pi-think-anchor 1.0.28 → 1.0.30

package/README.en.md

@@ -1,53 +1,9 @@
 # pi-think-anchor
 
-Chain-of-thought drift monitor: reads a word-frequency fingerprint from thinking, classifies focused / mixed / reactive, and nudges when drifting.
+Thinking-focus monitor with graded nudges, driven by a reproducible ratio test.
 
-## Install
+**English documentation:** [open the English section](README.md#english) · **中文文档：** [中文](README.md#zh)
 
-```bash
-pi install npm:pi-think-anchor
-```
+    pi install npm:pi-think-anchor
 
-## Configuration
-
-Config file `~/.pi/think-anchor.json` (optional; defaults apply, invalid JSON falls back):
-
-| Key | Default | Description |
-| --- | --- | --- |
-| `enabled` | `true` | Master switch |
-| `score.minHits` | `3` | Fewer hits in a turn: no scoring, no intervention |
-| `score.spec` | `1.2` | Focused threshold |
-| `score.react` | `0.5` | Reactive threshold |
-| `score.alpha` | `1` | Alpha smoothing |
-| `windowSize` | `5` | Window size in messages |
-| `weights.beta` | `0.4` | Neutral-word weight in the numerator |
-| `weights.eps` | `0.5` | Denominator floor |
-| `intervention.l1CooldownMin` | `10` | L1 cooldown (minutes) |
-| `intervention.l2CooldownMin` | `15` | L2 cooldown (minutes) |
-| `intervention.maxL2` | `2` | Max L2 per session |
-| `lexicon.positive/negative/neutral` | built-in Chinese lexicon (28 focus / 18 drift) | Lexicon override |
-| `bands.spec/react` | `1.2 / 0.5` | Legacy keys (used when `score` is absent) |
-
-Also: `THINK_ANCHOR=0` hard off; log `~/.pi/think-anchor.jsonl` (1 MiB truncation, tail-preserving).
-
-## Commands
-
-| Command | Purpose |
-| --- | --- |
-| `/think-anchor` | Current band, hits, drift share, last intervention result |
-| `/think-anchor stats` | Version, band distribution, actions, effectiveness, log size |
-
-## What it does (source facts)
-
-- `turn_end`: reads assistant thinking -> fingerprint (pos/neg/neu) -> sliding window -> band; when hits < `minHits` it only advances the window.
-- State machine `healthy -> warning -> critical`; L1/L2 nudges are injected as `nextTurn` trailing messages.
-- No escalation: when the previous `eff >= 0`, L2 degrades to L1.
-- `session_start`: resets state, keeping only the last 3 samples and pending (cross-session settlement).
-
-## Credits
-
-Distilled from the DSH plugin [`dsh-anchored-monitor`](https://dsh-plugin.org/zh/plugins/aik358/dsh-anchored-monitor) 0.3.1: the mechanism is upstream, the code is not (turn-level, no separate process).
-
-## License
-
-MIT
+Source: https://github.com/op15066212/pi-plugins-publish

package/README.md

@@ -7,6 +7,8 @@
 
 # pi-think-anchor
 
+> 适用 pi ≥ 0.74 · Node ≥ 20 · MIT · 无运行时依赖（仅 peer: @earendil-works/pi-coding-agent）
+
 思维链锚定监测：读取 thinking 的词频指纹，判定专注 / 过渡 / 发散，发散时分级提示。
 
 ## 快速事实
@@ -14,7 +16,7 @@
 | 项 | 说明 |
 | --- | --- |
 | 配置文件 | ~/.pi/think-anchor.json（可省略） |
-| 命令行 | /think-anchor [on|off|reset|stats] |
+| 命令行 | /think-anchor [on, off, reset, stats] |
 | 调用方式 | **全自动**：turn_end 评估 + 条件注入。on/off 只改当前会话内存（不写文件），重启后回到配置文件的值 |
 
 ## 安装
@@ -42,7 +44,7 @@ pi install npm:pi-think-anchor
 | `intervention.maxL2` | `2` | 每会话 L2 上限 |
 | `lexicon.positive/negative/neutral` | 专注 28 词 / 发散 18 词 / 中性 **默认空** | 词表覆盖 |
 
-其他：日志 `~/.pi/think-anchor.jsonl`（1 MiB 截断，保留尾部整行）。
+其他：日志 `~/.pi/think-anchor.jsonl`（1 MiB 截断，保留尾部整行）；日志里的 `ratio` 是**未平滑**的原始比率，而判定用的是 α 平滑后的比率（`(pos + beta·neu + alpha) / (neg + eps + alpha)`）——两者数值不同，别拿日志 ratio 当判定依据。
 **词表覆盖规则（源码事实）**
 
 - 三个键独立覆盖：只写想改的那个，其余保留默认（`lexicon` 按子键浅合并）。
@@ -139,7 +141,13 @@ pi install npm:pi-think-anchor
 ## 作用（源码事实）
 
 - `turn_end`：取 assistant 消息 thinking → 指纹（pos/neg/neu）→ 滑窗聚合 → 判波段；本轮命中 < `minHits` 时只推进窗口。
-- 状态机 `healthy → warning → critical`；L1/L2 干预以 `nextTurn` 尾部消息注入。
+- 状态机 `healthy → warning → critical`；L1/L2 干预以 `nextTurn` 尾部消息回灌（`display: false`），文本里带当前比率、命中量与触发阈值（如「比率 0.48 · 命中 22；react<0.5」）——让模型看到依据，而不是只收到一句指令。
+**注入样例（示意，字段可能随版本增删）**
+
+```text
+think-anchor 信号（词频启发式，非判据）：连续偏离。动作：停止探索，立即执行一个具体动作。（比率 0.48 · 命中 22；L2 1/2）
+```
+
 - 无效不升级：上次 `eff >= 0` 时 L2 降为 L1。
 - `session_start`：重置状态机，仅保留最近 3 条与 pending（跨会话结算）。
 
