Merge pull request '🐾 fix(api): cascade plain library delete; case-insensitive name conflict' (#9) from fix/library-api-delete-and-name-case into main

Reviewed-on: #9

docs/deploy.md

@@ -85,26 +85,21 @@ an explicit `when: mnemosyne_first_deploy` flag.
 
 ```bash
 # Apply Django ORM migrations (PostgreSQL schema)
-docker compose -f /srv/mnemosyne/docker-compose.yaml run --rm app migrate
+docker compose run --rm app migrate
 
 # Create Neo4j vector + full-text indexes and load library-type defaults
-docker compose -f /srv/mnemosyne/docker-compose.yaml \
-    run --rm app setup
+docker compose run --rm app setup
 
 # Seed the MCPSigningKey used to sign long-lived Pallas team JWTs.
 # --retire-other deactivates any previously-active key.  The hex
 # emitted to stdout is persisted in Mnemosyne's database and is
 # not re-injected from the vault — no operator action required
 # beyond running this command once per fresh deployment.
-docker compose -f /srv/mnemosyne/docker-compose.yaml \
-    run --rm app \
-    python manage.py seed_signing_key --kid daedalus-1 --retire-other
+docker compose run --rm app python manage.py seed_signing_key --kid daedalus-1 --retire-other
 
 # Create Django groups for SSO role mapping (View Only / Staff / SME / Admin).
 # Safe to re-run — idempotent.
-docker compose -f /srv/mnemosyne/docker-compose.yaml \
-    run --rm app \
-    python manage.py create_sso_groups
+docker compose run --rm app python manage.py create_sso_groups
 ```
 
 The `seed_signing_key` command prints the generated secret once to stdout — it

mnemosyne/library/api/serializers.py

@@ -37,6 +37,7 @@ class LibrarySerializer(serializers.Serializer):
         required=False, allow_blank=True, default=""
     )
     workspace_id = serializers.CharField(read_only=True)
+    managed_by = serializers.CharField(read_only=True)
     created_at = serializers.DateTimeField(read_only=True)
 
 
@@ -193,6 +194,7 @@ class WorkspaceStatusSerializer(serializers.Serializer):
     name = serializers.CharField()
     library_type = serializers.CharField()
     description = serializers.CharField(allow_blank=True)
+    managed_by = serializers.CharField(allow_null=True, required=False)
     item_count = serializers.IntegerField()
     chunk_count = serializers.IntegerField()
     created_at = serializers.DateTimeField()

mnemosyne/library/api/views.py

@@ -10,6 +10,7 @@ import os
 
 from django.core.files.base import ContentFile
 from django.core.files.storage import default_storage
+from neomodel.exceptions import UniqueProperty
 from rest_framework import status
 from rest_framework.decorators import api_view, parser_classes, permission_classes
 from rest_framework.parsers import FormParser, JSONParser, MultiPartParser
@@ -17,6 +18,8 @@ from rest_framework.permissions import IsAuthenticated
 from rest_framework.response import Response
 
 from library.content_types import get_library_type_config
+from library.services.library_delete import delete_library_cascade
+from mcp_server.drf_auth import request_token_label
 
 from .serializers import (
     CollectionSerializer,
@@ -50,7 +53,7 @@ def library_list_create(request):
     a per-library ``item_count``. Off by default because the count is a
     Cypher aggregate; on for the Daedalus-side registry poll.
     """
-    from library.models import Library
+    from library.models import Library, find_library_by_name_ci
 
     if request.method == "GET":
         include_workspace = request.GET.get("include_workspace", "true").lower() != "false"
@@ -84,6 +87,28 @@ def library_list_create(request):
     serializer.is_valid(raise_exception=True)
     data = serializer.validated_data
 
+    # Library names are unique. The Neo4j index is case-sensitive, but
+    # clients (Spelunker, humans) treat names case-insensitively, so the
+    # create-time check is case-insensitive too — otherwise "amazon connect"
+    # silently creates a near-duplicate of "Amazon Connect". Reject with a
+    # clean 409 (uid + managed_by included so the caller can say who owns
+    # the name) instead of letting the unique-index save raise a 500.
+    existing = find_library_by_name_ci(data["name"])
+    if existing is not None:
+        logger.warning(
+            "library_create name_conflict name=%s existing_uid=%s caller=%s",
+            data["name"], existing.uid, request.user.username,
+        )
+        return Response(
+            {
+                "detail": f"A library named '{existing.name}' already exists.",
+                "code": "name_conflict",
+                "uid": existing.uid,
+                "managed_by": existing.managed_by_display or None,
+            },
+            status=status.HTTP_409_CONFLICT,
+        )
+
     # Populate defaults from content-type config if not provided
     library_type = data["library_type"]
     defaults = get_library_type_config(library_type)
@@ -92,6 +117,7 @@ def library_list_create(request):
         name=data["name"],
         library_type=library_type,
         description=data.get("description", ""),
+        managed_by=request_token_label(request),
         chunking_config=data.get("chunking_config") or defaults["chunking_config"],
         embedding_instruction=(
             data.get("embedding_instruction") or defaults["embedding_instruction"]
@@ -103,7 +129,22 @@ def library_list_create(request):
             data.get("llm_context_prompt") or defaults["llm_context_prompt"]
         ),
     )
-    lib.save()
+    try:
+        lib.save()
+    except UniqueProperty:
+        # Race between the pre-check and save — an exact-name twin landed
+        # in between. Same 409 shape, minus the loser's uid/manager.
+        logger.warning(
+            "library_create name_conflict (save race) name=%s caller=%s",
+            data["name"], request.user.username,
+        )
+        return Response(
+            {
+                "detail": f"A library named '{data['name']}' already exists.",
+                "code": "name_conflict",
+            },
+            status=status.HTTP_409_CONFLICT,
+        )
     return Response(LibrarySerializer(lib).data, status=status.HTTP_201_CREATED)
 
