 可选 `remind`（提醒）或 `stop`（像用户点停止一样中止本轮并保留排队输入）；默认 0 关闭。
 4. **任务状态守卫**：本轮的 todo 列表里还有 `in_progress` 项、且连续 `maxStaleTodoSteps`（默认 6）步没有重写列表时提醒一次，之后再每隔同样步数提醒；本轮没建列表（短任务）时不打扰，全部完成时也不触发。`0` 关闭。
 5. **设置弹窗页面**：`设置 → 个性化规则` 与插件详情页是同一个页面，含「规则正文」和「守卫」两块；保存经官方 `settings` 机制写回 profile patch 并热更新。
@@ -49,7 +49,7 @@ dsh plugin --profile my-profile add "https://github.com/NingYuanLee/dsh-personal
     header: |
       # 用户全局系统层规则
       以下内容来自用户全局 PERSONAL-SYSTEM-RULES.md，是系统层指令，必须遵守。
-    maxSilentSteps: 2
+    maxSilentSeconds: 60
     silentReminder: ''
     repeatLimit: 0
     repeatAction: remind
@@ -58,7 +58,8 @@ dsh plugin --profile my-profile add "https://github.com/NingYuanLee/dsh-personal
     staleTodoReminder: ''
 ```
 
-- `maxSilentSteps`、`silentReminder`、`repeatLimit`、`repeatAction`、`repeatReminder`、`maxStaleTodoSteps`、`staleTodoReminder` 是 `volatile()` 字段，可在设置弹窗编辑并热更新。
+- `maxSilentSeconds`、`silentReminder`、`repeatLimit`、`repeatAction`、`repeatReminder`、`maxStaleTodoSteps`、`staleTodoReminder` 是 `volatile()` 字段，可在设置弹窗编辑并热更新。
+- 0.7.0 起静默提醒按秒判定，旧的 `maxSilentSteps` 键不再生效（仍可留在 patch 里，会被忽略）；把阈值换成期望的秒数即可。
 - `file`、`header`、`order` 是普通配置，只在 `cordis.patch.yml` 里改；改 `file` 后需要重载插件。
 - `repeatLimit: 0` 关闭循环检测；文本回复或新用户轮次会重置重复计数。
 - `maxStaleTodoSteps: 0` 关闭任务状态守卫；提醒文案可含 `{count}`（未更新的步数）与 `{inProgress}`（进行中的项数）。

package/client.js

@@ -8,7 +8,7 @@ const PKG = "dsh-personal-system-rules";
 const SETTINGS_NS = "system-agents-md";
 const READ = "systemAgentsMd/read";
 const WRITE = "systemAgentsMd/write";
-const DEFAULT_REMINDER = "\u4F60\u5DF2\u7ECF\u8FDE\u7EED {count} \u6B65\u6CA1\u6709\u5411\u7528\u6237\u8BF4\u660E\u8FDB\u5C55\u3002\u8BF7\u5728\u672C\u6B65\u5F00\u5934\u7528\u4E00\u53E5\u8BDD\u8BF4\u660E\u6B63\u5728\u505A\u4EC0\u4E48\u3001\u4E3A\u4EC0\u4E48\uFF0C\u518D\u7EE7\u7EED\u8C03\u7528\u5DE5\u5177\u3002";
+const DEFAULT_REMINDER = "\u4F60\u5DF2\u7ECF {count} \u79D2\u6CA1\u6709\u5411\u7528\u6237\u8BF4\u660E\u8FDB\u5C55\u3002\u8BF7\u5728\u672C\u6B65\u5F00\u5934\u7528\u4E00\u53E5\u8BDD\u8BF4\u660E\u6B63\u5728\u505A\u4EC0\u4E48\u3001\u4E3A\u4EC0\u4E48\uFF0C\u518D\u7EE7\u7EED\u8C03\u7528\u5DE5\u5177\u3002";
 const DEFAULT_REPEAT_REMINDER = "\u4F60\u5DF2\u7ECF\u8FDE\u7EED {count} \u6B65\u6267\u884C\u5B8C\u5168\u76F8\u540C\u7684\u64CD\u4F5C\u3002\u8BF7\u505C\u6B62\u91CD\u590D\uFF1A\u5148\u8BF4\u660E\u5F53\u524D\u963B\u585E\u548C\u5DF2\u5B8C\u6210\u7684\u90E8\u5206\uFF0C\u518D\u6362\u4E00\u79CD\u505A\u6CD5\uFF1B\u5982\u679C\u786E\u5B9E\u9700\u8981\u7EE7\u7EED\u540C\u4E00\u64CD\u4F5C\uFF0C\u8BF7\u5148\u5411\u7528\u6237\u8BF4\u660E\u539F\u56E0\u3002";
 const DEFAULT_STALE_TODO_REMINDER = "\u4EFB\u52A1\u5217\u8868\u91CC\u6709 {inProgress} \u9879\u4ECD\u6807\u7740\u8FDB\u884C\u4E2D\uFF0C\u4F46\u5DF2\u8FDE\u7EED {count} \u6B65\u6CA1\u6709\u66F4\u65B0\u3002\u8BF7\u91CD\u5199 todo \u5217\u8868\uFF1A\u628A\u5DF2\u5B8C\u6210\u7684\u9879\u6807\u6210 completed\uFF0C\u53EA\u7559\u4E0B\u771F\u6B63\u5728\u505A\u7684\u4E00\u9879\uFF0C\u518D\u7EE7\u7EED\u3002";
