@bryti/agent 0.0.1 → 0.1.0

package/dist/memory/embeddings.js

@@ -0,0 +1,139 @@
+/**
+ * Local embeddings via node-llama-cpp.
+ *
+ * Singleton pattern: one Llama instance, one model, one embedding context,
+ * shared across all calls for the lifetime of the process. The model weighs
+ * 300MB+, so loading it per-call or per-request is not viable. Instead,
+ * getEmbeddingContext() initialises everything on first call and caches the
+ * result; every subsequent call returns the cached context immediately.
+ *
+ * Model files live in <dataDir>/.models/.
+ */
+import { getLlama, LlamaLogLevel, resolveModelFile } from "node-llama-cpp";
+// Hugging Face URI in node-llama-cpp's "hf:<owner>/<repo>/<file>" format.
+// On first use, node-llama-cpp resolves this automatically: it locates (or
+// downloads) the file and caches it in the modelsDir supplied at init time.
+const EMBEDDING_MODEL_URI = "hf:ggml-org/embeddinggemma-300M-GGUF/embeddinggemma-300M-Q8_0.gguf";
+let llamaInstance = null;
+let llamaModel = null;
+let embeddingContext = null;
+// In-flight initialisation guard. Without this, concurrent calls to
+// getEmbeddingContext() before the first load completes would each kick off a
+// separate model load. Storing the Promise here means every concurrent caller
+// awaits the same in-progress load instead of starting a new one.
+let initPromise = null;
+/**
+ * Initialize and return the embedding context, loading the model on first call.
+ * Subsequent calls return the cached context.
+ */
+async function getEmbeddingContext(modelsDir) {
+    if (embeddingContext !== null) {
+        return embeddingContext;
+    }
+    if (initPromise !== null) {
+        return initPromise;
+    }
+    initPromise = (async () => {
+        const llama = await getLlama({
+            gpu: "auto",
+            logger(level, message) {
+                // 'special_eos_id is not in special_eog_ids' is a benign metadata quirk
+                // in the embedding model's GGUF file: the end-of-sequence token id is
+                // not listed in the end-of-generation set. It has no effect on embedding
+                // quality and is safe to ignore.
+                if (level === LlamaLogLevel.warn && message.includes("special_eos_id is not in special_eog_ids")) {
+                    return;
+                }
+                if (level === LlamaLogLevel.error || level === LlamaLogLevel.fatal) {
+                    console.error("[llama]", message);
+                }
+            },
+        });
+        llamaInstance = llama;
+        const modelPath = await resolveModelFile(EMBEDDING_MODEL_URI, {
+            directory: modelsDir,
+            cli: false,
+            onProgress({ totalSize, downloadedSize }) {
+                const pct = totalSize > 0 ? Math.round((downloadedSize / totalSize) * 100) : 0;
+                process.stdout.write(`\rDownloading embedding model: ${pct}%`);
+            },
+        });
+        process.stdout.write("\n");
+        const model = await llama.loadModel({ modelPath });
+        llamaModel = model;
+        const ctx = await model.createEmbeddingContext();
+        embeddingContext = ctx;
+        return ctx;
+    })();
+    return initPromise;
+}
+/**
+ * Generate an embedding for a single text.
+ *
+ * @param text Input text (must be non-empty)
+ * @param modelsDir Directory to store/load the model (defaults to node-llama-cpp global dir)
+ */
+export async function embed(text, modelsDir) {
+    if (!text.trim()) {
+        throw new Error("Embedding input is empty");
+    }
+    const ctx = await getEmbeddingContext(modelsDir);
+    const result = await ctx.getEmbeddingFor(text);
+    return Array.from(result.vector);
+}
+/**
+ * Generate embeddings for multiple texts. Sequential; the model is fast
+ * enough on CPU that batching provides no meaningful advantage.
+ */
+export async function embedBatch(texts, modelsDir) {
+    if (texts.length === 0) {
+        return [];
+    }
+    for (const text of texts) {
+        if (!text.trim()) {
+            throw new Error("Embedding input is empty");
+        }
+    }
+    const results = [];
+    for (const text of texts) {
+        results.push(await embed(text, modelsDir));
+    }
+    return results;
+}
+/**
+ * Pre-load the embedding model at startup.
+ *
+ * Calling this eagerly means a missing or corrupt model file surfaces as a
+ * startup error rather than failing silently on the first user message that
+ * triggers a memory operation. Also amortises the cold-start download/load
+ * time before any user is waiting.
+ *
+ * @param modelsDir Directory to store/load the model
+ */
+export async function warmupEmbeddings(modelsDir) {
+    await getEmbeddingContext(modelsDir);
+}
+/**
+ * Release the embedding context, model, and Llama instance.
+ *
+ * Call this on graceful shutdown. node-llama-cpp allocates native (non-GC)
+ * resources for model weights and inference threads; skipping dispose leaves
+ * those resources live until the OS reclaims them, and can prevent the Node
+ * process from exiting cleanly.
+ */
+export async function disposeEmbeddings() {
+    if (embeddingContext) {
+        await embeddingContext.dispose();
+        embeddingContext = null;
+    }
+    if (llamaModel) {
+        await llamaModel.dispose();
+        llamaModel = null;
+    }
+    if (llamaInstance) {
+        await llamaInstance.dispose();
+        llamaInstance = null;
+    }
+    initPromise = null;
+}
+//# sourceMappingURL=embeddings.js.map