 
@@ -141,8 +182,16 @@ def library_detail(request, uid):
         lib.save()
         return Response(LibrarySerializer(lib).data)
 
-    # DELETE
-    lib.delete()
+    # DELETE — use the shared cascade so child nodes (Collections/Items/
+    # Chunks/Images) and orphan Concepts are removed too; a bare
+    # lib.delete() would leak them all.
+    result = delete_library_cascade(lib)
+    logger.info(
+        "Library deleted via API library_uid=%s name=%s items=%d "
+        "orphans_deleted=%d caller=%s",
+        result["library_uid"], result["name"], result["item_count"],
+        result["orphans_deleted"], request.user.username,
+    )
     return Response(status=status.HTTP_204_NO_CONTENT)
 
 

mnemosyne/library/api/workspaces.py

@@ -25,6 +25,7 @@ from rest_framework.response import Response
 
 from library.content_types import get_library_type_config
 from library.services.library_delete import delete_library_cascade
+from mcp_server.drf_auth import request_token_label
 
 from .serializers import WorkspaceCreateSerializer, WorkspaceStatusSerializer
 
@@ -49,6 +50,7 @@ def _serialize_workspace(lib):
         "name": lib.name,
         "library_type": lib.library_type,
         "description": lib.description or "",
+        "managed_by": lib.managed_by,
         "item_count": item_count,
         "chunk_count": chunk_count,
         "created_at": lib.created_at,
@@ -65,7 +67,7 @@ def workspace_create(request):
     workspace (200) — not an error. The library_type is frozen at first
     create; subsequent calls are not allowed to change it.
     """
-    from library.models import Library
+    from library.models import Library, find_library_by_name_ci
 
     serializer = WorkspaceCreateSerializer(data=request.data)
     serializer.is_valid(raise_exception=True)
@@ -104,6 +106,18 @@ def workspace_create(request):
                 },
                 status=status.HTTP_409_CONFLICT,
             )
+        # Lazy backfill: pre-managed_by libraries pick up the label from
+        # the first idempotent re-POST. Null-only — an already-stamped
+        # library never changes manager.
+        if not existing.managed_by:
+            label = request_token_label(request)
+            if label:
+                existing.managed_by = label
+                existing.save()
+                logger.info(
+                    "Backfilled managed_by=%s workspace_id=%s library_uid=%s",
+                    label, existing.workspace_id, existing.uid,
+                )
         logger.info(
             "Workspace already exists workspace_id=%s library_uid=%s",
             data["workspace_id"], existing.uid,
@@ -113,6 +127,28 @@ def workspace_create(request):
             status=status.HTTP_200_OK,
         )
 
+    # New workspace: reject a name already taken by any other library,
+    # case-insensitively — the Neo4j index is case-sensitive, so without
+    # this check "amazon connect" would silently coexist with an existing
+    # "Amazon Connect" and confuse every name-matching client.
+    name_taken = find_library_by_name_ci(data["name"])
+    if name_taken is not None:
+        logger.warning(
+            "workspace_create name_conflict workspace_id=%s name=%s "
+            "existing_uid=%s",
+            data["workspace_id"], data["name"], name_taken.uid,
+        )
+        return Response(
+            {
+                "detail": (
+                    f"A library named '{name_taken.name}' already exists in "
+                    "Mnemosyne."
+                ),
+                "code": "name_conflict",
+            },
+            status=status.HTTP_409_CONFLICT,
+        )
+
     defaults = get_library_type_config(data["library_type"])
     lib = Library(
         name=data["name"],
@@ -120,6 +156,7 @@ def workspace_create(request):
         description=data.get("description", ""),
         workspace_id=data["workspace_id"],
         owner_username=request.user.username,
+        managed_by=request_token_label(request),
         chunking_config=defaults["chunking_config"],
         embedding_instruction=defaults["embedding_instruction"],
         reranker_instruction=defaults["reranker_instruction"],

