crispy-recall 0.1.1 → 0.1.3

package/dist/embed-pending.js

@@ -4594,6 +4594,11 @@ async function startServer() {
     "-ub",
     "8192",
     // micro-batch (ubatch) — also defaults to 512, must be raised
+    // Flash Attention: the attention compute buffer is otherwise O(n²) in the
+    // 8192-token batch (~3.7 GB on GPU), which OOMs cards <6 GB and forces CPU.
+    // -fa makes it O(n) (~0.8 GB → ~2.9 GB total), so 4 GB cards run on GPU,
+    // with numerically identical embeddings (verified cosine 1.0 vs non-FA).
+    "-fa",
     // nomic-embed-text-v1.5 is trained at 2048 ctx and uses Dynamic NTK-aware
     // RoPE scaling to extend to 8192. Without these flags, newer llama.cpp
     // (b9253+) refuses inputs >2048 tokens, and older versions silently
@@ -4801,6 +4806,9 @@ async function embedViaProcess(texts, modelPath) {
       "array",
       "-c",
       "8192",
+      // Flash Attention — see startServer: O(n) attention buffer instead of
+      // O(n²), so the one-shot path also fits cards <6 GB. Identical embeddings.
+      "-fa",
       // Match the YaRN flags on the llama-server batch path — nomic-embed-text-v1.5
       // is trained at 2048 ctx and extends to 8192 via Dynamic NTK-aware RoPE.
       // Without these, newer llama.cpp (b9253+) refuses inputs >2048 tokens here
@@ -5552,6 +5560,9 @@ function normalizePath(p) {
   return normalized;
 }
 
