@sdsrs/code-graph 0.69.0 → 0.71.0

package/claude-plugin/.claude-plugin/plugin.json

@@ -4,7 +4,7 @@
   "author": {
     "name": "sdsrs"
   },
-  "version": "0.69.0",
+  "version": "0.71.0",
   "keywords": [
     "code-graph",
     "ast",

package/claude-plugin/scripts/pre-grep-guide.js

@@ -48,7 +48,15 @@ const { runGrepAnswer, runShowAnswer, sanitizeSearchPath } = require('./cg-answe
 // v0.48: also match bare `KEY=VALUE grep` prefixes (no `env` verb) — the shape
 // the deny message itself teaches (`CODE_GRAPH_NO_BLOCK_GREP=1 grep …`). With
 // the old `env`-only form those commands failed gate 1 and were invisible.
-const GREP_HEAD = /^\s*(?:env\s+)?(?:[A-Za-z_][A-Za-z0-9_]*=\S*\s+)*(grep|rg|ag)\b/;
+// v0.71: `git grep` shares the verb set — its head is `git`, so it leaked past
+// the matcher until folded in here. cg grep is a SUPERSET (covers tracked AND
+// gitignored files), so routing `git grep` to it is sound. GREP_VERB is the
+// single source of truth for every parse site that recognizes the search verb.
+const GREP_VERB = 'git\\s+grep|grep|rg|ag';
+const GREP_HEAD = new RegExp(`^\\s*(?:env\\s+)?(?:[A-Za-z_][A-Za-z0-9_]*=\\S*\\s+)*(${GREP_VERB})\\b`);
+// Verb + prefix strip (kept in sync with GREP_HEAD via GREP_VERB; non-capturing).
+// Shared by extractPatterns and countNamedPaths so the verb is removed identically.
+const VERB_STRIP = new RegExp(`^\\s*(?:env\\s+)?(?:[A-Za-z_][A-Za-z0-9_]*=\\S*\\s+)*(?:${GREP_VERB})\\s+`);
 // Source-tree prefix list. Expanded v0.27+ Phase C: original `src/tests/lib/...`
 // missed real-world backend conventions where the prefix list term is preceded
 // by something else (`backend/app/...` — `app/` doesn't match because `/` isn't
@@ -63,7 +71,7 @@ const SRC_PREFIXES =
 const SRC_PATH = new RegExp(`(?:^|\\s|["'])(${SRC_PREFIXES})/`);
 // Anchored variant for whole-token matching in extractSearchPath.
 const SRC_PATH_TOKEN = new RegExp(`^(?:\\./)?(${SRC_PREFIXES})/`);
-const PIPE_INTO_GREP = /\|\s*(?:grep|rg|ag)\b/;
+const PIPE_INTO_GREP = new RegExp(`\\|\\s*(?:${GREP_VERB})\\b`);
 const CG_INVOKED = /\bcode-graph-mcp\b/;
 // File argument(s) that end in a config/lockfile/data extension. If, after removing
 // ALL of them, no source-tree path remains, the grep is searching config/data not code.
@@ -80,12 +88,41 @@ const CONFIG_TARGET_ONLY = new RegExp(`(?:^|\\s)[^\\s|<>]*\\.(?:${NON_SOURCE_EXT
 // data-file tokens both match; global so every one is peeled before the SRC_PATH re-check.
 const CONFIG_TARGET_STRIP = new RegExp(`(?:^|\\s)[^\\s|<>]*\\.(?:${NON_SOURCE_EXTS})(?=\\s|$)`, 'gi');
 