mnemosyne/library/management/commands/backfill_managed_by.py

@@ -0,0 +1,111 @@
+"""One-off backfill of ``Library.managed_by`` for pre-existing libraries.
+
+``managed_by`` is stamped from the creating API token's name, so
+libraries created before the property existed have it null. This
+command labels them:
+
+* Workspace-scoped libraries get :func:`infer_legacy_manager`'s answer —
+  ``Kairos`` for ``kairos-mail-*`` workspace ids, ``Daedalus`` otherwise.
+* Global libraries that Spelunker ingested into (any ``IngestJob`` with
+  ``source="spelunker"``) get the Spelunker label.
+* Everything else stays null (hand-made in the web UI).
+
+Label flags let the operator match the *actual* production token names
+so backfilled rows render identically to newly stamped ones.
+
+Idempotent: only null ``managed_by`` is ever written, and the lazy fill
+in ``workspace_create`` is also null-only, so re-runs and later API
+traffic never overwrite these labels.
+"""
+
+from __future__ import annotations
+
+import logging
+
+from django.core.management.base import BaseCommand, CommandError
+
+from library.models import IngestJob
+
+logger = logging.getLogger(__name__)
+
+
+class Command(BaseCommand):
+    help = (
+        "Backfill Library.managed_by for libraries created before the "
+        "property existed. Workspace libraries are labelled by inference "
+        "(kairos-mail-* → Kairos, else Daedalus); global libraries with "
+        "Spelunker ingest jobs get the Spelunker label."
+    )
+
+    def add_arguments(self, parser):
+        parser.add_argument(
+            "--daedalus-label", default="Daedalus",
+            help="Label for non-Kairos workspace libraries (default: Daedalus).",
+        )
+        parser.add_argument(
+            "--kairos-label", default="Kairos",
+            help="Label for kairos-mail-* workspace libraries (default: Kairos).",
+        )
+        parser.add_argument(
+            "--spelunker-label", default="Spelunker",
+            help="Label for global libraries with Spelunker ingest jobs "
+                 "(default: Spelunker).",
+        )
+        parser.add_argument(
+            "--dry-run", action="store_true",
+            help="Report what would be labelled, don't persist.",
+        )
+
+    def handle(self, *args, **options):
+        try:
+            from library.models import Library, infer_legacy_manager
+        except Exception as exc:  # pragma: no cover
+            raise CommandError(
+                f"Could not import library.models.Library (Neo4j unreachable?): {exc}"
+            ) from exc
+
+        overrides = {
+            "Daedalus": options["daedalus_label"],
+            "Kairos": options["kairos_label"],
+        }
+
+        spelunker_uids = set(
+            IngestJob.objects
+            .filter(source="spelunker")
+            .values_list("library_uid", flat=True)
+            .distinct()
+        )
+
+        candidates = list(Library.nodes.filter(managed_by__isnull=True))
+
+        to_label = []
+        for lib in candidates:
+            label = infer_legacy_manager(lib.workspace_id)
+            if label:
+                label = overrides[label]
+            elif lib.uid in spelunker_uids:
+                label = options["spelunker_label"]
+            if label:
+                to_label.append((lib, label))
+
+        self.stdout.write(f"Libraries with null managed_by: {len(candidates)}")
+        self.stdout.write(
+            self.style.SUCCESS(f"Will label: {len(to_label)} "
+                               f"(unmanaged, left null: {len(candidates) - len(to_label)})")
+        )
+
+        for lib, label in to_label:
+            self.stdout.write(f"  {lib.uid}  {lib.name!r} → {label}")
+
+        if options["dry_run"]:
+            self.stdout.write(self.style.WARNING("--dry-run: nothing written."))
+            return
+
+        for lib, label in to_label:
+            lib.managed_by = label
+            lib.save()
+            logger.info(
+                "backfill_managed_by uid=%s name=%s label=%s",
+                lib.uid, lib.name, label,
+            )
+        self.stdout.write(self.style.SUCCESS("Done."))

mnemosyne/library/models.py

@@ -51,6 +51,35 @@ class NearbyImageRel(StructuredRel):
 # --- Node models ---
 
 
+def infer_legacy_manager(workspace_id):
+    """Manager label for pre-``managed_by`` workspace libraries, else None.
+
+    Kairos mail workspaces are recognisable by their deterministic
+    ``kairos-mail-`` id prefix; every other workspace id is a Daedalus
+    workspace UUID.
+    """
+    if not workspace_id:
+        return None
+    return "Kairos" if workspace_id.startswith("kairos-mail-") else "Daedalus"
+
+
+def find_library_by_name_ci(name):
+    """Case-insensitively find a Library by name, or None.
+
+    Parameterised Cypher rather than neomodel's ``iexact``, which embeds
+    the value in a regex and so breaks on names containing regex
+    metacharacters (e.g. "C++ Notes").
+    """
+    from neomodel import db
+
+    rows, _ = db.cypher_query(
+        "MATCH (l:Library) WHERE toLower(l.name) = toLower($name) "
+        "RETURN l LIMIT 1",
+        {"name": name},
+    )
+    return Library.inflate(rows[0][0]) if rows else None
+
+
 class Library(StructuredNode):
     """
     Top-level container representing a content library.
@@ -63,6 +92,11 @@ class Library(StructuredNode):
     across the whole instance) or *workspace-scoped* (workspace_id set —
     visible only to agents inside that Daedalus workspace). Scoping is
     enforced structurally by every search query.
+
+    Independently of scoping, a library may be *app-managed*
+    (``managed_by`` set — created through the API by an external app such
+    as Daedalus, Kairos, or Spelunker, which owns its content lifecycle)
+    or unmanaged (created by hand in the web UI).
     """
 
     uid = UniqueIdProperty()
@@ -93,6 +127,12 @@ class Library(StructuredNode):
     # this user. Null for global libraries.
     owner_username = StringProperty(required=False, index=True)
 
+    # Name of the API token that created this library ("Daedalus",
+    # "Kairos", "Spelunker", ...). Null for libraries created in the
+    # web UI. Stamped at create time only — token rotation or edits by
+    # another token never change it.
+    managed_by = StringProperty(required=False, index=True)
+
     # Content-type configuration
     chunking_config = JSONProperty(default={})
     embedding_instruction = StringProperty(default="")
@@ -104,6 +144,16 @@ class Library(StructuredNode):
     # Relationships
     collections = RelationshipTo("Collection", "CONTAINS")
 
+    @property
+    def managed_by_display(self):
+        """Managing-app label for display; empty string when unmanaged.
+
+        Falls back to inference for workspace libraries created before
+        ``managed_by`` existed, so rendering is identical before and
+        after the backfill command runs.
+        """
+        return self.managed_by or infer_legacy_manager(self.workspace_id) or ""
+
     def __str__(self):
         return f"{self.name} ({self.library_type})"
 

mnemosyne/library/services/parsers.py

@@ -22,6 +22,17 @@ from .text_utils import remove_excessive_whitespace, sanitize_text
 
 logger = logging.getLogger(__name__)
 