package/dist/memory/embeddings.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"embeddings.js","sourceRoot":"","sources":["../../src/memory/embeddings.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,OAAO,EAAE,QAAQ,EAAE,aAAa,EAAE,gBAAgB,EAAE,MAAM,gBAAgB,CAAC;AAG3E,0EAA0E;AAC1E,2EAA2E;AAC3E,4EAA4E;AAC5E,MAAM,mBAAmB,GACvB,oEAAoE,CAAC;AAEvE,IAAI,aAAa,GAAiB,IAAI,CAAC;AACvC,IAAI,UAAU,GAAsB,IAAI,CAAC;AACzC,IAAI,gBAAgB,GAAiC,IAAI,CAAC;AAC1D,oEAAoE;AACpE,8EAA8E;AAC9E,8EAA8E;AAC9E,kEAAkE;AAClE,IAAI,WAAW,GAA0C,IAAI,CAAC;AAE9D;;;GAGG;AACH,KAAK,UAAU,mBAAmB,CAAC,SAAkB;IACnD,IAAI,gBAAgB,KAAK,IAAI,EAAE,CAAC;QAC9B,OAAO,gBAAgB,CAAC;IAC1B,CAAC;IAED,IAAI,WAAW,KAAK,IAAI,EAAE,CAAC;QACzB,OAAO,WAAW,CAAC;IACrB,CAAC;IAED,WAAW,GAAG,CAAC,KAAK,IAAI,EAAE;QACxB,MAAM,KAAK,GAAG,MAAM,QAAQ,CAAC;YAC3B,GAAG,EAAE,MAAM;YACX,MAAM,CAAC,KAAK,EAAE,OAAO;gBACnB,wEAAwE;gBACxE,sEAAsE;gBACtE,yEAAyE;gBACzE,iCAAiC;gBACjC,IAAI,KAAK,KAAK,aAAa,CAAC,IAAI,IAAI,OAAO,CAAC,QAAQ,CAAC,0CAA0C,CAAC,EAAE,CAAC;oBACjG,OAAO;gBACT,CAAC;gBACD,IAAI,KAAK,KAAK,aAAa,CAAC,KAAK,IAAI,KAAK,KAAK,aAAa,CAAC,KAAK,EAAE,CAAC;oBACnE,OAAO,CAAC,KAAK,CAAC,SAAS,EAAE,OAAO,CAAC,CAAC;gBACpC,CAAC;YACH,CAAC;SACF,CAAC,CAAC;QACH,aAAa,GAAG,KAAK,CAAC;QAEtB,MAAM,SAAS,GAAG,MAAM,gBAAgB,CAAC,mBAAmB,EAAE;YAC5D,SAAS,EAAE,SAAS;YACpB,GAAG,EAAE,KAAK;YACV,UAAU,CAAC,EAAE,SAAS,EAAE,cAAc,EAAE;gBACtC,MAAM,GAAG,GAAG,SAAS,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,cAAc,GAAG,SAAS,CAAC,GAAG,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;gBAC/E,OAAO,CAAC,MAAM,CAAC,KAAK,CAAC,kCAAkC,GAAG,GAAG,CAAC,CAAC;YACjE,CAAC;SACF,CAAC,CAAC;QACH,OAAO,CAAC,MAAM,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;QAE3B,MAAM,KAAK,GAAG,MAAM,KAAK,CAAC,SAAS,CAAC,EAAE,SAAS,EAAE,CAAC,CAAC;QACnD,UAAU,GAAG,KAAK,CAAC;QACnB,MAAM,GAAG,GAAG,MAAM,KAAK,CAAC,sBAAsB,EAAE,CAAC;QAEjD,gBAAgB,GAAG,GAAG,CAAC;QACvB,OAAO,GAAG,CAAC;IACb,CAAC,CAAC,EAAE,CAAC;IAEL,OAAO,WAAW,CAAC;AACrB,CAAC;AAED;;;;;GAKG;AACH,MAAM,CAAC,KAAK,UAAU,KAAK,CAAC,IAAY,EAAE,SAAkB;IAC1D,IAAI,CAAC,IAAI,CAAC,IAAI,EAAE,EAAE,CAAC;QACjB,MAAM,IAAI,KAAK,CAAC,0BAA0B,CAAC,CAAC;IAC9C,CAAC;IAED,MAAM,GAAG,GAAG,MAAM,mBAAmB,CAAC,SAAS,CAAC,CAAC;IACjD,MAAM,MAAM,GAAG,MAAM,GAAG,CAAC,eAAe,CAAC,IAAI,CAAC,CAAC;IAC/C,OAAO,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC;AACnC,CAAC;AAED;;;GAGG;AACH,MAAM,CAAC,KAAK,UAAU,UAAU,CAAC,KAAe,EAAE,SAAkB;IAClE,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACvB,OAAO,EAAE,CAAC;IACZ,CAAC;IAED,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;QACzB,IAAI,CAAC,IAAI,CAAC,IAAI,EAAE,EAAE,CAAC;YACjB,MAAM,IAAI,KAAK,CAAC,0BAA0B,CAAC,CAAC;QAC9C,CAAC;IACH,CAAC;IAED,MAAM,OAAO,GAAe,EAAE,CAAC;IAC/B,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;QACzB,OAAO,CAAC,IAAI,CAAC,MAAM,KAAK,CAAC,IAAI,EAAE,SAAS,CAAC,CAAC,CAAC;IAC7C,CAAC;IACD,OAAO,OAAO,CAAC;AACjB,CAAC;AAED;;;;;;;;;GASG;AACH,MAAM,CAAC,KAAK,UAAU,gBAAgB,CAAC,SAAkB;IACvD,MAAM,mBAAmB,CAAC,SAAS,CAAC,CAAC;AACvC,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,CAAC,KAAK,UAAU,iBAAiB;IACrC,IAAI,gBAAgB,EAAE,CAAC;QACrB,MAAM,gBAAgB,CAAC,OAAO,EAAE,CAAC;QACjC,gBAAgB,GAAG,IAAI,CAAC;IAC1B,CAAC;IACD,IAAI,UAAU,EAAE,CAAC;QACf,MAAM,UAAU,CAAC,OAAO,EAAE,CAAC;QAC3B,UAAU,GAAG,IAAI,CAAC;IACpB,CAAC;IACD,IAAI,aAAa,EAAE,CAAC;QAClB,MAAM,aAAa,CAAC,OAAO,EAAE,CAAC;QAC9B,aAAa,GAAG,IAAI,CAAC;IACvB,CAAC;IACD,WAAW,GAAG,IAAI,CAAC;AACrB,CAAC"}