+// src/recall/message-ingest.ts
+init_log();
+
 // src/adapters/claude/jsonl-reader.ts
 var fs = __toESM(require("fs"));
 init_log();
@@ -5962,8 +5973,7 @@ function emitFunctionCall(payload, base, outputIndex, _counter) {
   const outputRecord = outputIndex.get(callId);
   if (outputRecord) {
     const outputPayload = outputRecord.payload;
-    const rawOutput = outputPayload.output;
-    const { exitCode, body } = parseExecOutputHeader(rawOutput);
+    const { exitCode, body } = parseExecOutputHeader(outputPayload.output);
     const isError = exitCode !== 0;
     const toolResult = {
       type: "tool_result",
@@ -6101,7 +6111,7 @@ function emitApplyPatch(callId, input, base, outputIndex) {
   return entries;
 }
 function buildCustomToolResult(parentUuid, callId, outputRecord, base) {
-  const rawOutput = outputRecord.payload.output;
+  const rawOutput = coerceOutputText(outputRecord.payload.output);
   let content;
   let isError = false;
   try {
@@ -6182,7 +6192,7 @@ function emitOrphanedOutput(payload, subtype, base, outputIndex, _counter) {
     return [];
   if (!outputIndex.has(callId))
     return [];
-  const rawOutput = payload.output;
+  const rawOutput = coerceOutputText(payload.output);
   let content;
   let isError = false;
   if (subtype === "function_call_output") {
@@ -6295,7 +6305,18 @@ function mapFunctionCall(name, args) {
       return { toolName: name, toolInput: args };
   }
 }
-function parseExecOutputHeader(output) {
+function coerceOutputText(output) {
+  if (typeof output === "string")
+    return output;
+  if (Array.isArray(output)) {
+    return output.map(
+      (item) => item && typeof item === "object" && typeof item.text === "string" ? item.text : ""
+    ).filter(Boolean).join("\n");
+  }
+  return "";
+}
+function parseExecOutputHeader(rawOutput) {
+  const output = coerceOutputText(rawOutput);
   if (!output)
     return { exitCode: 0, body: "" };
   const currentMatch = output.match(/Process exited with code (\d+)/);
@@ -6460,6 +6481,39 @@ async function ingestSessionMessages(sessionId, transcriptPath, vendor, options)
 }
 var MAX_EMBED_CHARS = 14e3;
 var MAX_EMBED_BATCH = 10;
+var SAFE_EMBED_CHARS = 6e3;
+async function embedRowsResilient(rows) {
+  const { embedBatch: embedBatch2 } = await Promise.resolve().then(() => (init_embedder(), embedder_exports));
+  const { quantizeToQ8: quantizeToQ82, computeNorm: computeNorm2 } = await Promise.resolve().then(() => (init_quantize(), quantize_exports));
+  const toRecord = (messageId, f32) => {
+    const { q8, scale } = quantizeToQ82(f32);
+    return { messageId, embeddingQ8: q8, norm: computeNorm2(f32), quantScale: scale };
+  };
+  try {
+    const vectors = await embedBatch2(rows.map((r) => r.text));
+    return rows.map((r, j) => toRecord(r.messageId, vectors[j]));
+  } catch {
+    const records = [];
+    for (const r of rows) {
+      try {
+        const [v] = await embedBatch2([r.text]);
+        records.push(toRecord(r.messageId, v));
+      } catch {
+        try {
+          const [v] = await embedBatch2([r.text.slice(0, SAFE_EMBED_CHARS)]);
+          records.push(toRecord(r.messageId, v));
+        } catch (err) {
+          log({
+            source: "recall:embed",
+            level: "warn",
+            summary: `skipped message ${r.messageId} (embed failed even at ${SAFE_EMBED_CHARS} chars): ${err.message}`
+          });
+        }
+      }
+    }
+    return records;
+  }
+}
 async function embedSessionMessages(sessionId, force) {
   const d = getDb(dbPath());
   const rows = d.all(
@@ -6484,22 +6538,7 @@ async function embedSessionMessages(sessionId, force) {
   if (validRows.length > MAX_EMBED_BATCH) {
     validRows.length = MAX_EMBED_BATCH;
   }
-  const { embedBatch: embedBatch2 } = await Promise.resolve().then(() => (init_embedder(), embedder_exports));
-  const { quantizeToQ8: quantizeToQ82, computeNorm: computeNorm2 } = await Promise.resolve().then(() => (init_quantize(), quantize_exports));
-  const texts = validRows.map((r) => r.text);
-  const vectors = await embedBatch2(texts);
-  const records = [];
-  for (let j = 0; j < validRows.length; j++) {
-    const f32 = vectors[j];
-    const { q8, scale } = quantizeToQ82(f32);
-    const norm = computeNorm2(f32);
-    records.push({
-      messageId: validRows[j].messageId,
-      embeddingQ8: q8,
-      norm,
-      quantScale: scale
-    });
-  }
+  const records = await embedRowsResilient(validRows);
   insertMessageVectors(records);
   return records.length;
 }
@@ -6518,22 +6557,7 @@ async function embedMessageBatch(messages) {
   }
   if (truncated.length === 0)
     return 0;
-  const { embedBatch: embedBatch2 } = await Promise.resolve().then(() => (init_embedder(), embedder_exports));
-  const { quantizeToQ8: quantizeToQ82, computeNorm: computeNorm2 } = await Promise.resolve().then(() => (init_quantize(), quantize_exports));
-  const texts = truncated.map((r) => r.text);
-  const vectors = await embedBatch2(texts);
-  const records = [];
-  for (let j = 0; j < truncated.length; j++) {
-    const f32 = vectors[j];
-    const { q8, scale } = quantizeToQ82(f32);
-    const norm = computeNorm2(f32);
-    records.push({
-      messageId: truncated[j].messageId,
-      embeddingQ8: q8,
-      norm,
-      quantScale: scale
-    });
-  }
+  const records = await embedRowsResilient(truncated);
   insertMessageVectors(records);
   return records.length;
 }

