mcp-server/src/legal_mcp/services/digest_library.py

@@ -30,7 +30,7 @@ from typing import Awaitable, Callable
 from uuid import UUID
 
 from legal_mcp import config
-from legal_mcp.services import db, embeddings, extractor, ingest
+from legal_mcp.services import db, embeddings, extractor, ingest, storage
 
 logger = logging.getLogger(__name__)
 
@@ -144,12 +144,23 @@ async def create_pending_digest(
         raise ValueError(f"file not found: {file_path}")
 
     await progress("staging", 10, "מעתיק קובץ")
-    staged = await ingest._stage_file(src, DIGEST_LIBRARY_DIR, "incoming")
+    # ``_stage_file`` returns the storage KEY (DATA_DIR-relative path). Resolve a
-    rel_path = str(staged.relative_to(config.DATA_DIR)) \
+    # real local path to read from — on s3-only this downloads to a temp file we
-        if str(staged).startswith(str(config.DATA_DIR)) else str(staged)
+    # own and remove after extraction (INV-STG1; the key is not guaranteed to be
+    # on local disk).
+    rel_path = await ingest._stage_file(src, DIGEST_LIBRARY_DIR, "incoming")
+    local = await storage.ensure_local(rel_path, bucket=storage.Bucket.DOCUMENTS)
+    local_is_tmp = storage.local_path(rel_path, bucket=storage.Bucket.DOCUMENTS) is None
 
     await progress("extracting_text", 50, "מחלץ טקסט")
-    raw_text, _pc, _off = await extractor.extract_text(str(staged))
+    try:
+        raw_text, _pc, _off = await extractor.extract_text(str(local))
+    finally:
+        if local_is_tmp:
+            try:
+                local.unlink(missing_ok=True)
+            except OSError as e:  # noqa: BLE001 — temp cleanup, never fatal
+                logger.debug("could not remove temp digest file %s: %s", local, e)
     raw_text = (raw_text or "").strip()
     if not raw_text:
         raise ValueError("no text extracted from digest")

mcp-server/src/legal_mcp/services/ingest.py

@@ -66,13 +66,14 @@ def _safe_filename(name: str) -> str:
     return re.sub(r"[^\w.\-+א-ת ]", "_", base) or f"upload-{uuid4().hex[:8]}"
 
 
-async def _stage_file(src_path: Path, root: Path, subdir: str) -> Path:
+async def _stage_file(src_path: Path, root: Path, subdir: str) -> str:
     """Stage an intake file through the unified storage layer (INV-STG1).
 
-    Returns the DATA_DIR path the rest of the pipeline reads from — under the
+    Returns the storage KEY (DATA_DIR-relative path) the blob was written under.
-    filesystem/dual backends the bytes are on disk and the key is the
+    The caller resolves a readable local path via ``storage.ensure_local`` — the
-    DATA_DIR-relative path. The Hebrew original filename rides as object
+    key is NOT guaranteed to map to an existing on-disk file (under the s3-only
-    metadata, never as the key (INV-STG2)."""
+    backend the bytes live only in object storage). The Hebrew original filename
+    rides as object metadata, never as the key (INV-STG2)."""
     dest = root / (subdir or "other") / f"{uuid4().hex[:8]}_{_safe_filename(src_path.name)}"
     key = dest.relative_to(config.DATA_DIR).as_posix()
     await storage.put_file(
@@ -80,7 +81,7 @@ async def _stage_file(src_path: Path, root: Path, subdir: str) -> Path:
         content_type=mimetypes.guess_type(src_path.name)[0],
         metadata={"filename": src_path.name},
     )
-    return dest
+    return key
 
 
 def _validate_enums(spec: IntakeSpec, inputs: dict) -> None:
@@ -155,12 +156,24 @@ async def ingest_document(
     page_count = 0
     page_offsets = None
     staged: Path | None = None
+    staged_is_tmp = False
     if file_path:
         src = Path(file_path)
         if not src.is_file():
             raise FileNotFoundError(f"file not found: {src}")
         await progress("staging", 5, "מעתיק את הקובץ לאחסון")
-        staged = await _stage_file(src, spec.staging_root, spec.staging_subdir(inputs))
+        staged_key = await _stage_file(src, spec.staging_root, spec.staging_subdir(inputs))
+        # Resolve a real local path to read from. Under filesystem/dual this is
+        # the on-disk copy; under s3-only the blob lives only in object storage,
+        # so ensure_local downloads it to a temp file we own and must clean up
+        # (INV-STG1 — the pipeline must read through the storage layer, never
+        # assume the key maps to an existing DATA_DIR file).
+        staged_is_tmp = storage.local_path(
+            staged_key, bucket=storage.Bucket.DOCUMENTS) is None
+        staged = await storage.ensure_local(
+            staged_key, bucket=storage.Bucket.DOCUMENTS)
+    try:
+        if staged is not None:
             await progress("extracting", 15, "מחלץ טקסט מהקובץ")
             try:
                 raw_text, page_count, page_offsets = await extractor.extract_text(str(staged))
@@ -237,6 +250,14 @@ async def ingest_document(
             await db.set_case_law_extraction_status(case_law_id, "failed")
             await progress("failed", 100, f"כשל בעיבוד: {e}")
             raise
+    finally:
+        # Drop the temp download (s3-only); on filesystem/dual ``staged`` is the
+        # canonical on-disk copy and must NOT be removed.
+        if staged_is_tmp and staged is not None:
+            try:
+                staged.unlink(missing_ok=True)
+            except OSError as e:  # noqa: BLE001 — temp cleanup, never fatal
+                logger.debug("could not remove temp staged file %s: %s", staged, e)
 