 const zh = {
@@ -20,10 +20,10 @@ const zh = {
   rulesTitle: "\u89C4\u5219\u6B63\u6587",
   guardTitle: "\u5B88\u536B",
   silentTitle: "\u9759\u9ED8\u63D0\u9192",
-  guardStepsLabel: "\u8FDE\u7EED\u65E0\u53EF\u89C1\u6587\u672C\u7684\u6B65\u6570\u9608\u503C",
-  guardStepsHint: "\u8FBE\u5230\u8BE5\u6B65\u6570\u65F6\u63D0\u9192\u4E00\u6B21\uFF1B0 \u8868\u793A\u5173\u95ED\u3002\u9ED8\u8BA4 2\u3002",
+  guardStepsLabel: "\u591A\u4E45\u6CA1\u6709\u53EF\u89C1\u8F93\u51FA\u5C31\u63D0\u9192\uFF08\u79D2\uFF09",
+  guardStepsHint: "\u8FBE\u5230\u8BE5\u65F6\u957F\u65F6\u63D0\u9192\u4E00\u6B21\uFF0C\u4E4B\u540E\u6BCF\u518D\u6EE1\u540C\u6837\u65F6\u957F\u63D0\u9192\u4E00\u6B21\uFF1B0 \u8868\u793A\u5173\u95ED\u3002\u9ED8\u8BA4 60 \u79D2\u3002",
   guardReminderLabel: "\u63D0\u9192\u6587\u6848",
-  guardReminderHint: "\u53EF\u7528 {count} \u8868\u793A\u5F53\u524D\u8FDE\u7EED\u6B65\u6570\uFF1B\u7559\u7A7A\u4F7F\u7528\u9ED8\u8BA4\u6587\u6848\u3002",
+  guardReminderHint: "\u53EF\u7528 {count} \u8868\u793A\u5F53\u524D\u5DF2\u9759\u9ED8\u7684\u79D2\u6570\uFF1B\u7559\u7A7A\u4F7F\u7528\u9ED8\u8BA4\u6587\u6848\u3002",
   loopTitle: "\u5FAA\u73AF\u68C0\u6D4B",
   loopLimitLabel: "\u540C\u4E00\u64CD\u4F5C\u7684\u6700\u5927\u91CD\u590D\u6B21\u6570",
   loopLimitHint: "\u8FDE\u7EED\u8FD9\u4E48\u591A\u6B65\u6267\u884C\u5B8C\u5168\u76F8\u540C\u7684\u5DE5\u5177\u8C03\u7528\u65F6\u89E6\u53D1\uFF1B0 \u8868\u793A\u5173\u95ED\uFF0C\u9ED8\u8BA4\u5173\u95ED\u3002",
@@ -46,7 +46,7 @@ const zh = {
   optionsMissing: "\u6CA1\u6709\u627E\u5230\u8FD9\u4E2A\u63D2\u4EF6\u7684\u8BBE\u7F6E\u547D\u540D\u7A7A\u95F4\uFF0C\u5148\u6309\u9ED8\u8BA4\u503C\u663E\u793A\u3002",
   optionsLoadFailed: "\u8BFB\u53D6\u5DF2\u4FDD\u5B58\u914D\u7F6E\u5931\u8D25\uFF0C\u5148\u6309\u9ED8\u8BA4\u503C\u663E\u793A\uFF1A{message}",
   optionsFailed: "\u4FDD\u5B58\u5931\u8D25\uFF1A{message}",
-  invalidSteps: "\u6B65\u6570\u9608\u503C\u5FC5\u987B\u662F\u4E0D\u5C0F\u4E8E 0 \u7684\u6574\u6570\u3002",
+  invalidSteps: "\u9759\u9ED8\u9608\u503C\u5FC5\u987B\u662F\u4E0D\u5C0F\u4E8E 0 \u7684\u6574\u6570\uFF08\u79D2\uFF09\u3002",
   reload: "\u91CD\u65B0\u8BFB\u53D6",
   save: "\u4FDD\u5B58",
   saving: "\u4FDD\u5B58\u4E2D",
@@ -71,11 +71,11 @@ const en = {
   effective: "Effective file",
   rulesTitle: "Rules body",
   guardTitle: "Guards",
-  silentTitle: "Silent-step reminder",
-  guardStepsLabel: "Consecutive steps without visible text",
-  guardStepsHint: "Reminds once when reached; 0 disables it. Default 2.",
+  silentTitle: "Silent reminder",
+  guardStepsLabel: "Seconds without visible output before reminding",
+  guardStepsHint: "Reminds once when reached, then once per further interval; 0 disables it. Default 60 seconds.",
