scream-code 0.8.4 → 0.8.5

package/dist/{app-DYDC5U75.mjs → app-DFMT0YIn.mjs}

@@ -5,7 +5,7 @@ const __filename = __cjsShimFileURLToPath(import.meta.url);
 const __dirname = __cjsShimDirname(__filename);
 import { a as __toESM, i as __require, r as __exportAll, t as __commonJSMin } from "./chunk-apG1qJts.mjs";
 import "./suppress-sqlite-warning-C2VB0doZ.mjs";
-import { C as join$1, D as resolve$1, E as relative$1, S as isAbsolute$1, T as parse$7, a as isSupportedFile, b as basename$1, i as ingestFile, r as ingestDirectory, t as multiSearch, w as normalize, x as dirname$2, y as KnowledgeStore } from "./src-CptHxtUY.mjs";
+import { C as join$1, D as resolve$1, E as relative$1, S as isAbsolute$1, T as parse$7, a as isSupportedFile, b as basename$1, i as ingestFile, r as ingestDirectory, t as multiSearch, w as normalize, x as dirname$2, y as KnowledgeStore } from "./src-BHYcdFfg.mjs";
 import { createRequire } from "node:module";
 import { createHash, randomBytes, randomInt, randomUUID } from "node:crypto";
 import Jn, { access, appendFile, chmod, copyFile, cp, lstat, mkdir, mkdtemp, open, readFile, readdir, realpath, rename, rm, rmdir, stat, unlink, writeFile } from "node:fs/promises";
@@ -56624,10 +56624,31 @@ function buildEmbeddingText(memo) {
 	return `${memo.userNeed} ${memo.approach} ${memo.whatWorked}`;
 }
 /**
+* Cache of created engines keyed by cacheDir.
+* Guarantees that all callers sharing the same cacheDir reuse the same engine
+* instance and the same in-flight model download/load — avoiding duplicate
+* downloads and file-corruption races when /memory and /knowledge both start
+* before the model is cached.
+*/
+const engineCache = /* @__PURE__ */ new Map();
+/**
 * Create an embedding engine backed by fastembed.
 * Lazily loads the model on first use so startup is not blocked.
-*/
-function createFastEmbedEngine() {
+* Engines are cached by cacheDir so repeated calls with the same cacheDir
+* return the same instance, sharing model state and download progress.
+* @param cacheDir Absolute path for model cache (e.g. ~/.scream-code/cache/fastembed).
+*                 Defaults to "local_cache" (CWD-relative) if not provided — prefer
+*                 passing an explicit path so the cache doesn't duplicate across CWDs.
+*/
+function createFastEmbedEngine(cacheDir) {
+	const key = cacheDir ?? "";
+	const cached = engineCache.get(key);
+	if (cached !== void 0) return cached;
+	const engine = createFastEmbedEngineImpl(cacheDir);
+	engineCache.set(key, engine);
+	return engine;
+}
+function createFastEmbedEngineImpl(cacheDir) {
 	let embedder = null;
 	let initPromise = null;
 	let loadFailed = false;
@@ -56640,7 +56661,7 @@ function createFastEmbedEngine() {
 			if (texts.length === 0) return [];
 			try {
 				if (embedder === null) {
-					initPromise ??= loadEmbedder();
+					initPromise ??= loadEmbedder(cacheDir);
 					embedder = await initPromise;
 					if (embedder === null) {
 						loadFailed = true;
@@ -56670,9 +56691,12 @@ function createFastEmbedEngine() {
 			return denom === 0 ? 0 : dot / denom;
 		},
 		async ensureReady() {
-			if (embedder !== null) return true;
+			if (embedder !== null) {
+				loadFailed = false;
+				return true;
+			}
 			try {
-				initPromise ??= loadEmbedder();
+				if (loadFailed || initPromise === null) initPromise = loadEmbedder(cacheDir);
 				embedder = await initPromise;
 				if (embedder === null) {
 					initPromise = null;
@@ -56687,19 +56711,71 @@ function createFastEmbedEngine() {
 		}
 	};
 }
-async function loadEmbedder() {
+async function loadEmbedder(cacheDir) {
 	try {
 		const { FlagEmbedding, EmbeddingModel } = await import("./esm-Du2MwXei.mjs");
-		return await FlagEmbedding.init({ model: EmbeddingModel.BGESmallZH });
+		if (cacheDir !== void 0) mkdirSync(cacheDir, { recursive: true });
+		const model = EmbeddingModel.BGESmallZH;
+		const initOpts = cacheDir !== void 0 ? {
+			model,
+			cacheDir
+		} : { model };
+		try {
+			return await FlagEmbedding.init(initOpts);
+		} catch (initError) {
+			const msg = initError instanceof Error ? initError.message : "";
+			if (!/Config file not found|Tokenizer file not found|Tokens map file not found/iu.test(msg)) return null;
+			try {
+				await ensureFastembedModelSidecars(String(model), cacheDir);
+				return await FlagEmbedding.init(initOpts);
+			} catch {
+				return null;
+			}
+		}
 	} catch {
 		return null;
 	}
 }
+/**
+* Small config/tokenizer files that fastembed expects alongside model.onnx.
