@modusensus/dsh-mneme 0.2.4 → 0.2.6

package/test/inject.test.js

@@ -0,0 +1,82 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+import { createStore } from "../src/store.js";
+import { createService } from "../src/service.js";
+import { createInjector } from "../src/inject.js";
+import { createSettings } from "../src/settings.js";
+
+function setup(over = {}) {
+  const store = createStore(":memory:");
+  const service = createService({ store, mirror: null, config: {} });
+  const settings = createSettings(store.db);
+  const contexts = [];
+  const disposers = [];
+  const ctx = {
+    systemPrompt: {
+      context(def) {
+        contexts.push(def);
+        const dispose = () => disposers.push(def.name);
+        return dispose;
+      }
+    }
+  };
+  const config = { maxInjectedItems: 3, importanceThreshold: 3, ...over };
+  const injector = createInjector(ctx, service, settings, config);
+  return { store, service, contexts, injector, settings };
+}
+
+test("registers memory and user-settings contexts", () => {
+  const { contexts } = setup();
+  assert.equal(contexts.length, 2);
+  assert.equal(contexts[0].name, "memory");
+  assert.equal(contexts[1].name, "user-settings");
+});
+
+test("context text renders injected memories as markdown block", () => {
+  const { contexts, service } = setup();
+  service.saveWithDedupe({ type: "preference", title: "语言", content: "用户用中文交流", importance: 5 });
+  service.saveWithDedupe({ type: "project", title: "记忆插件", content: "SQLite+Markdown", importance: 4 });
+  const text = contexts[0].text({});
+  assert.ok(text.includes("[记忆库]"), "has header");
+  assert.ok(text.includes("语言"), "includes preference");
+  assert.ok(text.includes("记忆插件"), "includes high-importance project");
+});
+
+test("returns empty text when nothing qualifies", () => {
+  const { contexts } = setup();
+  const text = contexts[0].text({});
+  assert.equal(text, "");
+});
+
+test("renders summary block first when summary candidate exists", () => {
+  const { contexts, service } = setup();
+  service.saveWithDedupe({ type: "summary", title: "记忆库总览", content: "这是总览摘要", importance: 5 });
+  service.saveWithDedupe({ type: "preference", title: "语言", content: "中文", importance: 5 });
+  const text = contexts[0].text({});
+  const summaryIdx = text.indexOf("这是总览摘要");
+  const prefIdx = text.indexOf("语言");
+  assert.ok(summaryIdx !== -1, "summary present");
+  assert.ok(prefIdx !== -1, "preference present");
+  assert.ok(summaryIdx < prefIdx, "summary rendered first");
+});
+
+test("user-settings context renders profile and rules, empty when unset", () => {
+  const { contexts, settings } = setup();
+  const settingsCtx = contexts.find((c) => c.name === "user-settings");
+  assert.ok(settingsCtx, "user-settings context registered");
+  assert.equal(settingsCtx.text({}), "", "empty when no profile/rules");
+  settings.setProfile("我叫小明，是一名前端开发者");
+  settings.setRules(["回答时先给结论", "使用简体中文"]);
+  const text = settingsCtx.text({});
+  assert.ok(text.includes("用户画像"), "has profile header");
+  assert.ok(text.includes("前端开发者"), "includes profile");
+  assert.ok(text.includes("先给结论"), "includes rule");
+  assert.ok(text.includes("简体中文"), "includes second rule");
+});
+
+test("user-settings context precedes memory block (order 85 < 90)", () => {
+  const { contexts, settings } = setup();
+  settings.setProfile("画像");