+// v0.71 — `git grep --cached`/`--staged` searches the STAGED index, and a treeish
+// ref (`git grep "X" HEAD~3 -- src/`, `git grep "X" main -- src/`) searches another
+// commit/branch — a scope the working-tree inline answer (`code-graph-mcp grep`)
+// CANNOT honor. Folding them would substitute current-tree hits for a different
+// revision with no signal. These are NOT the working-tree source searches this hook
+// folds, so it stays out entirely (no hint, no deny) and the real git grep runs.
+// (`--no-index` is working-tree scope → cg covers it → NOT excluded; plain grep/rg/ag
+// have no revision concept.) A bare treeish without `--` (`git grep X main src/`) is
+// genuinely ambiguous with a pathspec → left as the residual minority.
+const GIT_GREP_HEAD = /^\s*(?:env\s+)?(?:[A-Za-z_][A-Za-z0-9_]*=\S*\s+)*git\s+grep\b/;
+const GIT_GREP_STAGED = /(?:^|\s)--(?:cached|staged)(?:\s|$)/;
+
+function isRevisionScopedGitGrep(cmd) {
+  if (typeof cmd !== 'string' || !GIT_GREP_HEAD.test(cmd)) return false;
+  if (GIT_GREP_STAGED.test(cmd)) return true;
+  // treeish before the `--` pathspec separator: git grep [flags] PATTERN <ref>... -- <path>
+  const sep = cmd.indexOf(' -- ');
+  if (sep === -1) return false;
+  const afterVerb = cmd.slice(0, sep).replace(GIT_GREP_HEAD, '').trimStart();
+  let seenPattern = false;
+  for (const tok of afterVerb.split(/\s+/)) {
+    if (!tok || tok.startsWith('-')) continue;          // a flag
+    if (!seenPattern) { seenPattern = true; continue; } // the search pattern
+    return true;                                        // a 2nd non-flag token before `--` = treeish
+  }
+  return false;
+}
+
 function shouldHint(cmd) {
   if (!cmd || typeof cmd !== 'string') return false;
   if (cmd.length > 1000) return false;             // sanity — oversize commands are noise
   if (CG_INVOKED.test(cmd)) return false;          // already using cg
   if (PIPE_INTO_GREP.test(cmd)) return false;      // `cargo test | grep FAILED` is output filter
   if (!GREP_HEAD.test(cmd)) return false;          // not a search command
+  if (isRevisionScopedGitGrep(cmd)) return false;  // v0.71 — git grep --cached/treeish: scope cg can't honor
   if (!SRC_PATH.test(cmd)) return false;           // not against indexed source tree
   // If a config file appears AND no source path remains after stripping it, skip.
   if (CONFIG_TARGET_ONLY.test(cmd)) {
@@ -128,7 +165,7 @@ const MARKER_ONLY =
 function extractPatterns(cmd) {
   if (!cmd || typeof cmd !== 'string') return [];
   // Strip leading verb + env/assignment prefix (kept in sync with GREP_HEAD)
-  const stripped = cmd.replace(/^\s*(?:env\s+)?(?:[A-Za-z_][A-Za-z0-9_]*=\S*\s+)*(?:grep|rg|ag)\s+/, '');
+  const stripped = cmd.replace(VERB_STRIP, '');
   // Collect every quoted argument — first one is the pattern in standard grep
   // usage; subsequent ones (e.g. `-e "second"`) are also patterns or filter
   // expressions and worth screening too.
@@ -166,6 +203,12 @@ function classifyBlock(cmd) {
     if (symbols.length === 0) return null;        // context read without named decls
     return { mode: 'show', symbols: symbols.slice(0, 3) };
   }
+  // v0.70 — only DENY when the inline grep answer can cover the SAME scope. It scopes to one
+  // path (extractSearchPath = first src-prefixed token), so a grep naming ≥2 file paths gets a
+  // first-path-only answer (the rest silently dropped) — an incomplete substitute that
+  // rationally teaches CODE_GRAPH_NO_BLOCK_GREP bypass. Downgrade to hint: the model's complete
+  // grep runs and the hint still nudges. (show mode above is symbol-scoped, not path → unaffected.)
+  if (countNamedPaths(cmd, patterns) >= 2) return null;
   return { mode: 'grep' };
 }
 
