pi-think-anchor 1.0.2 → 1.0.4

package/README.md

@@ -4,7 +4,7 @@
 
 ## 来源与借鉴
 
-蒸馏自 [dsh-anchored-monitor](https://dsh-plugin.org/zh/plugins/aik358/dsh-anchored-monitor) 0.3.1（DSH 生态）：机制思想来源，非代码移植（回合级而非流内实时、无独立进程）。判据 v2（证据门槛 + 发散占比）为本机 6151 条日志校准后的本地改进。
+蒸馏自 [dsh-anchored-monitor](https://dsh-plugin.org/zh/plugins/aik358/dsh-anchored-monitor) 0.3.1（DSH 生态）：机制思想来源，非代码移植（回合级而非流内实时、无独立进程）。判据经 6151 条日志三次校准：**v2.3 = 单轮证据门槛 + α 平滑专注/发散比**（α=1；react <0.5 / mixed <1.2 / spec ≥1.2）。命中不足的轮次不评分（避免把「长而散」判成反复试探），回放 react 11.3% → 3.4%。
 
 ## 功能
 
@@ -19,7 +19,8 @@
 | 项 | 默认 | 说明 |
 | --- | --- | --- |
 | 启用 | 开 | `THINK_ANCHOR=0` 或配置文件 `enabled: false` 关闭 |
-| 波段阈值 | 1.2 / 0.5 | 高专注 ≥1.2；发散 <0.5 |
+| 波段阈值 | 1.2 / 0.5 | α 平滑比率：高专注 ≥1.2；发散 <0.5 |
+| 证据门槛 | 3 命中/轮 | 本轮词频命中不足 3 个 → 不评分、不干预 |
 | 聚合窗口 | 5 轮 | 滑动窗口 |
 | L1 冷却 | 10 分钟 | 首次跌入发散 |
 | L2 冷却 | 15 分钟 | 持续发散，会话上限 2 次 |
@@ -28,7 +29,7 @@
 ## 安装
 
 ```bash
-pi install npm:pi-think-anchor@1.0.0
+pi install npm:pi-think-anchor
 pi install ./pi-think-anchor
 ```
 

package/index.ts

@@ -31,7 +31,7 @@ interface TAConfig {
   windowSize: number;
   bands: { spec: number; react: number };
   /** v2 主判据：证据门槛 + 发散词占比（与长度无关） */
-  score: { minHits: number; specRate: number; reactRate: number };
+  score: { minHits: number; spec: number; react: number; alpha: number };
   weights: { beta: number; eps: number };
   intervention: { l1CooldownMin: number; l2CooldownMin: number; maxL2: number };
 }
@@ -41,7 +41,7 @@ const DEF_CFG: TAConfig = {
   lexicon: { positive: DEF_POS, negative: DEF_NEG, neutral: DEF_NEU },
   windowSize: 5,
   bands: { spec: 1.2, react: 0.5 },
-  score: { minHits: 3, specRate: 0.25, reactRate: 0.5 },
+  score: { minHits: 3, spec: 1.2, react: 0.5, alpha: 1 },
   weights: { beta: 0.4, eps: 0.5 },
   intervention: { l1CooldownMin: 10, l2CooldownMin: 15, maxL2: 2 },
 };
@@ -58,6 +58,8 @@ function loadCfg(): TAConfig {
       ...DEF_CFG,
       ...raw,
       lexicon: { ...DEF_CFG.lexicon, ...(raw.lexicon ?? {}) },
+      // score 是 v2 判据；未配置 score 时兼容旧 bands（否则改 bands 静默无效）
+      score: { ...DEF_CFG.score, ...(raw.score ?? {}), ...(raw.score ? {} : (raw.bands ?? {})) },
       bands: { ...DEF_CFG.bands, ...(raw.bands ?? {}) },
       weights: { ...DEF_CFG.weights, ...(raw.weights ?? {}) },
       intervention: { ...DEF_CFG.intervention, ...(raw.intervention ?? {}) },
@@ -103,15 +105,17 @@ export function bandOf(ratio: number, bands: { spec: number; react: number } = D
 }
 