+  const settingsCtx = contexts.find((c) => c.name === "user-settings");
+  assert.ok(settingsCtx.order < contexts.find((c) => c.name === "memory").order);
+});

package/test/local-embedder.test.js

@@ -0,0 +1,227 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+import {
+  LocalEmbedder,
+  OllamaEmbedder,
+  OpenAIEmbedder,
+  createEmbedderByProvider
+} from "../src/local-embedder.js";
+
+/** Build a transformers.js-like Tensor: [rows x dim] flat data + dims. */
+function fakeTensor(rows) {
+  const flat = Float32Array.from(rows.flat());
+  return { data: flat, dims: rows.length ? [rows.length, rows[0].length] : [0] };
+}
+
+/** Fake extractor: deterministic [batch, dim], records option passthrough. */
+function makeFakeExtractor(dim, calls = []) {
+  const fn = async (texts, opts) => {
+    calls.push({ count: texts.length, opts });
+    return fakeTensor(
+      texts.map((_, i) => Array.from({ length: dim }, (__, j) => (i + 1) * 0.1 + j * 0.01))
+    );
+  };
+  fn.dispose = () => {
+    fn.disposed = true;
+  };
+  return fn;
+}
+
+/** Stub globalThis.fetch, returning the original on restore. */
+function stubFetch(handler) {
+  const real = globalThis.fetch;
+  globalThis.fetch = async (url, init) => handler(url, init);
+  return () => {
+    globalThis.fetch = real;
+  };
+}
+
+function okJson(status, body) {
+  return {
+    ok: status >= 200 && status < 300,
+    status,
+    json: async () => body
+  };
+}
+
+test("LocalEmbedder init loads via injected engineFactory with cache_dir", async () => {
+  let seen = null;
+  const loader = async (task, model, options) => {
+    seen = { task, model, options };
+    return makeFakeExtractor(512);
+  };
+  const e = new LocalEmbedder({ cacheDir: "/tmp/model-cache", engineFactory: loader });
+  await e.init();
+  assert.equal(seen.task, "feature-extraction");
+  assert.equal(seen.model, "Xenova/bge-small-zh-v1.5");
+  assert.equal(seen.options.cache_dir, "/tmp/model-cache");
+  assert.equal(seen.options.dtype, "q8");
+  assert.equal(seen.options.device, "cpu");
+});
+
+test("LocalEmbedder embed returns [n, dim] with mean pooling and chunking", async () => {
+  const calls = [];
+  const e = new LocalEmbedder({
+    model: "Xenova/bge-small-zh-v1.5",
+    batchSize: 2,
+    engineFactory: async () => makeFakeExtractor(512, calls)
+  });
+  await e.init();
+  const vecs = await e.embed(["你好", "hello world", "测试文本"]);
+  assert.equal(vecs.length, 3);
+  for (const v of vecs) {
+    assert.equal(v.length, 512);
+    assert.ok(v.every(Number.isFinite));
+  }
+  // 3 texts at batchSize 2 -> two extractor calls.
+  assert.deepEqual(calls.map((c) => c.count), [2, 1]);
+  assert.deepEqual(calls[0].opts, { pooling: "mean", normalize: true });
+});
+
+test("LocalEmbedder embedSingle returns a single vector", async () => {
+  const e = new LocalEmbedder({ engineFactory: async () => makeFakeExtractor(384) });
+  await e.init();
+  const v = await e.embedSingle("single");
+  assert.equal(v.length, 384);
+});
+
+test("LocalEmbedder dimension and modelHash are stable", () => {
+  const a = new LocalEmbedder({ model: "Xenova/bge-small-zh-v1.5" });
+  const b = new LocalEmbedder({ model: "Xenova/bge-small-zh-v1.5" });
+  assert.equal(a.dimension, 512);
+  assert.equal(a.modelHash, b.modelHash);
+  assert.match(a.modelHash, /^Xenova\/bge-small-zh-v1\.5#[0-9a-f]+$/);
+  // Different model -> different hash.
