@takara-ai/miru-code 0.4.12 → 0.6.0

package/README.md

@@ -6,6 +6,12 @@ Find code by meaning, not grep. Miru returns the best **chunks** (path, lines, s
 
 **Requires:** [Bun](https://bun.sh) 1.1+ · [Takara API key](https://takara.ai)
 
+## Privacy and API usage
+
+Miru sends **file contents** to the [Takara inference API](https://takara.ai) when building an index and when embedding search queries. Chunks from your repo are transmitted over HTTPS to generate embeddings. API usage may incur cost depending on your Takara plan.
+
+If you index proprietary code, make sure that sending snippets to Takara's endpoint fits your security and compliance requirements. Use `MIRU_WORKSPACE_ROOT` to restrict MCP indexing to a single workspace directory.
+
 ## Install
 
 ```bash
@@ -143,6 +149,9 @@ const results = await index.search({ query: "BM25 tokenize", topK: 10 });
 |----------|-------|
 | `TAKARA_API_KEY` | Required |
 | `MIRU_OPENAI_BASE_URL` | Default `https://infer.takara.ai/v1` |
+| `MIRU_WORKSPACE_ROOT` | Restrict MCP local `repo` paths to this directory |
+| `MIRU_MAX_INDEX_FILES` | Cap files indexed per operation |
+| `MIRU_ALLOW_HTTP_GIT` | Set `1` to allow plain `http://` git clones |
 | `MIRU_MCP_WATCH` | Set `0` to disable MCP file watch |
 | `NO_COLOR` | Disable CLI colors |
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@takara-ai/miru-code",
-  "version": "0.4.12",
+  "version": "0.6.0",
   "description": "Miru (見る) — hybrid code search for agents. Bun, TypeScript, Takara embeddings.",
   "type": "module",
   "license": "MIT",
@@ -50,6 +50,11 @@
     "lint": "biome check .",
     "lint:fix": "biome check --write .",
     "prepublishOnly": "bun run typecheck && bun test",
+    "benchmark:ab": "bun run scripts/search-ab-local.ts",
+    "benchmark:ab-multi": "bun run scripts/search-ab-multi.ts",
+    "benchmark:recall": "bun run scripts/benchmark-recall-tokens.ts",
+    "benchmark:vs-grep": "bun run scripts/benchmark-miru-vs-grep.ts",
+    "benchmark:workflow": "bun run scripts/benchmark-workflow-tokens.ts",
     "benchmark:quality": "bun run scripts/benchmark-quality.ts",
     "benchmark:quantization": "bun run scripts/benchmark-quantization.ts"
   },

package/src/agents/claude.md

@@ -3,30 +3,3 @@ name: miru-code
 description: Code search via Miru MCP. Use search, expand, and find_related. DO NOT use Grep, Glob, or Read for exploration when Miru MCP is available.
 tools: Bash, Read
 ---
-
-When Miru MCP is available, use MCP `search`, `expand`, and `find_related` — not Grep, Glob, or Read-for-exploration.
-
-| Task | Tool |
-|------|------|
-| How/where/what handles X? | Miru `search` (once) |
-| Same file, more context | Miru `expand` |
-| Similar implementations | Miru `find_related` |
-| Exact literal confirmation | Grep |
-| Edit a located file | Read |
-
-### MCP workflow
-
-1. Call `search` with `repo` set to the project root.
-2. If context is cut off, call `expand` with the hit's `file_path` and `line`.
-3. Use `find_related` for similar code in other files — not for same-file context.
-4. Read files only to edit or when `expand` still lacks context.
-
-### CLI fallback
-
-```bash
-miru search "authentication flow" .
-miru expand src/auth.ts 42 .
-miru find-related src/auth.ts 42 .
-```
-
-If `miru` is not on `$PATH`, use `bunx @takara-ai/miru-code`.

package/src/agents/copilot.md

@@ -2,15 +2,3 @@
 name: miru-code
 description: Code search via Miru MCP. Use search, expand, and find_related instead of grep_search or codebase_search when Miru is connected.
 ---
-
-When Miru MCP is available, use `search`, `expand`, and `find_related` instead of `grep_search`, `codebase_search`, or broad file reads.