+* If these are missing (e.g. GCS download partially failed), fastembed throws.
+* We download them from HuggingFace so the model can load — this covers the
+* case where the GCS tarball was incomplete but the HF repo is reachable.
+*/
+const FASTEMBED_SIDECARS = [
+	"config.json",
+	"tokenizer.json",
+	"tokenizer_config.json",
+	"special_tokens_map.json"
+];
+const FASTEMBED_HF_REPOS = { "fast-bge-small-zh-v1.5": "BAAI/bge-small-zh-v1.5" };
+async function ensureFastembedModelSidecars(model, cacheDir) {
+	const repo = FASTEMBED_HF_REPOS[model];
+	if (repo === void 0) return;
+	const modelDir = join(cacheDir ?? "local_cache", model);
+	mkdirSync(modelDir, { recursive: true });
+	for (const fileName of FASTEMBED_SIDECARS) {
+		const target = join(modelDir, fileName);
+		try {
+			const { access } = await import("node:fs/promises");
+			await access(target);
+			continue;
+		} catch {}
+		const hfUrl = `https://huggingface.co/${repo}/resolve/main/${fileName}`;
+		try {
+			const response = await fetch(hfUrl);
+			if (!response.ok) continue;
+			const { writeFile } = await import("node:fs/promises");
+			await writeFile(target, Buffer.from(await response.arrayBuffer()));
+		} catch {}
+	}
+}
 //#endregion
 //#region ../../packages/memory/src/store.ts
 const FILE_NAME = "entries.jsonl";
 const MIGRATION_MARKER = ".migrated";
 const SQLITE_MIGRATION_MARKER = ".migrated-to-sqlite";
+const EMBEDDING_LOAD_RETRY_MS = 300 * 1e3;
 var MemoryMemoStore = class MemoryMemoStore {
 	projectDir;
 	jsonlPath;
@@ -56714,6 +56790,7 @@ var MemoryMemoStore = class MemoryMemoStore {
 	embeddingDegraded = false;
 	embeddingConsecutiveFailures = 0;
 	lastEmbeddingError;
+	embeddingLoadTimer;
 	log;
 	constructor(projectDir, log) {
 		this.projectDir = projectDir;
@@ -56751,6 +56828,10 @@ var MemoryMemoStore = class MemoryMemoStore {
 	}
 	/** Close the database handle and checkpoint WAL. */
 	close() {
+		if (this.embeddingLoadTimer !== void 0) {
+			clearInterval(this.embeddingLoadTimer);
+			this.embeddingLoadTimer = void 0;
+		}
 		if (this.db !== void 0) {
 			this.db.exec("PRAGMA wal_checkpoint(TRUNCATE)");
 			this.db.close();
@@ -57128,6 +57209,34 @@ var MemoryMemoStore = class MemoryMemoStore {
 	/** Set the embedding engine. Call once after construction, before any writes. */
 	setEmbeddingEngine(engine) {
 		this.embeddingEngine = engine;
+		this.startEmbeddingLoadLoop();
+	}
+	/**
+	* Background loop that retries model loading every 5 minutes until success.