package/dist/memory/search.d.ts

@@ -0,0 +1,49 @@
+/**
+ * Hybrid search: keyword (FTS5) + vector, fused with Reciprocal Rank Fusion.
+ *
+ * RRF formula: score(doc) = w_vec * 1/(k + rank_vec) + w_kw * 1/(k + rank_kw)
+ * Defaults: 0.7 vector, 0.3 keyword, k = 60.
+ *
+ * Why RRF instead of a weighted average of raw scores?
+ * FTS5 BM25 scores and cosine similarity values live on completely different
+ * scales, so adding them directly produces garbage. RRF converts each result
+ * list to ranks first, making the fusion scale-invariant: it doesn't matter
+ * what units either scorer uses.
+ */
+import type { MemoryStore, ScoredResult } from "./store.js";
+export interface SearchResult extends ScoredResult {
+    /** Combined score from both methods. */
+    combinedScore: number;
+    /** Which methods matched (for debugging/display). */
+    matchedBy: ("keyword" | "vector")[];
+}
+export interface HybridSearchOptions {
+    /** Weight for vector search results (default: 0.7) */
+    vectorWeight?: number;
+    /** Weight for keyword search results (default: 0.3) */
+    keywordWeight?: number;
+    /**
+     * Fusion smoothing constant k (default: 60).
+     * Standard RRF value from the original paper. Lower values amplify the
+     * advantage of top-ranked documents; 60 is empirically good across most
+     * retrieval tasks and avoids over-rewarding a single strong signal.
+     */
+    k?: number;
+    /** Maximum results to return (default: 5) */
+    limit?: number;
+}
+/**
+ * Create a hybrid search function over the given memory store.
+ *
+ * Returns a closure that accepts a query string. The closure calls `embed` on
+ * every invocation to produce a fresh query embedding; embeddings are not
+ * cached here. Callers that need caching should wrap `embed` before passing
+ * it in.
+ *
+ * @param store   The memory store to search against (provides keyword + vector).
+ * @param embed   Embedding function — called once per query, not cached.
+ * @param options Weights, k, and result limit overrides.
+ * @returns       An async function `(query) => SearchResult[]` ready for use.
+ */
+export declare function createHybridSearch(store: MemoryStore, embed: (text: string) => Promise<number[]>, options?: HybridSearchOptions): (query: string) => Promise<SearchResult[]>;
+//# sourceMappingURL=search.d.ts.map

package/dist/memory/search.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"search.d.ts","sourceRoot":"","sources":["../../src/memory/search.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAEH,OAAO,KAAK,EAAE,WAAW,EAAE,YAAY,EAAE,MAAM,YAAY,CAAC;AAE5D,MAAM,WAAW,YAAa,SAAQ,YAAY;IAChD,wCAAwC;IACxC,aAAa,EAAE,MAAM,CAAC;IACtB,qDAAqD;IACrD,SAAS,EAAE,CAAC,SAAS,GAAG,QAAQ,CAAC,EAAE,CAAC;CACrC;AAED,MAAM,WAAW,mBAAmB;IAClC,sDAAsD;IACtD,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,uDAAuD;IACvD,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB;;;;;OAKG;IACH,CAAC,CAAC,EAAE,MAAM,CAAC;IACX,6CAA6C;IAC7C,KAAK,CAAC,EAAE,MAAM,CAAC;CAChB;AASD;;;;;;;;;;;;GAYG;AACH,wBAAgB,kBAAkB,CAChC,KAAK,EAAE,WAAW,EAClB,KAAK,EAAE,CAAC,IAAI,EAAE,MAAM,KAAK,OAAO,CAAC,MAAM,EAAE,CAAC,EAC1C,OAAO,GAAE,mBAAwB,GAChC,CAAC,KAAK,EAAE,MAAM,KAAK,OAAO,CAAC,YAAY,EAAE,CAAC,CAyE5C"}

package/dist/memory/search.js

@@ -0,0 +1,97 @@
+/**
+ * Hybrid search: keyword (FTS5) + vector, fused with Reciprocal Rank Fusion.
+ *
+ * RRF formula: score(doc) = w_vec * 1/(k + rank_vec) + w_kw * 1/(k + rank_kw)
+ * Defaults: 0.7 vector, 0.3 keyword, k = 60.
+ *
+ * Why RRF instead of a weighted average of raw scores?
+ * FTS5 BM25 scores and cosine similarity values live on completely different
+ * scales, so adding them directly produces garbage. RRF converts each result
+ * list to ranks first, making the fusion scale-invariant: it doesn't matter
+ * what units either scorer uses.
+ */
+const DEFAULT_OPTIONS = {
+    vectorWeight: 0.7,
+    keywordWeight: 0.3,
+    k: 60,
+    limit: 5,
+};
+/**
+ * Create a hybrid search function over the given memory store.
+ *
+ * Returns a closure that accepts a query string. The closure calls `embed` on
+ * every invocation to produce a fresh query embedding; embeddings are not
+ * cached here. Callers that need caching should wrap `embed` before passing
+ * it in.
+ *
+ * @param store   The memory store to search against (provides keyword + vector).
+ * @param embed   Embedding function — called once per query, not cached.
+ * @param options Weights, k, and result limit overrides.
+ * @returns       An async function `(query) => SearchResult[]` ready for use.
+ */
+export function createHybridSearch(store, embed, options = {}) {
+    const opts = { ...DEFAULT_OPTIONS, ...options };
+    return async function hybridSearch(query) {
+        // Handle empty query
+        if (!query.trim()) {
+            return [];
+        }
+        // Get embedding for query
+        const queryEmbedding = await embed(query);
+        // Run both searches in parallel
+        const [keywordResults, vectorResults] = await Promise.all([
+            store.searchKeyword(query, opts.limit * 2), // Get more to account for dedup
+            store.searchVector(queryEmbedding, opts.limit * 2),
+        ]);
+        // Handle empty results
+        if (keywordResults.length === 0 && vectorResults.length === 0) {
+            return [];
+        }
+        // Both searches can return the same document. Merge by id: the first time
+        // a document appears its entry is created; the second time its score is
+        // accumulated and the matching method is appended to matchedBy.
+        const resultsMap = new Map();
+        // Process keyword results — applies the RRF term: w_kw * 1/(k + rank_kw)
+        for (let i = 0; i < keywordResults.length; i++) {
+            const result = keywordResults[i];
+            const rank = i + 1; // RRF ranks are 1-based
+            const score = opts.keywordWeight * (1 / (opts.k + rank));
+            if (resultsMap.has(result.id)) {
+                const existing = resultsMap.get(result.id);
+                existing.combinedScore += score;
+                existing.matchedBy.push("keyword");
+            }
+            else {
+                resultsMap.set(result.id, {
+                    ...result,
+                    combinedScore: score,
+                    matchedBy: ["keyword"],
+                });
+            }
+        }
+        // Process vector results — applies the RRF term: w_vec * 1/(k + rank_vec)
+        for (let i = 0; i < vectorResults.length; i++) {
+            const result = vectorResults[i];
+            const rank = i + 1; // RRF ranks are 1-based
+            const score = opts.vectorWeight * (1 / (opts.k + rank));
+            if (resultsMap.has(result.id)) {
+                const existing = resultsMap.get(result.id);
+                existing.combinedScore += score;
+                existing.matchedBy.push("vector");
+            }
+            else {
+                resultsMap.set(result.id, {
+                    ...result,
+                    combinedScore: score,
+                    matchedBy: ["vector"],
+                });
+            }
+        }
+        // Convert to array and sort by combined score
+        const mergedResults = Array.from(resultsMap.values())
+            .sort((a, b) => b.combinedScore - a.combinedScore)
+            .slice(0, opts.limit);
+        return mergedResults;
+    };
+}
+//# sourceMappingURL=search.js.map