-
-| Task | Tool |
-|------|------|
-| How/where/what handles X? | Miru `search` (once) |
-| Same file, more context | Miru `expand` |
-| Similar implementations | Miru `find_related` |
-| Exact literal confirmation | grep_search |
-| Edit a located file | read_file |
-
-Pass project root as `repo` on every MCP call.

package/src/agents/cursor.md

@@ -2,29 +2,3 @@
 name: miru-code
 description: Code search via Miru MCP. Use search, expand, and find_related. DO NOT use Grep, Glob, or SemanticSearch when Miru MCP is available.
 ---
-
-Use Miru MCP tools when connected:
-
-- `search` — one call per question; always pass project root as `repo`
-- `expand` — adjacent chunks in the same file when a hit looks truncated (`file_path` + `line`)
-- `find_related` — similar code in other files after a search hit
-
-Do not re-search with paraphrases after a good first hit. Do not read whole files for exploration.
-
-| Task | Tool |
-|------|------|
-| How/where/what handles X? | Miru `search` (once) |
-| Same file, more context | Miru `expand` |
-| Similar implementations | Miru `find_related` |
-| Exact literal confirmation | Grep |
-| Edit a located file | Read |
-
-CLI fallback when MCP is unavailable:
-
-```bash
-miru search "authentication flow" .
-miru expand src/auth.ts 42 .
-miru find-related src/auth.ts 42 .
-```
-
-If `miru` is not on `$PATH`, use `bunx @takara-ai/miru-code`.

package/src/agents/gemini.md

@@ -2,17 +2,3 @@
 name: miru-code
 description: Code search via Miru MCP. Use search, expand, and find_related instead of grep_search or read_file for exploration.
 ---
-
-When Miru MCP is available, use `search`, `expand`, and `find_related` instead of `grep_search`, `glob`, or `read_file` for codebase exploration.
-
-| Task | Tool |
-|------|------|
-| How/where/what handles X? | Miru `search` (once) |
-| Same file, more context | Miru `expand` |
-| Similar implementations | Miru `find_related` |
-| Exact literal confirmation | grep_search |
-| Edit a located file | read_file |
-
-Pass project root as `repo` on every MCP call. Do not re-search with paraphrases after a good first hit.
-
-CLI fallback: `miru search "query" .` or `bunx @takara-ai/miru-code search "query" .`

package/src/agents/kiro.md

@@ -2,9 +2,3 @@
 name: miru-code
 description: Code search via Miru MCP. Use search, expand, and find_related for codebase exploration instead of built-in grep or file search.
 ---
-
-When Miru MCP is available, use `search`, `expand`, and `find_related` before any built-in search or broad file reads.
-
-Pass project root as `repo`. Use `expand` for truncated hits in the same file. Use Grep only for exact literal confirmation.
-
-CLI fallback: `miru search "query" .`

package/src/agents/opencode.md

@@ -3,9 +3,3 @@ name: miru-code
 description: Code search via Miru MCP. Use search, expand, and find_related instead of grep or glob for exploration.
 mode: subagent
 ---
-
-When Miru MCP is available, use `search`, `expand`, and `find_related` instead of grep, glob, or read-for-exploration.
-
-Pass project root as `repo` on every MCP call.