+	* Covers the case where the model download fails on first use (network issue,
+	* GCS blocked, etc.) — without this, the engine stays loadFailed forever and
+	* all future memos fall back to keyword search silently.
+	* Stops itself once the model loads successfully.
+	*/
+	startEmbeddingLoadLoop() {
+		if (this.embeddingEngine === void 0) return;
+		if (this.embeddingLoadTimer !== void 0) return;
+		const attempt = async () => {
+			if (this.embeddingEngine === void 0) return;
+			if (await this.embeddingEngine.ensureReady()) {
+				if (this.embeddingLoadTimer !== void 0) {
+					clearInterval(this.embeddingLoadTimer);
+					this.embeddingLoadTimer = void 0;
+				}
+				this.embeddingDegraded = false;
+				return;
+			}
+		};
+		this.embeddingLoadTimer = setInterval(() => {
+			attempt();
+		}, EMBEDDING_LOAD_RETRY_MS);
+		this.embeddingLoadTimer.unref?.();
+		attempt();
 	}
 	/** Check whether the store has any vector embeddings. */
 	hasEmbeddings() {
@@ -58232,7 +58341,7 @@ var KnowledgeLookupTool = class {
 				const llm = { generate: async (systemPrompt, userPrompt) => {
 					return this.agent.generateText(systemPrompt, userPrompt);
 				} };
-				const { multiSearchWithTrace } = await import("./src-DhT3eU2C.mjs");
+				const { multiSearchWithTrace } = await import("./src-CC8SSs-H.mjs");
 				const { results, trace } = await multiSearchWithTrace(store, llm, query, { topK });
 				if (results.length === 0) return {
 					isError: false,
@@ -98593,6 +98702,7 @@ var Agent = class {
 	dreamTracker;
 	replayBuilder;
 	lastLlmConfigLogSignature;
+	sharedEmbeddingEngine;
 	constructor(options) {
 		this.type = options.type ?? "main";
 		this.jian = options.jian;
@@ -98607,6 +98717,8 @@ var Agent = class {
 		this.subagentHost = options.subagentHost;
 		this.mcp = options.mcp;
 		this.hooks = options.hookEngine;
+		const embedCacheDir = options.screamHomeDir !== void 0 ? join$1(options.screamHomeDir, "cache", "fastembed") : void 0;
+		this.sharedEmbeddingEngine = createFastEmbedEngine(embedCacheDir);
 		this.log = options.log ?? log;
 		this.blobStore = options.homedir ? new BlobStore({ blobsDir: join$1(options.homedir, "blobs") }) : void 0;
 		this.records = new AgentRecords(this, options.persistence ?? (options.homedir ? new FileSystemAgentRecordPersistence(join$1(options.homedir, "wire.jsonl"), {
@@ -98665,7 +98777,7 @@ var Agent = class {
 				this.log.error("memory legacy migration failed", error);
 			}
 			try {
-				this.memoStore.setEmbeddingEngine(createFastEmbedEngine());
+				this.memoStore.setEmbeddingEngine(this.sharedEmbeddingEngine);
 			} catch (error) {
 				this.log.warn("embedding engine init failed; falling back to keyword search", error);
 			}
@@ -98680,7 +98792,7 @@ var Agent = class {
 				this.log.error("knowledge store init failed", error);
 			}
 			try {
-				store.setEmbeddingEngine(createFastEmbedEngine());
+				store.setEmbeddingEngine(this.sharedEmbeddingEngine);
 			} catch (error) {
 				this.log.warn("knowledge embedding engine init failed", error);
 			}
@@ -122269,7 +122381,7 @@ function optionalBuildString(value) {
 	return typeof value === "string" && value.length > 0 ? value : void 0;
 }
 const SCREAM_BUILD_INFO = {
-	version: optionalBuildString("0.8.4"),
+	version: optionalBuildString("0.8.5"),
 	channel: optionalBuildString(""),
 	commit: optionalBuildString(""),