package/dist/memory/search.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"search.js","sourceRoot":"","sources":["../../src/memory/search.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AA2BH,MAAM,eAAe,GAAkC;IACrD,YAAY,EAAE,GAAG;IACjB,aAAa,EAAE,GAAG;IAClB,CAAC,EAAE,EAAE;IACL,KAAK,EAAE,CAAC;CACT,CAAC;AAEF;;;;;;;;;;;;GAYG;AACH,MAAM,UAAU,kBAAkB,CAChC,KAAkB,EAClB,KAA0C,EAC1C,UAA+B,EAAE;IAEjC,MAAM,IAAI,GAAG,EAAE,GAAG,eAAe,EAAE,GAAG,OAAO,EAAE,CAAC;IAEhD,OAAO,KAAK,UAAU,YAAY,CAAC,KAAa;QAC9C,qBAAqB;QACrB,IAAI,CAAC,KAAK,CAAC,IAAI,EAAE,EAAE,CAAC;YAClB,OAAO,EAAE,CAAC;QACZ,CAAC;QAED,0BAA0B;QAC1B,MAAM,cAAc,GAAG,MAAM,KAAK,CAAC,KAAK,CAAC,CAAC;QAE1C,gCAAgC;QAChC,MAAM,CAAC,cAAc,EAAE,aAAa,CAAC,GAAG,MAAM,OAAO,CAAC,GAAG,CAAC;YACxD,KAAK,CAAC,aAAa,CAAC,KAAK,EAAE,IAAI,CAAC,KAAK,GAAG,CAAC,CAAC,EAAE,gCAAgC;YAC5E,KAAK,CAAC,YAAY,CAAC,cAAc,EAAE,IAAI,CAAC,KAAK,GAAG,CAAC,CAAC;SACnD,CAAC,CAAC;QAEH,uBAAuB;QACvB,IAAI,cAAc,CAAC,MAAM,KAAK,CAAC,IAAI,aAAa,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YAC9D,OAAO,EAAE,CAAC;QACZ,CAAC;QAED,0EAA0E;QAC1E,wEAAwE;QACxE,gEAAgE;QAChE,MAAM,UAAU,GAAG,IAAI,GAAG,EAAwB,CAAC;QAEnD,yEAAyE;QACzE,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,cAAc,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YAC/C,MAAM,MAAM,GAAG,cAAc,CAAC,CAAC,CAAC,CAAC;YACjC,MAAM,IAAI,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,wBAAwB;YAC5C,MAAM,KAAK,GAAG,IAAI,CAAC,aAAa,GAAG,CAAC,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC;YAEzD,IAAI,UAAU,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC,EAAE,CAAC;gBAC9B,MAAM,QAAQ,GAAG,UAAU,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,CAAE,CAAC;gBAC5C,QAAQ,CAAC,aAAa,IAAI,KAAK,CAAC;gBAChC,QAAQ,CAAC,SAAS,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;YACrC,CAAC;iBAAM,CAAC;gBACN,UAAU,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,EAAE;oBACxB,GAAG,MAAM;oBACT,aAAa,EAAE,KAAK;oBACpB,SAAS,EAAE,CAAC,SAAS,CAAC;iBACvB,CAAC,CAAC;YACL,CAAC;QACH,CAAC;QAED,0EAA0E;QAC1E,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,aAAa,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YAC9C,MAAM,MAAM,GAAG,aAAa,CAAC,CAAC,CAAC,CAAC;YAChC,MAAM,IAAI,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,wBAAwB;YAC5C,MAAM,KAAK,GAAG,IAAI,CAAC,YAAY,GAAG,CAAC,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC;YAExD,IAAI,UAAU,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC,EAAE,CAAC;gBAC9B,MAAM,QAAQ,GAAG,UAAU,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,CAAE,CAAC;gBAC5C,QAAQ,CAAC,aAAa,IAAI,KAAK,CAAC;gBAChC,QAAQ,CAAC,SAAS,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;YACpC,CAAC;iBAAM,CAAC;gBACN,UAAU,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,EAAE;oBACxB,GAAG,MAAM;oBACT,aAAa,EAAE,KAAK;oBACpB,SAAS,EAAE,CAAC,QAAQ,CAAC;iBACtB,CAAC,CAAC;YACL,CAAC;QACH,CAAC;QAED,8CAA8C;QAC9C,MAAM,aAAa,GAAG,KAAK,CAAC,IAAI,CAAC,UAAU,CAAC,MAAM,EAAE,CAAC;aAClD,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,aAAa,GAAG,CAAC,CAAC,aAAa,CAAC;aACjD,KAAK,CAAC,CAAC,EAAE,IAAI,CAAC,KAAK,CAAC,CAAC;QAExB,OAAO,aAAa,CAAC;IACvB,CAAC,CAAC;AACJ,CAAC"}