   guardReminderLabel: "Reminder text",
-  guardReminderHint: "Use {count} for the current streak; leave blank for the default.",
+  guardReminderHint: "Use {count} for the seconds of silence; leave blank for the default.",
   loopTitle: "Loop detection",
   loopLimitLabel: "Maximum repeats of the same action",
   loopLimitHint: "Triggers after this many consecutive identical tool-call steps; 0 disables it (default).",
@@ -98,7 +98,7 @@ const en = {
   optionsMissing: "This plugin has no settings namespace here; showing defaults.",
   optionsLoadFailed: "Failed to read the saved config; showing defaults: {message}",
   optionsFailed: "Save failed: {message}",
-  invalidSteps: "The step threshold must be an integer of 0 or more.",
+  invalidSteps: "The silence threshold must be an integer of 0 or more (seconds).",
   reload: "Reload",
   save: "Save",
   saving: "Saving",
@@ -177,7 +177,7 @@ function findSettingsRow(found) {
   if (exact) return exact;
   return rows.find((row) => {
     const value = row?.value && typeof row.value === "object" ? row.value : row;
-    return value !== null && typeof value === "object" && (typeof value.maxSilentSteps === "number" || typeof value.silentReminder === "string");
+    return value !== null && typeof value === "object" && (typeof value.maxSilentSeconds === "number" || typeof value.silentReminder === "string");
   }) ?? null;
 }
 function takeView(landed, fallbackRevision) {
@@ -282,7 +282,7 @@ function RulesSection(props) {
       }
       const value = row.value && typeof row.value === "object" && !Array.isArray(row.value) ? row.value : {};
       setConfigNs(typeof row.ns === "string" && row.ns ? row.ns : SETTINGS_NS);
-      setStepsDraft(Number.isFinite(value.maxSilentSteps) ? String(value.maxSilentSteps) : "");
+      setStepsDraft(Number.isFinite(value.maxSilentSeconds) ? String(value.maxSilentSeconds) : "");
       setReminderDraft(typeof value.silentReminder === "string" ? value.silentReminder : "");
       setRepeatLimitDraft(Number.isFinite(value.repeatLimit) && value.repeatLimit > 0 ? String(value.repeatLimit) : "");
       setRepeatActionDraft(value.repeatAction === "stop" ? "stop" : "remind");
@@ -402,7 +402,7 @@ function RulesSection(props) {
     const staleReminder = staleReminderDraft.trim() === "" ? "" : staleReminderDraft;
     const action = repeatActionDraft === "stop" ? "stop" : "remind";
     const ops = [
-      stepsText === "" ? { op: "unset", path: ["maxSilentSteps"] } : { op: "set", path: ["maxSilentSteps"], value: steps },
+      stepsText === "" ? { op: "unset", path: ["maxSilentSeconds"] } : { op: "set", path: ["maxSilentSeconds"], value: steps },
       reminder === "" ? { op: "unset", path: ["silentReminder"] } : { op: "set", path: ["silentReminder"], value: reminder },
       limitText === "" ? { op: "unset", path: ["repeatLimit"] } : { op: "set", path: ["repeatLimit"], value: limit },
