@modusensus/dsh-mneme 0.3.6 → 0.3.7

package/README.md

@@ -265,9 +265,6 @@ npm run sync       # 把 src/ 同步到 lib/（发布时由 prepack 钩子自动
 
 > 设计文档位于仓库根 `docs/`，链接以 `../docs/` 相对路径指向（GitHub 上从本目录打开可正常跳转）。
 
-- [记忆库设计](../docs/superpowers/specs/2026-08-13-dsh-mneme-design.md)
-- [autoDream 设计](../docs/superpowers/specs/2026-08-13-dsh-mneme-autodream-design.md)
-- [实施计划](../docs/superpowers/plans/2026-08-13-dsh-memory-autodream.md)
 - [实体结构化记忆设计](docs/ENTITIES.md)
 - [语义增强架构](docs/SEMANTIC.md)
 - [本地模型部署指南](docs/LOCAL_MODEL.md)

package/lib/index.js

@@ -80,11 +80,28 @@ export const apply = (ctx, config) => {
   const vectorIndex = createVectorIndex({ store, logger: ctx.logger });
   service.setVectorIndex(vectorIndex);
 
+  // Human edits in mirror files win on every sync; merge them back first.
+  // TYPE_FILE maps each memory type to its mirror filename. Read every type's
+  // edits up front: mergeHumanEdits re-renders ALL mirror files on success, so
+  // a per-type read-then-merge loop would overwrite edits in files not yet read
+  // (e.g. preferences.md merging would clobber unsynced projects.md edits).
+  const humanEdits = new Map();
+  for (const type of Object.keys(TYPE_FILE)) {
+    humanEdits.set(type, mirror.readHumanEdits(type));
+  }
+  const applyHumanEdits = () => {
+    for (const [type, edits] of humanEdits) {
+      if (edits.length) service.mergeHumanEdits(type, edits);
+    }
+  };
+
   let embedder = null;
   let reranker = null;
   if (cfg.embedProvider === "openai") {
     embedder = createEmbedder({ store, settings, logger: ctx.logger });
     service.setEmbedder(embedder);
+    // legacy OpenAI embedder is immediately usable
+    applyHumanEdits();
   } else {
     try {
       embedder = createEmbedderByProvider(cfg.embedProvider, {
@@ -97,12 +114,18 @@ export const apply = (ctx, config) => {
         logger: ctx.logger
       });
       service.setEmbedder(embedder);
-      embedder.init().catch((error) => {
-        ctx.logger?.warn?.(`[dsh-mneme] embedder init failed, search degrades to keyword: ${String(error)}`);
-        service.setEmbedder(null);
-      });
+      // issue #6: wait for extractor init before applying human edits, so
+      // scheduled embeddings see a ready embedder.
+      embedder.init()
+        .then(() => applyHumanEdits())
+        .catch((error) => {
+          ctx.logger?.warn?.(`[dsh-mneme] embedder init failed, search degrades to keyword: ${String(error)}`);
+          service.setEmbedder(null);
+          applyHumanEdits();
+        });
     } catch (error) {
       ctx.logger?.warn?.(`[dsh-mneme] embedder unavailable, search degrades to keyword: ${String(error)}`);
+      applyHumanEdits();
     }
   }
 
@@ -139,19 +162,6 @@ export const apply = (ctx, config) => {
     commands.sync();
   }
 
-  // Human edits in mirror files win on every sync; merge them back first.
-  // TYPE_FILE maps each memory type to its mirror filename. Read every type's
-  // edits up front: mergeHumanEdits re-renders ALL mirror files on success, so
-  // a per-type read-then-merge loop would overwrite edits in files not yet read
-  // (e.g. preferences.md merging would clobber unsynced projects.md edits).
-  const humanEdits = new Map();
-  for (const type of Object.keys(TYPE_FILE)) {
-    humanEdits.set(type, mirror.readHumanEdits(type));
-  }
-  for (const [type, edits] of humanEdits) {
-    if (edits.length) service.mergeHumanEdits(type, edits);
-  }
-
   // Dream scheduler: automatic consolidation + summary runs, triggered by
   // store growth. Writes through the service fire the dream hook, which asks
   // the scheduler to (re)schedule a run once absolute and since-last-run

package/lib/local-embedder.js

@@ -58,16 +58,21 @@ export class LocalEmbedder {
     // Test hook: replace the pipeline factory without touching modules.
     this.engineFactory = opts.engineFactory || defaultPipelineLoader;
     this.extractor = null;
+    // issue #6: readiness flag for the service's scheduleEmbed gate. False until
+    // init() succeeds, so "ready" in embedder is observable even pre-init.
+    this.ready = false;
   }
 
