test: Proposal A v3 - Unit Tests for Feedback Scoring (Phase 4)

src/reflection-slices.ts

@@ -316,3 +316,35 @@ export function extractReflectionSliceItems(reflectionText: string): ReflectionS
 export function extractInjectableReflectionSliceItems(reflectionText: string): ReflectionSliceItem[] {
   return buildReflectionSliceItemsFromSlices(extractInjectableReflectionSlices(reflectionText));
 }
+
+/**
+ * 判斷回應是否實際使用了注入的記憶 ID 或摘要。
+ * - 回應長度 <= 24：不視為使用（太短）
+ * - injectedIds + injectedSummaries 都為空：不視為使用
+ * - 有 usage marker（如「教練我記得」）+ 對應 ID：視為使用
+ * - 有 usage marker + verbatim summary match（>=10 字元）：視為使用
+ */
+export function isRecallUsed(
+  response: string,
+  injectedIds: string[],
+  injectedSummaries: string[]
+): boolean {
+  if (response.length <= 24) return false;
+  if (injectedIds.length === 0 && injectedSummaries.length === 0) return false;
+
+  const hasUsageMarker = /教練|教練我|教練我記得|記得|memory|id[-:]/i.test(response);
+
+  if (injectedIds.length > 0) {
+    const hasMatchingId = injectedIds.some(id => response.includes(id));
+    if (hasMatchingId && hasUsageMarker) return true;
+  }
+
+  if (injectedSummaries.length > 0) {
+    const hasMatchingSummary = injectedSummaries.some(
+      s => s.length >= 10 && response.includes(s)
+    );
+    if (hasMatchingSummary) return true;
+  }
+
+  return false;
+}

test/bad-recall-count.test.mjs

@@ -0,0 +1,33 @@
+// test/bad-recall-count.test.mjs
+// 測試 bad_recall_count 遞增邏輯（純邏輯測試，mock store 行為）
+import { describe, it } from 'node:test';
+import assert from 'node:assert';
+
+// 測試 bad_recall_count 的遞增邏輯（純邏輯測試，mock store）
+describe("bad_recall_count logic", () => {
+  function computeNextBadCount(current, isMiss, isConfirm, recallCount, minPenaltyThreshold) {
+    if (isConfirm) return 0;  // 確認使用，重置為 0
+    if (!isMiss) return current;  // 既不是 miss 也不是 confirm，保持現值
+    if (recallCount < minPenaltyThreshold) return current;  // recall 次數不夠，不 penalty
+    return current + 1;  // miss + 足夠次數，遞增
+  }
+
+  it("confirm resets count to 0", () => {
+    assert.strictEqual(computeNextBadCount(5, false, true, 3, 2), 0);
+  });
+  it("miss with insufficient recall count does not increment", () => {
+    assert.strictEqual(computeNextBadCount(0, true, false, 1, 2), 0);
+  });
+  it("miss with sufficient recall count increments", () => {
+    assert.strictEqual(computeNextBadCount(1, true, false, 2, 2), 2);
+  });
+  it("non-miss non-confirm keeps current value", () => {
+    assert.strictEqual(computeNextBadCount(3, false, false, 3, 2), 3);
+  });
+  it("reaches penalty threshold at badCount=2", () => {
+    // badCount >= 2 會觸發 penalty
+    const badCount = 2;
+    const isPenaltyTriggered = badCount >= 2;
+    assert.strictEqual(isPenaltyTriggered, true);
+  });
+});

