grepmax 0.26.14 → 0.26.15

package/.claude-plugin/marketplace.json

@@ -10,7 +10,7 @@
       "name": "grepmax",
       "source": "./plugins/grepmax",
       "description": "Semantic code search for Claude Code. Automatically indexes your project and provides intelligent search capabilities.",
-      "version": "0.26.14",
+      "version": "0.26.15",
       "author": {
         "name": "Robert Owens",
         "email": "robowens@me.com"

package/dist/lib/index/syncer.js

@@ -324,9 +324,13 @@ function initialSync(options) {
                         yield currentFlush;
                         (0, logger_1.debug)("index", `flush done: ${Date.now() - flushStart}ms`);
                         flushCount++;
-                        // Periodically compact during sync to prevent fragment accumulation
+                        // Periodically compact during sync to prevent fragment accumulation.
+                        // Uses the shared threshold deliberately: a hardcoded 30 here meant
+                        // the full-index path kept compacting at the old rate long after the
+                        // threshold was raised to 400, re-triggering the whole-table rewrite
+                        // that raising it exists to avoid.
                         if (flushCount % 10 === 0) {
-                            yield vectorDb.compactIfNeeded(30);
+                            yield vectorDb.compactIfNeeded(config_1.FRAGMENT_COMPACT_THRESHOLD);
                         }
                     }
                     catch (err) {