-
-CLI fallback: `miru search "query" .`

package/src/agents.ts

@@ -1,17 +1,53 @@
 import { mkdir, readFile } from "node:fs/promises";
 import { dirname, join } from "node:path";
+import { buildSubagentBody, type NativeToolNames } from "./installer/search-policy.ts";
 
 export type AgentId = "claude" | "copilot" | "cursor" | "gemini" | "kiro" | "opencode";
 
 const AGENTS_DIR = join(import.meta.dir, "agents");
 
+const AGENT_NATIVE_TOOLS: Record<AgentId, NativeToolNames> = {
+  claude: {
+    explorationDenied: "Grep, Glob, SemanticSearch, or Read",
+    grep: "Grep",
+    read: "Read",
+  },
+  cursor: {
+    explorationDenied: "Grep, Glob, SemanticSearch, or Read",
+    grep: "Grep",
+    read: "Read",
+  },
+  copilot: {
+    explorationDenied: "grep_search, codebase_search, glob, or read_file",
+    grep: "grep_search",
+    read: "read_file",
+  },
+  gemini: {
+    explorationDenied: "grep_search, glob, or read_file",
+    grep: "grep_search",
+    read: "read_file",
+  },
+  kiro: {
+    explorationDenied: "built-in grep, file search, or broad file reads",
+    grep: "Grep",
+    read: "Read",
+  },
+  opencode: {
+    explorationDenied: "grep, glob, or read-for-exploration",
+    grep: "grep",
+    read: "read",
+  },
+};
+
 export function agentDestination(agent: AgentId): string {
   const baseDir = agent === "copilot" ? ".github" : `.${agent}`;
   return join(baseDir, "agents", "miru-code.md");
 }
 
 export async function loadAgentTemplate(agent: AgentId): Promise<string> {
-  return readFile(join(AGENTS_DIR, `${agent}.md`), "utf-8");
+  const frontmatter = (await readFile(join(AGENTS_DIR, `${agent}.md`), "utf-8")).trim();
+  const body = buildSubagentBody(AGENT_NATIVE_TOOLS[agent]);
+  return `${frontmatter}\n\n${body}\n`;
 }
 
 export async function writeAgentFile(

package/src/cli-ui.ts

@@ -1,4 +1,5 @@
 import type { SearchResult } from "./types.ts";
+import { formatRelevanceScore } from "./utils.ts";
 
 const ANSI = {
   reset: "\x1b[0m",
@@ -145,6 +146,8 @@ export function formatSearchResultsPretty(query: string, results: SearchResult[]
   lines.push(bold(label) + dim(` for `) + cyan(`"${query}"`));
   lines.push("");
 
+  const maxScore = results.reduce((max, result) => Math.max(max, result?.score ?? 0), 0);
+
   for (let i = 0; i < results.length; i++) {
     const result = results[i];
     if (!result) {
@@ -154,7 +157,7 @@ export function formatSearchResultsPretty(query: string, results: SearchResult[]
     const location = `${chunk.file_path}:${chunk.start_line}-${chunk.end_line}`;
     const lang = chunk.language ? dim(`  ${chunk.language}`) : "";
     const rank = dim(`[${i + 1}]`);
-    const scoreText = magenta(score.toFixed(3));
+    const scoreText = magenta(formatRelevanceScore(score, maxScore));
 
     lines.push(`${rank} ${scoreText}  ${bold(location)}${lang}`);
     lines.push(dim("─".repeat(Math.min(PREVIEW_WIDTH, 52))));

package/src/cli.ts

@@ -397,15 +397,7 @@ async function runCli(argv: string[]): Promise<void> {
     }
     const line = Number(lineRaw);
     const path = resolveSearchPath(sizedRest[2] ?? process.cwd());
-    try {
-      await runFindRelated(path, filePath, line, topK, content, jsonFlag);
-    } catch (err) {
-      if (err instanceof RelatedChunkNotFoundError) {
-        fail(err.message);
-        process.exit(1);
-      }
-      throw err;
-    }
+    await runFindRelated(path, filePath, line, topK, content, jsonFlag);
     return;
   }
 

package/src/concurrency.ts

@@ -11,11 +11,7 @@ export function resolveCpuCount(): number {
 
 /** Worker pool size: logical CPUs minus 2, minimum 1. Override with MIRU_CONCURRENCY. */
 export function resolveWorkerConcurrency(): number {
-  const raw =
-    process.env.MIRU_CONCURRENCY ??
-    process.env.SEMBLE_CONCURRENCY ??
-    process.env.MIRU_WORKERS ??