+
+class UnsupportedFileTypeError(ValueError):
+    """Raised when a file's type cannot be parsed.
+
+    Deterministic and input-driven — re-parsing identical bytes can never
+    succeed — so ingest must treat it as a terminal failure and never retry.
+    Subclasses ``ValueError`` so existing ``except ValueError`` callers still
+    catch it.
+    """
+
+
 # File extensions supported by PyMuPDF
 PYMUPDF_EXTENSIONS = {
     "pdf", "epub", "xps", "mobi", "fb2", "cbz", "svg",
@@ -31,8 +42,11 @@ PYMUPDF_EXTENSIONS = {
 # Plain text extensions — read directly, no PyMuPDF needed
 PLAINTEXT_EXTENSIONS = {"txt", "md", "csv", "tsv", "log", "json", "yaml", "yml", "xml"}
 
-# Image extensions — store as Image nodes directly
-IMAGE_EXTENSIONS = {"jpg", "jpeg", "png", "gif", "bmp", "tiff", "tif", "webp", "svg"}
+# Image extensions — store as Image nodes directly.
+# SVG is deliberately absent: it is vector XML that Pillow cannot decode and
+# that the vision stage cannot send as a data URI, so it gets rasterized by
+# _parse_svg_file instead.
+IMAGE_EXTENSIONS = {"jpg", "jpeg", "png", "gif", "bmp", "tiff", "tif", "webp"}
 
 # Minimum image dimensions to extract (skip tiny icons/bullets)
 MIN_IMAGE_WIDTH = 50
@@ -85,7 +99,7 @@ class DocumentParser:
         :param file_path: Path to the document file.
         :param file_type: File extension (without dot), e.g. 'pdf', 'epub'.
         :returns: ParseResult with text blocks, images, and metadata.
-        :raises ValueError: If the file type is not supported.
+        :raises UnsupportedFileTypeError: If the file type is not supported.
         """
         file_type = file_type.lower().lstrip(".")
 
@@ -98,6 +112,12 @@ class DocumentParser:
         if file_type in PLAINTEXT_EXTENSIONS:
             return self._parse_plaintext(file_path, file_type)
 
+        # Checked before PYMUPDF_EXTENSIONS: PyMuPDF can open an SVG, but
+        # rendering a multi-page Write note as one document yields a blank or
+        # illegible image (see svg_raster), so it needs page-aware handling.
+        if file_type == "svg":
+            return self._parse_svg_file(file_path, file_type)
+
         if file_type in IMAGE_EXTENSIONS:
             return self._parse_image_file(file_path, file_type)
 
@@ -108,7 +128,7 @@ class DocumentParser:
         if file_type in ("html", "htm"):
             return self._parse_with_pymupdf(file_path, file_type)
 
-        raise ValueError(
+        raise UnsupportedFileTypeError(
             f"Unsupported file type '{file_type}'. "
             f"Supported: {sorted(PYMUPDF_EXTENSIONS | PLAINTEXT_EXTENSIONS | IMAGE_EXTENSIONS)}"
         )
@@ -309,6 +329,61 @@ class DocumentParser:
                 file_type=file_type,
             )
 
+    def _parse_svg_file(self, file_path: str, file_type: str) -> ParseResult:
+        """
+        Rasterize an SVG into one ExtractedImage per page.
+
+        SVG is vector XML: Pillow cannot decode it and the vision stage cannot
+        put it in a data URI, so it is rendered to PNG here. Multi-page Write
+        notes become one image per page so each page reaches vision/OCR at a
+        legible size.
+
+        :param file_path: Path to the SVG file.
+        :param file_type: Normalized file extension ("svg").
+        :returns: ParseResult with one image per rendered page.
+        """
+        with DOCUMENT_PARSE_DURATION.labels(file_type=file_type).time():
+            try:
+                from library.services.svg_raster import render_svg_pages
+
+                with open(file_path, "rb") as f:
+                    data = f.read()
+
+                pages = render_svg_pages(data)
+            except Exception as exc:
+                DOCUMENTS_PARSED_TOTAL.labels(file_type=file_type, status="error").inc()
+                logger.error("Failed to rasterize SVG file_type=%s: %s", file_type, exc)
+                raise
+
+            images = [
+                ExtractedImage(
+                    data=png,
+                    ext="png",
+                    width=width,
+                    height=height,
+                    source_page=index,
+                    source_index=0,
+                )
+                for index, (png, width, height) in enumerate(pages)
+            ]
+
+            DOCUMENTS_PARSED_TOTAL.labels(file_type=file_type, status="success").inc()
+            IMAGES_EXTRACTED_TOTAL.labels(file_type=file_type).inc(len(images))
+
+            logger.info(
+                "Parsed SVG file_type=%s pages=%d bytes=%d",
+                file_type,
+                len(images),
+                len(data),
+            )
+
+            return ParseResult(
+                text_blocks=[],
+                images=images,
+                metadata={"page_count": len(images)},
+                file_type=file_type,
+            )
+
     def _parse_image_file(self, file_path: str, file_type: str) -> ParseResult:
         """
         Handle a standalone image file — store as a single ExtractedImage.

