ltcai 0.2.1 → 0.2.2

package/README.md

@@ -154,14 +154,17 @@ Based on public product behavior as of 2026-05.
 |-------|----------|------|---------|
 | Qwen3-VL 4B | Multimodal / low spec | ~2.7 GB | 8 GB |
 | Qwen3-VL 8B | Multimodal / balanced | ~4.8 GB | 16 GB |
+| GPT-OSS 20B | Reasoning / open-weight | ~12.1 GB | 32 GB |
 | Gemma 4 26B | Multimodal / large | ~15.6 GB | 32 GB |
+| Gemma 4 31B | Multimodal / latest Gemma 4 | ~18.4 GB | 48 GB |
 | Qwen3-VL 30B A3B | Multimodal / top | ~18 GB | 48 GB |
+| GPT-OSS 120B | Reasoning / top open-weight | ~62.3 GB | 128 GB |
 | Phi 4 Mini | Coding (fast) | ~2.2 GB | 8 GB |
 | Llama 3.1 8B | General | ~4.7 GB | 8 GB |
 | Mistral 7B v0.3 | General / Apache | ~4.1 GB | 8 GB |
 
 **Cross-platform (Ollama / LM Studio / vLLM / llama.cpp):**
-Same models via Ollama pull, LM Studio download, or vLLM serve.
+Same models via Ollama pull, LM Studio download, vLLM serve, or llama.cpp GGUF where available.
 
 **Cloud (any platform):**
 OpenAI GPT-5.5 · Claude Opus 4.7 / Sonnet 4.6 / Haiku 4.5 via OpenRouter · Groq · Together · xAI · any OpenAI-compatible endpoint
@@ -360,7 +363,7 @@ Full reference: [docs/mcp-tools.md](docs/mcp-tools.md)
 | VS Code Marketplace | [marketplace.visualstudio.com](https://marketplace.visualstudio.com/items?itemName=parktaesoo.ltcai) |
 | Open VSX | [open-vsx.org](https://open-vsx.org/extension/parktaesoo/ltcai) |
 
-Current version: **0.2.1** — [Changelog](docs/CHANGELOG.md)
+Current version: **0.2.2** — [Changelog](docs/CHANGELOG.md)
 
 ---
 
@@ -416,8 +419,11 @@ LTCAI --tunnel                       # + Cloudflare 공개 URL 자동 발급
 |------|------|------|----------|
 | Qwen3-VL 4B | 멀티모달 / 저사양 | ~2.7GB | 8GB |
 | Qwen3-VL 8B | 멀티모달 / 균형 추천 | ~4.8GB | 16GB |
+| GPT-OSS 20B | 추론 / 오픈가중치 | ~12.1GB | 32GB |
 | Gemma 4 26B | 멀티모달 / 대형 | ~15.6GB | 32GB |
+| Gemma 4 31B | 멀티모달 / 최신 Gemma 4 | ~18.4GB | 48GB |
 | Qwen3-VL 30B A3B | 멀티모달 / 최고급 | ~18GB | 48GB |
+| GPT-OSS 120B | 추론 / 최고급 오픈가중치 | ~62.3GB | 128GB |
 
 자세한 내용: [docs/CHANGELOG.md](docs/CHANGELOG.md) · [보안](SECURITY.md) · [기여](CONTRIBUTING.md)
 

package/auto_setup.py

@@ -443,8 +443,16 @@ class Recommendation:
 _MODEL_CATALOG: List[Dict[str, Any]] = [
     # (min_ram_mb, min_vram_mb, model_id, quant, runtime_preference)
     # OS 오버헤드(~4-6 GB) + KV 캐시 여유를 감안한 보수적 RAM 임계값
+    {"ram": 128 * 1024, "vram": 48 * 1024,
+     "id": "mlx-community/gpt-oss-120b-MXFP4-Q4", "q": "mxfp4", "multimodal": False},
+    {"ram": 64 * 1024, "vram": 32 * 1024,
+     "id": "mlx-community/gemma-4-31b-it-4bit", "q": "4bit", "multimodal": True},
     {"ram": 64 * 1024, "vram": 32 * 1024,
      "id": "Qwen/Qwen3-VL-30B-A3B-Instruct", "q": "q4_K_M", "multimodal": True},
+    {"ram": 48 * 1024, "vram": 24 * 1024,
+     "id": "mlx-community/gemma-4-31b-it-4bit", "q": "4bit", "multimodal": True},
+    {"ram": 32 * 1024, "vram": 16 * 1024,
+     "id": "mlx-community/gpt-oss-20b-MXFP4-Q8", "q": "mxfp4", "multimodal": False},
     {"ram": 48 * 1024, "vram": 24 * 1024,
      "id": "Qwen/Qwen3-VL-30B-A3B-Instruct", "q": "q4_K_M", "multimodal": True},
     {"ram": 32 * 1024, "vram": 16 * 1024,
@@ -630,7 +638,13 @@ def plan(profile: SystemProfile, rec: Recommendation) -> InstallPlan:
     model_command = ["huggingface-cli", "download", rec.model_id, "--quiet"]
     if rec.runtime == "ollama":
         lower = rec.model_id.lower()
-        if "qwen3-vl-8b" in lower:
+        if "gpt-oss-120b" in lower:
+            model_command = ["ollama", "pull", "gpt-oss:120b"]
+        elif "gpt-oss-20b" in lower:
+            model_command = ["ollama", "pull", "gpt-oss:20b"]
+        elif "gemma-4-31b" in lower:
+            model_command = ["ollama", "pull", "hf.co/ggml-org/gemma-4-31B-it-GGUF:Q4_K_M"]
+        elif "qwen3-vl-8b" in lower:
             model_command = ["ollama", "pull", "qwen3-vl:8b"]
         elif "qwen3-vl-4b" in lower:
             model_command = ["ollama", "pull", "qwen3-vl:4b"]