@@ -303,6 +346,37 @@ function extractSearchPath(cmd) {
   return undefined;
 }
 
+// v0.70 — count the explicit file/dir path arguments a grep names (excluding flags and
+// the quoted search pattern). The deny's inline answer scopes to ONE path
+// (extractSearchPath returns only the first source-prefixed token), so a grep naming ≥2
+// paths gets an answer covering only the first — an incomplete substitute that rationally
+// drives CODE_GRAPH_NO_BLOCK_GREP bypass (2026-06-23: the dominant observed bypass was a
+// multi-file named grep whose deny silently dropped the other files). classifyBlock uses
+// this to downgrade those denies to a hint (which still nudges) so the complete grep runs.
+function countNamedPaths(cmd, patterns) {
+  if (!cmd || typeof cmd !== 'string') return 0;
+  const pats = new Set(patterns || []);
+  // Only the grep's OWN path args count. Stop at the first top-level command separator so a
+  // path in a compound tail (`grep X src/a.py | sed … src/b.py`) is NOT mistaken for a second
+  // grep target — that would wrongly downgrade a complete single-file grep to a hint.
+  let seg = cmd.replace(VERB_STRIP, '');
+  let quote = null;
+  for (let i = 0; i < seg.length; i++) {
+    const c = seg[i];
+    if (quote) { if (c === quote) quote = null; continue; }
+    if (c === '"' || c === "'") { quote = c; continue; }
+    if (c === ';' || c === '|' || c === '&' || c === '>' || c === '<' || c === '\n') { seg = seg.slice(0, i); break; }
+  }
+  let n = 0;
+  for (const raw of seg.split(/\s+/)) {
+    const tok = raw.replace(/^["']|["']$/g, '');
+    if (!tok || tok.startsWith('-')) continue;     // a flag
+    if (pats.has(tok)) continue;                    // the search pattern, not a path
+    if (tok.includes('/') || /\.[A-Za-z0-9]{1,6}$/.test(tok)) n++;  // dir-sep or file extension
+  }
+  return n;
+}
+
 // v0.47.0 — the pattern that justified the block: first identifier-like one.
 function pickBlockPattern(cmd) {
   return extractPatterns(cmd).find(p => IDENTIFIER_LIKE.test(p));
@@ -413,7 +487,8 @@ function buildShowDenyReason(answer, unansweredTail) {
 function translateBreToRg(cmd, pattern) {
   if (typeof pattern !== 'string' || !pattern) return pattern;
   const verb = (cmd.match(GREP_HEAD) || [])[1];
-  if (verb !== 'grep') return pattern;
+  // git grep speaks BRE like plain grep; rg/ag are already extended-regex.
+  if (!verb || !/grep$/.test(verb)) return pattern;
   if (/(?:^|\s)-[a-zA-Z]*[EP][a-zA-Z]*(?:\s|=|\d|$)|--(?:extended-regexp|perl-regexp)\b/.test(cmd)) {
     return pattern;
   }
@@ -607,6 +682,8 @@ module.exports = {
   extractSedReadTargets, // v0.49 — sed-range reads feed the read-fanout state
   extractUnansweredTail, // v0.50 — compound-tail honesty in answered denies
   extractPatterns,    // v0.32.1 — exposed for tests
+  countNamedPaths,    // v0.70 — multi-path deny→hint downgrade
+  isRevisionScopedGitGrep, // v0.71 — git grep --cached/treeish exclusion
   extractSearchPath,  // v0.47.0 — deny-with-answer
   normalizeCommandPaths, // v0.47.1 — abs-path matcher fix
   resolveProjectRoot,    // v0.48 — subdir-cwd dark fix

package/claude-plugin/scripts/pre-grep-guide.test.js

@@ -5,6 +5,7 @@ const {
   shouldHint,
   shouldBlock,
   classifyBlock,
+  countNamedPaths,
   extractDeclSymbols,
   translateBreToRg,
   buildShowDenyReason,
@@ -62,6 +63,71 @@ test('shouldHint: env-prefixed grep on src/', () => {
   assert.equal(shouldHint('env LANG=C grep -rn "Foo" src/'), true);
 });
 