-    process.env.SEMBLE_WORKERS;
+  const raw = process.env.MIRU_CONCURRENCY ?? process.env.MIRU_WORKERS;
   if (raw) {
     const n = Number(raw);
     if (Number.isFinite(n) && n >= 1) {

package/src/embeddings/openai.ts

@@ -32,14 +32,14 @@ export interface EmbeddingBackend {
 
 export function resolveEmbeddingModel(): string {
   return envFirstString(
-    ["MIRU_OPENAI_EMBEDDING_MODEL", "SEMBLE_OPENAI_EMBEDDING_MODEL", "OPENAI_EMBEDDING_MODEL"],
+    ["MIRU_OPENAI_EMBEDDING_MODEL", "OPENAI_EMBEDDING_MODEL"],
     DEFAULT_MODEL,
   );
 }
 
 export function resolveMaxEmbedChars(): number {
   return (
-    envOptionalInt(["MIRU_MAX_EMBED_CHARS", "SEMBLE_MAX_EMBED_CHARS"], 256) ??
+    envOptionalInt(["MIRU_MAX_EMBED_CHARS"], 256) ??
     DEFAULT_MAX_EMBED_CHARS
   );
 }
@@ -73,13 +73,26 @@ function stripLoneSurrogates(text: string): string {
 
 export function sanitizeEmbeddingInput(text: string): string {
   const out = stripLoneSurrogates(text);
-  // Some OpenAI-compatible gateways mis-handle backslash escapes in JSON text
-  // payloads. Keep a benchmark fallback mode to forcefully neutralize them.
-  const mode = process.env.MIRU_EMBED_ESCAPE_MODE ?? process.env.SEMBLE_EMBED_ESCAPE_MODE ?? "quad";
+  const mode = process.env.MIRU_EMBED_ESCAPE_MODE ?? "preserve";
+  if (mode === "quad") {
+    return out.replaceAll("\\", "\\\\\\\\");
+  }
   if (mode === "strip") {
     return out.replaceAll("\\", "/");
   }
-  return out.replaceAll("\\", "\\\\\\\\");
+  return out;
+}
+
+const TRANSIENT_EMBED_STATUSES = new Set([429, 500, 502, 503, 504]);
+const MAX_TRANSIENT_EMBED_RETRIES = 4;
+
+function isTransientEmbeddingError(err: unknown): boolean {
+  const status = (err as { status?: number }).status;
+  return status !== undefined && TRANSIENT_EMBED_STATUSES.has(status);
+}
+
+function transientEmbedBackoffMs(attempt: number): number {
+  return Math.min(250 * 2 ** attempt, 4000);
 }
 
 function isPayloadTooLargeError(err: unknown): boolean {
@@ -95,11 +108,7 @@ function isPayloadTooLargeError(err: unknown): boolean {
 }
 
 export function resolveEmbeddingDimensions(model?: string): number | undefined {
-  const fromEnv = envOptionalInt([
-    "MIRU_EMBEDDING_DIMENSIONS",
-    "SEMBLE_EMBEDDING_DIMENSIONS",
-    "OPENAI_EMBEDDING_DIMENSIONS",
-  ]);
+  const fromEnv = envOptionalInt(["MIRU_EMBEDDING_DIMENSIONS", "OPENAI_EMBEDDING_DIMENSIONS"]);
   if (fromEnv != null) {
     return fromEnv;
   }
@@ -108,17 +117,14 @@ export function resolveEmbeddingDimensions(model?: string): number | undefined {
 
 export function resolveEmbeddingBatchSize(): number {
   return (
-    envOptionalInt([
-      "MIRU_EMBEDDING_BATCH_SIZE",
-      "SEMBLE_EMBEDDING_BATCH_SIZE",
-      "OPENAI_EMBEDDING_BATCH_SIZE",
-    ]) ?? DEFAULT_BATCH_SIZE
+    envOptionalInt(["MIRU_EMBEDDING_BATCH_SIZE", "OPENAI_EMBEDDING_BATCH_SIZE"]) ??