@@ -194,7 +202,7 @@ Chain-of-thought drift monitor: reads a word-frequency fingerprint from thinking
 | Item | Description |
 | --- | --- |
 | Config file | ~/.pi/think-anchor.json (optional) |
-| Commands | /think-anchor [on|off|reset|stats] |
+| Commands | /think-anchor [on, off, reset, stats] |
 | Automation | **Fully automatic**: evaluated on turn_end, injects only when needed. on/off changes in-memory state only (not written to disk); a restart falls back to the config file |
 
 Full config (all keys and defaults):
@@ -224,7 +232,7 @@ Config file `~/.pi/think-anchor.json` (optional; defaults apply, invalid JSON fa
 | `intervention.maxL2` | `2` | Max L2 per session |
 | `lexicon.positive/negative/neutral` | 28 focus / 18 drift / neutral **empty by default** | Lexicon override |
 
-Also: log `~/.pi/think-anchor.jsonl` (1 MiB truncation, tail-preserving).
+Also: log `~/.pi/think-anchor.jsonl` (1 MiB truncation, tail-preserving). The `ratio` in the log is the **unsmoothed** raw ratio; classification uses the `alpha`-smoothed ratio (`(pos + beta·neu + alpha) / (neg + eps + alpha)`). They differ — do not use the logged ratio as the decision value.
 **Lexicon override rules (from source)**
 
 - Keys merge independently: write only what you change; the rest keep defaults (`lexicon` merges per sub-key).
@@ -321,7 +329,13 @@ Full config (copy-paste ready, includes the default lexicon):
 ## What it does (source facts)
 
 - `turn_end`: reads assistant thinking -> fingerprint (pos/neg/neu) -> sliding window -> band; when hits < `minHits` it only advances the window.
-- State machine `healthy -> warning -> critical`; L1/L2 nudges are injected as `nextTurn` trailing messages.
+- State machine `healthy -> warning -> critical`; L1/L2 nudges are fed back as `nextTurn` trailing messages (`display: false`), carrying the current ratio, hit count and the triggering threshold (e.g. ratio 0.48, 22 hits, react<0.5) so the model sees the evidence, not just an order.
+**Sample injection (illustrative; fields may change between versions)**
+
+```text
+think-anchor signal (heuristic, not a verdict): sustained drift. Action: stop exploring, execute one concrete step. (ratio 0.48 · 22 hits; L2 1/2)
+```
+
 - No escalation: when the previous `eff >= 0`, L2 degrades to L1.
 - `session_start`: resets state, keeping only the last 3 samples and pending (cross-session settlement).
 

package/index.ts

@@ -89,17 +89,12 @@ export function fingerprint(text: string, lex: { positive: string[]; negative: s
   };
 }
 
+
 /** 加权比率：专注度代理指标 */
 export function ratioOf(pos: number, neg: number, neu: number, w: { beta: number; eps: number } = DEF_CFG.weights): number {
   return (pos + w.beta * neu) / (neg + w.eps);
 }
 
-export function bandOf(ratio: number, t: { spec: number; react: number } = DEF_CFG.score): "spec" | "mixed" | "react" {
-  if (ratio >= t.spec) return "spec";
-  if (ratio >= t.react) return "mixed";
-  return "react";
-}
-
 /**
  * v2.1 判据：证据门槛 + α 平滑的专注/发散比。
  * 回放 2069 条可评分日志：旧判据 react 占 18.3%（其中 57% 是命中 <3 的长而散）；
@@ -328,7 +323,7 @@ export default function thinkAnchor(pi: ExtensionAPI): void {
       } else {
         // 被冷却/上限挡下：记 blocked，避免与「无需动作」混淆（实测 583 次 want=L2 只有 55 次真发）
         st.phase = prevPhase;
-        if (d.level !== "none") acted = st.autoMuted ? "muted-" + d.level : "blocked-" + d.level;
+        acted = st.autoMuted ? "muted-" + d.level : "blocked-" + d.level;
       }
       if (band === "spec" && st.phase === "healthy") st.l2Tried = 0; // 恢复后重置 L2 计数
       // v3 效果闭环：干预后第 3 轮结算（pre = 干预前最近 3 轮 ratio 均值）

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-think-anchor",
-  "version": "1.0.28",
+  "version": "1.0.30",
   "description": "Monitors thinking focus drift and nudges the model back on track.",
   "type": "module",
   "license": "MIT",
@@ -8,7 +8,11 @@
     "pi",
     "pi-extension",
     "pi-coding-agent",
-    "pi-package"
+    "pi-package",
+    "chain-of-thought",
+    "focus",
+    "monitoring",
+    "drift"
   ],
   "pi": {
     "extensions": [
@@ -25,7 +29,7 @@
     "@earendil-works/pi-coding-agent": "*"
   },
   "author": "15066212pp",
-  "homepage": "https://github.com/op15066212/pi-plugins-publish/tree/main/pi-think-anchor",
+  "homepage": "https://github.com/op15066212/pi-plugins-publish/tree/main/pi-think-anchor#readme",
   "bugs": {
     "url": "https://github.com/op15066212/pi-plugins-publish/issues"
   },
@@ -36,5 +40,8 @@
   },
   "scripts": {
     "prepublishOnly": "node ../publish-check.mjs"
+  },
+  "engines": {
+    "node": ">=20"
   }
 }