package/dist/lib/utils/file-utils.js

@@ -187,22 +187,115 @@ function isIndexableFile(filePath, size) {
 function isIndexablePath(filePath) {
     return isIndexableFile(filePath);
 }
-// Well-known machine-generated markers emitted at the top of codegen output.
-// Tools (protoc, graphql-codegen, Apollo, Relay, Go `stringer`, OpenAPI, etc.)
-// stamp one of these in a banner comment. Matched against only the first few KB
-// because the banner is always at the very top — this keeps a hand-written
-// "do not edit this section" note deep in a real file from being misclassified.
-const GENERATED_MARKER = /@generated\b|\bDO NOT EDIT\b|Code generated by |auto-?generated by |This (?:file|code) (?:is|was) (?:\w+[ -])?generated|machine generated|@autogenerated\b/i;
+// Markers that only ever appear in a codegen banner. Any one of these, on its
+// own, is enough. Tools (protoc, graphql-codegen, Apollo, Relay, Go `stringer`,
+// OpenAPI) stamp one of these; doc generators fence their output with the
+// `generated-body` / `AUTO-GENERATED-CONTENT` style directives.
+const STRONG_GENERATED_MARKER = /@generated\b|@autogenerated\b|\bCode generated by |\bauto-?generated by |\bThis (?:file|code) (?:is|was) (?:\w+[ -])?generated|\bmachine[ -]generated\b|\bauto-?generated[-_ ]?content\b|\bgenerated[-_](?:body|content|section|block)\b|\bBEGIN GENERATED\b/i;
+// "DO NOT EDIT" is not proof of codegen on its own — humans write it on prose
+// they want left alone (archived vendor snapshots, verbatim captures). Real
+// codegen banners pair it with a generation word, so require that pairing.
+const DO_NOT_EDIT_MARKER = /\bDO NOT EDIT\b/i;
+const GENERATION_CONTEXT = /generat|codegen|autogen/i;
 const GENERATED_SNIFF_BYTES = 2048;
+const DOC_LIKE_EXTENSIONS = /\.(?:md|mdx|markdown|txt|rst|adoc|asciidoc)$/i;
+/**
+ * Extract the leading banner: the run of shebang, frontmatter, blank, and
+ * comment lines at the very top, stopping at the first line of actual code.
+ *
+ * Scanning raw head bytes instead was the bug this replaces. Two ways it
+ * misfired, both silently erasing hand-written files from the index:
+ *
+ *  - A *generator* was classified as its own output, because the banner it
+ *    emits appears as a string literal in its body
+ *    (`lines.push('# Auto-generated by ...')`).
+ *  - A comment quoting another file's rules ("its README says do not edit")
+ *    counted as a banner.
+ *
+ * A real codegen banner precedes all code, so anchoring here removes both
+ * without weakening detection.
+ *
+ * In doc-like files `#` starts a heading, not a comment, so only frontmatter
+ * and HTML comments count — otherwise the whole document body would be read as
+ * banner.
+ */
+function leadingBanner(head, filePath) {
+    var _a, _b;
+    const docLike = filePath ? DOC_LIKE_EXTENSIONS.test(filePath) : false;
+    const lines = head.split("\n");
+    const banner = [];
+    let i = 0;
+    if (!docLike && ((_a = lines[0]) === null || _a === void 0 ? void 0 : _a.startsWith("#!"))) {
+        banner.push(lines[0]);
+        i = 1;
+    }
+    // YAML frontmatter is metadata, not code — a generated-doc marker often sits
+    // directly after it.
+    if (((_b = lines[i]) === null || _b === void 0 ? void 0 : _b.trim()) === "---") {
+        banner.push(lines[i]);
+        i++;
+        while (i < lines.length && lines[i].trim() !== "---")
+            banner.push(lines[i++]);
+        if (i < lines.length)
+            banner.push(lines[i++]);
+    }
+    let blockEnd;
+    for (; i < lines.length; i++) {
+        const line = lines[i];
+        const trimmed = line.trim();
+        if (blockEnd) {
+            banner.push(line);
+            if (trimmed.includes(blockEnd))
+                blockEnd = undefined;
+            continue;
+        }
+        if (trimmed === "") {
+            banner.push(line);
+            continue;
+        }
+        if (trimmed.startsWith("<!--")) {
+            banner.push(line);
+            if (!trimmed.includes("-->"))
+                blockEnd = "-->";
+            continue;
+        }
+        if (!docLike) {
+            if (trimmed.startsWith("/*")) {
+                banner.push(line);
+                if (!trimmed.includes("*/"))
+                    blockEnd = "*/";
+                continue;
+            }
+            if (trimmed.startsWith("//") ||
+                trimmed.startsWith("#") ||
+                trimmed.startsWith("--") ||
+                trimmed.startsWith(";") ||
+                trimmed.startsWith("*")) {
+                banner.push(line);
+                continue;
+            }
+        }
+        break; // first real line of content — the banner is over
+    }
+    return banner.join("\n");
+}
 /**
  * Detect machine-generated source by its banner. Generated files flood both the
  * vector index and the symbol graph (codegen types rank as god nodes), so they
  * are skipped at index time. Path globs in `ignore-patterns.ts` catch the
  * obvious filenames; this catches the rest by content.
+ *
+ * `filePath` is optional but worth passing: without it, doc-like files are
+ * scanned with code comment rules, so a markdown `# Heading` reads as a comment.
  */
-function isGeneratedContent(buffer) {
+function isGeneratedContent(buffer, filePath) {
     const head = buffer.subarray(0, GENERATED_SNIFF_BYTES).toString("utf-8");
-    return GENERATED_MARKER.test(head);
+    const banner = leadingBanner(head, filePath);
+    if (!banner.trim())
+        return false;
+    if (STRONG_GENERATED_MARKER.test(banner))
+        return true;
+    return DO_NOT_EDIT_MARKER.test(banner) && GENERATION_CONTEXT.test(banner);
 }
 function formatDenseSnippet(text, maxLength = 1500) {
     const clean = text !== null && text !== void 0 ? text : "";

package/dist/lib/workers/orchestrator.js

@@ -256,7 +256,7 @@ class WorkerOrchestrator {
                 (0, logger_1.debug)("orch", `skip ${input.path} (empty or binary)`);
                 return { vectors: [], hash, mtimeMs, size, shouldDelete: true };
             }
-            if ((0, file_utils_1.isGeneratedContent)(buffer)) {
+            if ((0, file_utils_1.isGeneratedContent)(buffer, absolutePath)) {
                 (0, logger_1.debug)("orch", `skip ${input.path} (machine-generated header)`);
                 return { vectors: [], hash, mtimeMs, size, shouldDelete: true };
             }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "grepmax",
-  "version": "0.26.14",
+  "version": "0.26.15",
   "author": "Robert Owens <78518764+reowens@users.noreply.github.com>",
   "homepage": "https://github.com/reowens/grepmax",
   "bugs": {

package/plugins/grepmax/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "grepmax",
-  "version": "0.26.14",
+  "version": "0.26.15",
   "description": "Semantic code search for Claude Code. Automatically indexes your project and provides intelligent search capabilities.",
   "author": {
     "name": "Robert Owens",