+    DEFAULT_BATCH_SIZE
   );
 }
 
 export function resolveEmbeddingBaseUrl(): string {
   return envFirstString(
-    ["MIRU_OPENAI_BASE_URL", "SEMBLE_OPENAI_BASE_URL", "OPENAI_BASE_URL"],
+    ["MIRU_OPENAI_BASE_URL", "OPENAI_BASE_URL"],
     DEFAULT_BASE_URL,
   ).replace(/\/$/, "");
 }
@@ -282,23 +288,42 @@ export class OpenAIEmbeddingBackend implements EmbeddingBackend {
     return out;
   }
 
+  private async requestEmbeddings(texts: string[]): Promise<EmbeddingResponse> {
+    let lastErr: unknown;
+    for (let attempt = 0; attempt <= MAX_TRANSIENT_EMBED_RETRIES; attempt++) {
+      try {
+        const started = performance.now();
+        const response = await this.client.createEmbeddings(
+          texts,
+          this.model,
+          this.requestedDimensions,
+        );
+        const elapsed = performance.now() - started;
+        this.stats.requests += 1;
+        this.stats.inputItems += texts.length;
+        this.stats.inputChars += texts.reduce((acc, text) => acc + text.length, 0);
+        this.stats.totalRttMs += elapsed;
+        this.stats.maxRttMs = Math.max(this.stats.maxRttMs, elapsed);
+        return response;
+      } catch (err: unknown) {
+        lastErr = err;
+        if (isTransientEmbeddingError(err) && attempt < MAX_TRANSIENT_EMBED_RETRIES) {
+          this.stats.retries += 1;
+          await Bun.sleep(transientEmbedBackoffMs(attempt));
+          continue;
+        }
+        throw err;
+      }
+    }
+    throw lastErr;
+  }
+
   private async embedBatchRawWithRetry(texts: string[]): Promise<Float32Array[]> {
     if (texts.length === 0) {
       return [];
     }
     try {
-      const started = performance.now();
-      const response = await this.client.createEmbeddings(
-        texts,
-        this.model,
-        this.requestedDimensions,
-      );
-      const elapsed = performance.now() - started;
-      this.stats.requests += 1;
-      this.stats.inputItems += texts.length;
-      this.stats.inputChars += texts.reduce((acc, text) => acc + text.length, 0);
-      this.stats.totalRttMs += elapsed;
-      this.stats.maxRttMs = Math.max(this.stats.maxRttMs, elapsed);
+      const response = await this.requestEmbeddings(texts);
       const vectors = vectorsFromResponse(response.data, texts.length);
       for (const vec of vectors) {
         if (this.requestedDimensions != null && vec.length !== this.requestedDimensions) {
@@ -341,7 +366,12 @@ export class OpenAIEmbeddingBackend implements EmbeddingBackend {
         const rightText = text.slice(mid);
         const left = await this.embedBatchRawWithRetry([leftText]);
         const right = await this.embedBatchRawWithRetry([rightText]);
-        return [...left, ...right];
+        const parts = [left[0], right[0]].filter((vec): vec is Float32Array => vec !== undefined);
+        if (parts.length === 0) {
+          this.stats.errors += 1;
+          throw err;
+        }
+        return [poolWindowVectors(parts)];
       }
       this.stats.errors += 1;
       throw err;

package/src/git.ts

@@ -7,7 +7,7 @@ const DEFAULT_CLONE_TIMEOUT_SEC = 60;
 
 export async function cloneGitRepository(url: string, ref?: string | null): Promise<string> {
   const timeoutSec =
-    envOptionalInt(["MIRU_CLONE_TIMEOUT", "SEMBLE_CLONE_TIMEOUT"], 1) ?? DEFAULT_CLONE_TIMEOUT_SEC;
+    envOptionalInt(["MIRU_CLONE_TIMEOUT"], 1) ?? DEFAULT_CLONE_TIMEOUT_SEC;
   const dir = await mkdtemp(join(tmpdir(), "miru-git-"));
   const args = ["clone", "--depth", "1"];
   if (ref) {

package/src/index/create.ts

@@ -3,9 +3,11 @@ import { chunkSource } from "../chunking/chunking.ts";
 import { resolveWorkerConcurrency } from "../concurrency.ts";
 import type { EmbeddingBackend } from "../embeddings/openai.ts";
 import { envOptionalInt } from "../env.ts";
+import { searchImprovementsEnabled } from "../ranking/features.ts";
 import { tokenize } from "../tokens.ts";
 import type { Chunk, ContentType } from "../types.ts";
 import { BM25Index } from "./bm25.ts";
+import { loadRootEntryChunks } from "./entry-chunks.ts";
 import { walkFiles } from "./file-walker.ts";
 import { detectLanguage, getExtensions, getFileStatus, readFileText } from "./files.ts";
 import type { SemanticIndex } from "./semantic-index.ts";
@@ -17,31 +19,36 @@ const DEFAULT_PIPELINE_EMBED_INFLIGHT = 8;
 
 function resolvePipelineEmbedBatch(): number {
   return (
-    envOptionalInt(["MIRU_PIPELINE_EMBED_BATCH", "SEMBLE_PIPELINE_EMBED_BATCH"], 1) ??