 	buildTarget: optionalBuildString("darwin-arm64")
@@ -133109,11 +133221,16 @@ function openUrl(url) {
 //#endregion
 //#region src/tui/commands/knowledge-store.ts
 let knowledgeStoreInstance;
+function getEmbeddingCacheDir() {
+	const dir = join(getDataDir(), "cache", "fastembed");
+	mkdirSync(dir, { recursive: true });
+	return dir;
+}
 async function getKnowledgeStore() {
 	if (knowledgeStoreInstance === void 0) {
 		knowledgeStoreInstance = new KnowledgeStore(getDataDir());
 		await knowledgeStoreInstance.init();
-		knowledgeStoreInstance.setEmbeddingEngine(createFastEmbedEngine());
+		knowledgeStoreInstance.setEmbeddingEngine(createFastEmbedEngine(getEmbeddingCacheDir()));
 	}
 	return knowledgeStoreInstance;
 }
@@ -133583,7 +133700,7 @@ function serveHTML(res) {
 	});
 	res.end(HTML);
 }
-async function handleWeb() {
+async function handleWeb(host) {
 	const store = await getKnowledgeStore();
 	const s = await store.stats();
 	if (s.entities === 0 && s.events === 0) throw new Error("知识库为空，请先用 /knowledge 摄入文档");
@@ -133604,7 +133721,7 @@ async function handleWeb() {
 	});
 	const url = `http://127.0.0.1:${server.address().port}`;
 	openUrl(url);
-	console.log(`\n  知识图谱: ${url}\n  按 Ctrl+C 停止服务器\n`);
+	host.showStatus(`知识图谱已打开: ${url}`);
 }
 //#endregion
 //#region src/tui/components/dialogs/knowledge-result-viewer.ts
@@ -134000,6 +134117,7 @@ function makeLlmCaller(host) {
 }
 function formatProgress(progress) {
 	switch (progress.stage) {
+		case "embedding-check": return progress.message;
 		case "chunking": return `切分文件中...`;
 		case "embedding-chunks": return `嵌入 chunks: ${progress.chunkIndex}/${progress.totalChunks}`;
 		case "extracting": return `抽取事件: ${progress.chunkIndex}/${progress.totalChunks}`;
@@ -134030,19 +134148,6 @@ async function handleIngest(host) {
 	}
 	const store = await getKnowledgeStore();
 	const llm = makeLlmCaller(host);
-	const checkSpinner = host.showProgressSpinner("检查向量模型...");
-	if (!(await store.ensureEmbeddingReady()).ok) {
-		checkSpinner.stop({
-			ok: false,
-			label: "向量模型下载失败"
-		});
-		host.showNotice("向量模型下载失败", "知识库需要中文向量模型 bge-small-zh-v1.5（约 95MB）才能进行语义检索。\n\n可能原因：\n• 网络不通，无法下载模型文件\n• onnxruntime native binding 加载失败\n\n请检查网络后重试。");
-		return;
-	}
-	checkSpinner.stop({
-		ok: true,
-		label: "向量模型就绪"
-	});
 	const spinner = host.showProgressSpinner("开始摄入...");
 	try {
 		if (stats.isDirectory()) {
@@ -134145,11 +134250,17 @@ async function handleSearch(host) {
 		ok: true,
 		label: "搜索完成"
 	});
+	const engine = store.getEmbeddingEngine();
+	const degraded = engine === void 0 || !engine.available;
 	if (results.length === 0) {
-		await showResultViewer(host, "搜索结果", `查询 "${query}" 未命中任何 chunk`);
+		await showResultViewer(host, "搜索结果", `查询 "${query}" 未命中任何 chunk${degraded ? "\n\n⚠️ 向量模型未就绪，当前为关键词检索模式。如下载失败，建议开启科学上网后重启。" : ""}`);
 		return;
 	}
 	const lines = [`查询: ${query}`, ""];
+	if (degraded) {
+		lines.push("⚠️ 向量模型未就绪，当前为关键词检索模式。如下载失败，建议开启科学上网后重启。");
+		lines.push("");
+	}
 	for (const [i, r] of results.entries()) {
 		lines.push(`#${i + 1} [score=${r.score.toFixed(3)}] ${r.heading ?? "(无标题)"}`);
 		lines.push(`   来源: ${r.sourceName}`);
@@ -134300,7 +134411,7 @@ async function handleKnowledgeCommand(host, _args) {
 						else if (value === "search") await handleSearch(host);
 						else if (value === "delete") await handleDelete(host);