       action === "remind" ? { op: "unset", path: ["repeatAction"] } : { op: "set", path: ["repeatAction"], value: action },
@@ -421,7 +421,7 @@ function RulesSection(props) {
       const view = takeView(landed, revision);
       setRevision(view.revision);
       if (view.value) {
-        setStepsDraft(Number.isFinite(view.value.maxSilentSteps) ? String(view.value.maxSilentSteps) : "");
+        setStepsDraft(Number.isFinite(view.value.maxSilentSeconds) ? String(view.value.maxSilentSeconds) : "");
         setReminderDraft(typeof view.value.silentReminder === "string" ? view.value.silentReminder : "");
         setRepeatLimitDraft(Number.isFinite(view.value.repeatLimit) && view.value.repeatLimit > 0 ? String(view.value.repeatLimit) : "");
         setRepeatActionDraft(view.value.repeatAction === "stop" ? "stop" : "remind");

package/index.js

@@ -4,7 +4,7 @@ import { homedir } from 'node:os'
 import { join } from 'node:path'
 import { pathToFileURL } from 'node:url'
 import { MAX_BYTES, readAgentsFile, writeAgentsFile } from './file.js'
-import { countRepeatSteps, staleTodoSteps } from './loop-guard.js'
+import { countRepeatSteps, silentMs, staleTodoSteps } from './loop-guard.js'
 
 /** Cordis plugin name. */
 export const name = 'system-agents-md'
@@ -14,7 +14,7 @@ export const inject = ['systemPrompt']
 
 const SECTION_NAME = 'user:agents-md'
 const DEFAULT_ORDER = 10150
-const DEFAULT_MAX_SILENT_STEPS = 2
+const DEFAULT_MAX_SILENT_SECONDS = 60
 const DEFAULT_MAX_STALE_TODO_STEPS = 6
 const DEFAULT_RULES_FILE = 'PERSONAL-SYSTEM-RULES.md'
 const DEFAULT_HEADER = [
@@ -23,8 +23,8 @@ const DEFAULT_HEADER = [
   '以下内容来自用户全局 PERSONAL-SYSTEM-RULES.md，是系统层指令，必须遵守；与项目内软性说明冲突时以本节为准。',
 ].join('\n')
 
-/** Default reminder used when consecutive steps carry no visible text. */
-const DEFAULT_SILENT_REMINDER = '你已经连续 {count} 步没有向用户说明进展。请在本步开头用一句话说明正在做什么、为什么，再继续调用工具。'
+/** Default reminder used when the model goes too long without visible text. */
+const DEFAULT_SILENT_REMINDER = '你已经 {count} 秒没有向用户说明进展。请在本步开头用一句话说明正在做什么、为什么，再继续调用工具。'
 
 /** Default reminder used when the same tool call repeats beyond the limit. */
 const DEFAULT_REPEAT_REMINDER = '你已经连续 {count} 步执行完全相同的操作。请停止重复：先说明当前阻塞和已完成的部分，再换一种做法；如果确实需要继续同一操作，请先向用户说明原因。'
@@ -41,7 +41,7 @@ const REMOTE_METHODS = '@deepseek-ai/dsh-typert-protocol/remote-methods'
 /**
  * Config schema projected into the DSH settings dialog.
  *
- * `file`, `maxSilentSteps` and `silentReminder` are `.volatile()`: that is what
+ * `file`, `maxSilentSeconds` and `silentReminder` are `.volatile()`: that is what
  * makes them editable from the GUI settings form (see volatileForm() in
  * @deepseek-ai/dsh-settings) and what lets a save apply without reloading the
  * plugin. The remaining fields stay ordinary config for cordis.patch.yml.