+  assert.notEqual(a.modelHash, new LocalEmbedder({ model: "other/model" }).modelHash);
+});
+
+test("LocalEmbedder throws before init and after dispose", async () => {
+  const e = new LocalEmbedder({ engineFactory: async () => makeFakeExtractor(512) });
+  await assert.rejects(() => e.embed(["x"]), /not initialized/);
+  await e.init();
+  const disposed = makeFakeExtractor(512);
+  e.dispose();
+  assert.ok(disposed.disposed || e.extractor === null);
+  await assert.rejects(() => e.embed(["x"]), /not initialized/);
+});
+
+test("OllamaEmbedder init probes server and infers dimension", async () => {
+  const restore = stubFetch(async (url, init) => {
+    assert.equal(url, "http://localhost:11434/api/embeddings");
+    assert.equal(init.method, "POST");
+    const body = JSON.parse(init.body);
+    assert.equal(body.model, "nomic-embed-text");
+    assert.equal(body.prompt, "ping");
+    return okJson(200, { embedding: Array.from({ length: 768 }, (_, i) => i / 768) });
+  });
+  try {
+    const e = new OllamaEmbedder({ baseUrl: "http://localhost:11434", model: "nomic-embed-text" });
+    await e.init();
+    assert.equal(e.dimension, 768);
+  } finally {
+    restore();
+  }
+});
+
+test("OllamaEmbedder embed loops single requests", async () => {
+  const prompts = [];
+  const restore = stubFetch(async (url, init) => {
+    const body = JSON.parse(init.body);
+    prompts.push(body.prompt);
+    return okJson(200, { embedding: [0.5, 0.25, 0.125] });
+  });
+  try {
+    const e = new OllamaEmbedder({});
+    await e.init();
+    const vecs = await e.embed(["one", "two"]);
+    assert.deepEqual(prompts, ["ping", "one", "two"]);
+    assert.equal(vecs.length, 2);
+    assert.deepEqual(vecs[0], [0.5, 0.25, 0.125]);
+    assert.equal(e.dimension, 3);
+  } finally {
+    restore();
+  }
+});
+
+test("OllamaEmbedder init throws when unreachable", async () => {
+  const restore = stubFetch(async () => {
+    throw new Error("ECONNREFUSED");
+  });
+  try {
+    const e = new OllamaEmbedder({});
+    await assert.rejects(() => e.init(), /ECONNREFUSED/);
+  } finally {
+    restore();
+  }
+});
+
+test("OllamaEmbedder embed throws on HTTP error", async () => {
+  const restore = stubFetch(async () => okJson(404, { error: "model not found" }));
+  try {
+    const e = new OllamaEmbedder({});
+    await assert.rejects(() => e.embedSingle("x"), /HTTP 404/);
+  } finally {
+    restore();
+  }
+});
+
+test("OllamaEmbedder modelHash is stable", () => {
+  const a = new OllamaEmbedder({ model: "nomic-embed-text" });
+  const b = new OllamaEmbedder({ model: "nomic-embed-text" });
+  assert.equal(a.modelHash, b.modelHash);
+  assert.notEqual(a.modelHash, new OllamaEmbedder({ model: "bge-m3" }).modelHash);
+});
+
+test("OpenAIEmbedder init requires config and calls /embeddings", async () => {
+  const e = new OpenAIEmbedder({ baseUrl: "http://localhost:8000/v1", apiKey: "k", model: "m" });
+  assert.equal(e._url, "http://localhost:8000/v1/embeddings");
+  await e.init();
+
+  const bad = new OpenAIEmbedder({ baseUrl: "", apiKey: "", model: "" });
+  await assert.rejects(() => bad.init(), /requires baseUrl, apiKey and model/);
+});
+
+test("OpenAIEmbedder embed batch posts array and parses data", async () => {
+  let captured = null;
+  const restore = stubFetch(async (url, init) => {
+    captured = { url, init };
+    return okJson(200, {
+      data: [
+        { embedding: [0.1, 0.2, 0.3] },
+        { embedding: [0.4, 0.5, 0.6] }
+      ]
+    });
+  });
+  try {
+    const e = new OpenAIEmbedder({ baseUrl: "https://x/v1", apiKey: "sk-abc", model: "text-embed" });
+    await e.init();
+    const vecs = await e.embed(["a", "b"]);
+    assert.equal(captured.url, "https://x/v1/embeddings");
+    assert.equal(captured.init.headers.Authorization, "Bearer sk-abc");