test/feedback-config.test.mjs

@@ -0,0 +1,100 @@
+// test/feedback-config.test.mjs
+// 測試 FeedbackConfigManager（Phase 3 回饋信號反饋配置管理器）
+import { describe, it } from 'node:test';
+import assert from 'node:assert';
+
+// FeedbackConfigManager mock（從 Phase 3 來的實作）
+class FeedbackConfigManager {
+  constructor(config) {
+    this.config = config;
+  }
+  computeImportanceDelta(event, recallCount = 1, badRecallCount = 0) {
+    if (event === 'use') {
+      if (recallCount < this.config.minRecallCountForBoost) return 0;
+      return this.config.importanceBoostOnUse;
+    }
+    if (event === 'confirm') {
+      return this.config.importanceBoostOnConfirm;
+    }
+    if (event === 'miss') {
+      if (recallCount < this.config.minRecallCountForPenalty) return 0;
+      return -this.config.importancePenaltyOnMiss;
+    }
+    if (event === 'error') {
+      return -this.config.importancePenaltyOnError;
+    }
+    return 0;
+  }
+  isConfirmKeyword(text) {
+    return this.config.confirmKeywords.some(k => text.toLowerCase().includes(k.toLowerCase()));
+  }
+  isErrorKeyword(text) {
+    return this.config.errorKeywords.some(k => text.toLowerCase().includes(k.toLowerCase()));
+  }
+  static defaultConfig() {
+    return {
+      importanceBoostOnUse: 0.05,
+      importanceBoostOnConfirm: 0.15,
+      importancePenaltyOnMiss: 0.03,
+      importancePenaltyOnError: 0.10,
+      minRecallCountForPenalty: 2,
+      minRecallCountForBoost: 1,
+      confirmKeywords: ["是對的", "確認", "正確", "right"],
+      errorKeywords: ["錯誤", "不對", "wrong", "not right"],
+    };
+  }
+}
+
+describe("FeedbackConfigManager", () => {
+  describe("computeImportanceDelta", () => {
+    it("use event with recallCount >= minRecallCountForBoost returns boostOnUse", () => {
+      const mgr = new FeedbackConfigManager(FeedbackConfigManager.defaultConfig());
+      const delta = mgr.computeImportanceDelta('use', 1, 0);
+      assert.strictEqual(delta, 0.05);
+    });
+    it("confirm event returns boostOnConfirm", () => {
+      const mgr = new FeedbackConfigManager(FeedbackConfigManager.defaultConfig());
+      const delta = mgr.computeImportanceDelta('confirm', 1, 0);
+      assert.strictEqual(delta, 0.15);
+    });
+    it("miss event with recallCount < minRecallCountForPenalty returns 0", () => {
+      const mgr = new FeedbackConfigManager(FeedbackConfigManager.defaultConfig());
+      const delta = mgr.computeImportanceDelta('miss', 1, 0);  // recallCount=1 < 2
+      assert.strictEqual(delta, 0);
+    });
+    it("miss event with recallCount >= minRecallCountForPenalty returns penalty", () => {
+      const mgr = new FeedbackConfigManager(FeedbackConfigManager.defaultConfig());
+      const delta = mgr.computeImportanceDelta('miss', 2, 0);  // recallCount=2 >= 2
+      assert.strictEqual(delta, -0.03);
+    });
+    it("error event returns error penalty", () => {
+      const mgr = new FeedbackConfigManager(FeedbackConfigManager.defaultConfig());
+      const delta = mgr.computeImportanceDelta('error', 1, 0);
+      assert.strictEqual(delta, -0.10);
+    });
+  });
+  describe("isConfirmKeyword", () => {
+    it("detects 是對的", () => {
+      const mgr = new FeedbackConfigManager(FeedbackConfigManager.defaultConfig());
+      assert.strictEqual(mgr.isConfirmKeyword("教練我覺得是對的"), true);
+    });
+    it("detects right", () => {
+      const mgr = new FeedbackConfigManager(FeedbackConfigManager.defaultConfig());
+      assert.strictEqual(mgr.isConfirmKeyword("that's right"), true);
+    });
+    it("rejects unrelated text", () => {
+      const mgr = new FeedbackConfigManager(FeedbackConfigManager.defaultConfig());
+      assert.strictEqual(mgr.isConfirmKeyword("今天天氣很好"), false);
+    });
+  });
+  describe("isErrorKeyword", () => {
+    it("detects 錯誤", () => {
+      const mgr = new FeedbackConfigManager(FeedbackConfigManager.defaultConfig());
+      assert.strictEqual(mgr.isErrorKeyword("教練這是錯誤的"), true);
+    });
+    it("rejects unrelated text", () => {
+      const mgr = new FeedbackConfigManager(FeedbackConfigManager.defaultConfig());
+      assert.strictEqual(mgr.isErrorKeyword("今天天氣很好"), false);
+    });
+  });
+});

test/isRecallUsed.test.mjs

@@ -0,0 +1,42 @@
+// test/isRecallUsed.test.mjs
+// 測試 isRecallUsed() 函式 - 判斷回應是否實際使用了注入的記憶 ID 或摘要
+import { describe, it } from 'node:test';
+import assert from 'node:assert';
+import { isRecallUsed } from '../src/reflection-slices.ts';
+
+describe("isRecallUsed", () => {
+  it("returns false for short response (<=24 chars)", () => {
+    assert.strictEqual(isRecallUsed("hi", ["id1"], []), false);
+  });
+  it("returns false when both injectedIds and injectedSummaries are empty", () => {
+    assert.strictEqual(isRecallUsed("這是一個很長的回應內容這是", [], []), false);
+  });
+  it("returns true when injected ID is present AND usage marker is present", () => {
+    const response = "教練我記得這件事 memory id-abc123";
+    const injectedIds = ["id-abc123"];
+    assert.strictEqual(isRecallUsed(response, injectedIds, []), true);
+  });
+  it("returns false when only ID is present but no usage marker", () => {
+    const response = "我提到了id-abc123這個項目";
+    const injectedIds = ["id-abc123"];
+    assert.strictEqual(isRecallUsed(response, injectedIds, []), false);
+  });
+  it("returns false when only usage marker is present but no ID", () => {
+    const response = "教練我記得這件事但沒有提到任何ID";
+    const injectedIds = ["id-abc123"];
+    assert.strictEqual(isRecallUsed(response, injectedIds, []), false);
+  });
+  it("returns true for verbatim summary match (>=10 chars)", () => {
+    // 回應長度 > 24，且包含已注入摘要（摘要為回應的子字串，且 >= 10 字元）
+    const response = "教練xx這是關於Python的import機制的詳細說明";
+    const injectedIds = [];
+    const injectedSummaries = ["這是關於Python的import機制的詳細說明"];
+    assert.strictEqual(isRecallUsed(response, injectedIds, injectedSummaries), true);
+  });
+  it("returns false for short summary (<10 chars)", () => {
+    const response = "教練提到了test這個詞";
+    const injectedIds = [];
+    const injectedSummaries = ["test"];
+    assert.strictEqual(isRecallUsed(response, injectedIds, injectedSummaries), false);
+  });
+});