 						else if (value === "stats") await handleStats(host);
-						else if (value === "web") await handleWeb();
+						else if (value === "web") await handleWeb(host);
 					} catch (error) {
 						const msg = error instanceof Error ? error.message : String(error);
 						host.showError(`操作失败: ${msg}`);

package/dist/main.mjs

@@ -6,7 +6,7 @@ const __dirname = __cjsShimDirname(__filename);
 import "./suppress-sqlite-warning-C2VB0doZ.mjs";
 //#region src/main.ts
 try {
-	(await import("./app-DYDC5U75.mjs")).main();
+	(await import("./app-DFMT0YIn.mjs")).main();
 } catch (error) {
 	process.stderr.write(`${error instanceof Error ? error.stack ?? error.message : String(error)}\n`);
 	process.exit(1);

package/dist/{src-CptHxtUY.mjs → src-BHYcdFfg.mjs}

@@ -1392,6 +1392,19 @@ Only include entities that explicitly appear in the query. If no clear entities,
 }
 //#endregion
 //#region ../../packages/knowledge/src/ingest.ts
+/**
+* Embed texts, returning null-array if embeddings are skipped or on failure.
+* Caller should set skipEmbedding=true when the model didn't load in time,
+* so the entire ingestion is consistent (all or nothing for embeddings).
+* Search falls back to FTS5 keyword match for chunks without embeddings.
+*/
+async function embedOrNulls(engine, texts, skipEmbedding) {
+	if (engine === void 0 || texts.length === 0 || skipEmbedding) return texts.map(() => null);
+	if (!engine.available) return texts.map(() => null);
+	const vectors = await engine.embedBatch(texts);
+	if (vectors === null || vectors.length !== texts.length) return texts.map(() => null);
+	return vectors;
+}
 /** LLM concurrency for extraction — kept low to avoid rate-limit bursts. */
 const LLM_CONCURRENCY = 3;
 const SUPPORTED_EXTENSIONS = new Set([
@@ -1454,12 +1467,25 @@ async function mapWithConcurrency(items, limit, fn, onProgress) {
 */
 async function ingestFile(store, llm, filePath, onProgress) {
 	const engine = store.getEmbeddingEngine();
-	if (engine === void 0 || !engine.available) {
+	let skipEmbed = engine === void 0;
+	if (engine !== void 0) {
 		onProgress?.({
-			stage: "error",
-			message: "向量模型未就绪"
+			stage: "embedding-check",
+			message: "检查向量模型..."
 		});
-		throw new Error("向量模型未就绪，请检查网络后重试");
+		try {
+			skipEmbed = !await Promise.race([engine.ensureReady(), new Promise((resolve) => setTimeout(() => resolve(false), 15e3))]);
+			if (skipEmbed) onProgress?.({
+				stage: "embedding-check",
+				message: "向量模型未就绪，本次摄入跳过向量嵌入"
+			});
+		} catch {
+			skipEmbed = true;
+			onProgress?.({
+				stage: "embedding-check",
+				message: "向量模型未就绪，本次摄入跳过向量嵌入"
+			});
+		}
 	}
 	const existing = await store.findSourceByFilePath(filePath);
 	if (existing !== void 0) {
@@ -1501,14 +1527,7 @@ async function ingestFile(store, llm, filePath, onProgress) {
 			totalChunks: sections.length,
 			message: `嵌入 chunks: 0/${sections.length}`
 		});
-		const chunkEmbeddings = await engine.embedBatch(sections.map((s) => s.heading !== null ? `${s.heading}\n${s.content}` : s.content));
-		if (chunkEmbeddings === null) {
-			onProgress?.({
-				stage: "error",
-				message: "chunk 向量嵌入失败"
-			});
-			throw new Error("chunk 向量嵌入失败");
-		}
+		const chunkEmbeddings = await embedOrNulls(engine, sections.map((s) => s.heading !== null ? `${s.heading}\n${s.content}` : s.content), skipEmbed);
 		const chunks = [];