 /**
- * v2 判据：先过证据门槛，再看发散词占比 negRate = neg/(pos+neg+neu)。
- * 数据依据（本机 6151 条实测）：react 行 thinking 长度中位 1998、词命中中位 2，43% 的 neg<=1
- * —— 旧 ratio（pos/(neg+eps)）把「长而散」误判成「反复试探」：571 次 react 里 326 次证据不足。
- * negRate 分离度：spec 中位 0.14 / mixed 0.50 / react 1.00。
+ * 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）；
+ * α=1（伪计数）+ minHits=3 后 react 占 9.7%，且选中样本 negRate 中位 0.86（真发散）。
+ * α 的作用：低命中时比率趋向 1（mixed），不再因「命中少」被误判成 react。
  */
-export function classify(hits: number, negRate: number, c: { minHits: number; specRate: number; reactRate: number }): TABand {
+export function classify(pos: number, neg: number, neu: number, c: { minHits: number; spec: number; react: number; alpha: number }): TABand {
+  const hits = pos + neg + neu;
   if (hits < c.minHits) return "none";
-  if (negRate >= c.reactRate) return "react";
-  if (negRate >= c.specRate) return "mixed";
+  const r = (pos + 0.4 * neu + c.alpha) / (neg + 0.5 + c.alpha);
+  if (r < c.react) return "react";
+  if (r < c.spec) return "mixed";
   return "spec";
 }
 
@@ -164,6 +168,11 @@ export function decideNext(phase: TAPhase, band: TABand, l2Tried: number, maxL2:
 }
 
 // ── 干预文案（中性建议式——原版措辞纪律：不指责、不自称审查者） ──
+export function downgradeIfIneffective(level: "none" | "L1" | "L2", lastEff: number | null): "none" | "L1" | "L2" {
+  if (level === "L2" && lastEff !== null && lastEff >= 0) return "L1";
+  return level;
+}
+
 const L1_TEXT = "「think-anchor」信号（词频启发式，非判据）：思考偏离收敛。动作：给结论 + 下一步，只留依据。";
 const L2_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。";
 
@@ -176,11 +185,19 @@ interface TAState {
   l1Count: number;
   l2Count: number;
   turns: number;
-  bandHist: { spec: number; mixed: number; react: number };
+  bandHist: { spec: number; mixed: number; react: number; none: number };
   lastL1: number;
   lastL2: number;
   lastRatio: number;
   lastBand: TABand;
+  /** v3：效果闭环与自适应（借鉴 ptc-meter 的数据自校正 + patch-keeper 的条件触发） */
+  recent: number[];
+  pending: { turn: number; pre: number; ts: number; seen: number } | null;
+  effSamples: number;
+  effImproved: number;
+  autoMuted: boolean;
+  lastLoggedBand: TABand;
+  skipped: number;
 }
 
 function newState(cfg: TAConfig): TAState {
@@ -192,14 +209,29 @@ function newState(cfg: TAConfig): TAState {
     l1Count: 0,
     l2Count: 0,
     turns: 0,
-    bandHist: { spec: 0, mixed: 0, react: 0 },
+    bandHist: { spec: 0, mixed: 0, react: 0, none: 0 },
     lastL1: 0,
     lastL2: 0,
     lastRatio: 0,
     lastBand: "mixed",
+    recent: [],
+    pending: null,
+    effSamples: 0,
+    effImproved: 0,
+    autoMuted: false,
+    lastLoggedBand: "mixed",
+    skipped: 0,
   };
 }
 
+/** 采样写盘（v3）：非事件轮每 5 轮留一行——实测 6151 行 / 1.0 MB 逼近 1 MiB 上限，降为 1/5 仍够统计 */
+const SAMPLE_EVERY = 5;
+/** 自适应（v3）：有效样本 ≥ AUTO_MIN 且下降率 < AUTO_RATE（不及抛硬币）即转静默，只记录不注入 */
+const AUTO_MIN = 8;
+const AUTO_RATE = 0.5;
+/** 上次干预的实测效果（eff，负 = 改善）；用于「无效则不升级」 */
+let lastEffValue: number | null = null;
+
 /** 日志上限 1 MiB（与 ptc-meter 同口径）：超限保留尾部，避免无界增长。 */
 const MAX_LOG = 1 << 20;
 