 
 async def _chunk_embed_store(case_law_id, text, page_offsets, page_count, progress) -> int:

mcp-server/tests/test_storage_staging.py

@@ -30,20 +30,24 @@ def test_stage_file_lands_under_datadir(_tmp_datadir):
     src.parent.mkdir(parents=True)
     src.write_bytes(b"%PDF-1.4 ...")
     root = _tmp_datadir / "precedent-library"
-    dest = run(ingest._stage_file(src, root, "court_ruling"))
+    # _stage_file returns the DATA_DIR-relative storage KEY (what the DB stores);
-    # dest is under the staging subdir, prefixed with a uuid, original suffix kept
+    # the caller resolves a local path via storage.ensure_local.
+    key = run(ingest._stage_file(src, root, "court_ruling"))
+    assert isinstance(key, str)
+    assert key.startswith("precedent-library/court_ruling/")
+    assert key.endswith(".pdf")
+    # under the filesystem backend the key maps to the exact legacy on-disk path
+    dest = _tmp_datadir / key
     assert dest.parent == root / "court_ruling"
     assert dest.exists()
     assert dest.read_bytes() == b"%PDF-1.4 ..."
-    assert dest.suffix == ".pdf"
-    # and the key is DATA_DIR-relative (what the DB column will store)
-    assert dest.relative_to(_tmp_datadir).as_posix().startswith("precedent-library/court_ruling/")
 
 
 def test_stage_file_default_subdir(_tmp_datadir):
     src = _tmp_datadir / "x.docx"
     src.write_bytes(b"doc")
-    dest = run(ingest._stage_file(src, _tmp_datadir / "digests", ""))
+    key = run(ingest._stage_file(src, _tmp_datadir / "digests", ""))
+    dest = _tmp_datadir / key
     assert dest.parent == _tmp_datadir / "digests" / "other"
     assert dest.exists()
 

mcp-server/tests/test_unified_ingest.py

@@ -172,3 +172,58 @@ def test_display_name_fallback_uses_canonical_id(patched, tmp_path):
         case_number="8046/24", text="t", chair_name="x", practice_area="betterment_levy"))
     kind, kw = patched["create"][0]
     assert kw["case_name"] == "8046/24", "missing case_name falls back to canonical id"
+
+
+def test_ingest_reads_via_ensure_local_when_no_disk_copy(patched, tmp_path, monkeypatch):
+    """Regression: under the s3-only backend the staged key has NO on-disk file,
+    so reading the DATA_DIR path directly 500'd ('Package not found at …').
+    ingest must resolve a readable local path via storage.ensure_local (a temp
+    download) and clean it up afterwards.
+    """
+    from legal_mcp.services import storage
+
+    store: dict[str, bytes] = {}
+
+    class _MemBackend(storage.StorageBackend):
+        """In-memory backend with NO local copy — mimics s3-only."""
+        name = "mem"
+
+        async def put_bytes(self, key, data, *, bucket=storage.Bucket.DOCUMENTS,
+                            content_type=None, metadata=None):
+            store[storage.normalize_key(key)] = bytes(data)
+            return f"mem://{key}"
+
+        async def put_file(self, src, key, *, bucket=storage.Bucket.DOCUMENTS,
+                          content_type=None, metadata=None):
+            store[storage.normalize_key(key)] = Path(src).read_bytes()
+            return f"mem://{key}"
+
+        async def get_bytes(self, key, *, bucket=storage.Bucket.DOCUMENTS):
+            return store[storage.normalize_key(key)]
+
+        async def exists(self, key, *, bucket=storage.Bucket.DOCUMENTS):
+            return storage.normalize_key(key) in store
+
+        def local_path(self, key, *, bucket=storage.Bucket.DOCUMENTS):
+            return None  # the s3-only condition: nothing on disk
+
+    monkeypatch.setattr(storage, "_singleton", _MemBackend())
+
+    # An extract_text that actually READS the path it is handed — proves ingest
+    # passes a real, readable local file rather than a phantom DATA_DIR path.
+    seen = []
+
+    async def _extract_reads(path):
+        seen.append(path)
+        assert Path(path).read_bytes() == b"%PDF-1.4 fake"
+        return ("full decision text", 1, [0])
+
+    monkeypatch.setattr(extractor, "extract_text", _extract_reads)
+
+    out = _run(precedent_library.ingest_precedent(
+        file_path=_make_pdf(tmp_path), citation="עע\"מ 1234/20",
+        practice_area="rishuy_uvniya", source_type="court_ruling",
+    ))
+    assert out["status"] == "completed"
+    # the temp download was cleaned up (no /tmp leak)
+    assert seen and not Path(seen[0]).exists()