+// ── git grep coverage (v0.71): `git grep` is raw BRE search on the tracked
+//    source tree — same foldable intent as `grep`, but its command HEAD is
+//    `git`, so it leaked past GREP_HEAD until v0.71. cg grep is a superset
+//    (tracked AND gitignored), so folding `git grep` into it is sound. The verb
+//    set is shared across GREP_HEAD / VERB_STRIP / PIPE_INTO_GREP — these lock
+//    each parse site that touches the verb.
+
+test('git grep: shouldHint fires on `git grep` against src/', () => {
+  assert.equal(shouldHint('git grep -n "fts5_search" src/storage/'), true);
+});
+
+test('git grep: shouldHint fires with the `--` pathspec separator', () => {
+  assert.equal(shouldHint('git grep "FooBar" -- src/lib.rs'), true);
+});
+
+test('git grep: identifier search is a deny (block tier, same as grep)', () => {
+  assert.equal(shouldBlock('git grep "FooBar" src/'), true);
+});
+
+test('git grep: context flag + decl anchor → show mode', () => {
+  assert.deepEqual(
+    classifyBlock('git grep "fn handle_message" -A 5 src/'),
+    { mode: 'show', symbols: ['handle_message'] });
+});
+
+test('git grep: multi-file named search downgrades to hint (v0.70 parity)', () => {
+  // inline answer scopes to ONE path; ≥2 named files → hint so the full grep runs.
+  assert.equal(classifyBlock('git grep "FooBar" src/a.rs src/b.rs'), null);
+});
+
+test('git grep: BRE alternation is translated to rust-regex dialect', () => {
+  // git grep speaks BRE like plain grep → an escaped \| must unescape for cg grep.
+  assert.equal(translateBreToRg('git grep "a\\|b" src/', 'a\\|b'), 'a|b');
+});
+
+test('git grep: `| git grep` is an output-filter pipe (no fire)', () => {
+  assert.equal(shouldHint('grep -rn "Foo" src/ | git grep "Bar"'), false);
+});
+
+test('git grep: rebaseRelativePaths rebases the real subdir path, not the `grep` word', () => {
+  // shell sits in backend/; `app` is subdir-relative → rebased. `grep` is the
+  // git subcommand and is existence-gated so it never masquerades as a path.
+  const exists = (p) => p.endsWith('/root/backend/app');
+  const out = rebaseRelativePaths('git grep "Foo" app', 'backend', '/root', exists);
+  assert.match(out, /git grep "Foo" backend\/app/);
+});
+
+// v0.71 — git grep at a scope the working-tree cg answer can't honor (staged
+// index / another revision) must NOT deny: folding it would substitute
+// current-tree hits for a different revision. The hook stays out entirely.
+test('git grep: --cached (staged index) is not denied — cg cannot honor that scope', () => {
+  assert.equal(shouldHint('git grep --cached "FooBar" src/'), false);
+  assert.equal(shouldBlock('git grep --cached "FooBar" src/'), false);
+});
+
+test('git grep: a treeish ref before `--` (another revision) is not denied', () => {
+  assert.equal(shouldHint('git grep "FooBar" HEAD~3 -- src/'), false);
+  assert.equal(shouldBlock('git grep "cascade_failure" main -- src/'), false);
+});
+
+test('git grep: a bare `-- path` (no ref, working-tree scope) STILL denies', () => {
+  // guard: the revision-scope exclusion must not over-catch a plain pathspec sep.
+  assert.equal(shouldBlock('git grep "FooBar" -- src/lib.rs'), true);
+});
+
 // ── Should NOT fire: pipe-grep (output filter, not search) ──────────
 
 test('shouldHint: pipe-grep on cargo test output', () => {
@@ -372,6 +438,45 @@ test('shouldBlock: --exclude=tests → hint only (answer cannot honor exclusion)
   assert.equal(shouldBlock('grep -rn --exclude=tests "EmbeddingModel" src/'), false);
 });
 