+    const body = JSON.parse(captured.init.body);
+    assert.equal(body.model, "text-embed");
+    assert.deepEqual(body.input, ["a", "b"]);
+    assert.equal(vecs.length, 2);
+    assert.equal(e.dimension, 3);
+  } finally {
+    restore();
+  }
+});
+
+test("OpenAIEmbedder throws on non-ok response", async () => {
+  const restore = stubFetch(async () => okJson(500, {}));
+  try {
+    const e = new OpenAIEmbedder({ baseUrl: "https://x/v1", apiKey: "k", model: "m" });
+    await assert.rejects(() => e.embed(["x"]), /HTTP 500/);
+  } finally {
+    restore();
+  }
+});
+
+test("createEmbedderByProvider returns the right class per provider", () => {
+  assert.ok(createEmbedderByProvider("local") instanceof LocalEmbedder);
+  assert.ok(createEmbedderByProvider("ollama") instanceof OllamaEmbedder);
+  assert.ok(createEmbedderByProvider("openai") instanceof OpenAIEmbedder);
+  assert.ok(createEmbedderByProvider("LOCAL") instanceof LocalEmbedder);
+  assert.throws(() => createEmbedderByProvider("watson"), /Unknown embedding provider/);
+});

package/test/mirror.test.js

@@ -0,0 +1,249 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+import { mkdtempSync, readFileSync, writeFileSync, rmSync, existsSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { createMirror } from "../src/mirror.js";
+import { createStore } from "../src/store.js";
+import { createService } from "../src/service.js";
+
+const TYPE_FILE = { preference: "preferences.md", project: "projects.md", decision: "decisions.md", history: "history.md" };
+
+function tempDir() {
+  return mkdtempSync(join(tmpdir(), "dsh-mneme-mirror-"));
+}
+
+function sampleMemory(type, over = {}) {
+  return {
+    id: "m1", type, title: "标题", content: "内容",
+    tags: ["a"], importance: 3, forgotten: false,
+    source: undefined, created_at: "2026-01-01T00:00:00.000Z",
+    updated_at: "2026-01-01T00:00:00.000Z", ...over
+  };
+}
+
+test("mirror writes one markdown file per type on sync", () => {
+  const dir = tempDir();
+  try {
+    const mirror = createMirror(dir);
+    mirror.sync([
+      sampleMemory("preference"),
+      sampleMemory("project")
+    ]);
+    assert.ok(mirror.filePath("preference").endsWith("preferences.md"));
+    for (const type of ["preference", "project"]) {
+      const text = readFileSync(join(dir, TYPE_FILE[type]), "utf8");
+      assert.ok(text.includes("m1"), `${type} file contains id`);
+      assert.ok(text.includes("标题"), `${type} file contains title`);
+      assert.ok(text.includes("内容"), `${type} file contains content`);
+    }
+    assert.ok(mirror.filePath("decision").includes("decisions"));
+  } finally {
+    rmSync(dir, { recursive: true, force: true });
+  }
+});
+
+test("mirror groups multiple memories newest-first with header", () => {
+  const dir = tempDir();
+  try {
+    const mirror = createMirror(dir);
+    mirror.sync([
+      sampleMemory("project", { id: "old", updated_at: "2026-01-01T00:00:00.000Z" }),
+      sampleMemory("project", { id: "new", updated_at: "2026-02-01T00:00:00.000Z" })
+    ]);
+    const text = readFileSync(join(dir, "projects.md"), "utf8");
+    const iNew = text.indexOf("new");
+    const iOld = text.indexOf("old");
+    assert.ok(iNew !== -1 && iOld !== -1 && iNew < iOld, "newest first");
+  } finally {
+    rmSync(dir, { recursive: true, force: true });
+  }
+});
+
+test("human edit wins on next sync (bidirectional, human-first)", () => {
+  const dir = tempDir();
+  try {
+    const mirror = createMirror(dir);
+    mirror.sync([sampleMemory("preference", { id: "m1", content: "机器内容" })]);
+    // human edits the mirror file
+    const file = join(dir, "preferences.md");
+    const edited = readFileSync(file, "utf8").replace("机器内容", "人类编辑内容");
+    writeFileSync(file, edited, "utf8");
+    const humanEdits = mirror.readHumanEdits();