+    envOptionalInt(["MIRU_PIPELINE_EMBED_BATCH"], 1) ??
     DEFAULT_PIPELINE_EMBED_BATCH
   );
 }
 
 function resolvePipelineEmbedInflight(): number {
   return (
-    envOptionalInt(["MIRU_PIPELINE_EMBED_INFLIGHT", "SEMBLE_PIPELINE_EMBED_INFLIGHT"], 1) ??
+    envOptionalInt(["MIRU_PIPELINE_EMBED_INFLIGHT"], 1) ??
     DEFAULT_PIPELINE_EMBED_INFLIGHT
   );
 }
 
+function resolveMaxIndexFiles(): number | undefined {
+  return envOptionalInt(["MIRU_MAX_INDEX_FILES"], 1);
+}
+
 export async function createIndexFromPath(
   path: string,
   embeddings: EmbeddingBackend,
   content: ContentType[] = ["code"],
   displayRoot?: string,
 ): Promise<{ bm25: BM25Index; semantic: SemanticIndex; chunks: Chunk[] }> {
-  const profile = process.env.MIRU_PROFILE === "1" || process.env.SEMBLE_PROFILE === "1";
+  const profile = process.env.MIRU_PROFILE === "1";
   const started = performance.now();
   let fileProcessMs = 0;
   let embedBackpressureWaitMs = 0;
   let fileEnumerated = 0;
   let fileTasks = 0;
   let emptyFileTasks = 0;
+  let fileErrors = 0;
 
   if (profile && "resetStats" in embeddings && typeof embeddings.resetStats === "function") {
     embeddings.resetStats();
@@ -53,6 +60,7 @@ export async function createIndexFromPath(
   const fileConcurrency = resolveWorkerConcurrency();
   const embedBatchSize = resolvePipelineEmbedBatch();
   const maxEmbedInflight = resolvePipelineEmbedInflight();
+  const maxIndexFiles = resolveMaxIndexFiles();
   const readyBySeq = new Map<number, Chunk[]>();
   const pendingEmbedChunks: Chunk[] = [];
   const emittedChunkBatches: Chunk[][] = [];
@@ -74,7 +82,7 @@ export async function createIndexFromPath(
       const chunkPath = displayRoot ? relative(root, filePath).replace(/\\/g, "/") : filePath;
       return chunkSource(source, chunkPath, language);
     } catch {
-      emptyFileTasks++;
+      fileErrors++;
       return [];
     } finally {
       fileProcessMs += performance.now() - t;
@@ -134,6 +142,9 @@ export async function createIndexFromPath(
 
   for await (const filePath of walkFiles(resolved, extensions)) {
     fileEnumerated++;
+    if (maxIndexFiles != null && fileEnumerated > maxIndexFiles) {
+      throw new Error(`Index file budget exceeded: max ${maxIndexFiles} files per operation`);
+    }
     while (runningFiles.size >= fileConcurrency) {
       await Promise.race(runningFiles);
     }
@@ -145,6 +156,14 @@ export async function createIndexFromPath(
     await Promise.race(runningFiles);
   }
 
+  if (searchImprovementsEnabled()) {
+    const entryChunks = await loadRootEntryChunks(resolved, displayRoot ? root : undefined);
+    if (entryChunks.length > 0) {
+      readyBySeq.set(fileSeq++, entryChunks);
+      flushReadySequentialChunks();
+    }
+  }
+
   flushReadySequentialChunks();
   await maybeScheduleEmbed(true);
   await Promise.all(embedPromises);
@@ -174,6 +193,7 @@ export async function createIndexFromPath(
         file_enumerated: fileEnumerated,
         file_tasks: fileTasks,
         empty_or_skipped_files: emptyFileTasks,
+        file_errors: fileErrors,
         chunks: chunks.length,
         vectors: vectors.length,
         pipeline: {

package/src/index/entry-chunks.ts