package/dist/memory/store.d.ts

@@ -0,0 +1,32 @@
+/**
+ * Per-user memory store backed by SQLite with FTS5 for keyword search and
+ * binary Float32Array blobs for embeddings. Vector search runs in-memory
+ * via cosine similarity.
+ */
+export interface ScoredResult {
+    id: string;
+    content: string;
+    source: string;
+    timestamp: number;
+    score: number;
+}
+export interface MemoryStore {
+    /** Add a fact to the store. Returns the fact ID. */
+    addFact(content: string, source: string, embedding: number[]): string;
+    /** Remove a fact by ID. */
+    removeFact(id: string): void;
+    /** Search using keyword (FTS5). */
+    searchKeyword(query: string, limit: number): ScoredResult[];
+    /** Search using vector similarity (in-memory cosine similarity). */
+    searchVector(embedding: number[], limit: number): ScoredResult[];
+    /** Close the database connection. */
+    close(): void;
+}
+/**
+ * Create a per-user memory store.
+ *
+ * @param userId User ID for isolation
+ * @param dataDir Base data directory
+ */
+export declare function createMemoryStore(userId: string, dataDir: string): MemoryStore;
+//# sourceMappingURL=store.d.ts.map

package/dist/memory/store.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"store.d.ts","sourceRoot":"","sources":["../../src/memory/store.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAQH,MAAM,WAAW,YAAY;IAC3B,EAAE,EAAE,MAAM,CAAC;IACX,OAAO,EAAE,MAAM,CAAC;IAChB,MAAM,EAAE,MAAM,CAAC;IACf,SAAS,EAAE,MAAM,CAAC;IAClB,KAAK,EAAE,MAAM,CAAC;CACf;AAED,MAAM,WAAW,WAAW;IAC1B,oDAAoD;IACpD,OAAO,CAAC,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,SAAS,EAAE,MAAM,EAAE,GAAG,MAAM,CAAC;IAEtE,2BAA2B;IAC3B,UAAU,CAAC,EAAE,EAAE,MAAM,GAAG,IAAI,CAAC;IAE7B,mCAAmC;IACnC,aAAa,CAAC,KAAK,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,GAAG,YAAY,EAAE,CAAC;IAE5D,oEAAoE;IACpE,YAAY,CAAC,SAAS,EAAE,MAAM,EAAE,EAAE,KAAK,EAAE,MAAM,GAAG,YAAY,EAAE,CAAC;IAEjE,qCAAqC;IACrC,KAAK,IAAI,IAAI,CAAC;CACf;AAuBD;;;;;GAKG;AACH,wBAAgB,iBAAiB,CAAC,MAAM,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,GAAG,WAAW,CA6M9E"}