package/docs/CHANGELOG.md

@@ -1,5 +1,29 @@
 # Changelog
 
+## [0.2.2] - 2026-05-26
+
+### 모델 카탈로그
+
+- `GPT-OSS 20B`, `GPT-OSS 120B`, `Gemma 4 31B 4-bit`를 MLX/Ollama/vLLM/LM Studio/llama.cpp 모델 선택 및 다운로드/로드 흐름에 추가
+- 엔진별 모델 목록에서 같은 패밀리의 최신 major/minor 버전이 있으면 낮은 버전 항목을 숨기도록 정리
+- 설정 마법사 추천표와 RAM 티어에 새 모델을 반영
+
+### 지식 그래프
+
+- 로컬 폴더 스캔 시 PDF, Word, PowerPoint, Excel, CSV, 텍스트/코드, OCR 이미지 등 지원 파일은 실제 본문 텍스트가 추출된 경우에만 그래프 노드로 생성
+- 빈 PDF/Word/PowerPoint/Excel 파일이나 OCR이 비어 있는 파일은 `skipped_empty_text`로 기록하고 그래프에는 표시하지 않도록 변경
+- 기존 버전에서 파일명/상대경로만으로 만들어진 로컬 파일 노드는 다음 스캔에서 재추출 검증 후 자동 정리
+- Word 표 셀, PowerPoint 슬라이드 텍스트, Excel 실제 셀 값 추출을 보강하고 파일명 기반 개념 추출을 제거
+
+### UX
+
+- 지식 그래프 오른쪽 사이드바의 하단 잘림 문제를 수정하고 데스크톱/모바일에서 패널, 메타데이터, 긴 경로가 자연스럽게 스크롤/줄바꿈되도록 조정
+
+### Release
+
+- 배포 버전을 `0.2.2`로 상향
+- 대상 채널: `npm` · `PyPI` · `VS Code Marketplace` · `Open VSX`
+
 ## [0.2.1] - 2026-05-25
 
 ### 버그 수정

package/knowledge_graph.py

@@ -1223,21 +1223,44 @@ class KnowledgeGraphStore:
             from docx import Document
             doc = Document(str(path))
             paragraphs = [p.text for p in doc.paragraphs if p.text.strip()]
+            table_lines = []
+            for table in doc.tables:
+                for row in table.rows:
+                    cells = [_clean_text(cell.text) for cell in row.cells]
+                    if any(cells):
+                        table_lines.append("\t".join(cells))
             meta["paragraphs"] = len(paragraphs)
             meta["tables"] = len(doc.tables)
-            text = "\n\n".join(paragraphs)
+            meta["table_rows"] = len(table_lines)
+            text = "\n\n".join([*paragraphs, *table_lines])
         elif ext == ".xlsx":
             from openpyxl import load_workbook
             wb = load_workbook(str(path), read_only=True, data_only=True)
             rows_all = []
+            non_empty_rows = 0
+            non_empty_cells = 0
+            char_count = 0
             for ws in wb.worksheets:
-                rows_all.append(f"[Sheet: {ws.title}]")
+                sheet_rows = []
                 for row in ws.iter_rows(values_only=True):
-                    cells = [str(cell) if cell is not None else "" for cell in row]
-                    rows_all.append("\t".join(cells))
-                    if len("\n".join(rows_all)) > 200_000:
+                    cells = [str(cell).strip() if cell is not None else "" for cell in row]
+                    if not any(cells):
+                        continue
+                    line = "\t".join(cells)
+                    non_empty_rows += 1
+                    non_empty_cells += sum(1 for cell in cells if cell)
+                    sheet_rows.append(line)
+                    char_count += len(line) + 1
+                    if char_count > 200_000:
                         break
+                if sheet_rows:
+                    rows_all.append(f"[Sheet: {ws.title}]")
+                    rows_all.extend(sheet_rows)
+                if char_count > 200_000:
+                    break
             meta["sheets"] = len(wb.worksheets)
+            meta["rows"] = non_empty_rows
+            meta["cells"] = non_empty_cells
             text = "\n".join(rows_all)
         elif ext == ".pptx":
             from pptx import Presentation
@@ -1247,9 +1270,13 @@ class KnowledgeGraphStore:
                 parts = []
                 for shape in slide.shapes:
                     if getattr(shape, "has_text_frame", False):
-                        parts.append(shape.text_frame.text)
-                slides_text.append(f"[Slide {index}]\n" + "\n".join(parts))
+                        slide_text = shape.text_frame.text.strip()
+                        if slide_text:
+                            parts.append(slide_text)
+                if parts:
+                    slides_text.append(f"[Slide {index}]\n" + "\n".join(parts))
             meta["slides"] = len(prs.slides)
+            meta["text_slides"] = len(slides_text)
             text = "\n\n".join(slides_text)
         elif category == "image":
             from PIL import Image
@@ -1362,13 +1389,13 @@ class KnowledgeGraphStore:
               extension=excluded.extension,
               size_bytes=excluded.size_bytes,
               modified_at=excluded.modified_at,
-              sha256=COALESCE(excluded.sha256, local_file_index.sha256),
+              sha256=excluded.sha256,
               last_scanned_at=excluded.last_scanned_at,
-              last_indexed_at=COALESCE(excluded.last_indexed_at, local_file_index.last_indexed_at),
+              last_indexed_at=excluded.last_indexed_at,
               parser_type=excluded.parser_type,
               status=excluded.status,
               error_message=excluded.error_message,
-              graph_node_id=COALESCE(excluded.graph_node_id, local_file_index.graph_node_id),
+              graph_node_id=excluded.graph_node_id,
               deleted=excluded.deleted,
               metadata_json=excluded.metadata_json
             """,
@@ -1381,6 +1408,113 @@ class KnowledgeGraphStore:
         )
         return index_id
 