mnemosyne/library/services/svg_raster.py

@@ -0,0 +1,263 @@
+"""SVG rasterization for the ingest pipeline.
+
+The vision stage sends each extracted image to a vision LLM as a ``data:`` URI,
+which cannot carry ``image/svg+xml`` — so an SVG has to become raster before it
+can be described, OCR'd, or embedded. PyMuPDF (already a dependency for PDF
+parsing) renders SVG natively, so this needs no cairo/rsvg system libraries.
+
+The non-obvious part is page splitting. Handwritten notes from the Write
+(Stylus Labs) app are a single SVG document holding one or more
+``<svg class="write-page">`` children stacked vertically via x/y offsets. Two
+root formats exist in the wild and *both* must be split per page:
+
+- Older files carry no width/height on the root ``<svg>``. Rendering the
+  document as-is makes PyMuPDF fall back to US-Letter and emit the top-left
+  corner only — ruled lines and no handwriting, in a perfectly valid PNG.
+- Newer files (Write commit eeab021) do carry root width/height, spanning the
+  full stacked extent. Rendering those as-is yields one very tall strip; capped
+  to a sane longest edge, a 5-page note squashes to ~209px wide, well past
+  illegible.
+
+Write never rewrites existing files, so the old format is permanent rather than
+a migration window. Page geometry lives on the ``write-page`` element and is
+identical across both formats, so splitting ignores the root dimensions
+entirely and needs no format detection.
+
+.. note::
+   Daedalus carries a twin of this module at
+   ``backend/daedalus/extraction/svg.py``, which returns base64 for direct
+   chat attachment. The two are deliberately duplicated rather than shared —
+   the repos ship no common package — so fixes belong in both.
+"""
+
+from __future__ import annotations
+
+import copy
+import logging
+import re
+
+from lxml import etree
+
+logger = logging.getLogger(__name__)
+
+SVG_NS = "http://www.w3.org/2000/svg"
+XLINK_NS = "http://www.w3.org/1999/xlink"
+
+_SVG = f"{{{SVG_NS}}}"
+_PAGE_CLASS = "write-page"
+
+# Write documents render on a grey backdrop; pages themselves are transparent,
+# so each rendered page needs an explicit white underlay or strokes land on
+# black in the flattened PNG.
+_PAGE_BACKGROUND = "#ffffff"
+
+#: Longest edge of a rendered page, in pixels. Enough for a vision model to
+#: read handwriting without spending tokens on unusable resolution.
+DEFAULT_TARGET_PX = 1568
+
+#: Cap on pages rendered from one document.
+DEFAULT_MAX_PAGES = 20
+
+# Unquoted attribute value in the root tag, e.g. Write's `width=auto`. Matches
+# only bare alphabetic values so numeric or already-quoted attributes are left
+# alone.
+_UNQUOTED_ATTR = re.compile(rb"(\s[-\w:]+)=([A-Za-z][-\w]*)(?=[\s>])")
+_ROOT_TAG = re.compile(rb"<svg[^>]*>")
+
+
+class SvgRenderError(Exception):
+    """Raised when an SVG cannot be parsed or contains no renderable page."""
+
+
+def _repair_root_tag(data: bytes) -> bytes:
+    """Quote unquoted attribute values in the root ``<svg>`` tag.
+
+    Some Write files emit ``width=auto height=auto``, which is not valid XML.
+    A malformation *inside* the root tag defeats ``recover=True`` differently
+    from one in the body: rather than dropping a subtree, libxml2 abandons the
+    whole document and yields a bare root, so every page becomes invisible.
+    Quoting the values first recovers the full tree.
+    """
+    match = _ROOT_TAG.search(data)
+    if not match:
+        return data
+    repaired = _UNQUOTED_ATTR.sub(rb'\1="\2"', match.group(0))
+    if repaired == match.group(0):
+        return data
+    return data[: match.start()] + repaired + data[match.end() :]
+
+
+def _parser() -> etree.XMLParser:
+    """Build the hardened parser used for all untrusted SVG input.
+
+    ``resolve_entities=False`` blocks XXE — ingest content is untrusted.
+    ``huge_tree`` is required because handwriting path data runs to megabytes.
+    ``recover`` salvages the handful of Write files that emit unescaped
+    attribute content and are not well-formed XML.
+    """
+    return etree.XMLParser(
+        huge_tree=True,
+        resolve_entities=False,
+        no_network=True,
+        recover=True,
+    )
+
+
+def _dimension(element: etree._Element, name: str) -> float | None:
+    """Read a CSS-pixel dimension attribute, tolerating a ``px`` suffix."""
+    raw = element.get(name)
+    if not raw:
+        return None
+    try:
+        return float(raw.strip().removesuffix("px"))
+    except ValueError:
+        return None
+
+
+def _viewbox_size(element: etree._Element) -> tuple[float, float] | None:
+    """Derive width/height from a viewBox extent."""
+    raw = element.get("viewBox")
+    if not raw:
+        return None
+    parts = raw.replace(",", " ").split()
+    if len(parts) != 4:
+        return None
+    try:
+        width, height = float(parts[2]), float(parts[3])
+    except ValueError:
+        return None
+    return (width, height) if width > 0 and height > 0 else None
+
+
+def _element_size(element: etree._Element) -> tuple[float, float] | None:
+    """Resolve an element's rendered size from width/height, else viewBox."""
+    width = _dimension(element, "width")
+    height = _dimension(element, "height")
+    if width and height:
+        return width, height
+    return _viewbox_size(element)
+
+
+def _standalone_page(
+    page: etree._Element,
+    defs: etree._Element | None,
+    width: float,
+    height: float,
+) -> bytes:
+    """Wrap one ``write-page`` element as its own renderable SVG document.
+
+    The page is repositioned to the origin (its x/y place it within the stacked
+    parent) and given an explicit viewBox so the renderer has an unambiguous
+    size. ``defs`` is copied in because pen and ruling definitions live on the
+    root and are referenced by page content.
+    """
+    root = etree.Element(f"{_SVG}svg", nsmap={None: SVG_NS, "xlink": XLINK_NS})
+    root.set("width", str(width))
+    root.set("height", str(height))
+    root.set("viewBox", f"0 0 {width} {height}")
+
+    background = etree.SubElement(root, f"{_SVG}rect")
+    background.set("width", "100%")
+    background.set("height", "100%")
+    background.set("fill", _PAGE_BACKGROUND)
+
+    if defs is not None:
+        root.append(copy.deepcopy(defs))
+
+    element = copy.deepcopy(page)
+    for positional in ("x", "y"):
+        element.attrib.pop(positional, None)
+    element.set("viewBox", f"0 0 {width} {height}")
+    root.append(element)
+
+    return etree.tostring(root)
+
+
+def split_svg_pages(data: bytes) -> list[bytes]:
+    """Split an SVG into one standalone document per renderable page.
+
+    Write multi-page notes yield one document per ``write-page`` child. Any
+    other SVG (a diagram, a logo) yields a single document — the input itself,
+    which the renderer sizes from its own width/height or viewBox.
+
+    :param data: Raw SVG bytes.
+    :returns: One or more standalone SVG documents, in page order.
+    :raises SvgRenderError: If the SVG cannot be parsed or has no usable size.
+    """
+    try:
+        root = etree.fromstring(_repair_root_tag(data), _parser())
+    except etree.XMLSyntaxError as exc:
+        raise SvgRenderError(f"Could not parse SVG: {exc}") from exc
+
+    if root is None:
+        raise SvgRenderError("Could not parse SVG: no root element.")
+
+    defs = root.find(f"{_SVG}defs")
+
+    pages: list[bytes] = []
+    for element in root.iter(f"{_SVG}svg"):
+        if element is root:
+            continue
+        if _PAGE_CLASS not in (element.get("class") or "").split():
+            continue
+        size = _element_size(element)
+        if not size:
+            continue
+        pages.append(_standalone_page(element, defs, *size))
+
+    if pages:
+        return pages
+
+    # Generic SVG: render as-is. It must still be sizeable, or the renderer
+    # would silently substitute a default page box.
+    if not _element_size(root):
+        raise SvgRenderError(
+            "SVG has no width/height or viewBox, so its size is undefined."
+        )
+    return [data]
+
+
+def render_svg_pages(
+    data: bytes,
+    max_pages: int = DEFAULT_MAX_PAGES,
+    target_px: int = DEFAULT_TARGET_PX,
+) -> list[tuple[bytes, int, int]]:
+    """Rasterize an SVG to one PNG per page.
+
+    :param data: Raw SVG bytes.
+    :param max_pages: Cap on pages rendered.
+    :param target_px: Longest edge of each rendered page, in pixels.
+    :returns: One ``(png_bytes, width, height)`` tuple per page, in page order.
+    :raises SvgRenderError: If nothing could be rendered.
+    """
+    import fitz
+
+    pages = split_svg_pages(data)
+
+    rendered: list[tuple[bytes, int, int]] = []
+    for index, page_svg in enumerate(pages[:max_pages]):
+        try:
+            with fitz.open(stream=page_svg, filetype="svg") as document:
+                page = document[0]
+                longest = max(page.rect.width, page.rect.height)
+                zoom = target_px / longest if longest else 1.0
+                pixmap = page.get_pixmap(matrix=fitz.Matrix(zoom, zoom))
+                rendered.append(
+                    (pixmap.tobytes("png"), pixmap.width, pixmap.height)
+                )
+        except Exception as exc:
+            raise SvgRenderError(
+                f"Could not render SVG page {index + 1}: {exc}"
+            ) from exc
+
+    if not rendered:
+        raise SvgRenderError("No renderable pages in SVG.")
+
+    logger.info(
+        "Rasterized SVG pages_rendered=%d total_pages=%d target_px=%d",
+        len(rendered),
+        len(pages),
+        target_px,
+    )
+
+    return rendered

