fix(increment): stop endless full graph rebuild from stale hash entries

build_ast_graph.py

@@ -3668,10 +3668,17 @@ def incremental_rebuild(
 
 
 def _init_hash_tracker(source_root: Path, ladybug_path: Path) -> int:
-    """Initialize hash tracker for all Java files. Returns number of files hashed."""
+    """Initialize hash tracker for all Java files. Returns number of files hashed.
+
+    Called right after a full graph rebuild (``write_ladybug``), so the store must
+    mirror exactly the files that were just indexed. We deliberately do NOT
+    ``load()`` the existing store: ``update`` re-hashes every current file anyway,
+    and preserving old entries would leave stale hashes for files that no longer
+    exist (deleted or now-ignored). Those ghosts would be re-detected as "removed"
+    on every subsequent ``increment``, sustaining an endless full-rebuild loop.
+    """
     index_dir = ladybug_path.parent
     tracker = FileHashTracker(index_dir)
-    tracker.load()
     ignore = LayeredIgnore(source_root)
     all_files: set[str] = set()
     source_root_resolved = source_root.resolve()
@@ -3742,7 +3749,7 @@ def _write_clients_producers_and_calls(conn: ladybug.Connection, tables: GraphTa
 
     # Write declares_client edges
     for row in tables.declares_client_rows:
-        source_file = member_by_id.get(row.symbol_id, MemberEntry(kind="", decl=None, parent_id="", parent_fqn="", file_path="", module="", microservice="")).file_path
+        source_file = member_by_id.get(row.symbol_id, MemberEntry(kind="", decl=None, parent_id="", parent_fqn="", file_path="", module="", microservice="", node_id="")).file_path
         conn.execute(_CREATE_DECLARES_CLIENT, {
             "sid": row.symbol_id,
             "cid": row.client_id,
@@ -3753,7 +3760,7 @@ def _write_clients_producers_and_calls(conn: ladybug.Connection, tables: GraphTa
 
     # Write declares_producer edges
     for row in tables.declares_producer_rows:
-        source_file = member_by_id.get(row.symbol_id, MemberEntry(kind="", decl=None, parent_id="", parent_fqn="", file_path="", module="", microservice="")).file_path
+        source_file = member_by_id.get(row.symbol_id, MemberEntry(kind="", decl=None, parent_id="", parent_fqn="", file_path="", module="", microservice="", node_id="")).file_path
         conn.execute(_CREATE_DECLARES_PRODUCER, {
             "sid": row.symbol_id,
             "pid": row.producer_id,

tests/test_incremental_graph.py

@@ -4,6 +4,8 @@
 """
 from __future__ import annotations
 
+import json
+import shutil
 from pathlib import Path
 
 import ladybug
@@ -979,3 +981,141 @@ def test_incremental_preserves_incoming_edges_to_dependent(self, tmp_path: Path)
         assert cb_after_count > 0, "out-of-scope C->B CALLS edge must be preserved"
 