+    def _delete_local_file_graph(self, conn: sqlite3.Connection, file_node_id: Optional[str]) -> None:
+        if not file_node_id:
+            return
+
+        file_row = conn.execute(
+            "SELECT metadata_json FROM nodes WHERE id=?",
+            (file_node_id,),
+        ).fetchone()
+        source_id = None
+        if file_row:
+            source_id = _safe_loads(file_row["metadata_json"]).get("source_id")
+
+        linked_rows = conn.execute(
+            """
+            SELECT n.id, n.type, n.metadata_json
+            FROM edges e
+            JOIN nodes n ON n.id=e.to_node
+            WHERE e.from_node=?
+            """,
+            (file_node_id,),
+        ).fetchall()
+        owned_ids: set = set()
+        auto_candidate_ids: set = set()
+        for row in linked_rows:
+            metadata = _safe_loads(row["metadata_json"])
+            if row["type"] in {"Chunk", "ImageText", "Section"} or metadata.get("source_node") == file_node_id:
+                owned_ids.add(row["id"])
+            elif metadata.get("auto_extracted") and metadata.get("source") == "local_folder":
+                auto_candidate_ids.add(row["id"])
+
+        conn.execute("DELETE FROM chunks WHERE source_node=?", (file_node_id,))
+        conn.execute("DELETE FROM edges WHERE from_node=? OR to_node=?", (file_node_id, file_node_id))
+        conn.execute("DELETE FROM nodes WHERE id=?", (file_node_id,))
+
+        def delete_nodes(node_ids: set) -> None:
+            if not node_ids:
+                return
+            placeholders = ",".join("?" * len(node_ids))
+            params = list(node_ids)
+            conn.execute(f"DELETE FROM chunks WHERE source_node IN ({placeholders})", params)
+            conn.execute(f"DELETE FROM edges WHERE from_node IN ({placeholders}) OR to_node IN ({placeholders})", params * 2)
+            conn.execute(f"DELETE FROM nodes WHERE id IN ({placeholders})", params)
+
+        delete_nodes(owned_ids)
+
+        removable_auto_ids: set = set()
+        for node_id in auto_candidate_ids:
+            remaining_edges = conn.execute(
+                "SELECT from_node, to_node FROM edges WHERE from_node=? OR to_node=?",
+                (node_id, node_id),
+            ).fetchall()
+            if all(
+                (row["from_node"] in auto_candidate_ids and row["to_node"] in auto_candidate_ids)
+                for row in remaining_edges
+            ):
+                removable_auto_ids.add(node_id)
+        delete_nodes(removable_auto_ids)
+        if source_id:
+            self._cleanup_local_graph_orphans(conn, str(source_id))
+
+    def _cleanup_local_graph_orphans(self, conn: sqlite3.Connection, source_id: str) -> None:
+        while True:
+            folder_rows = conn.execute(
+                "SELECT id, metadata_json FROM nodes WHERE type='Folder'"
+            ).fetchall()
+            leaf_ids = []
+            for row in folder_rows:
+                metadata = _safe_loads(row["metadata_json"])
+                if metadata.get("source_id") != source_id:
+                    continue
+                has_children = conn.execute(
+                    "SELECT 1 FROM edges WHERE from_node=? LIMIT 1",
+                    (row["id"],),
+                ).fetchone()
+                if not has_children:
+                    leaf_ids.append(row["id"])
+            if not leaf_ids:
+                break
+            placeholders = ",".join("?" * len(leaf_ids))
+            conn.execute(f"DELETE FROM edges WHERE from_node IN ({placeholders}) OR to_node IN ({placeholders})", leaf_ids * 2)
+            conn.execute(f"DELETE FROM nodes WHERE id IN ({placeholders})", leaf_ids)
+
+        for node_type in ("Drive", "Computer"):
+            rows = conn.execute("SELECT id FROM nodes WHERE type=?", (node_type,)).fetchall()
+            removable = []
+            for row in rows:
+                has_children = conn.execute(
+                    "SELECT 1 FROM edges WHERE from_node=? LIMIT 1",
+                    (row["id"],),
+                ).fetchone()
+                if not has_children:
+                    removable.append(row["id"])
+            if removable:
+                placeholders = ",".join("?" * len(removable))
+                conn.execute(f"DELETE FROM edges WHERE from_node IN ({placeholders}) OR to_node IN ({placeholders})", removable * 2)
+                conn.execute(f"DELETE FROM nodes WHERE id IN ({placeholders})", removable)
+
+    def _local_file_index_has_extracted_text(self, row: sqlite3.Row) -> bool:
+        metadata = _safe_loads(row["metadata_json"])
+        parser = metadata.get("parser") if isinstance(metadata, dict) else {}
+        if not isinstance(parser, dict):
+            return False
+        try:
+            return int(parser.get("extracted_chars") or 0) > 0
+        except (TypeError, ValueError):
+            return False
+
     def _upsert_local_file_node(
         self,
         conn: sqlite3.Connection,
@@ -1397,6 +1531,9 @@ class KnowledgeGraphStore:
         text: str,
         parser_meta: Dict[str, Any],
     ) -> str:
+        text = _clean_text(text)
+        if not text:
+            raise ValueError("텍스트 추출 결과가 비어 있습니다.")
         try:
             relative_path = file_path.relative_to(root).as_posix()
         except ValueError:
@@ -1446,7 +1583,7 @@ class KnowledgeGraphStore:
             file_node_id,
             _node_type_for_category(category),
             file_path.name,
-            summary=(_clean_text(text) or relative_path)[:700],
+            summary=text[:700],
             metadata=metadata,
             raw=metadata,
         )
@@ -1488,7 +1625,7 @@ class KnowledgeGraphStore:
             )
             self._upsert_edge(conn, file_node_id, chunk_id, "포함함", weight=0.7, metadata={"source": "local_scan"})
 
-        concepts = _extract_concepts(f"{file_path.name}\n{target_for_concepts}", limit=18)
+        concepts = _extract_concepts(target_for_concepts, limit=18)
         concept_ids: Dict[str, str] = {}
         for concept in concepts:
             node_t = _classify_node_type(concept, target_for_concepts)
@@ -1620,10 +1757,21 @@ class KnowledgeGraphStore:
                 except ValueError:
                     relative_path = file_path.name
                 seen_relative_paths.add(relative_path)
+                modified_at = _safe_iso_from_stat_mtime(stat.st_mtime)
+                existing = conn.execute(
+                    """
+                    SELECT size_bytes, modified_at, sha256, graph_node_id, status, metadata_json
+                    FROM local_file_index
+                    WHERE source_id=? AND relative_path=?
+                    """,
+                    (source_id, relative_path),
+                ).fetchone()
                 decision = self._local_file_decision(file_path, root, stat)
                 parser_type = decision["parser_type"]
                 if not decision["indexable"]:
                     counts[decision["status"]] += 1
+                    if existing and existing["graph_node_id"]:
+                        self._delete_local_file_graph(conn, existing["graph_node_id"])
                     self._upsert_local_file_index(
                         conn,
                         source_id=source_id,
@@ -1638,19 +1786,11 @@ class KnowledgeGraphStore:
                     )
                     continue
 
-                modified_at = _safe_iso_from_stat_mtime(stat.st_mtime)
-                existing = conn.execute(
-                    """
-                    SELECT size_bytes, modified_at, sha256, graph_node_id, status
-                    FROM local_file_index
-                    WHERE source_id=? AND relative_path=?
-                    """,
-                    (source_id, relative_path),
-                ).fetchone()
                 if (
                     existing
                     and existing["status"] == "indexed"
                     and existing["graph_node_id"]
+                    and self._local_file_index_has_extracted_text(existing)
                     and existing["size_bytes"] == stat.st_size
                     and existing["modified_at"] == modified_at
                 ):
@@ -1667,7 +1807,7 @@ class KnowledgeGraphStore:
                         parser_type=parser_type,
                         sha256=existing["sha256"],
                         graph_node_id=existing["graph_node_id"],
-                        metadata={"category": decision["category"], "unchanged": True},
+                        metadata={**_safe_loads(existing["metadata_json"]), "category": decision["category"], "unchanged": True},
                     )
                     continue
 