@@ -70,7 +70,7 @@ async function loadConfigSchema() {
     file: z.string().default(''),
     order: z.number().default(DEFAULT_ORDER),
     header: z.string().default(DEFAULT_HEADER),
-    maxSilentSteps: z.number().min(0).default(DEFAULT_MAX_SILENT_STEPS).volatile(),
+    maxSilentSeconds: z.number().min(0).default(DEFAULT_MAX_SILENT_SECONDS).volatile(),
     silentReminder: z.string().default('').volatile(),
     repeatLimit: z.number().min(0).default(0).volatile(),
     repeatAction: z.union([z.const('remind'), z.const('stop')]).default('remind').volatile(),
@@ -136,8 +136,8 @@ function resolveSettings(config = {}) {
   const file = typeof fileValue === 'string' && fileValue.trim() ? fileValue.trim() : join(dshHome(), DEFAULT_RULES_FILE)
   const order = Number.isFinite(Number(config.order)) ? Number(config.order) : DEFAULT_ORDER
   const header = typeof config.header === 'string' ? config.header : DEFAULT_HEADER
-  const stepsValue = Number(unwrapVolatile(config.maxSilentSteps))
-  const maxSilentSteps = Number.isFinite(stepsValue) ? Math.max(0, Math.floor(stepsValue)) : DEFAULT_MAX_SILENT_STEPS
+  const silentValue = Number(unwrapVolatile(config.maxSilentSeconds))
+  const maxSilentSeconds = Number.isFinite(silentValue) ? Math.max(0, Math.floor(silentValue)) : DEFAULT_MAX_SILENT_SECONDS
   const reminderValue = unwrapVolatile(config.silentReminder)
   const silentTemplate = typeof reminderValue === 'string' && reminderValue.trim() ? reminderValue : DEFAULT_SILENT_REMINDER
   const repeatLimitValue = Number(unwrapVolatile(config.repeatLimit))
@@ -153,7 +153,7 @@ function resolveSettings(config = {}) {
     file,
     order,
     header,
-    maxSilentSteps,
+    maxSilentSeconds,
     silentTemplate,
     repeatLimit,
     repeatAction,
@@ -279,39 +279,13 @@ function readRules(file) {
   }
 }
 
-/** Whether one assistant message carries visible text the user can read. */
-function hasVisibleText(message) {
-  const content = message?.content
-  if (!Array.isArray(content)) return false
-  return content.some((part) => part?.type === 'text' && String(part.text ?? '').trim() !== '')
-}
-
-/** Count trailing assistant steps without visible text since the last user turn. */
-function countSilentSteps(agent) {
-  const nodes = agent?.session?.surface?.nodes
-  if (!Array.isArray(nodes)) return 0
-  let count = 0
-  for (let index = nodes.length - 1; index >= 0; index -= 1) {
-    const event = agent.session.eventAt(nodes[index])
-    if (event === undefined) continue
-    if (event.type === 'assistant/message') {
-      if (hasVisibleText(event.data?.message)) return count
-      count += 1
-      continue
-    }
-    if (event.type === 'turn/start') return count
-    if (event.type === 'user/message' && event.data?.source?.kind === 'user') return count
-  }
-  return count
-}
-
 /**
  * Register the rules file as a global system-prompt section, keep it in sync
  * when the file or the volatile config changes, expose it to the settings and
- * detail pages, and nudge the model when it runs several steps without visible
- * text.
+ * detail pages, and nudge the model when it leaves the user without a word for
+ * too long.
  * @param {object} ctx - host plugin context.
- * @param {{file?: string, order?: number, header?: string, maxSilentSteps?: number, silentReminder?: string}} [config]
+ * @param {{file?: string, order?: number, header?: string, maxSilentSeconds?: number, silentReminder?: string}} [config]
  */
 export function apply(ctx, config = {}) {
   let current = resolveSettings(config)
@@ -367,11 +341,13 @@ export function apply(ctx, config = {}) {
   }, 'system-agents-md.section')
 
   // One pre-step hook drives three guards:
-  // - silent guard: consecutive steps without visible text;
+  // - silent guard: elapsed time since the last readable assistant text;
   // - loop guard: consecutive identical tool-call steps (remind, or stop the
   //   turn the way the user's own stop does, keeping queued input);
   // - stale-todo guard: a task list left with `in_progress` items that the
-  //   model has not rewritten for several steps.
+  //   model has not rewritten for several steps. The list is read from the
+  //   model's `todo_write` tool call: the `todo/write` session event the tool
+  //   appends is log-only and never reaches the surface the guards scan.