package/dist/memory/store.js

@@ -0,0 +1,205 @@
+/**
+ * Per-user memory store backed by SQLite with FTS5 for keyword search and
+ * binary Float32Array blobs for embeddings. Vector search runs in-memory
+ * via cosine similarity.
+ */
+import Database from "better-sqlite3";
+import { cosineSimilarity } from "../util/math.js";
+import fs from "node:fs";
+import path from "node:path";
+import crypto from "node:crypto";
+/**
+ * Serialize an embedding vector to a raw binary buffer.
+ * Stored as Float32Array bytes (4 bytes per dimension) rather than JSON,
+ * which avoids the significant parse/stringify overhead for ~300-dim vectors
+ * and halves the storage footprint compared to text representation.
+ */
+function serializeEmbedding(embedding) {
+    return Buffer.from(new Float32Array(embedding).buffer);
+}
+/**
+ * Deserialize a raw binary buffer back to a number array.
+ * Inverse of serializeEmbedding: reads raw Float32 bytes from the blob column.
+ */
+function deserializeEmbedding(buffer) {
+    const array = new Float32Array(buffer.buffer, buffer.byteOffset, buffer.byteLength / 4);
+    return Array.from(array);
+}
+/**
+ * Create a per-user memory store.
+ *
+ * @param userId User ID for isolation
+ * @param dataDir Base data directory
+ */
+export function createMemoryStore(userId, dataDir) {
+    // Schema overview:
+    //   facts          — main table; one row per stored fact (id, content, source, timestamp, hash)
+    //   facts_fts      — FTS5 virtual table shadowing facts.content for BM25 keyword search
+    //   fact_embeddings — binary blob table keyed by the same id as facts
+    //
+    // Three triggers (facts_ai, facts_ad, facts_au) keep facts_fts in sync with
+    // facts automatically on insert, delete, and update.
+    const userDir = path.join(dataDir, "users", userId);
+    fs.mkdirSync(userDir, { recursive: true });
+    const dbPath = path.join(userDir, "memory.db");
+    const db = new Database(dbPath);
+    // Enable WAL mode for concurrent reads
+    db.pragma("journal_mode = WAL");
+    // Create tables
+    db.exec(`
+    -- Main facts table
+    CREATE TABLE IF NOT EXISTS facts (
+      id TEXT PRIMARY KEY,
+      content TEXT NOT NULL,
+      source TEXT NOT NULL,
+      timestamp INTEGER NOT NULL,
+      hash TEXT NOT NULL
+    );
+
+    -- FTS5 virtual table for keyword search
+    CREATE VIRTUAL TABLE IF NOT EXISTS facts_fts USING fts5(
+      content,
+      content='facts',
+      content_rowid='rowid'
+    );
+
+    -- Triggers to keep FTS in sync
+    CREATE TRIGGER IF NOT EXISTS facts_ai AFTER INSERT ON facts BEGIN
+      INSERT INTO facts_fts(rowid, content) VALUES (NEW.rowid, NEW.content);
+    END;
+
+    CREATE TRIGGER IF NOT EXISTS facts_ad AFTER DELETE ON facts BEGIN
+      INSERT INTO facts_fts(facts_fts, rowid, content) VALUES('delete', OLD.rowid, OLD.content);
+    END;
+
+    CREATE TRIGGER IF NOT EXISTS facts_au AFTER UPDATE ON facts BEGIN
+      INSERT INTO facts_fts(facts_fts, rowid, content) VALUES('delete', OLD.rowid, OLD.content);
+      INSERT INTO facts_fts(rowid, content) VALUES (NEW.rowid, NEW.content);
+    END;
+
+    -- Table to store embeddings (binary Float32Array)
+    CREATE TABLE IF NOT EXISTS fact_embeddings (
+      id TEXT PRIMARY KEY,
+      embedding BLOB NOT NULL,
+      FOREIGN KEY (id) REFERENCES facts(id) ON DELETE CASCADE
+    );
+  `);
+    // Prepared statements for efficiency
+    const insertFact = db.prepare(`
+    INSERT INTO facts (id, content, source, timestamp, hash)
+    VALUES (?, ?, ?, ?, ?)
+  `);
+    const insertEmbedding = db.prepare(`
+    INSERT INTO fact_embeddings (id, embedding) VALUES (?, ?)
+  `);
+    const deleteFact = db.prepare(`
+    DELETE FROM facts WHERE id = ?
+  `);
+    const deleteEmbedding = db.prepare(`
+    DELETE FROM fact_embeddings WHERE id = ?
+  `);
+    const selectEmbeddings = db.prepare(`
+    SELECT f.id, f.content, f.source, f.timestamp, fe.embedding
+    FROM facts f
+    JOIN fact_embeddings fe ON f.id = fe.id
+  `);
+    return {
+        /**
+         * Add a fact to the store and return its generated UUID.
+         *
+         * @param content  The fact text to store and index.
+         * @param source   Provenance label (e.g. "reflection", "user").
+         * @param embedding Pre-computed embedding vector for this content.
+         *
+         * The `hash` field is a truncated SHA-256 of the content. It is used to
+         * deduplicate facts during bulk imports, not as an integrity check — the
+         * store does not verify it on read.
+         */
+        addFact(content, source, embedding) {
+            const id = crypto.randomUUID();
+            const timestamp = Date.now();
+            const hash = crypto.createHash("sha256").update(content).digest("hex").slice(0, 16);
+            // Insert into facts table
+            insertFact.run(id, content, source, timestamp, hash);
+            // Insert embedding as binary blob
+            insertEmbedding.run(id, serializeEmbedding(embedding));
+            return id;
+        },
+        removeFact(id) {
+            deleteEmbedding.run(id);
+            deleteFact.run(id);
+        },
+        /**
+         * Search facts by keyword using SQLite FTS5.
+         *
+         * Scoring uses FTS5's built-in BM25 implementation (via the `bm25()`
+         * function), which ranks results by term frequency and inverse document
+         * frequency. Returns at most `limit` results ordered by relevance.
+         */
+        searchKeyword(query, limit) {
+            if (!query.trim()) {
+                return [];
+            }
+            // Use FTS5 match with BM25 ranking.
+            // Wrapping the query in double quotes makes it a phrase query, which
+            // neutralises FTS5 operators (OR, AND, NOT, NEAR, *, etc.).
+            // Double quotes inside the phrase would break out, so we escape them
+            // by doubling ("" is the FTS5 escape for a literal double quote).
+            // Single quotes are stripped to avoid tokenizer surprises.
+            const escapedQuery = query.replace(/'/g, "").replace(/"/g, '""');
+            const stmt = db.prepare(`
+        SELECT f.id, f.content, f.source, f.timestamp,
+               bm25(facts_fts) as score
+        FROM facts_fts
+        JOIN facts f ON facts_fts.rowid = f.rowid
+        WHERE facts_fts MATCH ?
+        ORDER BY bm25(facts_fts)
+        LIMIT ?
+      `);
+            const results = stmt.all(`"${escapedQuery}"`, limit);
+            return results.map((row) => ({
+                id: row.id,
+                content: row.content,
+                source: row.source,
+                timestamp: row.timestamp,
+                score: row.score,
+            }));
+        },
+        /**
+         * Search facts by vector similarity.
+         *
+         * This is a full table scan: all embeddings are loaded from SQLite into
+         * memory, cosine similarity is computed for each one, and the top `limit`
+         * results are returned. This is acceptable up to roughly 100K facts.
+         * Beyond that, an approximate nearest-neighbour index would be needed
+         * (hnswlib-node or the sqlite-vec extension are natural fits here).
+         *
+         * TODO: add ANN index when fact count regularly exceeds ~100K per user.
+         */
+        searchVector(embedding, limit) {
+            const rows = selectEmbeddings.all();
+            if (rows.length === 0) {
+                return [];
+            }
+            // Compute cosine similarity for each fact
+            const scored = rows
+                .map((row) => ({
+                ...row,
+                similarity: cosineSimilarity(embedding, deserializeEmbedding(row.embedding)),
+            }))
+                .sort((a, b) => b.similarity - a.similarity)
+                .slice(0, limit);
+            return scored.map((row) => ({
+                id: row.id,
+                content: row.content,
+                source: row.source,
+                timestamp: row.timestamp,
+                score: row.similarity,
+            }));
+        },
+        close() {
+            db.close();
+        },
+    };
+}
+//# sourceMappingURL=store.js.map