@@ -1677,6 +1817,8 @@ class KnowledgeGraphStore:
                 except Exception as exc:
                     counts["failed"] += 1
                     errors.append({"path": str(file_path), "error": str(exc)})
+                    if existing and existing["graph_node_id"]:
+                        self._delete_local_file_graph(conn, existing["graph_node_id"])
                     self._upsert_local_file_index(
                         conn,
                         source_id=source_id,
@@ -1692,7 +1834,12 @@ class KnowledgeGraphStore:
                     )
                     continue
 
-                if existing and existing["sha256"] == digest and existing["graph_node_id"]:
+                if (
+                    existing
+                    and existing["sha256"] == digest
+                    and existing["graph_node_id"]
+                    and self._local_file_index_has_extracted_text(existing)
+                ):
                     counts["skipped_unchanged"] += 1
                     self._upsert_local_file_index(
                         conn,
@@ -1706,7 +1853,7 @@ class KnowledgeGraphStore:
                         parser_type=parser_type,
                         sha256=digest,
                         graph_node_id=existing["graph_node_id"],
-                        metadata={"category": decision["category"], "sha256_unchanged": True},
+                        metadata={**_safe_loads(existing["metadata_json"]), "category": decision["category"], "sha256_unchanged": True},
                     )
                     continue
 
@@ -1716,6 +1863,27 @@ class KnowledgeGraphStore:
                         decision["category"],
                         include_ocr=include_ocr,
                     )
+                    text = _clean_text(text)
+                    parser_meta = {**parser_meta, "extracted_chars": len(text)}
+                    if not text:
+                        counts["skipped_empty_text"] += 1
+                        if existing and existing["graph_node_id"]:
+                            self._delete_local_file_graph(conn, existing["graph_node_id"])
+                        self._upsert_local_file_index(
+                            conn,
+                            source_id=source_id,
+                            root=root,
+                            file_path=file_path,
+                            stat=stat,
+                            os_type=os_type,
+                            drive_id=drive_id,
+                            status="skipped_empty_text",
+                            parser_type=parser_type,
+                            sha256=digest,
+                            error_message="텍스트 추출 결과가 비어 있습니다.",
+                            metadata={"category": decision["category"], "parser": parser_meta},
+                        )
+                        continue
                     graph_node_id = self._upsert_local_file_node(
                         conn,
                         source_id=source_id,
@@ -1749,6 +1917,8 @@ class KnowledgeGraphStore:
                 except Exception as exc:
                     counts["failed"] += 1
                     errors.append({"path": str(file_path), "error": str(exc)})
+                    if existing and existing["graph_node_id"]:
+                        self._delete_local_file_graph(conn, existing["graph_node_id"])
                     self._upsert_local_file_index(
                         conn,
                         source_id=source_id,
@@ -1765,19 +1935,20 @@ class KnowledgeGraphStore:
                     )
 
             if not limit_reached:
-                existing_paths = {
-                    row["relative_path"]
+                existing_rows = {
+                    row["relative_path"]: row["graph_node_id"]
                     for row in conn.execute(
-                        "SELECT relative_path FROM local_file_index WHERE source_id=?",
+                        "SELECT relative_path, graph_node_id FROM local_file_index WHERE source_id=?",
                         (source_id,),
                     )
                 }
-                deleted_paths = existing_paths - seen_relative_paths
+                deleted_paths = set(existing_rows) - seen_relative_paths
                 for relative_path in deleted_paths:
+                    self._delete_local_file_graph(conn, existing_rows.get(relative_path))
                     conn.execute(
                         """
                         UPDATE local_file_index
-                        SET status='deleted', deleted=1, last_scanned_at=?, error_message=NULL
+                        SET status='deleted', deleted=1, last_scanned_at=?, error_message=NULL, graph_node_id=NULL
                         WHERE source_id=? AND relative_path=?
                         """,
                         (_now(), source_id, relative_path),

package/llm_router.py

@@ -329,7 +329,7 @@ class LLMRouter:
             return f"Cached: {cache_key}"
 
         self._enforce_local_model_limit(cache_key)
-        print(f"⏳ Loading Gemma 4 Stack: {cache_key}...")
+        print(f"⏳ Loading local model stack: {cache_key}...")
         loop = asyncio.get_event_loop()
         target_model_id = _resolve_local_hf_model(model_id)
         target_draft_model_id = _resolve_local_hf_model(draft_model_id) if draft_model_id else None

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ltcai",
-  "version": "0.2.1",
+  "version": "0.2.2",
   "description": "Lattice AI local MLX/cloud LLM workspace server",
   "homepage": "https://github.com/TaeSooPark-PTS/LatticeAI#readme",
   "repository": {
@@ -54,7 +54,7 @@
     "tools.py",
     "codex_telegram_bot.py",
     "mcp_registry.py",
-    "latticeai/",
+    "latticeai/**/*.py",
     "skills/",
     "static/account.html",
     "static/chat.html",

package/server.py

@@ -1103,7 +1103,7 @@ async def lifespan(app: FastAPI):
             except Exception:
                 pass
 
-app = FastAPI(title=f"Lattice AI Server ({APP_MODE})", version="2.1.0", lifespan=lifespan)
+app = FastAPI(title=f"Lattice AI Server ({APP_MODE})", version="0.2.2", lifespan=lifespan)
 