@@ -207,9 +239,10 @@ export function appendLog(row: Record<string, unknown>): void {
   try {
     appendFileSync(LOG, JSON.stringify(row) + "\n");
     if (statSync(LOG).size > MAX_LOG) {
-      const keep = readFileSync(LOG, "utf8").slice(-MAX_LOG);
-      const cut = keep.indexOf("\n");
-      writeFileSync(LOG, cut >= 0 ? keep.slice(cut + 1) : keep);
+      const buf = readFileSync(LOG);
+      const keep = buf.subarray(Math.max(0, buf.length - MAX_LOG));
+      const cut = keep.indexOf(10);
+      writeFileSync(LOG, cut >= 0 ? keep.subarray(cut + 1) : keep);
     }
   } catch {
     /* 静默 */
@@ -220,7 +253,12 @@ export default function thinkAnchor(pi: ExtensionAPI): void {
   let st = newState(loadCfg());
 
   pi.on("session_start", () => {
+    // 跨会话保住效果闭环：干预常发生在会话末尾，3 轮结算窗口会跨过 session 边界
+    const carryRecent = st.recent.slice(-3);
+    const carryPending = st.pending;
     st = newState(loadCfg());
+    st.recent = carryRecent;
+    st.pending = carryPending;
   });
 
   pi.on("turn_end", async (event, ctx) => {
@@ -234,7 +272,18 @@ export default function thinkAnchor(pi: ExtensionAPI): void {
       const fp = fingerprint(thinking, st.cfg.lexicon);
       if (fp.pos + fp.neg + fp.neu === 0) {
         st.turns++;
-        appendLog({ ts: new Date().toISOString(), turn: st.turns, v: 2, len: thinking.length, pos: 0, neg: 0, neu: 0, ratio: null, hits: 0, negRate: null, band: "none", phase: st.phase, want: "none", acted: "none" });
+        if (st.turns % SAMPLE_EVERY === 0) appendLog({ ts: new Date().toISOString(), turn: st.turns, v: 3, len: thinking.length, pos: 0, neg: 0, neu: 0, ratio: null, hits: 0, negRate: null, band: "none", phase: st.phase, want: "none", acted: "none", eff: null });
+        return;
+      }
+      // v2.3：证据门槛作用于**本轮**。窗口聚合 hits 中位 27（仅 1.4% 低于门槛）→ 门槛形同虚设；
+      // 而 42.7% 的轮次单轮命中 <3（多为长而散），正是旧口径下 react 偏高的来源（回放 11.3% → 3.4%）。
+      const turnHits = fp.pos + fp.neg + fp.neu;
+      if (turnHits < st.cfg.score.minHits) {
+        st.bandHist.none++;  // 会话内计数：v3 采样下 none 轮通常不写日志，只有这里能观测 v2.3 生效
+        st.window.push(fp);
+        if (st.window.length > st.cfg.windowSize) st.window.shift();
+        st.turns++;
+        if (st.turns % SAMPLE_EVERY === 0) appendLog({ ts: new Date().toISOString(), turn: st.turns, v: 3, len: thinking.length, pos: fp.pos, neg: fp.neg, neu: fp.neu, ratio: null, hits: turnHits, negRate: null, band: "none", phase: st.phase, want: "none", acted: "none", eff: null });
         return;
       }
       st.window.push(fp);
@@ -245,65 +294,95 @@ export default function thinkAnchor(pi: ExtensionAPI): void {
       lastHits = hits;
       lastNegRate = negRate;
       const ratio = ratioOf(agg.pos, agg.neg, agg.neu, st.cfg.weights);
-      const band = classify(hits, negRate, st.cfg.score);
+      const band = classify(agg.pos, agg.neg, agg.neu, st.cfg.score);
       const prevPhase = st.phase;
-      const d = decideNext(st.phase, band, st.l2Tried, st.cfg.intervention.maxL2);
+      const r = (agg.pos + 0.4 * agg.neu + st.cfg.score.alpha) / (agg.neg + 0.5 + st.cfg.score.alpha);
+      const d0 = decideNext(st.phase, band, st.l2Tried, st.cfg.intervention.maxL2);
+      // 升级前先看上次干预有没有用：无效则本次不升 L2（实测 L2 无收益）
+      const d = { ...d0, level: downgradeIfIneffective(d0.level, lastEffValue) };
       st.turns++;
-      st.lastRatio = ratio;
+      st.lastRatio = r;
       st.lastBand = band;
       if (band !== "none") st.bandHist[band]++; // none = 证据不足，不计波段（否则 NaN）
       let acted = "none";
       const now = Date.now();
       if (d.level === "none") {
         st.phase = d.phase;
-      } else if (d.level === "L1" && now - st.lastL1 > st.cfg.intervention.l1CooldownMin * 60_000) {
+      } else if (d.level === "L1" && !st.autoMuted && now - st.lastL1 > st.cfg.intervention.l1CooldownMin * 60_000) {
         st.lastL1 = now;
         st.l1Count++;