+// ── B (v0.70): deny only when the inline answer covers the FULL scope ──
+// The deny scopes to ONE path (extractSearchPath = first src-prefixed token). A grep naming
+// ≥2 file paths would get a first-path-only answer (the rest silently dropped) — an incomplete
+// substitute that rationally teaches CODE_GRAPH_NO_BLOCK_GREP bypass. Downgrade those to HINT;
+// single file / directory greps (which the answer fully covers) still deny.
+
+test('B: ≥2 named files downgrade deny→hint (deny would drop all but the first)', () => {
+  const cmd = 'grep -n "CLAUDE_MEM_DIR" scripts/setup.sh hook-shared.mjs';
+  assert.equal(shouldHint(cmd), true);     // still nudges
+  assert.equal(shouldBlock(cmd), false);   // but does NOT deny (answer can't cover hook-shared.mjs)
+  assert.equal(classifyBlock(cmd), null);
+});
+
+test('B: two source files also downgrade (deny would cover only the first)', () => {
+  assert.equal(shouldBlock('grep -rn "set_hook" src/main.rs src/lib.rs'), false);
+  assert.equal(shouldHint('grep -rn "set_hook" src/main.rs src/lib.rs'), true);
+});
+
+test('B: single file still DENIES (inline answer fully covers it)', () => {
+  assert.deepEqual(classifyBlock('grep -n "handleMessage" src/server.mjs'), { mode: 'grep' });
+});
+
+test('B: single directory (recursive) still DENIES (cg grep covers the whole dir)', () => {
+  assert.equal(shouldBlock('grep -rn "EmbeddingModel" src/'), true);
+});
+
+test('B: --include on a single dir still DENIES (one path, fully scoped)', () => {
+  assert.equal(shouldBlock('grep -rn --include="*.rs" "EmbeddingModel" src/'), true);
+});
+
+test('countNamedPaths: counts paths, excludes flags and the quoted pattern', () => {
+  assert.equal(countNamedPaths('grep -n "Foo" src/a.rs src/b.rs', ['Foo']), 2);
+  assert.equal(countNamedPaths('grep -rn "Foo" src/', ['Foo']), 1);
+  // a path-shaped pattern is the pattern, not a second path token
+  assert.equal(countNamedPaths('grep "config.json" src/app.rs', ['config.json']), 1);
+  // a path in a compound tail (sed/pipe target) is NOT a 2nd grep target → stays 1 (deny)
+  assert.equal(countNamedPaths("grep -n \"Foo\" src/foo.rs | head; sed -n '1,5p' src/bar.rs", ['Foo']), 1);
+});
+
 test('shouldBlock: -L / -v inverted intents → hint only', () => {
   assert.equal(shouldBlock('grep -rL "EmbeddingModel" src/'), false);
   assert.equal(shouldBlock('grep -rnv "EmbeddingModel" src/'), false);
@@ -945,6 +1050,22 @@ test('e2e: denied grep with stub hits → deny JSON embeds the answer + records
   }
 });
 
+test('e2e: `git grep` identifier on src/ → deny with the embedded answer', () => {
+  const uniq = `GitHit${Date.now()}`;
+  const fixture = e2eFixture(
+    `process.stdout.write('src/foo.rs:9  fn ' + process.argv[3] + '()\\n');`);
+  const cmd = `git grep -n "${uniq}" src/`;
+  try {
+    const res = runHook(cmd, fixture);
+    assert.equal(res.status, 0);
+    const out = JSON.parse(res.stdout);
+    assert.equal(out.hookSpecificOutput.permissionDecision, 'deny');
+    assert.match(out.hookSpecificOutput.permissionDecisionReason, new RegExp(uniq));
+  } finally {
+    cleanupFixture(fixture, cmd);
+  }
+});
+
 test('e2e: denied grep records the denied pattern (fingerprint for verbatim re-grep detection)', () => {
   // The Rust funnel (aggregate_recommendations_jsonl) scores a follow-up search
   // carrying the SAME pattern as the armed answered deny as fall-through, not a

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@sdsrs/code-graph",
-  "version": "0.69.0",
+  "version": "0.71.0",
   "description": "MCP server that indexes codebases into an AST knowledge graph with semantic search, call graph traversal, and HTTP route tracing",
   "license": "MIT",
   "repository": {
@@ -35,10 +35,10 @@
     "node": ">=16"
   },
   "optionalDependencies": {
-    "@sdsrs/code-graph-linux-x64": "0.69.0",
-    "@sdsrs/code-graph-linux-arm64": "0.69.0",
-    "@sdsrs/code-graph-darwin-x64": "0.69.0",
-    "@sdsrs/code-graph-darwin-arm64": "0.69.0",
-    "@sdsrs/code-graph-win32-x64": "0.69.0"
+    "@sdsrs/code-graph-linux-x64": "0.71.0",
+    "@sdsrs/code-graph-linux-arm64": "0.71.0",
+    "@sdsrs/code-graph-darwin-x64": "0.71.0",
+    "@sdsrs/code-graph-darwin-arm64": "0.71.0",
+    "@sdsrs/code-graph-win32-x64": "0.71.0"
   }
 }