 CORS_ALLOWED_ORIGINS = [
     f"http://localhost:{DEFAULT_PORT}",
@@ -1620,6 +1620,9 @@ ENGINE_MODEL_CATALOG = {
         {"id": "mlx-community/Qwen3-VL-30B-A3B-Instruct-4bit", "name": "Qwen3-VL 30B A3B", "family": "Qwen3-VL", "tag": "local-vlm", "size": "18GB", "pullable": True},
         {"id": "mlx-community/gemma-3-27b-it-4bit", "name": "Gemma 3 27B", "family": "Gemma 3", "tag": "local-vlm", "size": "17GB", "pullable": True},
         {"id": "mlx-community/gemma-4-26b-a4b-it-4bit", "name": "Gemma 4 26B A4B Instruct", "family": "Gemma 4", "tag": "local-vlm", "size": "15.6GB", "pullable": True},
+        {"id": "mlx-community/gemma-4-31b-it-4bit", "name": "Gemma 4 31B Instruct", "family": "Gemma 4", "tag": "local-vlm", "size": "18.4GB", "pullable": True},
+        {"id": "mlx-community/gpt-oss-20b-MXFP4-Q8", "name": "GPT-OSS 20B", "family": "GPT-OSS", "tag": "local-reasoning", "size": "12.1GB", "pullable": True},
+        {"id": "mlx-community/gpt-oss-120b-MXFP4-Q4", "name": "GPT-OSS 120B", "family": "GPT-OSS", "tag": "local-large", "size": "62.3GB", "pullable": True},
         {"id": "mlx-community/Llama-3.3-70B-Instruct-4bit", "name": "Llama 3.3 70B", "family": "Llama 3.x", "tag": "local-general", "size": "40GB+", "pullable": True},
         {"id": "mlx-community/Llama-3.1-70B-Instruct-4bit", "name": "Llama 3.1 70B", "family": "Llama 3.1", "tag": "local-general", "size": "40GB+", "pullable": True},
     ],
@@ -1627,6 +1630,9 @@ ENGINE_MODEL_CATALOG = {
         {"id": "ollama:qwen3-vl:4b", "name": "Qwen3-VL 4B via Ollama", "family": "Qwen3-VL", "tag": "local-vlm", "size": "pull required", "pullable": True},
         {"id": "ollama:qwen3-vl:8b", "name": "Qwen3-VL 8B via Ollama", "family": "Qwen3-VL", "tag": "local-vlm", "size": "pull required", "pullable": True},
         {"id": "ollama:qwen3-vl:30b", "name": "Qwen3-VL 30B via Ollama", "family": "Qwen3-VL", "tag": "local-vlm", "size": "pull required", "pullable": True},
+        {"id": "ollama:gpt-oss:20b", "name": "GPT-OSS 20B via Ollama", "family": "GPT-OSS", "tag": "local-reasoning", "size": "pull required", "pullable": True},
+        {"id": "ollama:gpt-oss:120b", "name": "GPT-OSS 120B via Ollama", "family": "GPT-OSS", "tag": "local-large", "size": "pull required", "pullable": True},
+        {"id": "ollama:hf.co/ggml-org/gemma-4-31B-it-GGUF:Q4_K_M", "name": "Gemma 4 31B Q4 via Ollama", "family": "Gemma 4", "tag": "local-vlm", "size": "18.7GB", "pullable": True},
         {"id": "ollama:qwen3:8b", "name": "Qwen3 8B via Ollama", "family": "Qwen", "tag": "local-server", "size": "pull required", "pullable": True},
         {"id": "ollama:qwen2.5-coder:14b", "name": "Qwen2.5 Coder 14B via Ollama", "family": "Qwen", "tag": "local-coding", "size": "pull required", "pullable": True},
         {"id": "ollama:gemma3:1b", "name": "Gemma 3 1B via Ollama", "family": "Gemma", "tag": "local-light", "size": "pull required", "pullable": True},
@@ -1649,6 +1655,8 @@ ENGINE_MODEL_CATALOG = {
         {"id": "ollama:smollm2:1.7b", "name": "SmolLM2 1.7B via Ollama", "family": "SmolLM", "tag": "local-light", "size": "pull required", "pullable": True},
     ],
     "vllm": [
+        {"id": "vllm:openai/gpt-oss-20b", "name": "GPT-OSS 20B via vLLM", "family": "GPT-OSS", "tag": "local-reasoning", "size": "server model", "pullable": True},
+        {"id": "vllm:openai/gpt-oss-120b", "name": "GPT-OSS 120B via vLLM", "family": "GPT-OSS", "tag": "local-large", "size": "server model", "pullable": True},
         {"id": "vllm:Qwen/Qwen3-VL-4B-Instruct", "name": "Qwen3-VL 4B via vLLM", "family": "Qwen3-VL", "tag": "local-vlm", "size": "server model", "pullable": True},
         {"id": "vllm:Qwen/Qwen3-VL-8B-Instruct", "name": "Qwen3-VL 8B via vLLM", "family": "Qwen3-VL", "tag": "local-vlm", "size": "server model", "pullable": True},
         {"id": "vllm:Qwen/Qwen3-VL-30B-A3B-Instruct", "name": "Qwen3-VL 30B A3B via vLLM", "family": "Qwen3-VL", "tag": "local-vlm", "size": "server model", "pullable": True},
@@ -1671,6 +1679,9 @@ ENGINE_MODEL_CATALOG = {
         {"id": "vllm:meta-llama/Llama-3.1-70B-Instruct", "name": "Llama 3.1 70B via vLLM", "family": "Llama 3.1", "tag": "local-server", "size": "server model", "pullable": True},
     ],
     "lmstudio": [
+        {"id": "lmstudio:openai/gpt-oss-20b", "name": "GPT-OSS 20B via LM Studio", "family": "GPT-OSS", "tag": "local-reasoning", "size": "server model", "pullable": True},
+        {"id": "lmstudio:openai/gpt-oss-120b", "name": "GPT-OSS 120B via LM Studio", "family": "GPT-OSS", "tag": "local-large", "size": "server model", "pullable": True},
+        {"id": "lmstudio:ggml-org/gemma-4-31B-it-GGUF", "name": "Gemma 4 31B 4-bit via LM Studio", "family": "Gemma 4", "tag": "local-vlm", "size": "server model", "pullable": True},
         {"id": "lmstudio:Qwen/Qwen3-VL-4B-Instruct", "name": "Qwen3-VL 4B via LM Studio", "family": "Qwen3-VL", "tag": "local-vlm", "size": "server model", "pullable": True},
         {"id": "lmstudio:Qwen/Qwen3-VL-8B-Instruct", "name": "Qwen3-VL 8B via LM Studio", "family": "Qwen3-VL", "tag": "local-vlm", "size": "server model", "pullable": True},
         {"id": "lmstudio:Qwen/Qwen3-VL-30B-A3B-Instruct", "name": "Qwen3-VL 30B A3B via LM Studio", "family": "Qwen3-VL", "tag": "local-vlm", "size": "server model", "pullable": True},
@@ -1691,6 +1702,9 @@ ENGINE_MODEL_CATALOG = {
         {"id": "lmstudio:meta-llama/Llama-3.1-70B-Instruct", "name": "Llama 3.1 70B via LM Studio", "family": "Llama 3.1", "tag": "local-server", "size": "server model", "pullable": True},
     ],
     "llamacpp": [
+        {"id": "llamacpp:ggml-org/gpt-oss-20b-GGUF", "name": "GPT-OSS 20B GGUF via llama.cpp", "family": "GPT-OSS", "tag": "gguf-q4", "size": "gguf", "pullable": True},
+        {"id": "llamacpp:ggml-org/gpt-oss-120b-GGUF", "name": "GPT-OSS 120B GGUF via llama.cpp", "family": "GPT-OSS", "tag": "gguf-q4", "size": "gguf", "pullable": True},
+        {"id": "llamacpp:ggml-org/gemma-4-31B-it-GGUF", "name": "Gemma 4 31B GGUF via llama.cpp", "family": "Gemma 4", "tag": "gguf-q4", "size": "gguf", "pullable": True},
         {"id": "llamacpp:Qwen/Qwen3-VL-4B-Instruct-GGUF", "name": "Qwen3-VL 4B GGUF via llama.cpp", "family": "Qwen3-VL", "tag": "gguf-vlm", "size": "gguf", "pullable": True},
         {"id": "llamacpp:Qwen/Qwen3-VL-8B-Instruct-GGUF", "name": "Qwen3-VL 8B GGUF via llama.cpp", "family": "Qwen3-VL", "tag": "gguf-vlm", "size": "gguf", "pullable": True},
         {"id": "llamacpp:unsloth/gemma-2-2b-it-GGUF", "name": "Gemma 2 2B GGUF via llama.cpp", "family": "Gemma", "tag": "gguf-q4", "size": "gguf", "pullable": True},
@@ -1706,6 +1720,97 @@ ENGINE_MODEL_CATALOG = {
     ],
 }
 