         (pi as unknown as { sendMessage?: (m: unknown, o?: unknown) => void }).sendMessage?.(
-          { customType: "think-anchor", content: L1_TEXT + `（发散占比 ${negRate.toFixed(2)} · 命中 ${hits}）`, display: false },
+          { customType: "think-anchor", content: L1_TEXT + `（比率 ${r.toFixed(2)} · 命中 ${hits}；react<0.5）`, display: false },
           { deliverAs: "nextTurn" },
         );
         acted = "L1";
+        st.pending = { turn: st.turns, pre: st.recent.slice(-3).reduce((a, b) => a + b, 0) / Math.max(1, Math.min(3, st.recent.length)), ts: Date.now(), seen: 0 };
         st.phase = d.phase;
-      } else if (d.level === "L2" && now - st.lastL2 > st.cfg.intervention.l2CooldownMin * 60_000) {
+      } else if (d.level === "L2" && !st.autoMuted && now - st.lastL2 > st.cfg.intervention.l2CooldownMin * 60_000) {
         st.lastL2 = now;
         st.l2Tried++;
         st.l2Count++;
         (pi as unknown as { sendMessage?: (m: unknown, o?: unknown) => void }).sendMessage?.(
-          { customType: "think-anchor", content: L2_TEXT + `（发散占比 ${negRate.toFixed(2)} · 命中 ${hits}，L2 ${st.l2Tried}/${st.cfg.intervention.maxL2}）`, display: false },
+          { customType: "think-anchor", content: L2_TEXT + `（比率 ${r.toFixed(2)} · 命中 ${hits}；L2 ${st.l2Tried}/${st.cfg.intervention.maxL2}）`, display: false },
           { deliverAs: "nextTurn" },
         );
         acted = "L2";
+        st.pending = { turn: st.turns, pre: st.recent.slice(-3).reduce((a, b) => a + b, 0) / Math.max(1, Math.min(3, st.recent.length)), ts: Date.now(), seen: 0 };
         st.phase = d.phase;
       } 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 = "blocked-" + d.level;
+        if (d.level !== "none") 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
-      appendLog({
-        ts: new Date().toISOString(),
-        turn: st.turns,
-        v: 2,
-        len: thinking.length,
-        pos: fp.pos,
-        neg: fp.neg,
-        neu: fp.neu,
-        hits,
-        negRate: Math.round(negRate * 100) / 100,
-        ratio: Math.round(ratio * 100) / 100,
-        band,
-        phase: st.phase,
-        want: d.level,
-        acted,
-      });
+      // v3 效果闭环：干预后第 3 轮结算（pre = 干预前最近 3 轮 ratio 均值）
+      st.recent.push(ratio);
+      if (st.recent.length > 6) st.recent.shift();
+      if (st.pending) st.pending.seen++;  // 用独立计数而非 turns（新会话 turns 归零，差值会为负）
+      let eff: number | null = null;
+      // 跨会话只在 30 分钟内结算（避免隔天比率混算）
+      const pendingFresh = st.pending !== null && Date.now() - st.pending.ts < 30 * 60_000;
+      if (st.pending && pendingFresh && st.pending.seen >= 3) {
+        const tail = st.recent.slice(-3);
+        const post = tail.reduce((a, b) => a + b, 0) / Math.max(1, tail.length);
+        eff = Math.round((st.pending.pre - post) * 100) / 100; // 负 = 改善（比率上升）
+        lastEffValue = eff;
+        st.effSamples++;
+        if (eff < 0) st.effImproved++;
+        st.pending = null;
+        // 自适应：样本够且实测下降率不及抛硬币 → 转静默（只记录，不注入）
+        st.autoMuted = st.effSamples >= AUTO_MIN && st.effImproved / st.effSamples < AUTO_RATE;
+      }
+      // 采样写盘：事件轮（有干预/有结算/波段变化）必写，其余每 SAMPLE_EVERY 轮一行
+      const eventful = acted !== "none" || eff !== null || band !== st.lastLoggedBand;
+      if (!eventful && st.turns % SAMPLE_EVERY !== 0) { st.skipped++; }
+      else {
+        st.lastLoggedBand = band;
+        appendLog({
+          ts: new Date().toISOString(),
+          turn: st.turns,
+          v: 3,
+          len: thinking.length,
+          pos: fp.pos,
+          neg: fp.neg,
+          neu: fp.neu,
+          hits,
+          negRate: Math.round(negRate * 100) / 100,
+          ratio: Math.round(r * 100) / 100,
+          band,
+          phase: st.phase,
+          want: d.level,
+          acted,
+          eff,
+        });
+      }
     } catch {
       /* 绝不抛错 */
     }
   });
 