+    assert.ok(Array.isArray(humanEdits));
+    const m1 = humanEdits.find((e) => e.id === "m1");
+    assert.ok(m1, "detects human edit for m1");
+    assert.equal(m1.content, "人类编辑内容");
+  } finally {
+    rmSync(dir, { recursive: true, force: true });
+  }
+});
+
+test("readHumanEdits content is not polluted by file header", () => {
+  const dir = tempDir();
+  try {
+    const mirror = createMirror(dir);
+    mirror.sync([sampleMemory("preference", { id: "m1", content: "机器内容" })]);
+    const m1 = mirror.readHumanEdits("preference").find((e) => e.id === "m1");
+    assert.ok(m1, "detects m1");
+    assert.equal(m1.content, "机器内容");
+    assert.ok(!m1.content.includes("#"), "no H1 header in content");
+    assert.ok(!m1.content.includes("dsh-mneme 镜像"), "no mirror banner in content");
+    assert.ok(!m1.content.includes("<!--"), "no html comment in content");
+  } finally {
+    rmSync(dir, { recursive: true, force: true });
+  }
+});
+
+test("body containing '---' survives round-trip (no silent truncation)", () => {
+  const dir = tempDir();
+  try {
+    const mirror = createMirror(dir);
+    const content = "第一段\n\n---\n\n第二段";
+    mirror.sync([sampleMemory("project", { id: "p1", content })]);
+    const p1 = mirror.readHumanEdits("project").find((e) => e.id === "p1");
+    assert.ok(p1, "detects p1");
+    assert.equal(p1.content, content);
+  } finally {
+    rmSync(dir, { recursive: true, force: true });
+  }
+});
+
+test("body lines resembling metadata are preserved", () => {
+  const dir = tempDir();
+  try {
+    const mirror = createMirror(dir);
+    const content = "- **ID**: 假条目\n- **重要性**: 5（正文里写的）";
+    mirror.sync([sampleMemory("preference", { id: "m1", content })]);
+    const m1 = mirror.readHumanEdits("preference").find((e) => e.id === "m1");
+    assert.ok(m1, "detects m1");
+    assert.equal(m1.content, content);
+  } finally {
+    rmSync(dir, { recursive: true, force: true });
+  }
+});
+
+test("body line in machine ID format does not split entry or create ghost entry", () => {
+  const dir = tempDir();
+  try {
+    const mirror = createMirror(dir);
+    const content = "- **ID**: `phantom`\n- **重要性**: 5（正文里写的）";
+    mirror.sync([sampleMemory("preference", { id: "m1", content })]);
+    const edits = mirror.readHumanEdits("preference");
+    const m1 = edits.find((e) => e.id === "m1");
+    assert.ok(m1, "detects m1");
+    assert.equal(m1.content, content);
+    assert.ok(!edits.some((e) => e.id === "phantom"), "no ghost entry for phantom");
+  } finally {
+    rmSync(dir, { recursive: true, force: true });
+  }
+});
+
+test("sync with empty array deletes stale mirror files", () => {
+  const dir = tempDir();
+  try {
+    const mirror = createMirror(dir);
+    mirror.sync([
+      sampleMemory("preference"),
+      sampleMemory("project"),
+      sampleMemory("decision"),
+      sampleMemory("history")
+    ]);
+    for (const type of Object.keys(TYPE_FILE)) {
+      assert.ok(existsSync(join(dir, TYPE_FILE[type])), `${type} file exists`);
+    }
+    mirror.sync([]);
+    for (const type of Object.keys(TYPE_FILE)) {
+      assert.ok(!existsSync(join(dir, TYPE_FILE[type])), `${type} file removed`);
+    }
+  } finally {
+    rmSync(dir, { recursive: true, force: true });
+  }
+});
+
+test("startup merge keeps human edits across multiple type files (read-all-then-merge)", () => {
+  const dir = tempDir();
+  try {
+    const store = createStore(join(dir, "memory.db"));
+    const mirror = createMirror(dir);
+    const service = createService({ store, mirror, config: {} });
+    service.saveWithDedupe({ type: "preference", title: "语言", content: "机器内容" });
+    service.saveWithDedupe({ type: "project", title: "项目", content: "机器内容" });
+    // human edits BOTH mirror files before the startup merge
+    for (const [file, to] of [["preferences.md", "人工偏好"], ["projects.md", "人工项目"]]) {
