fix(cli): erase removes graph/cocoindex.db/.graph_hashes.json by type (#346) (#348)

* fix(cli): erase removes graph/cocoindex.db/.graph_hashes.json by type (#346)

`erase` reported success but left code_graph.lbug on disk because its
deletion was type-blind: shutil.rmtree silently no-ops on a regular file
(code_graph.lbug) and Path.unlink raises IsADirectoryError on a directory
(cocoindex.db), both swallowed; .graph_hashes.json was never targeted.
The next init then refused (exit 2), deadlooping the documented
`erase --yes` -> `init` clean-slate workflow.

Replace the type-blind deletes with a _rm_any helper that dispatches on
type (file/dir/symlink — a symlinked dir is unlinked, never recursed into,
so the target is not followed), so both the file-backed and dir-backed
LadybugDB layouts are handled. erase now also removes .graph_hashes.json
and lists it in the "Will delete:" preview. Deletion failures are warned
to stderr instead of swallowed, so erase no longer reports success while
leaving an artifact behind (the same silent-failure class as #346).

`reprocess` is unaffected: its full rebuild opens the existing .lbug and
_drop_all()s every node + edge table in place, and _init_hash_tracker
resets .graph_hashes.json — it never relies on the broken deletion.

Tests: add an always-on regression that creates a real lbug-file /
cocoindex.db-dir / hash-store layout and asserts erase removes all three;
convert the false-green test_init_after_erase_succeeds into a real
build -> erase -> re-init lifecycle check.

Co-Authored-By: Claude <noreply@anthropic.com>

* bump version to 0.6.6

Co-Authored-By: Claude <noreply@anthropic.com>

---------

Co-authored-by: Claude <noreply@anthropic.com>

Author: HumanBean17

docs/JAVA-CODEBASE-RAG-CLI.md

@@ -240,7 +240,7 @@ java-codebase-rag reprocess --source-root /path/to/java/repo --index-dir /path/t
 