   pi.registerCommand("think-anchor", {
-    description: "思维锚定监测：状态 / on / off / reset",
+    description: "思维锚定监测：状态 / stats（日志统计）/ on / off / reset",
     handler: async (args, ctx) => {
       const sub = (args || "").trim().toLowerCase();
       if (sub === "on") {
@@ -321,14 +400,38 @@ export default function thinkAnchor(pi: ExtensionAPI): void {
         if (ctx.hasUI) ctx.ui.notify("think-anchor: 会话状态已重置。", "info");
         return;
       }
+      if (sub === "stats") {
+        try {
+          const txt = readFileSync(LOG, "utf8");
+          const rows = txt.split("\n").filter((l) => l.trim()).map((l) => { try { return JSON.parse(l) as Record<string, unknown>; } catch { return null; } }).filter(Boolean) as Array<Record<string, unknown>>;
+          const group = (k: string) => {
+            const m = new Map<string, number>();
+            for (const r of rows) { const key = String(r[k] ?? "?"); m.set(key, (m.get(key) ?? 0) + 1); }
+            return [...m.entries()].map(([a, b]) => a + "=" + b).join(" ") || "-";
+          };
+          const effRows = rows.filter((r) => typeof r.eff === "number");
+          const improved = effRows.filter((r) => (r.eff as number) < 0).length;
+          const size = (statSync(LOG).size / 1048576).toFixed(2);
+          const lines = [
+            "think-anchor 统计：" + size + " MB / 上限 1.00 · " + rows.length + " 行",
+            " 版本 " + group("v") + " · 波段 " + group("band"),
+            " 动作 " + group("acted") + " · want " + group("want"),
+            " 效果结算 " + improved + "/" + effRows.length + (effRows.length ? "（下降 " + Math.round((100 * improved) / effRows.length) + "%）" : "（样本未满：干预后需再走 3 轮）"),
+          ];
+          if (ctx.hasUI) ctx.ui.notify(lines.join("\n"), "info");
+        } catch { if (ctx.hasUI) ctx.ui.notify("think-anchor: 统计失败（日志读取异常）", "warning"); }
+        return;
+      }
       const line =
         "think-anchor：" +
         (st.cfg.enabled ? "ON" : "OFF") +
         " · phase=" + st.phase +
         " · 消息 " + st.turns +
-        " · 波段 spec/mixed/react=" + st.bandHist.spec + "/" + st.bandHist.mixed + "/" + st.bandHist.react +
+        " · 波段 spec/mixed/react/none=" + st.bandHist.spec + "/" + st.bandHist.mixed + "/" + st.bandHist.react + "/" + st.bandHist.none +
         " · 干预 L1×" + st.l1Count + " L2×" + st.l2Count +
-        " · 上轮 命中" + lastHits + " 发散占比" + (Math.round(lastNegRate * 100) / 100) + "（" + st.lastBand + "）";
+        " · 有效率 " + st.effImproved + "/" + st.effSamples + (st.autoMuted ? "（不及抛硬币→已静默）" : st.effSamples === 0 ? "（样本未满：干预后 3 轮结算）" : st.effSamples < AUTO_MIN ? "（样本 " + st.effSamples + "/" + AUTO_MIN + "）" : "") +
+        " · 上轮 命中" + lastHits + " 发散占比" + (Math.round(lastNegRate * 100) / 100) + "（" + st.lastBand + "）" +
+        " · 采样跳过 " + st.skipped;
       if (ctx.hasUI) ctx.ui.notify(line, "info");
     },
   });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-think-anchor",
-  "version": "1.0.2",
+  "version": "1.0.4",
   "description": "Monitors thinking focus drift and nudges the model back on track.",
   "type": "module",
   "license": "MIT",
@@ -22,5 +22,6 @@
   ],
   "peerDependencies": {
     "@earendil-works/pi-coding-agent": "*"
-  }
+  },
+  "author": "15066212pp"
 }