         conn.close()
+
+
+class TestIncrementalRegressions:
+    """Regression tests for the ``increment`` always-fully-reprocesses loop.
+
+    Two bugs fed the loop:
+      1. ``_write_clients_producers_and_calls`` built a default ``MemberEntry``
+         missing the required ``node_id`` field. Because ``dict.get(k, default)``
+         evaluates ``default`` eagerly, the TypeError fired whenever ANY
+         ``declares_client`` / ``declares_producer`` row existed — crashing every
+         client-bearing incremental rebuild into a full-rebuild fallback.
+      2. ``_init_hash_tracker`` (run by every full reprocess AND by that fallback)
+         did ``load()`` + ``update()`` and never pruned hashes for files no longer
+         on disk, so ghost entries persisted and re-triggered the loop every run.
+    """
+
+    def test_init_hash_tracker_prunes_stale_entries(self, tmp_path: Path) -> None:
+        """A full rebuild drops hashes for files no longer on disk (ghost pruning).
+
+        Without pruning, a stale entry is re-detected as 'removed' on every
+        ``increment``, sustaining an endless full-rebuild loop.
+        """
+        from build_ast_graph import write_ladybug
+
+        source_root = tmp_path / "src"
+        source_root.mkdir()
+        (source_root / "A.java").write_text("package pkg; class A {}", encoding="utf-8")
+        index_dir = tmp_path / "index"
+        index_dir.mkdir()
+        ladybug_path = index_dir / "code_graph.lbug"
+
+        tables = GraphTables()
+        pass1_parse(source_root, tables, verbose=False)
+        write_ladybug(ladybug_path, tables, source_root=source_root, verbose=False)
+
+        # Inject a ghost hash for a file that does not exist on disk.
+        hash_file = index_dir / ".graph_hashes.json"
+        data = json.loads(hash_file.read_text(encoding="utf-8"))
+        data["ghost/Deleted.java"] = "0" * 64
+        hash_file.write_text(json.dumps(data), encoding="utf-8")
+
+        # A second full rebuild (what `reprocess --graph-only` does) re-runs
+        # _init_hash_tracker, which must drop the ghost.
+        tables2 = GraphTables()
+        pass1_parse(source_root, tables2, verbose=False)
+        write_ladybug(ladybug_path, tables2, source_root=source_root, verbose=False)
+
+        after = json.loads(hash_file.read_text(encoding="utf-8"))
+        assert "ghost/Deleted.java" not in after
+        assert "A.java" in after
+
+    def test_incremental_with_http_clients_does_not_fall_back(self, tmp_path: Path) -> None:
+        """A corpus with Feign/Kafka clients/producers rebuilds incrementally.
+
+        ``http_caller_smoke`` emits DECLARES_CLIENT / DECLARES_PRODUCER rows, so
+        the buggy eager ``MemberEntry`` default in
+        ``_write_clients_producers_and_calls`` crashed here before the fix
+        (forcing full_fallback). After the fix: mode is "incremental".
+        """
+        from _builders import build_ladybug_full_into
+        from build_ast_graph import incremental_rebuild
+
+        corpus = Path(__file__).parent / "fixtures" / "http_caller_smoke"
+        source_root = tmp_path / "src"
+        shutil.copytree(corpus, source_root)
+        index_dir = tmp_path / "index"
+        index_dir.mkdir()
+        ladybug_path = index_dir / "code_graph.lbug"
+
+        # Full build seeds .graph_hashes.json via write_ladybug -> _init_hash_tracker.
+        build_ladybug_full_into(source_root, ladybug_path)
+
+        # Mutate one file unrelated to the clients/producers.
+        target = source_root / "src" / "main" / "java" / "smoke" / "http" / "TopicNames.java"
+        target.write_text(target.read_text(encoding="utf-8") + "\n// edit\n", encoding="utf-8")
+
+        result = incremental_rebuild(source_root, ladybug_path, verbose=False)
+        assert result.mode == "incremental", (
+            f"expected incremental, got {result.mode!r} (the node_id crash in "
+            "_write_clients_producers_and_calls forces a full fallback)"
+        )
+        assert result.files_changed == 1
+
+    def test_reprocess_graph_only_then_increment_is_noop(self, tmp_path: Path) -> None:
+        """The reported scenario at the builder level: a full graph rebuild (what
+        ``reprocess --graph-only`` does) followed by ``increment`` with no source
+        changes must be a no-op, not a second full rebuild."""
+        from _builders import build_ladybug_full_into
+        from build_ast_graph import incremental_rebuild
+
+        corpus = Path(__file__).parent / "fixtures" / "http_caller_smoke"
+        source_root = tmp_path / "src"
+        shutil.copytree(corpus, source_root)
+        index_dir = tmp_path / "index"
+        index_dir.mkdir()
+        ladybug_path = index_dir / "code_graph.lbug"
+
+        # Simulate `reprocess --graph-only`: full rebuild seeds the hash store.
+        build_ladybug_full_into(source_root, ladybug_path)
+
+        # `increment` with no source changes.
+        result = incremental_rebuild(source_root, ladybug_path, verbose=False)
+        assert result.mode == "incremental"
+        assert (result.files_changed, result.files_added, result.files_removed) == (0, 0, 0)
+
+    def test_incremental_ghost_entry_then_next_run_is_noop(self, tmp_path: Path) -> None:
+        """A ghost hash entry is detected as 'removed' once, processed by the
+        scoped path (which prunes it), so the following run is a clean no-op.
+
+        Guards both fixes together: the node_id fix lets the scoped path
+        complete, and that path prunes the ghost (lines that delete `removed`
+        hashes). Before the fixes this fell back to full and preserved the ghost.
+        """
+        from _builders import build_ladybug_full_into
+        from build_ast_graph import incremental_rebuild
+
+        corpus = Path(__file__).parent / "fixtures" / "http_caller_smoke"
+        source_root = tmp_path / "src"
+        shutil.copytree(corpus, source_root)
+        index_dir = tmp_path / "index"
+        index_dir.mkdir()
+        ladybug_path = index_dir / "code_graph.lbug"
+        build_ladybug_full_into(source_root, ladybug_path)
+
+        # Inject a ghost (no source change).
+        hash_file = index_dir / ".graph_hashes.json"
+        data = json.loads(hash_file.read_text(encoding="utf-8"))
+        data["ghost/Gone.java"] = "0" * 64
+        hash_file.write_text(json.dumps(data), encoding="utf-8")
+
+        first = incremental_rebuild(source_root, ladybug_path, verbose=False)
+        assert first.mode == "incremental", f"expected incremental, got {first.mode!r}"
+        assert first.files_removed == 1
+
+        # The ghost must be gone, so the next run detects nothing.
+        second = incremental_rebuild(source_root, ladybug_path, verbose=False)
+        assert second.mode == "incremental"
+        assert (second.files_changed, second.files_added, second.files_removed) == (0, 0, 0)

