pi-ptc-meter 1.0.33 → 1.0.35

package/README.md

@@ -52,34 +52,39 @@ pi install npm:pi-ptc-meter
 
 | 命令 | 作用 |
 | --- | --- |
-| `/ptc [n]` | 最近 n 轮：次数、时长、中位、超标 |
+| `/ptc [n]` | 状态（阈值 / inject / notify）+ 最近 8 轮明细（次数 · 时长）+ 统计（中位 / 超标轮次） |
 
 ## 模拟测试
 
 pi-ptc-meter 的改动在**隔离副本**上做回放验证后才发版；当前版本通过 **10,011 条断言**。
 
-| 测试面 | 规模 |
-| --- | --- |
-| 真实日志回放 | 4,349 轮 |
-| 随机状态 | 10,000 条 |
-| 统计函数边界 | 11 项 |
+| 测试面 | 规模 | 结果 |
+| --- | --- | --- |
+| 真实日志回放 | 4,349 轮 | 🟢 |
+| 随机状态 | 10,000 条 | 🟢 |
+| 统计函数边界 | 11 项 | 🟢 |
+
+🟢 **结论**：预估功能已删除——412 个声明点上 6 种预测器全部落在 71–94% 误差，预估恰好在它被需要的那一刻不可预测。
 
 改动：删调用数预估 + 耗时预估。依据：预估在声明点不可预测：412 个声明点上 6 种预测器全部 71–94% 误差；耗时中位误差 199%。
+
 ## 作用（源码事实）
 
 - `tool_call`：顶层调用计数 +1（codemode 内层调用不另计；不拦截、不改调用）。
 - `turn_end`：记录本轮次数与耗时（不落盘）；超阈值时每 run 提示一次（goal 自动续轮共享同一 run 时 `agent_end` 可能整段不触发，避免回路失效）。
 - `agent_end`：每 run 落盘一行日志。
 - `before_agent_start`：结算上一轮并以 `nextTurn` 尾部消息回灌——计数、耗时、**缓存命中率**（仅当该轮有 usage 数据时出现；< 90% 标「低」）、会话中位、阈值判定（`display: false`）。
-- **1.0.32 起不再回灌「预计 / 实际 / 偏差」**：声明率虽升到 65%，中位误差仍为 94%（调用数）与 199%（耗时），且 412 个声明点上 6 种预测器（常量 12/16、1.1× 上轮、2.0× 上轮、全局中位、区间 3~40）全部落在 71–94% 误差。
+- **1.0.32 起不回灌「预计 / 实际 / 偏差」**：412 个声明点上 6 种预测器全部 71–94% 误差（耗时 199%），预估不可用。
 
 **注入样例（示意，字段可能随版本增删）**
 
 ```text
 PTC 计量：上一轮 9 次工具调用 · 实际 34s · 命中 92% · 会话中位 25 · 阈值 50 · 判据 = 本轮计数不超阈值且无同类调用可并
 ```
+- 耗时预估仍写入日志并在 `/ptc` 中显示（历史值），但不再回灌给模型。
 - `session_start`：模块级计量归零。
 - system 内 section `ptc_meter` 只含逐字不变的骨架。
+- 只写自己的 section，**永不返回 `systemPrompt`**（返回即覆盖式替换，会静默丢弃其它扩展的 section 修改）。
 ## 原理（专业）
 
 ```text
@@ -162,34 +167,38 @@ Also: log `~/.pi/ptc-meter.jsonl` (1 MiB truncation, tail-preserving); no enviro
 
 | Command | Purpose |
 | --- | --- |
-| `/ptc [n]` | Last n turns: counts, duration, median, over-threshold |
+| `/ptc [n]` | State (threshold / inject / notify) + last 8 turns (calls · duration) + stats (median / over-threshold) |
 
 ## Verification
 
 Changes to pi-ptc-meter were verified on an isolated copy before release; the current version passes **10,011 assertions**.
 
-| Surface | Scale |
-| --- | --- |
-| Real log replay | 4,349 turns |
-| Random states | 10,000 |
-| Statistic edges | 11 |
+| Surface | Scale | Result |
+| --- | --- | --- |
+| Real log replay | 4,349 turns | 🟢 |
+| Random states | 10,000 | 🟢 |
+| Statistic edges | 11 | 🟢 |
 
+🟢 **Result**: both estimates removed - on 412 declaration points all six predictors landed at 71-94% error, so estimating is unpredictable exactly when it is needed.
 Change: both estimates removed. Why: estimates are unpredictable at declaration points: six predictors all landed at 71-94% error; duration median error 199%.
+
 ## What it does (source facts)
 
 - `tool_call`: increments the counter for top-level calls only (nested codemode calls are not counted separately; never intercepts or rewrites).
 - `turn_end`: records calls and duration (no disk write); when goal auto-continuation shares one run and `agent_end` never fires, it warns once above threshold (once per run).
 - `agent_end`: writes one log line per run.
 - `before_agent_start`: settles the previous turn and feeds back counts, duration, the **cache hit rate** (only when that turn reported usage; flagged low below 90%), the session median and the threshold verdict as a `nextTurn` trailing message (`display: false`).
-- **1.0.32 no longer feeds back estimate / actual / deviation**: declaration rate rose to 65% but the median error stayed at 94% (call count) and 199% (duration); on 412 declaration points all six predictors (constant 12/16, 1.1x previous, 2.0x previous, global median, interval 3-40) landed between 71% and 94% error.
+- **1.0.32 drops estimate feedback**: on 412 declaration points all six predictors landed at 71-94% error (duration 199%) - estimating is not viable.
 
 **Sample injection (illustrative; fields may change between versions)**
 
 ```text
 PTC metering: last turn 9 tool calls · 34s actual · 92% cache hit · session median 25 · threshold 50 · verdict: under threshold, no obvious merge candidates
 ```
+- The duration estimate is still logged and shown by `/ptc` as a historical value, but is no longer fed back to the model.
 - `session_start`: resets module-level counters.
 - The `ptc_meter` system section holds only a byte-stable skeleton.
+- Writes only its own section and **never returns `systemPrompt`** (a returned prompt replaces everything and silently drops other extensions' sections).
 ## How it works (technical)
 
 ```text

package/index.ts

@@ -166,7 +166,7 @@ const EST_RE = /(?:预计|预估|eta)[^0-9\n]{0,8}(\d+(?:\.\d+)?)\s*(秒|s(?!\w)
 /** 从我的回复文本里解析预计耗时（秒）；无声明或非法值返回 null。 */
 function parseEstimate(text: string): number | null {
   try {
-    // 扫**尾部窗口**：声明写在回复末尾；前 6000 字符从未覆盖到它（实测 est 填充率仅 1.5%）
+    // 扫**尾部窗口**：声明写在回复末尾；前 6000 字符从未覆盖到它（实测 est 填充率 9.7%，2026-10-18 全量 4373 轮）
     // 取最后一个匹配 = 末尾的声明（同时谈格式时，正文中部的不算，因窗口在尾部）
     const m = [...text.slice(-15000).matchAll(new RegExp(EST_RE.source, "gi"))].pop();
     if (!m) return null;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-ptc-meter",
-  "version": "1.0.33",
+  "version": "1.0.35",
   "description": "Measures tool-call batching per turn and feeds the numbers back to the model.",
   "type": "module",
   "license": "MIT",