package/dist/memory/store.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"store.js","sourceRoot":"","sources":["../../src/memory/store.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAEH,OAAO,QAAQ,MAAM,gBAAgB,CAAC;AACtC,OAAO,EAAE,gBAAgB,EAAE,MAAM,iBAAiB,CAAC;AACnD,OAAO,EAAE,MAAM,SAAS,CAAC;AACzB,OAAO,IAAI,MAAM,WAAW,CAAC;AAC7B,OAAO,MAAM,MAAM,aAAa,CAAC;AA6BjC;;;;;GAKG;AACH,SAAS,kBAAkB,CAAC,SAAmB;IAC7C,OAAO,MAAM,CAAC,IAAI,CAAC,IAAI,YAAY,CAAC,SAAS,CAAC,CAAC,MAAM,CAAC,CAAC;AACzD,CAAC;AAED;;;GAGG;AACH,SAAS,oBAAoB,CAAC,MAAc;IAC1C,MAAM,KAAK,GAAG,IAAI,YAAY,CAAC,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,UAAU,EAAE,MAAM,CAAC,UAAU,GAAG,CAAC,CAAC,CAAC;IACxF,OAAO,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;AAC3B,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,iBAAiB,CAAC,MAAc,EAAE,OAAe;IAC/D,mBAAmB;IACnB,gGAAgG;IAChG,wFAAwF;IACxF,sEAAsE;IACtE,EAAE;IACF,4EAA4E;IAC5E,qDAAqD;IACrD,MAAM,OAAO,GAAG,IAAI,CAAC,IAAI,CAAC,OAAO,EAAE,OAAO,EAAE,MAAM,CAAC,CAAC;IACpD,EAAE,CAAC,SAAS,CAAC,OAAO,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAC;IAE3C,MAAM,MAAM,GAAG,IAAI,CAAC,IAAI,CAAC,OAAO,EAAE,WAAW,CAAC,CAAC;IAC/C,MAAM,EAAE,GAAG,IAAI,QAAQ,CAAC,MAAM,CAAC,CAAC;IAEhC,uCAAuC;IACvC,EAAE,CAAC,MAAM,CAAC,oBAAoB,CAAC,CAAC;IAEhC,gBAAgB;IAChB,EAAE,CAAC,IAAI,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAqCP,CAAC,CAAC;IAEH,qCAAqC;IACrC,MAAM,UAAU,GAAG,EAAE,CAAC,OAAO,CAAC;;;GAG7B,CAAC,CAAC;IAEH,MAAM,eAAe,GAAG,EAAE,CAAC,OAAO,CAAC;;GAElC,CAAC,CAAC;IAEH,MAAM,UAAU,GAAG,EAAE,CAAC,OAAO,CAAC;;GAE7B,CAAC,CAAC;IAEH,MAAM,eAAe,GAAG,EAAE,CAAC,OAAO,CAAC;;GAElC,CAAC,CAAC;IAEH,MAAM,gBAAgB,GAAG,EAAE,CAAC,OAAO,CAAC;;;;GAInC,CAAC,CAAC;IAEH,OAAO;QACL;;;;;;;;;;WAUG;QACH,OAAO,CAAC,OAAe,EAAE,MAAc,EAAE,SAAmB;YAC1D,MAAM,EAAE,GAAG,MAAM,CAAC,UAAU,EAAE,CAAC;YAC/B,MAAM,SAAS,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;YAC7B,MAAM,IAAI,GAAG,MAAM,CAAC,UAAU,CAAC,QAAQ,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;YAEpF,0BAA0B;YAC1B,UAAU,CAAC,GAAG,CAAC,EAAE,EAAE,OAAO,EAAE,MAAM,EAAE,SAAS,EAAE,IAAI,CAAC,CAAC;YAErD,kCAAkC;YAClC,eAAe,CAAC,GAAG,CAAC,EAAE,EAAE,kBAAkB,CAAC,SAAS,CAAC,CAAC,CAAC;YAEvD,OAAO,EAAE,CAAC;QACZ,CAAC;QAED,UAAU,CAAC,EAAU;YACnB,eAAe,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;YACxB,UAAU,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;QACrB,CAAC;QAED;;;;;;WAMG;QACH,aAAa,CAAC,KAAa,EAAE,KAAa;YACxC,IAAI,CAAC,KAAK,CAAC,IAAI,EAAE,EAAE,CAAC;gBAClB,OAAO,EAAE,CAAC;YACZ,CAAC;YAED,oCAAoC;YACpC,qEAAqE;YACrE,4DAA4D;YAC5D,qEAAqE;YACrE,kEAAkE;YAClE,2DAA2D;YAC3D,MAAM,YAAY,GAAG,KAAK,CAAC,OAAO,CAAC,IAAI,EAAE,EAAE,CAAC,CAAC,OAAO,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC;YAEjE,MAAM,IAAI,GAAG,EAAE,CAAC,OAAO,CAAC;;;;;;;;OAQvB,CAAC,CAAC;YAEH,MAAM,OAAO,GAAG,IAAI,CAAC,GAAG,CAAC,IAAI,YAAY,GAAG,EAAE,KAAK,CAMjD,CAAC;YAEH,OAAO,OAAO,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC;gBAC3B,EAAE,EAAE,GAAG,CAAC,EAAE;gBACV,OAAO,EAAE,GAAG,CAAC,OAAO;gBACpB,MAAM,EAAE,GAAG,CAAC,MAAM;gBAClB,SAAS,EAAE,GAAG,CAAC,SAAS;gBACxB,KAAK,EAAE,GAAG,CAAC,KAAK;aACjB,CAAC,CAAC,CAAC;QACN,CAAC;QAED;;;;;;;;;;WAUG;QACH,YAAY,CAAC,SAAmB,EAAE,KAAa;YAC7C,MAAM,IAAI,GAAG,gBAAgB,CAAC,GAAG,EAM/B,CAAC;YAEH,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;gBACtB,OAAO,EAAE,CAAC;YACZ,CAAC;YAED,0CAA0C;YAC1C,MAAM,MAAM,GAAG,IAAI;iBAChB,GAAG,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC;gBACb,GAAG,GAAG;gBACN,UAAU,EAAE,gBAAgB,CAAC,SAAS,EAAE,oBAAoB,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC;aAC7E,CAAC,CAAC;iBACF,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,UAAU,GAAG,CAAC,CAAC,UAAU,CAAC;iBAC3C,KAAK,CAAC,CAAC,EAAE,KAAK,CAAC,CAAC;YAEnB,OAAO,MAAM,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC;gBAC1B,EAAE,EAAE,GAAG,CAAC,EAAE;gBACV,OAAO,EAAE,GAAG,CAAC,OAAO;gBACpB,MAAM,EAAE,GAAG,CAAC,MAAM;gBAClB,SAAS,EAAE,GAAG,CAAC,SAAS;gBACxB,KAAK,EAAE,GAAG,CAAC,UAAU;aACtB,CAAC,CAAC,CAAC;QACN,CAAC;QAED,KAAK;YACH,EAAE,CAAC,KAAK,EAAE,CAAC;QACb,CAAC;KACF,CAAC;AACJ,CAAC"}