package/dist/recall.js

@@ -3268,6 +3268,11 @@ async function startServer() {
     "-ub",
     "8192",
     // micro-batch (ubatch) — also defaults to 512, must be raised
+    // Flash Attention: the attention compute buffer is otherwise O(n²) in the
+    // 8192-token batch (~3.7 GB on GPU), which OOMs cards <6 GB and forces CPU.
+    // -fa makes it O(n) (~0.8 GB → ~2.9 GB total), so 4 GB cards run on GPU,
+    // with numerically identical embeddings (verified cosine 1.0 vs non-FA).
+    "-fa",
     // nomic-embed-text-v1.5 is trained at 2048 ctx and uses Dynamic NTK-aware
     // RoPE scaling to extend to 8192. Without these flags, newer llama.cpp
     // (b9253+) refuses inputs >2048 tokens, and older versions silently
@@ -3475,6 +3480,9 @@ async function embedViaProcess(texts, modelPath) {
       "array",
       "-c",
       "8192",
+      // Flash Attention — see startServer: O(n) attention buffer instead of
+      // O(n²), so the one-shot path also fits cards <6 GB. Identical embeddings.
+      "-fa",
       // Match the YaRN flags on the llama-server batch path — nomic-embed-text-v1.5
       // is trained at 2048 ctx and extends to 8192 via Dynamic NTK-aware RoPE.
       // Without these, newer llama.cpp (b9253+) refuses inputs >2048 tokens here
@@ -13157,8 +13165,7 @@ function emitFunctionCall(payload, base, outputIndex, _counter) {
   const outputRecord = outputIndex.get(callId);
   if (outputRecord) {
     const outputPayload = outputRecord.payload;
-    const rawOutput = outputPayload.output;
-    const { exitCode, body } = parseExecOutputHeader(rawOutput);
+    const { exitCode, body } = parseExecOutputHeader(outputPayload.output);
     const isError = exitCode !== 0;
     const toolResult = {
       type: "tool_result",
@@ -13296,7 +13303,7 @@ function emitApplyPatch(callId, input, base, outputIndex) {
   return entries;
 }
 function buildCustomToolResult(parentUuid, callId, outputRecord, base) {
-  const rawOutput = outputRecord.payload.output;
+  const rawOutput = coerceOutputText(outputRecord.payload.output);
   let content;
   let isError = false;
   try {
@@ -13377,7 +13384,7 @@ function emitOrphanedOutput(payload, subtype, base, outputIndex, _counter) {
     return [];
   if (!outputIndex.has(callId))
     return [];
-  const rawOutput = payload.output;
+  const rawOutput = coerceOutputText(payload.output);
   let content;
   let isError = false;
   if (subtype === "function_call_output") {
@@ -13490,7 +13497,18 @@ function mapFunctionCall(name, args) {
       return { toolName: name, toolInput: args };
   }
 }
-function parseExecOutputHeader(output) {
+function coerceOutputText(output) {
+  if (typeof output === "string")
+    return output;
+  if (Array.isArray(output)) {
+    return output.map(
+      (item) => item && typeof item === "object" && typeof item.text === "string" ? item.text : ""
+    ).filter(Boolean).join("\n");
+  }
+  return "";
+}
+function parseExecOutputHeader(rawOutput) {
+  const output = coerceOutputText(rawOutput);
   if (!output)
     return { exitCode: 0, body: "" };
   const currentMatch = output.match(/Process exited with code (\d+)/);
@@ -13658,6 +13676,38 @@ async function ingestSessionMessages(sessionId, transcriptPath, vendor, options)
     skipped: false
   };
 }
+async function embedRowsResilient(rows) {
+  const { embedBatch: embedBatch2 } = await Promise.resolve().then(() => (init_embedder(), embedder_exports));
+  const { quantizeToQ8: quantizeToQ82, computeNorm: computeNorm2 } = await Promise.resolve().then(() => (init_quantize(), quantize_exports));
+  const toRecord = (messageId, f32) => {
+    const { q8, scale } = quantizeToQ82(f32);
+    return { messageId, embeddingQ8: q8, norm: computeNorm2(f32), quantScale: scale };
+  };
+  try {
+    const vectors = await embedBatch2(rows.map((r2) => r2.text));
+    return rows.map((r2, j2) => toRecord(r2.messageId, vectors[j2]));
+  } catch {
+    const records = [];
+    for (const r2 of rows) {
+      try {
+        const [v2] = await embedBatch2([r2.text]);
+        records.push(toRecord(r2.messageId, v2));
+      } catch {
+        try {
+          const [v2] = await embedBatch2([r2.text.slice(0, SAFE_EMBED_CHARS)]);
+          records.push(toRecord(r2.messageId, v2));
+        } catch (err) {
+          log({
+            source: "recall:embed",
+            level: "warn",
+            summary: `skipped message ${r2.messageId} (embed failed even at ${SAFE_EMBED_CHARS} chars): ${err.message}`
+          });
+        }
+      }
+    }
+    return records;
+  }
+}
 async function embedMessageBatch(messages) {
   if (messages.length === 0)
     return 0;
@@ -13673,26 +13723,11 @@ async function embedMessageBatch(messages) {
   }
   if (truncated.length === 0)
     return 0;
-  const { embedBatch: embedBatch2 } = await Promise.resolve().then(() => (init_embedder(), embedder_exports));
-  const { quantizeToQ8: quantizeToQ82, computeNorm: computeNorm2 } = await Promise.resolve().then(() => (init_quantize(), quantize_exports));
-  const texts = truncated.map((r2) => r2.text);
-  const vectors = await embedBatch2(texts);
-  const records = [];
-  for (let j2 = 0; j2 < truncated.length; j2++) {
-    const f32 = vectors[j2];
-    const { q8, scale } = quantizeToQ82(f32);
-    const norm = computeNorm2(f32);
-    records.push({
-      messageId: truncated[j2].messageId,
-      embeddingQ8: q8,
-      norm,
-      quantScale: scale
-    });
-  }
+  const records = await embedRowsResilient(truncated);
   insertMessageVectors(records);
   return records.length;
 }
-var MAX_EMBED_CHARS;
+var MAX_EMBED_CHARS, SAFE_EMBED_CHARS;
 var init_message_ingest = __esm({
   "src/recall/message-ingest.ts"() {
     "use strict";
@@ -13701,11 +13736,13 @@ var init_message_ingest = __esm({
     init_db();
     init_paths();
     init_url_path_resolver();
+    init_log();
     init_jsonl_reader();
     init_claude_entry_adapter();
     init_codex_jsonl_reader();
     init_codex_jsonl_adapter();
     MAX_EMBED_CHARS = 14e3;
+    SAFE_EMBED_CHARS = 6e3;
   }
 });
 