@@ -0,0 +1,71 @@
+import { join, relative } from "node:path";
+import type { Chunk } from "../types.ts";
+
+function indexedPath(absPath: string, repoRoot: string, displayRoot?: string): string {
+  if (displayRoot) {
+    return relative(displayRoot, absPath).replace(/\\/g, "/");
+  }
+  return absPath.replace(/\\/g, "/");
+}
+
+async function packageJsonEntryChunks(
+  absPath: string,
+  repoRoot: string,
+  displayRoot?: string,
+): Promise<Chunk[]> {
+  let pkg: Record<string, unknown>;
+  try {
+    pkg = JSON.parse(await Bun.file(absPath).text()) as Record<string, unknown>;
+  } catch {
+    return [];
+  }
+
+  const entryLines = ["[package entry]"];
+
+  const bin = pkg.bin;
+  if (typeof bin === "string") {
+    entryLines.push(`bin: ${bin}`);
+  } else if (bin && typeof bin === "object") {
+    for (const [name, target] of Object.entries(bin)) {
+      entryLines.push(`bin ${name}: ${String(target)}`);
+    }
+  }
+
+  if (typeof pkg.main === "string") {
+    entryLines.push(`main: ${pkg.main}`);
+  }
+
+  const types = pkg.types ?? pkg.typings;
+  if (typeof types === "string") {
+    entryLines.push(`types: ${types}`);
+  }
+
+  if (entryLines.length <= 1) {
+    return [];
+  }
+
+  const filePath = indexedPath(absPath, repoRoot, displayRoot);
+  const content = `${entryLines.join("\n")}\n`;
+
+  return [
+    {
+      content,
+      file_path: filePath,
+      start_line: 1,
+      end_line: entryLines.length,
+      language: "json",
+    },
+  ];
+}
+
+/** Synthetic chunks for root config entry points (bin/main) when indexing code repos. */
+export async function loadRootEntryChunks(
+  repoRoot: string,
+  displayRoot?: string,
+): Promise<Chunk[]> {
+  const pkgPath = join(repoRoot, "package.json");
+  if (!(await Bun.file(pkgPath).exists())) {
+    return [];
+  }
+  return packageJsonEntryChunks(pkgPath, repoRoot, displayRoot);
+}

package/src/index/vector-storage.ts

@@ -6,7 +6,7 @@ export type SemanticStorage = "int8" | "float32";
 
 /** Default: int8 quantized vectors (~4x less RAM). Set MIRU_FLOAT_VECTORS=1 for float32. */
 export function resolveSemanticStorage(): SemanticStorage {
-  return process.env.MIRU_FLOAT_VECTORS === "1" || process.env.SEMBLE_FLOAT_VECTORS === "1"
+  return process.env.MIRU_FLOAT_VECTORS === "1"
     ? "float32"
     : "int8";
 }

package/src/installer/hooks/search-guard.ts

@@ -1,3 +1,5 @@
+import { SEARCH_GUARD_EXPAND_HINT } from "../search-policy.ts";
+
 const REDIRECT_PREFIX =
   "Use Miru MCP instead of built-in search tools (pass project root as `repo`). " +
   "Grep/Glob/Shell are only for exact literal confirmation or non-code tasks.";
@@ -119,10 +121,7 @@ export function searchGuardBlockReason(
   toolInput: Record<string, unknown>,
 ): string {
   const query = suggestedMiruQuery(toolName, toolInput);
-  return (
-    `${REDIRECT_PREFIX} Try Miru \`search\` with query "${query}". ` +
-    "If the hit looks truncated, call `expand` on its file_path + line — do not re-search or read the whole file."
-  );
+  return `${REDIRECT_PREFIX} Try Miru \`search\` with query "${query}". ${SEARCH_GUARD_EXPAND_HINT}`;
 }
 
 /** Map host-specific stdin JSON into a common evaluation shape. */