   // Settings edits are volatile, so `current` is read on every step. A guard
   // disabled with 0 resets its own per-agent state.
   const silentRemindedAt = new WeakMap()
@@ -384,18 +360,23 @@ export function apply(ctx, config = {}) {
     if (agent === undefined) return decision
     const additions = Array.isArray(decision.messages) ? [...decision.messages] : []
 
-    const silentLimit = current.maxSilentSteps
-    if (silentLimit <= 0) {
+    // Silent guard: wall-clock time since the last readable assistant text, so
+    // one long tool step cannot hide ten minutes of silence behind a single
+    // step, and quick steps do not nag before the user has waited a while.
+    // One reminder per `maxSilentSeconds` of further silence.
+    const silentLimitMs = current.maxSilentSeconds * 1000
+    if (silentLimitMs <= 0) {
       silentRemindedAt.delete(agent)
     } else {
-      const streak = countSilentSteps(agent)
-      if (streak < silentLimit) {
+      const silentFor = silentMs(agent)
+      if (silentFor < silentLimitMs) {
         silentRemindedAt.delete(agent)
-      } else if ((silentRemindedAt.get(agent) ?? 0) < streak) {
-        silentRemindedAt.set(agent, streak)
+      } else if (silentFor >= (silentRemindedAt.get(agent) ?? 0) + silentLimitMs) {
+        silentRemindedAt.set(agent, silentFor)
+        const seconds = Math.round(silentFor / 1000)
         additions.push(createUserMessage({
-          content: [{ type: 'text', text: current.silentTemplate.replaceAll('{count}', String(streak)) }],
-          source: { kind: 'system-agents-md', form: 'notice', summary: streak + ' steps without visible text' },
+          content: [{ type: 'text', text: current.silentTemplate.replaceAll('{count}', String(seconds)) }],
+          source: { kind: 'system-agents-md', form: 'notice', summary: seconds + ' seconds without visible text' },
         }))
       }
     }

package/locale/en.json

@@ -1,6 +1,6 @@
 {
   "meta": {
     "title": "Personalized rules",
-    "description": "Inject a user-chosen rules file as a system-prompt section, and configure its body plus silent-step, repeated-action and stale-todo guards from the DSH settings dialog or the plugin detail page. Inactive in the minimal preset; nothing is injected when the plugin is disabled or fails to load."
+    "description": "Inject a user-chosen rules file as a system-prompt section, and configure its body plus the silent (elapsed-time) reminder, repeated-action and stale-todo guards from the DSH settings dialog or the plugin detail page. Inactive in the minimal preset; nothing is injected when the plugin is disabled or fails to load."
   }
 }

package/loop-guard.js

@@ -1,5 +1,6 @@
 /**
- * Pure helpers for the repeated-action loop guard.
+ * Pure helpers for the pre-step guards: the silent reminder, the repeated-action
+ * loop guard, and the stale-todo guard.
  *
  * A "repeat step" is one assistant message whose ordered tool calls are
  * identical to the previous assistant message's calls: same tool names and the
@@ -89,14 +90,88 @@ export function countRepeatSteps(agent) {
   return repeatStreak(recentToolSteps(agent))
 }
 
+/** Name of the tool whose call carries the complete task list. */
+const TODO_TOOL_NAME = 'todo_write'
+
+/** Whether one assistant message carries visible text the user can read. */
+export function hasVisibleText(message) {
+  const content = message?.content
+  if (!Array.isArray(content)) return false
+  return content.some((part) => part?.type === 'text' && String(part.text ?? '').trim() !== '')
+}
+
+/**
+ * Milliseconds since the model last showed the user something readable.
+ *
+ * Measured from the newest assistant message with visible text; when the turn
+ * has shown none yet, from the newest turn boundary. A step-count threshold
+ * cannot express this: one long tool step can hold the user in silence for ten
+ * minutes while still counting as a single step.