@@ -14924,8 +14961,8 @@ function cudaAssetUrl() {
   const tag = version ? `download/v${version}` : "latest/download";
   return `https://github.com/${owner}/${repo}/releases/${tag}/${CUDA_ASSET_NAME}`;
 }
-function binCudaDir() {
-  return (0, import_node_path9.join)(recallRoot(), "bin-cuda");
+function cudaBackendLib() {
+  return (0, import_node_path9.join)(binDir(), "libggml-cuda.so");
 }
 async function readVram() {
   try {
@@ -14957,7 +14994,7 @@ async function detectGpu(opts = {}) {
   if (p === "win32" && a3 === "x64") {
     return { ...base, cudaAvailable: "prebuilt", plannedMode: "gpu" };
   }
-  if (p === "linux" && a3 === "x64" && (0, import_node_fs9.existsSync)(binCudaDir())) {
+  if (p === "linux" && a3 === "x64" && (0, import_node_fs9.existsSync)(cudaBackendLib())) {
     return { ...base, cudaAvailable: "reuse-existing", plannedMode: "gpu" };
   }
   if (p === "linux" && a3 === "x64") {
@@ -14966,10 +15003,7 @@ async function detectGpu(opts = {}) {
   return { ...base, cudaAvailable: "none", plannedMode: "cpu" };
 }
 function stderrIndicatesOffload(stderr) {
-  if (/ggml_cuda_init/.test(stderr))
-    return true;
-  const m2 = /offloaded\s+(\d+)\s*\/\s*\d+\s+layers?\s+to\s+GPU/i.exec(stderr);
-  return m2 ? Number(m2[1]) > 0 : false;
+  return /ggml_cuda_init|found\s+\d+\s+CUDA\s+devices|loaded CUDA backend|using device CUDA|CUDA0[^\n]*buffer/i.test(stderr);
 }
 async function defaultProbe(args) {
   const { binaryPath: binaryPath2, modelPath, libDir, ngl, platform: platform2 } = args;
@@ -14990,6 +15024,12 @@ async function defaultProbe(args) {
     "array",
     "-c",
     "8192",
+    // Probe with the SAME flags the real embed server uses (esp. -fa). Without
+    // Flash Attention the probe allocates the ~3.7 GB O(n²) attention buffer and
+    // OOMs on a 4 GB card — failing the probe and falling back to CPU even
+    // though the actual -fa server (~2.9 GB) would have fit. Probe must mirror
+    // runtime to gate GPU adoption correctly.
+    "-fa",
     "--rope-scaling",
     "yarn",
     "--rope-freq-scale",
@@ -15044,7 +15084,7 @@ async function runGpuPhase(opts = {}) {
     return { mode, libDir: libDir2, ngl: mode === "gpu" ? DEFAULT_NGL : 0, cudaAvailable: info.cudaAvailable, ...reason ? { reason } : {} };
   };
   if (info.plannedMode === "cpu") {
-    const reason = info.detected && info.vendor === "nvidia" ? "NVIDIA GPU detected but no usable CUDA libs (no prebuilt for this platform; no ~/.recall/bin-cuda/ build present)" : void 0;
+    const reason = info.detected && info.vendor === "nvidia" ? "NVIDIA GPU detected but no usable CUDA libs (no prebuilt for this platform; no ~/.recall/bin/libggml-cuda.so present)" : void 0;
     return persist("cpu", null, reason);
   }
   if (info.cudaAvailable === "metal") {
@@ -15052,7 +15092,7 @@ async function runGpuPhase(opts = {}) {
   }
   let libDir;
   if (info.cudaAvailable === "reuse-existing") {
-    libDir = binCudaDir();
+    libDir = binDir();
   } else if (info.cudaAvailable === "prebuilt" && p === "win32") {
     libDir = binDir();
   } else if (info.cudaAvailable === "prebuilt") {
@@ -15087,8 +15127,8 @@ async function runGpuPhase(opts = {}) {
   }
 }
 async function defaultStage(args) {
-  const target = binCudaDir();
-  const libPath = (0, import_node_path9.join)(target, "libggml-cuda.so");
+  const target = binDir();
+  const libPath = cudaBackendLib();
   if ((0, import_node_fs9.existsSync)(libPath))
     return target;
   if (args.offline)