-  /** Load the model; throws when it cannot be loaded. */
+  /** Load the model; throws when it cannot be loaded. Idempotent. */
   async init() {
+    if (this.extractor) return this; // already initialized: no-op
     const options = {
       dtype: this.useDtype,
       device: this.device
     };
     if (this.cacheDir) options.cache_dir = this.cacheDir;
     this.extractor = await this.engineFactory("feature-extraction", this.model, options);
+    this.ready = true; // service reads this to flush queued re-embeds
     this.logger?.info?.(
       `[dsh-mneme] local embedder ready: ${this.model} (dim=${this._dimension}, device=${this.device})`
     );
@@ -107,6 +112,7 @@ export class LocalEmbedder {
       // best-effort: some engines free resources on GC
     }
     this.extractor = null;
+    this.ready = false;
   }
 }
 

package/lib/service.js

@@ -37,11 +37,69 @@ export function createService({ store, mirror, config, onWrite, logger }) {
   // replays them exactly once against the committed state.
   let txDepth = 0;
 
+  // issue #6 (part 2): startup race defense. A local embedder (LocalEmbedder /
+  // Ollama) exposes an async init(), so between `setEmbedder` and init()
+  // resolving there is a window where embedSingle would throw "not initialized"
+  // and the re-embed would be silently dropped. When the embedder carries a
+  // `ready` flag we queue writes in embedPending until init sets ready=true,
+  // then flush them through the embedder's real interface. Embedders without a
+  // `ready` flag (legacy OpenAI, instantly usable) keep their old behavior.
+  let embedPending = [];
+  let embedReadyTimer = null;
+  const EMBED_PENDING_MAX = 100; // bound the queue; drop oldest beyond this
+  const EMBED_READY_POLL_MS = 100;
+  const EMBED_READY_POLL_LIMIT = 30; // ~3s ceiling; never poll forever
+
+  /** Flush the queued re-embeds once the embedder is ready. Fail-safe. */
+  function flushEmbedPending() {
+    if (!embedder || embedPending.length === 0) return;
+    const batch = embedPending.splice(0, embedPending.length);
+    for (const memory of batch) {
+      try {
+        if (!memory?.id) continue;
+        if (typeof embedder.schedule === "function") {
+          embedder.schedule(memory);
+        } else if (typeof embedder.embedSingle === "function") {
+          const text = [memory.title, memory.content].filter(Boolean).join("\n");
+          if (!text) continue;
+          embedder
+            .embedSingle(text)
+            .then((vec) => {
+              if (Array.isArray(vec) && vec.length) {
+                store.setEmbedding(memory.id, vec);
+              }
+            })
+            .catch((err) => {
+              logger?.warn?.("flushEmbedPending embedSingle failed:", err);
+            });
+        }
+      } catch (err) {
+        logger?.warn?.("flushEmbedPending failed:", err);
+      }
+    }
+  }
+
+  function stopEmbedReadyPolling() {
+    if (embedReadyTimer) {
+      clearInterval(embedReadyTimer);
+      embedReadyTimer = null;
+    }
+  }
+
   function scheduleEmbed(memory) {
     try {
       if (txDepth > 0) return; // deferred to the transaction's commit
       if (!embedder || !memory?.id) return;
 
+      // Readiness gate: embedder exposes `ready` (async init) and is not ready
+      // yet — queue instead of firing embedSingle into a half-built extractor.
+      const hasReady = "ready" in embedder;
+      if (hasReady && embedder.ready !== true) {
+        if (embedPending.length >= EMBED_PENDING_MAX) embedPending.shift();
+        embedPending.push(memory);
+        return;
+      }
+
       if (typeof embedder.schedule === "function") {
         embedder.schedule(memory);
         return;
@@ -681,7 +739,32 @@ export function createService({ store, mirror, config, onWrite, logger }) {
     toApiList,
     transaction,
     setDreamHook(fn) { dreamHook = fn; },
-    setEmbedder(emb) { embedder = emb; },
+    setEmbedder(emb) {
+      embedder = emb;
+      if (!emb) {
+        // embedder removed (init failed in index.js): stop polling and drop
+        // queued re-embeds — search just degrades to keyword.
+        stopEmbedReadyPolling();
+        embedPending = [];
+        return;
+      }
+      if (emb.ready === true) {
+        flushEmbedPending();
+        return;
+      }
+      // Async-initializing embedder: poll `ready` until it flips, then flush.
+      if ("ready" in emb && embedReadyTimer === null) {
+        let attempts = 0;
+        embedReadyTimer = setInterval(() => {
+          attempts++;
+          if (emb.ready === true || attempts >= EMBED_READY_POLL_LIMIT) {
+            stopEmbedReadyPolling();
+            if (emb.ready === true) flushEmbedPending();
+            else embedPending = []; // init never landed: drop the queue
+          }
+        }, EMBED_READY_POLL_MS);
+      }
+    },
     setEntityExtractor(fn) { entityExtractor = fn; },
     setVectorIndex(vi) { vectorIndex = vi; },
     setReranker(rn) { reranker = rn; },