 ### `erase`
 
-Deletes cocoindex state, the LadybugDB directory, and Lance tables under the index dir. Requires **`--yes`** or interactive confirmation on a TTY. Non-TTY without `--yes` exits **2**.
+Deletes cocoindex state, the LadybugDB graph (`code_graph.lbug`), the graph builder's content-hash store (`.graph_hashes.json`), and Lance tables under the index dir. Requires **`--yes`** or interactive confirmation on a TTY. Non-TTY without `--yes` exits **2**.
 
 ```bash
 java-codebase-rag erase --source-root /path/to/java/repo --index-dir /path/to/.java-codebase-rag --yes

java_codebase_rag/cli.py

@@ -580,11 +580,33 @@ def _cmd_update(args: argparse.Namespace) -> int:
     )
 
 
+def _rm_any(path: Path) -> None:
+    """Remove ``path`` whether it is a regular file, directory, or symlink.
+
+    ``code_graph.lbug`` is a single regular file in this repo, but kuzu may lay
+    the graph out as a directory; ``cocoindex.db`` is always a directory.
+    ``shutil.rmtree`` is a silent no-op on a regular file and ``Path.unlink``
+    raises ``IsADirectoryError`` on a directory, so a type-blind delete left
+    index artifacts on disk (issue #346). A symlinked directory is unlinked, not
+    recursed into, so the link target is never followed. Failures are warned to
+    stderr rather than swallowed, so erase does not report success while leaving
+    an artifact behind (the exact failure mode issue #346 reported).
+    """
+    try:
+        if path.is_dir() and not path.is_symlink():
+            shutil.rmtree(path)
+        elif path.exists() or path.is_symlink():
+            path.unlink(missing_ok=True)
+    except OSError as exc:
+        print(f"warning: failed to remove {path}: {exc}", file=sys.stderr)
+
+
 def _cmd_erase(args: argparse.Namespace) -> int:
     cfg = _resolved_from_ns(args)
     _startup_hints(cfg)
     cfg.apply_to_os_environ()
-    to_describe: list[Path] = [cfg.ladybug_path, cfg.cocoindex_db]
+    graph_hashes_path = cfg.ladybug_path.parent / ".graph_hashes.json"
+    to_describe: list[Path] = [cfg.ladybug_path, cfg.cocoindex_db, graph_hashes_path]
     if cfg.index_dir.is_dir():
         try:
             import lancedb
@@ -621,13 +643,15 @@ def work(progress: "PipelineProgress | None") -> int:
             )
         elif drop.returncode != 0:
             print(clip(drop.stderr, 4000), file=sys.stderr)
-        if cfg.ladybug_path.exists():
-            shutil.rmtree(cfg.ladybug_path, ignore_errors=True)
-        if cfg.cocoindex_db.exists():
-            try:
-                cfg.cocoindex_db.unlink()
-            except OSError:
-                pass
+        # Remove the LadybugDB graph, the cocoindex state store, and the graph
+        # builder's content-hash store. Each is removed by type (see _rm_any):
+        # code_graph.lbug is a file here but may be a dir under kuzu, while
+        # cocoindex.db is a directory — a type-blind delete silently no-oped on
+        # one or the other, and .graph_hashes.json was never targeted at all
+        # (issue #346).
+        _rm_any(cfg.ladybug_path)
+        _rm_any(cfg.cocoindex_db)
+        _rm_any(graph_hashes_path)
         if cfg.index_dir.is_dir():
             try:
                 import lancedb

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "java-codebase-rag"
-version = "0.6.5"
+version = "0.6.6"
 description = "MCP server for semantic + structural search over Java codebases"
 readme = "README.md"
 requires-python = ">=3.11"

tests/test_java_codebase_rag_cli.py

@@ -109,6 +109,35 @@ def test_cli_erase_succeeds_with_yes_flag(tmp_path: Path) -> None:
     assert proc.returncode == 0, proc.stderr + proc.stdout
 
 
+def test_erase_removes_graph_file_cocoindex_dir_and_hash_store(tmp_path: Path) -> None:
+    """erase must delete code_graph.lbug (file), cocoindex.db (dir), .graph_hashes.json.
+
+    Regression for issue #346: a type-blind delete left both on disk.
+    shutil.rmtree is a silent no-op on a regular file (code_graph.lbug), and
+    Path.unlink raises IsADirectoryError on cocoindex.db (a directory) — both
+    swallowed — and .graph_hashes.json was never targeted. The follow-up init
+    then refused because code_graph.lbug survived.
+    """
+    idx = tmp_path / "erase_artifacts"
+    idx.mkdir()
+    # Real on-disk layout: graph is a single FILE, cocoindex state is a DIR.
+    (idx / "code_graph.lbug").write_bytes(b"fake-kuzu-db")
+    (idx / "cocoindex.db").mkdir()
+    (idx / "cocoindex.db" / "state.json").write_text("{}", encoding="utf-8")
+    (idx / ".graph_hashes.json").write_text("{}", encoding="utf-8")
+    env = os.environ.copy()
+    env["JAVA_CODEBASE_RAG_INDEX_DIR"] = str(idx)
+    env["JAVA_CODEBASE_RAG_SOURCE_ROOT"] = str(tmp_path)
+    proc = _run_cli(
+        ["erase", "--source-root", str(tmp_path), "--index-dir", str(idx), "--yes"],
+        env=env,
+    )
+    assert proc.returncode == 0, proc.stderr + proc.stdout
+    assert not (idx / "code_graph.lbug").exists(), "erase left code_graph.lbug on disk"
+    assert not (idx / "cocoindex.db").exists(), "erase left cocoindex.db/ on disk"
+    assert not (idx / ".graph_hashes.json").exists(), "erase left .graph_hashes.json on disk"
+
+
 def test_embedding_model_precedence_cli_over_env_over_yaml_over_default(
     tmp_path: Path, monkeypatch: pytest.MonkeyPatch
 ) -> None:
@@ -377,21 +406,34 @@ def test_legacy_env_var_set_emits_stderr_hint(monkeypatch: pytest.MonkeyPatch, t
 
 @pytest.mark.skipif(not _cocoindex_available(), reason="cocoindex not installed in venv")
 def test_init_after_erase_succeeds(corpus_root: Path, tmp_path: Path) -> None:
+    """Build a real index, erase it, then init again from a clean slate.
+
+    Regression for issue #346: the previous body erased an *empty* index dir and
+    then inited, so it never exercised "erase a real graph -> re-init" and stayed
+    green while erase silently left code_graph.lbug on disk.
+    """
     idx = tmp_path / "lifecycle_idx"
     idx.mkdir(parents=True)
     env = os.environ.copy()
     env["JAVA_CODEBASE_RAG_INDEX_DIR"] = str(idx)
     env["JAVA_CODEBASE_RAG_SOURCE_ROOT"] = str(corpus_root.resolve())
+    init1 = _run_cli(
+        ["init", "--source-root", str(corpus_root), "--index-dir", str(idx), "--quiet"],
+        env=env,
+    )
+    assert init1.returncode == 0, init1.stdout + init1.stderr
+    assert (idx / "code_graph.lbug").exists(), "init did not build code_graph.lbug"
     e1 = _run_cli(
         ["erase", "--source-root", str(corpus_root), "--index-dir", str(idx), "--yes"],
         env=env,
     )
     assert e1.returncode == 0, e1.stderr
-    init = _run_cli(
+    assert not (idx / "code_graph.lbug").exists(), "erase left code_graph.lbug on disk"
+    init2 = _run_cli(
         ["init", "--source-root", str(corpus_root), "--index-dir", str(idx), "--quiet"],
         env=env,
     )
-    assert init.returncode == 0, init.stdout + init.stderr
+    assert init2.returncode == 0, init2.stdout + init2.stderr
 
 
 @pytest.mark.skipif(not _cocoindex_available(), reason="cocoindex not installed in venv")