mnemosyne/library/tasks.py

@@ -347,6 +347,7 @@ def ingest_from_daedalus(self, job_id: str):
     from datetime import datetime, timezone
 
     from library.models import IngestJob, Item, Library
+    from library.services.parsers import UnsupportedFileTypeError
     from library.services.source_s3 import (
         copy_into_mnemosyne,
         fetch_from_source,
@@ -465,6 +466,24 @@ def ingest_from_daedalus(self, job_id: str):
             **result,
         }
 
+    except UnsupportedFileTypeError as exc:
+        # Deterministic, input-driven — re-parsing identical bytes can never
+        # succeed. Terminal client-data failure, never retried, and logged at
+        # WARNING (not ERROR) because an unparseable input is not a server fault.
+        logger.warning(
+            "Task ingest_from_daedalus rejected job_id=%s: %s", job_id, exc,
+        )
+        job.status = "failed"
+        job.error = str(exc)
+        job.completed_at = datetime.now(timezone.utc)
+        job.save(update_fields=["status", "error", "completed_at"])
+        return {
+            "success": False,
+            "job_id": job_id,
+            "error": str(exc),
+            "reason": "unsupported_file_type",
+        }
+
     except Exception as exc:
         logger.error(
             "Task ingest_from_daedalus failed job_id=%s: %s",