package/dist/stop-hook.js

@@ -1807,6 +1807,9 @@ function normalizePath(p) {
   return normalized;
 }
 
+// src/recall/message-ingest.ts
+init_log();
+
 // src/adapters/claude/jsonl-reader.ts
 var fs = __toESM(require("fs"));
 init_log();
@@ -2217,8 +2220,7 @@ function emitFunctionCall(payload, base, outputIndex, _counter) {
   const outputRecord = outputIndex.get(callId);
   if (outputRecord) {
     const outputPayload = outputRecord.payload;
-    const rawOutput = outputPayload.output;
-    const { exitCode, body } = parseExecOutputHeader(rawOutput);
+    const { exitCode, body } = parseExecOutputHeader(outputPayload.output);
     const isError = exitCode !== 0;
     const toolResult = {
       type: "tool_result",
@@ -2356,7 +2358,7 @@ function emitApplyPatch(callId, input, base, outputIndex) {
   return entries;
 }
 function buildCustomToolResult(parentUuid, callId, outputRecord, base) {
-  const rawOutput = outputRecord.payload.output;
+  const rawOutput = coerceOutputText(outputRecord.payload.output);
   let content;
   let isError = false;
   try {
@@ -2437,7 +2439,7 @@ function emitOrphanedOutput(payload, subtype, base, outputIndex, _counter) {
     return [];
   if (!outputIndex.has(callId))
     return [];
-  const rawOutput = payload.output;
+  const rawOutput = coerceOutputText(payload.output);
   let content;
   let isError = false;
   if (subtype === "function_call_output") {
@@ -2550,7 +2552,18 @@ function mapFunctionCall(name, args) {
       return { toolName: name, toolInput: args };
   }
 }
-function parseExecOutputHeader(output) {
+function coerceOutputText(output) {
+  if (typeof output === "string")
+    return output;
+  if (Array.isArray(output)) {
+    return output.map(
+      (item) => item && typeof item === "object" && typeof item.text === "string" ? item.text : ""
+    ).filter(Boolean).join("\n");
+  }
+  return "";
+}
+function parseExecOutputHeader(rawOutput) {
+  const output = coerceOutputText(rawOutput);
   if (!output)
     return { exitCode: 0, body: "" };
   const currentMatch = output.match(/Process exited with code (\d+)/);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "crispy-recall",
-  "version": "0.1.1",
+  "version": "0.1.3",
   "description": "Local session transcript memory for Claude Code and Codex — search past sessions with FTS5 + semantic vectors.",
   "license": "MIT",
   "author": "Sylvester Wong",