+ * @param {object} agent live agent (session surface + eventAt)
+ * @param {number} [now] reference clock in epoch milliseconds
+ * @returns {number} elapsed milliseconds, 0 when nothing datable is on the surface
+ */
+export function silentMs(agent, now = Date.now()) {
+  const nodes = agent?.session?.surface?.nodes
+  if (!Array.isArray(nodes)) return 0
+  for (let index = nodes.length - 1; index >= 0; index -= 1) {
+    const event = agent.session.eventAt(nodes[index])
+    if (event === undefined) continue
+    const time = Number(event.time)
+    if (!Number.isFinite(time)) continue
+    if (event.type === 'assistant/message') {
+      if (hasVisibleText(event.data?.message)) return Math.max(0, now - time)
+      continue
+    }
+    if (event.type === 'turn/start') return Math.max(0, now - time)
+    if (event.type === 'user/message' && event.data?.source?.kind === 'user') return Math.max(0, now - time)
+  }
+  return 0
+}
+
 /**
- * Steps taken since the newest `todo/write` in the current turn, plus how many
- * items that list left `in_progress`.
+ * The task list a `todo_write` call carries, or null when the message made none.
  *
- * The todo projection is reset by `turn/start`, so a list that is still
- * non-empty was necessarily written in this turn. `hasList` is false when the
- * turn wrote no list yet: the model may be skipping todos on purpose (short
- * tasks), so callers must not nag in that case.
+ * The `todo/write` session event the tool appends is log-only: the surface
+ * carries message-producing events only (`system`/`developer`/`user`/
+ * `assistant` messages and `tool/result`), so a guard reading
+ * `session.surface.nodes` can never see that event. The model's own tool call
+ * is the surface-visible record of the same list.
+ * @param {object} message one assistant message
+ * @returns {object[] | null}
+ */
+export function todoWriteList(message) {
+  const content = message?.content
+  if (!Array.isArray(content)) return null
+  let todos = null
+  for (const part of content) {
+    if (part?.type !== 'tool-call' || part.name !== TODO_TOOL_NAME) continue
+    todos = parseTodoList(part.arguments) ?? todos
+  }
+  return todos
+}
+
+/** Parse a `todo_write` argument string into its todo array, or null. */
+function parseTodoList(raw) {
+  if (typeof raw !== 'string' || raw.trim() === '') return null
+  let parsed
+  try {
+    parsed = JSON.parse(raw)
+  } catch {
+    return null
+  }
+  if (parsed === null || typeof parsed !== 'object' || !Array.isArray(parsed.todos)) return null
+  return parsed.todos
+}
+
+/**
+ * Steps taken since the newest `todo_write` call in the current turn, plus how
+ * many items that list left `in_progress`.
+ *
+ * A list written in an earlier turn is not counted: the scan stops at the turn
+ * boundary, and the todos projection is itself reset by `turn/start`.
+ * `hasList` is false when the turn wrote no list yet: the model may be skipping
+ * todos on purpose (short tasks), so callers must not nag in that case.
  * @param {object} agent live agent (session surface + eventAt)
  * @returns {{ steps: number, inProgress: number, hasList: boolean }}
  */
@@ -107,15 +182,15 @@ export function staleTodoSteps(agent) {
   for (let index = nodes.length - 1; index >= 0; index -= 1) {
     const event = agent.session.eventAt(nodes[index])
     if (event === undefined) continue
-    if (event.type === 'todo/write') {
-      const todos = Array.isArray(event.data?.todos) ? event.data.todos : []
-      return {
-        steps,
-        inProgress: todos.filter((todo) => todo?.status === 'in_progress').length,
-        hasList: true,
-      }
-    }
     if (event.type === 'assistant/message') {
+      const todos = todoWriteList(event.data?.message)
+      if (todos !== null) {
+        return {
+          steps,
+          inProgress: todos.filter((todo) => todo?.status === 'in_progress').length,
+          hasList: true,
+        }
+      }
       steps += 1
       continue
     }

package/package.json

@@ -1,8 +1,8 @@
 {
   "name": "dsh-personal-system-rules",
-  "version": "0.6.0",
+  "version": "0.7.0",
   "displayName": "个性化规则",
-  "description": "Inject a user-chosen rules file as a system-prompt section, and configure its body plus silent-step, repeated-action and stale-todo guards from the DSH settings dialog or the plugin detail page. Inactive in the minimal preset; nothing is injected when the plugin is disabled or fails to load.",
+  "description": "Inject a user-chosen rules file as a system-prompt section, and configure its body plus the silent (elapsed-time) reminder, repeated-action and stale-todo guards from the DSH settings dialog or the plugin detail page. Inactive in the minimal preset; nothing is injected when the plugin is disabled or fails to load.",
   "license": "MIT",
   "author": "NingYuanLee",
   "type": "module",