+      const p = join(dir, file);
+      writeFileSync(p, readFileSync(p, "utf8").replace("机器内容", to), "utf8");
+    }
+    // replicate index.js startup: read EVERY type's edits first, then merge
+    // (a per-type read-then-merge loop would let the first syncMirror
+    // overwrite the unread type's file and lose the edit)
+    const byType = new Map();
+    for (const type of Object.keys(TYPE_FILE)) byType.set(type, mirror.readHumanEdits(type));
+    for (const [type, edits] of byType) if (edits.length) service.mergeHumanEdits(type, edits);
+    const prefs = service.list({ type: "preference", includeForgotten: true });
+    const projects = service.list({ type: "project", includeForgotten: true });
+    assert.ok(prefs.some((m) => m.content.includes("人工偏好")), "preference edit survives");
+    assert.ok(projects.some((m) => m.content.includes("人工项目")), "project edit survives");
+    store.close();
+  } finally {
+    rmSync(dir, { recursive: true, force: true });
+  }
+});
+
+// --- item ④: runtime human edits are reconciled on sync (not only at startup)
+
+test("runtime human edit survives the next sync (human wins, no silent overwrite)", () => {
+  const dir = tempDir();
+  try {
+    const store = createStore(join(dir, "memory.db"));
+    const mirror = createMirror(dir);
+    const service = createService({ store, mirror, config: {} });
+    service.saveWithDedupe({ type: "preference", title: "语言", content: "机器内容" });
+    // human edits the mirror file AFTER the initial sync
+    const file = join(dir, "preferences.md");
+    writeFileSync(file, readFileSync(file, "utf8").replace("机器内容", "人类编辑内容"), "utf8");
+    // the next unrelated store write triggers syncMirror → must merge the edit back
+    service.saveWithDedupe({ type: "project", title: "无关", content: "x" });
+    const prefs = service.list({ type: "preference", includeArchived: true });
+    const m = prefs.find((p) => p.title === "语言");
+    assert.equal(m.content, "人类编辑内容", "human edit merged back into the store");
+    store.close();
+  } finally {
+    rmSync(dir, { recursive: true, force: true });
+  }
+});
+
+test("concurrent human edit + store update produces a three-way conflict marker, no side lost", () => {
+  const dir = tempDir();
+  try {
+    const store = createStore(join(dir, "memory.db"));
+    const mirror = createMirror(dir);
+    const service = createService({ store, mirror, config: {} });
+    const { memory: m } = service.saveWithDedupe({ type: "preference", title: "语言", content: "机器内容" });
+    // human edits the file…
+    const file = join(dir, "preferences.md");
+    writeFileSync(file, readFileSync(file, "utf8").replace("机器内容", "人类编辑内容"), "utf8");
+    // …while the store is concurrently updated → next sync sees both sides changed
+    service.update(m.id, { content: "并发机器版本" });
+    const updated = service.getById(m.id);
+    assert.ok(updated.content.includes("人类编辑内容"), "human edit is kept as the head");
+    assert.ok(updated.content.includes("并发机器版本"), "store's concurrent version is preserved");
+    assert.ok(updated.content.includes("并发冲突"), "conflict marker appended");
+    store.close();
+  } finally {
+    rmSync(dir, { recursive: true, force: true });
+  }
+});
+
+test("readHumanEdits records the machine-written updated_at as the version token", () => {
+  const dir = tempDir();
+  try {
+    const mirror = createMirror(dir);
+    mirror.sync([sampleMemory("preference", { id: "m1", content: "机器内容", updated_at: "2026-03-01T00:00:00.000Z" })]);
+    const m1 = mirror.readHumanEdits("preference").find((e) => e.id === "m1");
+    assert.ok(m1, "detects m1");
+    assert.equal(m1.updated_at, "2026-03-01T00:00:00.000Z", "updated_at captured for three-way merge");
+  } finally {
+    rmSync(dir, { recursive: true, force: true });
+  }
+});