+MODEL_ENGINE_ALIASES = {
+    "gpt-oss-20b": {
+        "local_mlx": "mlx-community/gpt-oss-20b-MXFP4-Q8",
+        "ollama": "gpt-oss:20b",
+        "vllm": "openai/gpt-oss-20b",
+        "lmstudio": "openai/gpt-oss-20b",
+        "llamacpp": "ggml-org/gpt-oss-20b-GGUF",
+    },
+    "openai/gpt-oss-20b": {
+        "local_mlx": "mlx-community/gpt-oss-20b-MXFP4-Q8",
+        "ollama": "gpt-oss:20b",
+        "vllm": "openai/gpt-oss-20b",
+        "lmstudio": "openai/gpt-oss-20b",
+        "llamacpp": "ggml-org/gpt-oss-20b-GGUF",
+    },
+    "gpt-oss-120b": {
+        "local_mlx": "mlx-community/gpt-oss-120b-MXFP4-Q4",
+        "ollama": "gpt-oss:120b",
+        "vllm": "openai/gpt-oss-120b",
+        "lmstudio": "openai/gpt-oss-120b",
+        "llamacpp": "ggml-org/gpt-oss-120b-GGUF",
+    },
+    "openai/gpt-oss-120b": {
+        "local_mlx": "mlx-community/gpt-oss-120b-MXFP4-Q4",
+        "ollama": "gpt-oss:120b",
+        "vllm": "openai/gpt-oss-120b",
+        "lmstudio": "openai/gpt-oss-120b",
+        "llamacpp": "ggml-org/gpt-oss-120b-GGUF",
+    },
+    "gemma-4-31b-it-4bit": {
+        "local_mlx": "mlx-community/gemma-4-31b-it-4bit",
+        "ollama": "hf.co/ggml-org/gemma-4-31B-it-GGUF:Q4_K_M",
+        "vllm": "suitch/gemma-4-31B-it-4bit",
+        "lmstudio": "ggml-org/gemma-4-31B-it-GGUF",
+        "llamacpp": "ggml-org/gemma-4-31B-it-GGUF",
+    },
+    "suitch/gemma-4-31b-it-4bit": {
+        "local_mlx": "mlx-community/gemma-4-31b-it-4bit",
+        "ollama": "hf.co/ggml-org/gemma-4-31B-it-GGUF:Q4_K_M",
+        "vllm": "suitch/gemma-4-31B-it-4bit",
+        "lmstudio": "ggml-org/gemma-4-31B-it-GGUF",
+        "llamacpp": "ggml-org/gemma-4-31B-it-GGUF",
+    },
+    "mlx-community/gemma-4-31b-it-4bit": {
+        "local_mlx": "mlx-community/gemma-4-31b-it-4bit",
+        "ollama": "hf.co/ggml-org/gemma-4-31B-it-GGUF:Q4_K_M",
+        "vllm": "suitch/gemma-4-31B-it-4bit",
+        "lmstudio": "ggml-org/gemma-4-31B-it-GGUF",
+        "llamacpp": "ggml-org/gemma-4-31B-it-GGUF",
+    },
+}
+
+_VERSIONED_MODEL_PATTERNS = (
+    ("gemma", re.compile(r"\bgemma[-\s]?(\d+(?:\.\d+)?)", re.IGNORECASE)),
+    ("qwen", re.compile(r"\bqwen[-\s]?(\d+(?:\.\d+)?)", re.IGNORECASE)),
+    ("llama", re.compile(r"\bllama[-\s]?(\d+(?:\.\d+)?)", re.IGNORECASE)),
+    ("phi", re.compile(r"\bphi[-\s]?(\d+(?:\.\d+)?)", re.IGNORECASE)),
+)
+
+
+def _version_tuple(raw: str) -> tuple[int, ...]:
+    return tuple(int(part) for part in raw.split(".") if part.isdigit())
+
+
+def _model_family_version(model: Dict[str, object]) -> Optional[tuple[str, tuple[int, ...]]]:
+    text = " ".join(str(model.get(key) or "") for key in ("family", "name", "id"))
+    for family, pattern in _VERSIONED_MODEL_PATTERNS:
+        match = pattern.search(text)
+        if match:
+            version = _version_tuple(match.group(1))
+            if version:
+                return family, version
+    return None
+
+
+def filter_lower_family_versions(models: List[Dict[str, object]]) -> List[Dict[str, object]]:
+    max_versions: Dict[str, tuple[int, ...]] = {}
+    detected: List[tuple[Dict[str, object], Optional[tuple[str, tuple[int, ...]]]]] = []
+    for model in models:
+        version_info = _model_family_version(model)
+        detected.append((model, version_info))
+        if not version_info:
+            continue
+        family, version = version_info
+        if version > max_versions.get(family, (0,)):
+            max_versions[family] = version
+    return [
+        model for model, version_info in detected
+        if not version_info or version_info[1] >= max_versions.get(version_info[0], version_info[1])
+    ]
+
 def _update_env_file(env_file: Path, key: str, value: str) -> None:
     lines = []
     found = False
@@ -2525,17 +2630,20 @@ def engine_status() -> List[Dict]:
     for m in ENGINE_MODEL_CATALOG["ollama"]:
         pull_name = m["id"].removeprefix("ollama:")
         ollama_models.append({**m, "pulled": pull_name in pulled})
+    ollama_models = filter_lower_family_versions(ollama_models)
 