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@modusensus/dsh-mneme",
   "description": "Cross-session memory plugin for DeepSeek Harness with autoDream consolidation: SQLite store, Markdown mirrors, 6 model tools, automatic injection, session summarization, user profile/rules, custom slash commands, vector (semantic) search, and a Web GUI panel",
-  "version": "0.3.6",
+  "version": "0.3.7",
   "license": "MIT",
   "type": "module",
   "main": "lib/index.js",

package/src/index.js

@@ -80,11 +80,28 @@ export const apply = (ctx, config) => {
   const vectorIndex = createVectorIndex({ store, logger: ctx.logger });
   service.setVectorIndex(vectorIndex);
 
+  // Human edits in mirror files win on every sync; merge them back first.
+  // TYPE_FILE maps each memory type to its mirror filename. Read every type's
+  // edits up front: mergeHumanEdits re-renders ALL mirror files on success, so
+  // a per-type read-then-merge loop would overwrite edits in files not yet read
+  // (e.g. preferences.md merging would clobber unsynced projects.md edits).
+  const humanEdits = new Map();
+  for (const type of Object.keys(TYPE_FILE)) {
+    humanEdits.set(type, mirror.readHumanEdits(type));
+  }
+  const applyHumanEdits = () => {
+    for (const [type, edits] of humanEdits) {
+      if (edits.length) service.mergeHumanEdits(type, edits);
+    }
+  };
+
   let embedder = null;
   let reranker = null;
   if (cfg.embedProvider === "openai") {
     embedder = createEmbedder({ store, settings, logger: ctx.logger });
     service.setEmbedder(embedder);
+    // legacy OpenAI embedder is immediately usable
+    applyHumanEdits();
   } else {
     try {
       embedder = createEmbedderByProvider(cfg.embedProvider, {
@@ -97,12 +114,18 @@ export const apply = (ctx, config) => {
         logger: ctx.logger
       });
       service.setEmbedder(embedder);
-      embedder.init().catch((error) => {
-        ctx.logger?.warn?.(`[dsh-mneme] embedder init failed, search degrades to keyword: ${String(error)}`);
-        service.setEmbedder(null);
-      });
+      // issue #6: wait for extractor init before applying human edits, so
+      // scheduled embeddings see a ready embedder.
+      embedder.init()
+        .then(() => applyHumanEdits())
+        .catch((error) => {
+          ctx.logger?.warn?.(`[dsh-mneme] embedder init failed, search degrades to keyword: ${String(error)}`);
+          service.setEmbedder(null);
+          applyHumanEdits();
+        });
     } catch (error) {
       ctx.logger?.warn?.(`[dsh-mneme] embedder unavailable, search degrades to keyword: ${String(error)}`);
+      applyHumanEdits();
     }
   }
 
@@ -139,19 +162,6 @@ export const apply = (ctx, config) => {
     commands.sync();
   }
 
-  // Human edits in mirror files win on every sync; merge them back first.
-  // TYPE_FILE maps each memory type to its mirror filename. Read every type's
-  // edits up front: mergeHumanEdits re-renders ALL mirror files on success, so
-  // a per-type read-then-merge loop would overwrite edits in files not yet read
-  // (e.g. preferences.md merging would clobber unsynced projects.md edits).
-  const humanEdits = new Map();
-  for (const type of Object.keys(TYPE_FILE)) {
-    humanEdits.set(type, mirror.readHumanEdits(type));
-  }
-  for (const [type, edits] of humanEdits) {
-    if (edits.length) service.mergeHumanEdits(type, edits);
-  }
-
   // Dream scheduler: automatic consolidation + summary runs, triggered by
   // store growth. Writes through the service fire the dream hook, which asks
   // the scheduler to (re)schedule a run once absolute and since-last-run

package/src/local-embedder.js

@@ -58,16 +58,21 @@ export class LocalEmbedder {
     // Test hook: replace the pipeline factory without touching modules.
     this.engineFactory = opts.engineFactory || defaultPipelineLoader;
     this.extractor = null;
+    // issue #6: readiness flag for the service's scheduleEmbed gate. False until
+    // init() succeeds, so "ready" in embedder is observable even pre-init.
+    this.ready = false;
   }
 