mnemosyne/library/templates/library/library_confirm_delete.html

@@ -12,15 +12,22 @@
     <div class="alert alert-warning mb-6">
         <span>Are you sure you want to delete <strong>{{ library.name }}</strong>? This action cannot be undone.</span>
     </div>
-    {% if library.workspace_id %}
+    {% if library.managed_by_display %}
     <div class="alert alert-error mb-6">
         <span>
-            <strong>This Library is managed by Daedalus</strong>
-            (workspace <code>{{ library.workspace_id }}</code>).
+            <strong>This Library is managed by {{ library.managed_by_display }}</strong>{% if library.workspace_id %}
+            (workspace <code>{{ library.workspace_id }}</code>){% endif %}.
+            {% if library.workspace_id %}
             Deleting it here removes its embedded content from Mnemosyne, but the
-            source files still live in Daedalus — it will be <strong>recreated and
-            re-embedded on the next Daedalus sync</strong>. Use this to clear an
-            orphaned Library that is blocking workspace re-registration.
+            source files still live in {{ library.managed_by_display }} — it will
+            be <strong>recreated and re-embedded on the next sync</strong>. Use
+            this to clear an orphaned Library that is blocking workspace
+            re-registration.
+            {% else %}
+            Deleting it here removes its embedded content from Mnemosyne;
+            {{ library.managed_by_display }} may recreate and re-embed it on its
+            next sync.
+            {% endif %}
         </span>
     </div>
     {% endif %}

mnemosyne/library/templates/library/library_detail.html

@@ -12,10 +12,10 @@
         <h1 class="text-3xl font-bold">{{ library.name }}</h1>
         <div class="flex flex-wrap gap-2 mt-2">
             <div class="badge badge-primary">{{ library.library_type }}</div>
-            {% if library.workspace_id %}
+            {% if library.managed_by_display %}
             <div class="badge badge-warning gap-1"
-                 title="Workspace {{ library.workspace_id }}">
-                Daedalus workspace
+                 {% if library.workspace_id %}title="Workspace {{ library.workspace_id }}"{% endif %}>
+                Managed by {{ library.managed_by_display }}
             </div>
             {% endif %}
         </div>
@@ -29,18 +29,25 @@
     </div>
 </div>
 
-{% if library.workspace_id %}
+{% if library.managed_by_display %}
 <div class="alert alert-warning mb-6">
     <div>
-        <div class="font-semibold">Managed by Daedalus</div>
+        <div class="font-semibold">Managed by {{ library.managed_by_display }}</div>
         <div class="text-sm opacity-80">
-            This library was created for Daedalus workspace
+            {% if library.workspace_id %}
+            This library was created for workspace
             <code class="font-mono">{{ library.workspace_id }}</code>.
-            Normally you manage it from Daedalus. Deleting it here removes its
-            embedded content from Mnemosyne, but the source files still live in
-            Daedalus — it will be recreated and re-embedded on the next sync.
+            Normally you manage it from {{ library.managed_by_display }}.
+            Deleting it here removes its embedded content from Mnemosyne, but
+            the source files still live in {{ library.managed_by_display }} —
+            it will be recreated and re-embedded on the next sync.
             Use Delete to clear an orphaned library that is blocking workspace
             re-registration.
+            {% else %}
+            Content in this library is pushed by
+            {{ library.managed_by_display }}. Edits made here may be
+            overwritten or re-created on its next sync.
+            {% endif %}
         </div>
     </div>
 </div>

mnemosyne/library/templates/library/library_list.html

@@ -19,9 +19,9 @@
     <div class="form-control">
         <label class="label"><span class="label-text">Scope</span></label>
         <select name="scope" class="select select-bordered select-sm">
-            <option value="all"      {% if scope == "all" %}selected{% endif %}>All libraries</option>
-            <option value="global"   {% if scope == "global" %}selected{% endif %}>Global only</option>
-            <option value="daedalus" {% if scope == "daedalus" %}selected{% endif %}>Daedalus workspaces only</option>
+            <option value="all"       {% if scope == "all" %}selected{% endif %}>All libraries</option>
+            <option value="unmanaged" {% if scope == "unmanaged" %}selected{% endif %}>Unmanaged only</option>
+            <option value="managed"   {% if scope == "managed" %}selected{% endif %}>App-managed only</option>
         </select>
     </div>
     <button type="submit" class="btn btn-sm btn-outline">Filter</button>
@@ -45,9 +45,9 @@
             </h2>
             <div class="flex flex-wrap gap-1">
                 <div class="badge badge-outline">{{ lib.library_type }}</div>
-                {% if lib.workspace_id %}
-                <div class="badge badge-warning gap-1" title="Managed by Daedalus workspace {{ lib.workspace_id }} — do not delete from Mnemosyne.">
-                    Daedalus workspace
+                {% if lib.managed_by_display %}
+                <div class="badge badge-warning gap-1" title="Managed by {{ lib.managed_by_display }}{% if lib.workspace_id %} (workspace {{ lib.workspace_id }}){% endif %} — do not delete from Mnemosyne.">
+                    Managed by {{ lib.managed_by_display }}
                 </div>
                 {% endif %}
             </div>