package/dist/message-queue.d.ts

@@ -0,0 +1,73 @@
+/**
+ * Per-channel FIFO message queue with merge and backpressure.
+ *
+ * Two core guarantees:
+ *
+ * 1. Per-channel serialization: only one message is processed at a time per
+ *    channel. Subsequent messages queue up behind it rather than racing into
+ *    the agent loop in parallel.
+ *
+ * 2. Burst merging: rapid-fire messages that arrive within MERGE_WINDOW_MS of
+ *    each other are joined into a single prompt before being dispatched. This
+ *    handles the common "user sends three quick messages" pattern without the
+ *    agent seeing three separate incomplete thoughts.
+ *
+ * New messages queue up to MAX_DEPTH; beyond that the caller gets a rejection
+ * callback (backpressure signal, not silent drop).
+ */
+import type { IncomingMessage } from "./channels/types.js";
+type ProcessFn = (msg: IncomingMessage) => Promise<void>;
+type RejectFn = (msg: IncomingMessage) => Promise<void>;
+/**
+ * Serialises processing per channel and merges rapid follow-up messages.
+ */
+export declare class MessageQueue {
+    private readonly queues;
+    private readonly processFn;
+    private readonly rejectFn;
+    private readonly maxDepth;
+    private readonly mergeWindowMs;
+    private readonly rateLimiter;
+    constructor(processFn: ProcessFn, rejectFn: RejectFn, maxDepth?: number, mergeWindowMs?: number);
+    /**
+     * Enqueue a message. If nothing is processing for this channel, starts
+     * draining immediately. Rate-limited per user (10 messages/minute).
+     */
+    enqueue(msg: IncomingMessage): void;
+    /**
+     * Drain the queue for a channel sequentially, merging close messages.
+     */
+    private drain;
+    /**
+     * Take a batch of entries that should be merged together.
+     *
+     * The first entry is always included. Additional entries are included if they
+     * arrived within mergeWindowMs of the FIRST entry in the batch. This is a
+     * fixed window anchored to the first message, not a sliding window — an
+     * entry that arrives 1ms after the previous one is still excluded if it
+     * falls outside the window from the first entry.
+     */
+    private takeMergeBatch;
+    /**
+     * Merge multiple queue entries into a single IncomingMessage by joining
+     * their text with newlines. Metadata (userId, channelId, platform, etc.) is
+     * taken from the first entry.
+     *
+     * Note: images (and other non-text attachments) from subsequent burst
+     * entries are currently dropped — only their text is merged.
+     * TODO: carry images from all burst entries into the merged message.
+     */
+    private mergeEntries;
+    /**
+     * Number of queued (not-yet-processing) messages for a channel.
+     * Exposed for monitoring dashboards and unit tests.
+     */
+    queueDepth(channelId: string): number;
+    /**
+     * Whether the channel is currently mid-process (drain loop running).
+     * Exposed for monitoring dashboards and unit tests.
+     */
+    isProcessing(channelId: string): boolean;
+}
+export {};
+//# sourceMappingURL=message-queue.d.ts.map

package/dist/message-queue.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"message-queue.d.ts","sourceRoot":"","sources":["../src/message-queue.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAEH,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,qBAAqB,CAAC;AAmB3D,KAAK,SAAS,GAAG,CAAC,GAAG,EAAE,eAAe,KAAK,OAAO,CAAC,IAAI,CAAC,CAAC;AACzD,KAAK,QAAQ,GAAG,CAAC,GAAG,EAAE,eAAe,KAAK,OAAO,CAAC,IAAI,CAAC,CAAC;AA+CxD;;GAEG;AACH,qBAAa,YAAY;IACvB,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAmC;IAC1D,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAY;IACtC,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAW;IACpC,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAS;IAClC,OAAO,CAAC,QAAQ,CAAC,aAAa,CAAS;IACvC,OAAO,CAAC,QAAQ,CAAC,WAAW,CAAc;gBAGxC,SAAS,EAAE,SAAS,EACpB,QAAQ,EAAE,QAAQ,EAClB,QAAQ,SAAY,EACpB,aAAa,SAAkB;IASjC;;;OAGG;IACH,OAAO,CAAC,GAAG,EAAE,eAAe,GAAG,IAAI;IAkCnC;;OAEG;YACW,KAAK;IAmBnB;;;;;;;;OAQG;IACH,OAAO,CAAC,cAAc;IAiBtB;;;;;;;;OAQG;IACH,OAAO,CAAC,YAAY;IAYpB;;;OAGG;IACH,UAAU,CAAC,SAAS,EAAE,MAAM,GAAG,MAAM;IAIrC;;;OAGG;IACH,YAAY,CAAC,SAAS,EAAE,MAAM,GAAG,OAAO;CAGzC"}