brain-cache 0.3.3 → 0.3.5

package/dist/{askCodebase-OAAG35G5.js → askCodebase-NYPWHDEZ.js}

@@ -1,11 +1,11 @@
 #!/usr/bin/env node
-import {
-  formatTokenSavings
-} from "./chunk-GGOUKACO.js";
 import {
   runBuildContext
-} from "./chunk-QNPB2UD2.js";
+} from "./chunk-AEJKKJDU.js";
 import "./chunk-5FXXZBZV.js";
+import {
+  formatTokenSavings
+} from "./chunk-GGOUKACO.js";
 import "./chunk-BF5UDEIF.js";
 import "./chunk-GR6QXZ4J.js";
 import "./chunk-V4ARVFRG.js";

package/dist/{buildContext-QIMH5LQY.js → buildContext-IND6F67G.js}

@@ -1,7 +1,7 @@
 #!/usr/bin/env node
 import {
   runBuildContext
-} from "./chunk-QNPB2UD2.js";
+} from "./chunk-AEJKKJDU.js";
 import "./chunk-5FXXZBZV.js";
 import "./chunk-BF5UDEIF.js";
 import "./chunk-GR6QXZ4J.js";

package/dist/{chunk-QNPB2UD2.js → chunk-AEJKKJDU.js}

@@ -1,6 +1,7 @@
 #!/usr/bin/env node
 import {
-  assembleContext
+  assembleContext,
+  countChunkTokens
 } from "./chunk-5FXXZBZV.js";
 import {
   RETRIEVAL_STRATEGIES,
@@ -26,6 +27,7 @@ import {
 } from "./chunk-3SFDFUEX.js";
 
 // src/workflows/buildContext.ts
+import { readFile } from "fs/promises";
 import { resolve } from "path";
 async function runBuildContext(query, opts) {
   const profile = await readProfile();
@@ -62,8 +64,20 @@ async function runBuildContext(query, opts) {
   const results = await searchChunks(table, queryVector, strategy);
   const deduped = deduplicateChunks(results);
   const assembled = assembleContext(deduped, { maxTokens });
-  const estimatedWithoutBraincache = indexState.totalTokens;
-  const reductionPct = estimatedWithoutBraincache > 0 ? Math.round((1 - assembled.tokenCount / estimatedWithoutBraincache) * 100) : 0;
+  const uniqueFiles = [...new Set(assembled.chunks.map((c) => c.filePath))];
+  let estimatedWithoutBraincache = 0;
+  for (const filePath of uniqueFiles) {
+    try {
+      const fileContent = await readFile(filePath, "utf-8");
+      estimatedWithoutBraincache += countChunkTokens(fileContent);
+    } catch {
+    }
+  }
+  const rawChunkTokens = assembled.chunks.reduce(
+    (sum, c) => sum + countChunkTokens(c.content),
+    0
+  );
+  const reductionPct = estimatedWithoutBraincache > 0 ? Math.round((1 - rawChunkTokens / estimatedWithoutBraincache) * 100) : 0;
   const result = {
     content: assembled.content,
     chunks: assembled.chunks,

package/dist/claude-md-section-YPMDPAPW.js

@@ -0,0 +1,34 @@
+#!/usr/bin/env node
+
+// src/lib/claude-md-section.ts
+var CLAUDE_MD_SECTION = `
+## Brain-Cache MCP Tools
+
+**IMPORTANT: Use brain-cache tools as your FIRST action before reading files or using Grep/Glob.**
+
+### build_context (use for understanding)
+
+Call \`mcp__brain-cache__build_context\` with the user's question BEFORE reading individual files. This retrieves semantically relevant code from across the entire repo, deduplicates it, and returns a token-budgeted context block.
+
+**Always use for questions like:**
+- "How does X work?" / "Explain X end to end"
+- "Walk me through the flow of X"
+- "What does this page/feature/component do?"
+- "Explain the architecture" / "How is the project structured?"
+- "What happens when Y is called?"
+- Any question that requires understanding code across multiple files
+- Debugging unfamiliar code paths or understanding error flows
+
+**Do NOT skip this tool and jump to reading files.** build_context returns better results with fewer tokens than manually reading files.
+
+### search_codebase (use for locating)
+
+Call \`mcp__brain-cache__search_codebase\` instead of Grep or Glob when locating functions, symbols, types, definitions, or implementations. It uses semantic search \u2014 finds code by meaning, not just keyword match.
+
+### doctor (use for diagnosing)
+
+Call \`mcp__brain-cache__doctor\` first when any brain-cache tool fails or returns unexpected results.
+`;
+export {
+  CLAUDE_MD_SECTION
+};