tests/test_java_codebase_rag_cli.py

@@ -544,6 +544,61 @@ def test_increment_first_run_falls_back_to_full(
 
 
 
+def test_reprocess_graph_only_then_increment_graph_is_noop(
+    corpus_root: Path, tmp_path: Path, monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """The reported scenario, exercised through the real CLI wiring.
+
+    ``reprocess --graph-only`` rebuilds the graph and seeds ``.graph_hashes.json``
+    (via ``write_ladybug`` -> ``_init_hash_tracker``); the next ``increment``'s
+    graph stage must be a no-op, NOT a second full rebuild.
+
+    cocoindex is stubbed so ``increment`` runs only its real graph stage (no
+    embedding model needed). ``reprocess --graph-only`` needs no cocoindex
+    regardless, so this test runs in the normal (non-heavy) suite.
+    """
+    idx = tmp_path / "idx_reprocess_then_increment"
+    idx.mkdir()
+    monkeypatch.setenv("JAVA_CODEBASE_RAG_INDEX_DIR", str(idx))
+    monkeypatch.setenv("JAVA_CODEBASE_RAG_SOURCE_ROOT", str(corpus_root))
+
+    rc = cli_mod.main(
+        ["reprocess", "--graph-only", "--source-root", str(corpus_root), "--index-dir", str(idx), "--quiet"],
+    )
+    assert rc == 0, "reprocess --graph-only must succeed"
+    # reprocess --graph-only must seed the hash store.
+    hash_file = idx / ".graph_hashes.json"
+    assert hash_file.exists(), "hash store not seeded by reprocess --graph-only"
+
+    # Inject a ghost entry for a file that does not exist — the exact "N removed
+    # files every run" symptom. On bank-chat (which has Feign/Kafka clients) the
+    # scoped path this triggers reaches _write_clients_producers_and_calls, so a
+    # missing-field MemberEntry default here used to crash into a full fallback.
+    data = json.loads(hash_file.read_text(encoding="utf-8"))
+    data["ghost/DoesNotExist.java"] = "0" * 64
+    hash_file.write_text(json.dumps(data), encoding="utf-8")
+
+    # Stub cocoindex so increment exercises ONLY its graph stage.
+    def _noop_coco(env, *, full_reprocess, quiet, verbose=True, lance_project_root=None):
+        return subprocess.CompletedProcess(args=[], returncode=0, stdout="", stderr="")
+
+    monkeypatch.setattr(cli_mod, "run_cocoindex_update", _noop_coco)
+
+    buf_out = io.StringIO()
+    buf_err = io.StringIO()
+    with contextlib.redirect_stdout(buf_out), contextlib.redirect_stderr(buf_err):
+        rc2 = cli_mod.main(
+            ["increment", "--source-root", str(corpus_root), "--index-dir", str(idx), "--quiet"],
+        )
+    assert rc2 == 0
+    # The graph stage must NOT have fallen back to a full rebuild.
+    assert "fell back to full graph rebuild" not in buf_err.getvalue()
+    assert "increment completed (Lance + graph updated)" in buf_out.getvalue()
+    # The ghost must be pruned, so the next increment is clean.
+    after = json.loads(hash_file.read_text(encoding="utf-8"))
+    assert "ghost/DoesNotExist.java" not in after
+
+
 @pytest.mark.skipif(not _cocoindex_available(), reason="cocoindex not installed in venv")
 def test_increment_updates_lance_after_touch_java_file(corpus_root: Path, tmp_path: Path) -> None:
     import lancedb  # noqa: PLC0415