     HF_MODELS_ROOT.mkdir(parents=True, exist_ok=True)
     mlx_models = []
     for m in ENGINE_MODEL_CATALOG.get("local_mlx", []):
         repo_id = m["id"]
         mlx_models.append({**m, "pulled": hf_model_ready(repo_id, "local_mlx")})
+    mlx_models = filter_lower_family_versions(mlx_models)
 
     vllm_models = []
     for m in ENGINE_MODEL_CATALOG.get("vllm", []):
         repo_id = m["id"].removeprefix("vllm:")
         vllm_models.append({**m, "pulled": hf_model_ready(repo_id, "vllm")})
+    vllm_models = filter_lower_family_versions(vllm_models)
 
     lmstudio_models = []
     downloaded_lmstudio = get_lmstudio_models()
@@ -2567,11 +2675,13 @@ def engine_status() -> List[Dict]:
             repo_id = m["id"].removeprefix("lmstudio:")
             if f"lmstudio:{repo_id}" not in known_ids and repo_id not in downloaded_by_key:
                 lmstudio_models.append({**m, "pulled": False})
+    lmstudio_models = filter_lower_family_versions(lmstudio_models)
 
     llamacpp_models = []
     for m in ENGINE_MODEL_CATALOG.get("llamacpp", []):
         repo_id = m["id"].removeprefix("llamacpp:")
         llamacpp_models.append({**m, "pulled": hf_model_ready(repo_id, "llamacpp")})
+    llamacpp_models = filter_lower_family_versions(llamacpp_models)
 
     local_server_specs = [
         {
@@ -2768,8 +2878,29 @@ def install_engine(engine: str) -> Dict:
     return result
 
 
+def _resolve_model_alias(model_id: str, engine: Optional[str] = None) -> str:
+    raw = model_id.strip()
+    engine_hint = (engine or "").strip().lower()
+    provider: Optional[str] = None
+    model_name = raw
+    if ":" in raw:
+        prefix, rest = raw.split(":", 1)
+        prefix = prefix.strip().lower()
+        if prefix in {"ollama", "vllm", "lmstudio", "llamacpp", "local_mlx", "mlx"}:
+            provider = "local_mlx" if prefix in {"local_mlx", "mlx"} else prefix
+            model_name = rest.strip()
+    provider = provider or ("local_mlx" if engine_hint in {"", "local_mlx", "mlx"} else engine_hint)
+    aliases = MODEL_ENGINE_ALIASES.get(model_name.lower())
+    if not aliases:
+        return raw
+    mapped = aliases.get(provider)
+    if not mapped:
+        return raw
+    return mapped if provider == "local_mlx" else f"{provider}:{mapped}"
+
+
 def normalize_local_model_request(model_id: str, engine: Optional[str] = None) -> str:
-    model_id = model_id.strip()
+    model_id = _resolve_model_alias(model_id, engine)
     engine = (engine or "").strip().lower()
     if engine in {"local_mlx", "mlx"} and model_id.startswith(("local_mlx:", "mlx:")):
         return model_id.split(":", 1)[1].strip()
@@ -3165,7 +3296,7 @@ async def verify_cloud_models(force: bool = False, provider_filter: Optional[str
 
 @app.get("/health")
 async def health(request: Request):
-    base = {"status": "ok", "version": "2.1.0", "mode": APP_MODE}
+    base = {"status": "ok", "version": "0.2.2", "mode": APP_MODE}
     if not get_current_user(request) and REQUIRE_AUTH:
         return base
     engines = await asyncio.to_thread(engine_status)
@@ -3206,7 +3337,7 @@ async def engines_verify_cloud(req: VerifyCloudRequest, request: Request):
 @app.post("/engines/pull-model")
 async def pull_ollama_model(req: PullModelRequest, request: Request):
     require_user(request)
-    model_ref = req.model.strip()
+    model_ref = normalize_local_model_request(req.model, None)
     if not model_ref:
         raise HTTPException(status_code=400, detail="모델 식별자가 비어 있습니다.")
 