-  /** Load the model; throws when it cannot be loaded. */
+  /** Load the model; throws when it cannot be loaded. Idempotent. */
   async init() {
+    if (this.extractor) return this; // already initialized: no-op
     const options = {
       dtype: this.useDtype,
       device: this.device
     };
     if (this.cacheDir) options.cache_dir = this.cacheDir;
     this.extractor = await this.engineFactory("feature-extraction", this.model, options);
+    this.ready = true; // service reads this to flush queued re-embeds
     this.logger?.info?.(
       `[dsh-mneme] local embedder ready: ${this.model} (dim=${this._dimension}, device=${this.device})`
     );
@@ -107,6 +112,7 @@ export class LocalEmbedder {
       // best-effort: some engines free resources on GC
     }
     this.extractor = null;
+    this.ready = false;
   }
 }
 

package/src/service.js

@@ -37,11 +37,69 @@ export function createService({ store, mirror, config, onWrite, logger }) {
   // replays them exactly once against the committed state.
   let txDepth = 0;
 
+  // issue #6 (part 2): startup race defense. A local embedder (LocalEmbedder /
+  // Ollama) exposes an async init(), so between `setEmbedder` and init()
+  // resolving there is a window where embedSingle would throw "not initialized"
+  // and the re-embed would be silently dropped. When the embedder carries a
+  // `ready` flag we queue writes in embedPending until init sets ready=true,
+  // then flush them through the embedder's real interface. Embedders without a
+  // `ready` flag (legacy OpenAI, instantly usable) keep their old behavior.
+  let embedPending = [];
+  let embedReadyTimer = null;
+  const EMBED_PENDING_MAX = 100; // bound the queue; drop oldest beyond this
+  const EMBED_READY_POLL_MS = 100;
+  const EMBED_READY_POLL_LIMIT = 30; // ~3s ceiling; never poll forever
+
+  /** Flush the queued re-embeds once the embedder is ready. Fail-safe. */
+  function flushEmbedPending() {
+    if (!embedder || embedPending.length === 0) return;
+    const batch = embedPending.splice(0, embedPending.length);
+    for (const memory of batch) {
+      try {
+        if (!memory?.id) continue;
+        if (typeof embedder.schedule === "function") {
+          embedder.schedule(memory);
+        } else if (typeof embedder.embedSingle === "function") {
+          const text = [memory.title, memory.content].filter(Boolean).join("\n");
+          if (!text) continue;
+          embedder
+            .embedSingle(text)
+            .then((vec) => {
+              if (Array.isArray(vec) && vec.length) {
+                store.setEmbedding(memory.id, vec);
+              }
+            })
+            .catch((err) => {
+              logger?.warn?.("flushEmbedPending embedSingle failed:", err);
+            });
+        }
+      } catch (err) {
+        logger?.warn?.("flushEmbedPending failed:", err);
+      }
+    }
+  }
+
+  function stopEmbedReadyPolling() {
+    if (embedReadyTimer) {
+      clearInterval(embedReadyTimer);
+      embedReadyTimer = null;
+    }
+  }
+
   function scheduleEmbed(memory) {
     try {
       if (txDepth > 0) return; // deferred to the transaction's commit
       if (!embedder || !memory?.id) return;
 
+      // Readiness gate: embedder exposes `ready` (async init) and is not ready
+      // yet — queue instead of firing embedSingle into a half-built extractor.
+      const hasReady = "ready" in embedder;
+      if (hasReady && embedder.ready !== true) {
+        if (embedPending.length >= EMBED_PENDING_MAX) embedPending.shift();
+        embedPending.push(memory);
+        return;
+      }
+
       if (typeof embedder.schedule === "function") {
         embedder.schedule(memory);
         return;
@@ -681,7 +739,32 @@ export function createService({ store, mirror, config, onWrite, logger }) {
     toApiList,
     transaction,
     setDreamHook(fn) { dreamHook = fn; },
-    setEmbedder(emb) { embedder = emb; },
+    setEmbedder(emb) {
+      embedder = emb;
+      if (!emb) {
+        // embedder removed (init failed in index.js): stop polling and drop
+        // queued re-embeds — search just degrades to keyword.
+        stopEmbedReadyPolling();
+        embedPending = [];
+        return;
+      }
+      if (emb.ready === true) {
+        flushEmbedPending();
+        return;
+      }
+      // Async-initializing embedder: poll `ready` until it flips, then flush.
+      if ("ready" in emb && embedReadyTimer === null) {
+        let attempts = 0;
+        embedReadyTimer = setInterval(() => {
+          attempts++;
+          if (emb.ready === true || attempts >= EMBED_READY_POLL_LIMIT) {
+            stopEmbedReadyPolling();
+            if (emb.ready === true) flushEmbedPending();
+            else embedPending = []; // init never landed: drop the queue
+          }
+        }, EMBED_READY_POLL_MS);
+      }
+    },
     setEntityExtractor(fn) { entityExtractor = fn; },
     setVectorIndex(vi) { vectorIndex = vi; },
     setReranker(rn) { reranker = rn; },