 		for (let i = 0; i < sections.length; i++) {
 			const section = sections[i];
@@ -1552,14 +1571,7 @@ async function ingestFile(store, llm, filePath, onProgress) {
 		});
 		const titleTexts = extractedEvents.map((e) => e.title);
 		const contentTexts = extractedEvents.map((e) => `${e.title}\n\n${e.content}`);
-		const [titleEmbeddings, contentEmbeddings] = await Promise.all([engine.embedBatch(titleTexts), engine.embedBatch(contentTexts)]);
-		if (titleEmbeddings === null || contentEmbeddings === null) {
-			onProgress?.({
-				stage: "error",
-				message: "event 向量嵌入失败"
-			});
-			throw new Error("event 向量嵌入失败");
-		}
+		const [titleEmbeddings, contentEmbeddings] = await Promise.all([embedOrNulls(engine, titleTexts, skipEmbed), embedOrNulls(engine, contentTexts, skipEmbed)]);
 		const events = [];
 		for (let i = 0; i < extractedEvents.length; i++) {
 			const extracted = extractedEvents[i];
@@ -1595,18 +1607,11 @@ async function ingestFile(store, llm, filePath, onProgress) {
 			if (!entityMap.has(key)) entityMap.set(key, entity);
 		}
 		const uniqueEntities = Array.from(entityMap.values());
-		const entityEmbeddings = uniqueEntities.length > 0 ? await engine.embedBatch(uniqueEntities.map((e) => e.name)) : [];
-		if (uniqueEntities.length > 0 && entityEmbeddings === null) {
-			onProgress?.({
-				stage: "error",
-				message: "entity 向量嵌入失败"
-			});
-			throw new Error("entity 向量嵌入失败");
-		}
+		const entityEmbeddings = uniqueEntities.length > 0 ? await embedOrNulls(engine, uniqueEntities.map((e) => e.name), skipEmbed) : [];
 		const entityIdByEntityKey = /* @__PURE__ */ new Map();
 		for (let i = 0; i < uniqueEntities.length; i++) {
 			const entity = uniqueEntities[i];
-			const embedding = entityEmbeddings?.[i] ?? null;
+			const embedding = entityEmbeddings[i] ?? null;
 			const stored = await store.upsertEntity({
 				sourceId: source.id,
 				type: entity.type,
@@ -1632,14 +1637,7 @@ async function ingestFile(store, llm, filePath, onProgress) {
 			const event = events[eventIndex];
 			return entity.description.length > 0 ? entity.description : `${event.title} ${entity.name}`;
 		});
-		const relationEmbeddings = relationPairs.length > 0 ? await engine.embedBatch(relationTexts) : [];
-		if (relationPairs.length > 0 && relationEmbeddings === null) {
-			onProgress?.({
-				stage: "error",
-				message: "relation 向量嵌入失败"
-			});
-			throw new Error("relation 向量嵌入失败");
-		}
+		const relationEmbeddings = relationPairs.length > 0 ? await embedOrNulls(engine, relationTexts, skipEmbed) : [];
 		for (let i = 0; i < relationPairs.length; i++) {
 			const pair = relationPairs[i];
 			const event = events[pair.eventIndex];

package/dist/{src-DhT3eU2C.mjs → src-CC8SSs-H.mjs}

@@ -3,5 +3,5 @@ import { fileURLToPath as __cjsShimFileURLToPath } from 'node:url';
 import { dirname as __cjsShimDirname } from 'node:path';
 const __filename = __cjsShimFileURLToPath(import.meta.url);
 const __dirname = __cjsShimDirname(__filename);
-import { n as multiSearchWithTrace } from "./src-CptHxtUY.mjs";
+import { n as multiSearchWithTrace } from "./src-BHYcdFfg.mjs";
 export { multiSearchWithTrace };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "scream-code",
-  "version": "0.8.4",
+  "version": "0.8.5",
   "description": "A terminal-native AI agent for builders",
   "license": "MIT",
   "author": "ScreamCli",