package/dist/cli.js

@@ -5,11 +5,11 @@ import {
 
 // src/cli/index.ts
 import { Command } from "commander";
-var version = "0.3.3";
+var version = "0.3.5";
 var program = new Command();
 program.name("brain-cache").description("Local AI runtime \u2014 GPU cache layer for Claude").version(version);
 program.command("init").description("Detect hardware, pull embedding model, create config directory").action(async () => {
-  const { runInit } = await import("./init-ILR3VJNO.js");
+  const { runInit } = await import("./init-L4CMLYVY.js");
   await runInit();
 });
 program.command("doctor").description("Report system health: GPU, VRAM tier, Ollama status").action(async () => {
@@ -17,7 +17,7 @@ program.command("doctor").description("Report system health: GPU, VRAM tier, Oll
   await runDoctor();
 });
 program.command("index").description("Index a codebase: parse, chunk, embed, and store in LanceDB").argument("[path]", "Directory to index (defaults to current directory)").option("-f, --force", "Force full reindex, ignoring cached file hashes").action(async (path, opts) => {
-  const { runIndex } = await import("./workflows-SYWKVVOK.js");
+  const { runIndex } = await import("./workflows-ZC24RRNX.js");
   await runIndex(path, { force: opts.force });
 });
 program.command("search").description("Search indexed codebase with a natural language query").argument("<query>", "Natural language query string").option("-n, --limit <n>", "Maximum number of results", "10").option("-p, --path <path>", "Project root directory").action(async (query, opts) => {
@@ -35,7 +35,7 @@ program.command("status").description(
 });
 program.command("context").description("Build token-budgeted context from codebase for a query").argument("<query>", "Natural language query string").option("-n, --limit <n>", "Maximum number of search results", "10").option("-b, --budget <tokens>", "Token budget for assembled context", "4096").option("-p, --path <path>", "Project root directory").option("--raw", "Output raw JSON (MCP transport compatible)").action(
   async (query, opts) => {
-    const { runBuildContext } = await import("./buildContext-QIMH5LQY.js");
+    const { runBuildContext } = await import("./buildContext-IND6F67G.js");
     const result = await runBuildContext(query, {
       limit: parseInt(opts.limit, 10),
       maxTokens: parseInt(opts.budget, 10),
@@ -61,7 +61,7 @@ ${formatTokenSavings({ tokensSent: result.metadata.tokensSent, estimatedWithout,
 program.command("ask").description(
   "Ask a natural language question about the codebase \u2014 retrieves context locally, reasons via Claude"
 ).argument("<question>", "Natural language question about the codebase").option("-b, --budget <tokens>", "Token budget for context retrieval", "4096").option("-p, --path <path>", "Project root directory").action(async (question, opts) => {
-  const { runAskCodebase } = await import("./askCodebase-OAAG35G5.js");
+  const { runAskCodebase } = await import("./askCodebase-NYPWHDEZ.js");
   const result = await runAskCodebase(question, {
     path: opts.path,
     maxContextTokens: parseInt(opts.budget, 10)

package/dist/{init-ILR3VJNO.js → init-L4CMLYVY.js}

@@ -86,13 +86,7 @@ async function runInit() {
     process.stderr.write("brain-cache: created .mcp.json with brain-cache MCP server.\n");
   }
   const claudeMdPath = "CLAUDE.md";
-  const brainCacheSection = `
-## Brain-Cache MCP Tools
-
-When answering "how does X work", "explain the architecture", or any question requiring multi-file reasoning, call \`mcp__brain-cache__build_context\` with the question BEFORE reading individual files. It returns pre-built, relevant context with token savings.
-
-When locating functions, symbols, types, or definitions, call \`mcp__brain-cache__search_codebase\` instead of using Grep or file-find tools. It searches the local vector index and returns ranked results.
-`;
+  const { CLAUDE_MD_SECTION: brainCacheSection } = await import("./claude-md-section-YPMDPAPW.js");
   if (existsSync(claudeMdPath)) {
     const content = readFileSync(claudeMdPath, "utf-8");
     if (content.includes("## Brain-Cache MCP Tools")) {

package/dist/mcp.js

@@ -1087,6 +1087,7 @@ async function runSearch(query, opts) {
 }
 
 // src/workflows/buildContext.ts
+import { readFile as readFile5 } from "fs/promises";
 import { resolve as resolve3 } from "path";
 async function runBuildContext(query, opts) {
   const profile = await readProfile();
@@ -1123,8 +1124,20 @@ async function runBuildContext(query, opts) {
   const results = await searchChunks(table, queryVector, strategy);
   const deduped = deduplicateChunks(results);
   const assembled = assembleContext(deduped, { maxTokens });
-  const estimatedWithoutBraincache = indexState.totalTokens;
-  const reductionPct = estimatedWithoutBraincache > 0 ? Math.round((1 - assembled.tokenCount / estimatedWithoutBraincache) * 100) : 0;
+  const uniqueFiles = [...new Set(assembled.chunks.map((c) => c.filePath))];
+  let estimatedWithoutBraincache = 0;
+  for (const filePath of uniqueFiles) {
+    try {
+      const fileContent = await readFile5(filePath, "utf-8");
+      estimatedWithoutBraincache += countChunkTokens(fileContent);
+    } catch {
+    }
+  }
+  const rawChunkTokens = assembled.chunks.reduce(
+    (sum, c) => sum + countChunkTokens(c.content),
+    0
+  );
+  const reductionPct = estimatedWithoutBraincache > 0 ? Math.round((1 - rawChunkTokens / estimatedWithoutBraincache) * 100) : 0;
   const result = {
     content: assembled.content,
     chunks: assembled.chunks,
@@ -1144,7 +1157,7 @@ async function runBuildContext(query, opts) {
 }
 
 // src/mcp/index.ts
-var version = "0.3.3";
+var version = "0.3.5";
 var log9 = childLogger("mcp");
 var server = new McpServer({ name: "brain-cache", version });
 server.registerTool(

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "brain-cache",
-  "version": "0.3.3",
+  "version": "0.3.5",
   "description": "Local MCP-first context engine for Claude. Index your codebase, retrieve only what matters, and cut token usage.",
   "license": "MIT",
   "type": "module",

package/dist/{workflows-SYWKVVOK.js → workflows-ZC24RRNX.js}

@@ -1,10 +1,10 @@
 #!/usr/bin/env node
-import {
-  formatTokenSavings
-} from "./chunk-GGOUKACO.js";
 import {
   countChunkTokens
 } from "./chunk-5FXXZBZV.js";
+import {
+  formatTokenSavings
+} from "./chunk-GGOUKACO.js";
 import {
   embedBatchWithRetry
 } from "./chunk-GR6QXZ4J.js";