@@ -3324,23 +3455,8 @@ async def set_api_key(req: SetApiKeyRequest, request: Request):
 async def list_models():
     """HuggingFace 추천 모델 목록 및 로드 상태 반환"""
     recommended = [
-        {"id": "mlx-community/Qwen3-VL-4B-Instruct-4bit",       "name": "Qwen3-VL 4B",      "tag": "multimodal",  "size": "2.7GB"},
-        {"id": "mlx-community/Qwen3-VL-8B-Instruct-4bit",       "name": "Qwen3-VL 8B",      "tag": "multimodal",  "size": "4.8GB"},
-        {"id": "mlx-community/Qwen3-VL-30B-A3B-Instruct-4bit",  "name": "Qwen3-VL 30B A3B","tag": "multimodal",  "size": "18GB"},
-        {"id": "mlx-community/SmolLM-1.7B-Instruct-4bit",       "name": "SmolLM 1.7B",       "tag": "ultra-light", "size": "963MB"},
-        {"id": "mlx-community/gemma-3-1b-it-4bit",              "name": "Gemma 3 1B",        "tag": "ultra-light", "size": "733MB"},
-        {"id": "mlx-community/Llama-3.2-1B-Instruct-4bit",      "name": "Llama 3.2 1B",      "tag": "light",       "size": "1.3GB"},
-        {"id": "mlx-community/Llama-3.2-3B-Instruct-4bit",      "name": "Llama 3.2 3B",      "tag": "light",       "size": "2.0GB"},
-        {"id": "mlx-community/Phi-4-mini-instruct-4bit",        "name": "Phi 4 Mini",        "tag": "coding",      "size": "2.2GB"},
-        {"id": "mlx-community/Qwen2.5-VL-7B-Instruct-4bit",     "name": "Qwen2.5-VL 7B",    "tag": "multimodal",  "size": "4.4GB"},
-        {"id": "mlx-community/Mistral-7B-Instruct-v0.3-4bit",   "name": "Mistral 7B v0.3",   "tag": "general",     "size": "4.1GB"},
-        {"id": "mlx-community/Llama-3.1-8B-Instruct-4bit",      "name": "Llama 3.1 8B",      "tag": "general",     "size": "4.7GB"},
-        {"id": "mlx-community/gemma-4-e4b-it-4bit",             "name": "Gemma 4 E4B",       "tag": "multimodal",  "size": "5.2GB"},
-        {"id": "mlx-community/gemma-3-12b-it-4bit",             "name": "Gemma 3 12B",       "tag": "balanced",    "size": "8.0GB"},
-        {"id": "mlx-community/phi-4-4bit",                      "name": "Phi 4",             "tag": "coding",      "size": "8.3GB"},
-        {"id": "mlx-community/Mistral-Small-24B-Instruct-2501-4bit", "name": "Mistral Small 24B", "tag": "large", "size": "13.3GB"},
-        {"id": "mlx-community/Qwen2.5-Coder-32B-Instruct-4bit", "name": "Qwen2.5 Coder 32B","tag": "coding",      "size": "18.5GB"},
-        {"id": "mlx-community/gemma-4-26b-a4b-it-4bit",         "name": "Gemma 4 26B A4B",   "tag": "multimodal",  "size": "15.6GB"},
+        {"id": item["id"], "name": item["name"], "tag": item["tag"], "size": item["size"]}
+        for item in filter_lower_family_versions(ENGINE_MODEL_CATALOG.get("local_mlx", []))
     ]
     return {
         "recommended": recommended,

package/static/lattice-reference.css

@@ -2608,6 +2608,28 @@ body.lattice-ref-graph {
       font-family: "SF Pro Display", "Inter", ui-sans-serif, system-ui, -apple-system, sans-serif;
     }
 
+    body.lattice-ref-graph .app {
+      min-height: 0;
+      height: 100dvh;
+      overflow: hidden;
+    }
+
+    body.lattice-ref-graph .stage {
+      height: calc(100dvh - 138px);
+      max-height: calc(100dvh - 138px);
+      min-height: 0;
+    }
+
+    body.lattice-ref-graph .app > aside:not(.reference-rail) {
+      height: calc(100dvh - 138px);
+      max-height: calc(100dvh - 138px);
+      min-height: 0;
+      overflow-y: auto;
+      overscroll-behavior: contain;
+      padding-bottom: max(28px, env(safe-area-inset-bottom));
+      scrollbar-gutter: stable;
+    }
+
     .app {
       display: grid;
       grid-template-columns: minmax(0, 1fr) 360px;
@@ -2993,7 +3015,7 @@ body.lattice-ref-graph {
       display: flex;
       flex-direction: column;
       gap: 6px;
-      max-height: 150px;
+      max-height: min(150px, 22dvh);
       overflow-y: auto;
       padding-right: 2px;
     }
@@ -3110,7 +3132,7 @@ body.lattice-ref-graph {
       color: var(--muted);
       font-size: 12px;
       line-height: 1.3;
-      white-space: nowrap;
+      white-space: normal;
     }
 
     .local-option-row input {
@@ -3174,6 +3196,7 @@ body.lattice-ref-graph {
       gap: 9px;
       min-height: 26px;
       font-size: 12px;
+      min-width: 0;
     }
 
     .filter-item {
@@ -3208,9 +3231,11 @@ body.lattice-ref-graph {
     .filter-name,
     .legend-name {
       flex: 1;
+      min-width: 0;
       color: #14162c;
       font-size: 13px;
       font-weight: 500;
+      overflow-wrap: anywhere;
     }
 
     .filter-count,
@@ -3223,9 +3248,10 @@ body.lattice-ref-graph {
     }
 
     .detail-wrap {
-      flex: 1;
-      overflow-y: auto;
-      padding: 18px;
+      flex: 0 0 auto;
+      min-height: auto;
+      overflow: visible;
+      padding: 18px 18px max(24px, env(safe-area-inset-bottom));
     }
 
     .type-badge {
@@ -3245,6 +3271,7 @@ body.lattice-ref-graph {
       font-weight: 750;
       margin-bottom: 8px;
       letter-spacing: -0.01em;
+      overflow-wrap: anywhere;
     }
 
     .detail-summary {
@@ -3296,6 +3323,9 @@ body.lattice-ref-graph {
       line-height: 1.65;
       white-space: pre-wrap;
       word-break: break-word;
+      overflow-wrap: anywhere;
+      max-height: min(42dvh, 460px);
+      overflow: auto;
     }
 
     .jump-btn {
@@ -3323,6 +3353,65 @@ body.lattice-ref-graph {
     }
 
     @media (max-width: 900px) {
+      body.lattice-ref-graph {
+        overflow-y: auto;
+      }
+
+      body.lattice-ref-graph .reference-rail {
+        display: none;
+      }
+
+      body.lattice-ref-graph {
+        grid-template-columns: 1fr;
+      }
+
+      body.lattice-ref-graph .app {
+        height: auto;
+        min-height: 100dvh;
+        overflow: visible;
+        grid-template-columns: 1fr;
+        grid-template-rows: minmax(420px, 58dvh) auto;
+      }
+
+      body.lattice-ref-graph .stage {
+        height: auto;
+        max-height: none;
+        min-height: 420px;
+        margin: 84px 12px 12px;
+      }
+
+      body.lattice-ref-graph .stage::before {
+        top: 26px;
+        left: 58px;
+        font-size: 24px;
+      }
+
+      body.lattice-ref-graph .stage::after {
+        top: 22px;
+        left: 14px;
+      }
+
+      body.lattice-ref-graph .search-shell {
+        top: -60px;
+        left: 60px;
+        right: 12px;
+        width: auto;
+      }
+
+      body.lattice-ref-graph .app > aside:not(.reference-rail) {
+        height: auto;
+        max-height: none;
+        min-height: 0;
+        margin: 0 12px 24px;
+        overflow: visible;
+        padding-bottom: max(18px, env(safe-area-inset-bottom));
+      }
+
+      body.lattice-ref-graph .detail-wrap {
+        flex: none;
+        overflow: visible;
+      }
+
       .app {
         grid-template-columns: 1fr;
         grid-template-rows: 1fr 360px;