mnemosyne/library/tests/test_library_api.py

@@ -0,0 +1,70 @@
+"""Tests for the plain library REST endpoints beyond create.
+
+Currently covers the DELETE cascade: ``DELETE /library/api/libraries/{uid}/``
+must go through ``delete_library_cascade`` (shared with the HTML and
+workspace delete paths) — a bare ``lib.delete()`` leaks Collections, Items,
+Chunks, and Images and skips orphan-Concept GC. Neo4j is stubbed via
+``sys.modules``, same style as ``test_managed_by.py``.
+"""
+
+from __future__ import annotations
+
+from types import SimpleNamespace
+from unittest.mock import MagicMock, patch
+
+from django.contrib.auth import get_user_model
+from django.test import TestCase
+from rest_framework.test import APIClient
+
+User = get_user_model()
+
+
+class LibraryApiDeleteTests(TestCase):
+    """DELETE on the plain library endpoint cascades."""
+
+    def setUp(self):
+        self.user = User.objects.create_user(username="op", password="pw")
+        self.client = APIClient()
+        self.client.force_authenticate(user=self.user)
+
+    def _fake_models_module(self, lib):
+        fake_nodes = MagicMock()
+        if lib is None:
+            class DoesNotExist(Exception):
+                pass
+
+            fake_library = SimpleNamespace(nodes=fake_nodes, DoesNotExist=DoesNotExist)
+            fake_nodes.get.side_effect = DoesNotExist()
+        else:
+            fake_library = SimpleNamespace(nodes=fake_nodes, DoesNotExist=Exception)
+            fake_nodes.get.return_value = lib
+        return SimpleNamespace(Library=fake_library)
+
+    def test_delete_uses_shared_cascade(self):
+        lib = SimpleNamespace(uid="lib-1", name="Docs")
+        cascade_result = {
+            "library_uid": "lib-1",
+            "name": "Docs",
+            "item_count": 3,
+            "item_s3_keys": [],
+            "orphans_deleted": 1,
+        }
+        with patch.dict(
+            "sys.modules", {"library.models": self._fake_models_module(lib)}
+        ), patch(
+            "library.api.views.delete_library_cascade",
+            return_value=cascade_result,
+        ) as mock_cascade:
+            response = self.client.delete("/library/api/libraries/lib-1/")
+
+        self.assertEqual(response.status_code, 204)
+        mock_cascade.assert_called_once_with(lib)
+
+    def test_delete_missing_library_returns_404(self):
+        with patch.dict(
+            "sys.modules", {"library.models": self._fake_models_module(None)}
+        ), patch("library.api.views.delete_library_cascade") as mock_cascade:
+            response = self.client.delete("/library/api/libraries/nope/")
+
+        self.assertEqual(response.status_code, 404)
+        mock_cascade.assert_not_called()

mnemosyne/library/tests/test_managed_by.py

@@ -0,0 +1,207 @@
+"""Tests for the per-app ``managed_by`` concept.
+
+Covers the pure helpers (``infer_legacy_manager``,
+``Library.managed_by_display``), the token-derived stamping and
+duplicate-name rejection on the plain create endpoint (Neo4j stubbed
+via ``sys.modules``, same style as ``test_views.py``), and the
+``WorkspaceStatusSerializer`` surface. Cypher-touching paths are
+covered by the manual end-to-end plan, not these unit tests.
+"""
+
+from __future__ import annotations
+
+from types import SimpleNamespace
+from unittest.mock import patch
+
+from django.contrib.auth import get_user_model
+from django.test import TestCase
+from rest_framework.test import APIClient
+
+from library.api.serializers import WorkspaceStatusSerializer
+from library.models import Library, infer_legacy_manager
+from mcp_server.models import UserToken
+
+User = get_user_model()
+
+
+class InferLegacyManagerTests(TestCase):
+    """Truth table for the pre-``managed_by`` inference."""
+
+    def test_null_workspace_is_unmanaged(self):
+        self.assertIsNone(infer_legacy_manager(None))
+        self.assertIsNone(infer_legacy_manager(""))
+
+    def test_kairos_mail_prefix_is_kairos(self):
+        self.assertEqual(
+            infer_legacy_manager("kairos-mail-abc123-7"), "Kairos"
+        )
+
+    def test_other_workspace_is_daedalus(self):
+        self.assertEqual(
+            infer_legacy_manager("2f9c4a1e-uuid-ish"), "Daedalus"
+        )
+
+
+class ManagedByDisplayTests(TestCase):
+    """``managed_by_display`` on in-memory (unsaved) Library nodes."""
+
+    def test_stamped_label_wins(self):
+        lib = Library(name="x", managed_by="Spelunker")
+        self.assertEqual(lib.managed_by_display, "Spelunker")
+
+    def test_stamped_label_wins_over_inference(self):
+        lib = Library(
+            name="x", managed_by="My Token", workspace_id="kairos-mail-a-1"
+        )
+        self.assertEqual(lib.managed_by_display, "My Token")
+
+    def test_legacy_workspace_falls_back_to_inference(self):
+        lib = Library(name="x", workspace_id="ws-uuid")
+        self.assertEqual(lib.managed_by_display, "Daedalus")
+
+    def test_unmanaged_is_empty_string(self):
+        lib = Library(name="x")
+        self.assertEqual(lib.managed_by_display, "")
+
+
+class _FakeLibrary:
+    """Stand-in for the neomodel Library on the plain create endpoint."""
+
+    class DoesNotExist(Exception):
+        pass
+
+    existing = None          # what find_library_by_name_ci returns
+    instances = []           # constructor kwargs, in order
+    save_raises = None       # exception instance save() should raise, if any
+
+    def __init__(self, **kwargs):
+        type(self).instances.append(kwargs)
+        self.__dict__.update(kwargs)
+        self.uid = "lib-new"
+        self.workspace_id = None
+        self.created_at = None
+
+    def save(self):
+        if type(self).save_raises is not None:
+            raise type(self).save_raises
+        return self
+
+
+def _fake_find_ci(name):
+    _FakeLibrary.ci_queries.append(name)
+    return _FakeLibrary.existing
+
+
+def _fake_models_module():
+    return SimpleNamespace(
+        Library=_FakeLibrary, find_library_by_name_ci=_fake_find_ci
+    )
+
+
+class LibraryCreateStampingTests(TestCase):
+    """POST /library/api/libraries/ stamps ``managed_by`` and rejects dupes."""
+
+    def setUp(self):
+        self.user = User.objects.create_user(username="op", password="pw")
+        self.client = APIClient()
+        _FakeLibrary.existing = None
+        _FakeLibrary.instances = []
+        _FakeLibrary.save_raises = None
+        _FakeLibrary.ci_queries = []
+
+    def _post(self, token=None, name="Docs"):
+        self.client.force_authenticate(user=self.user, token=token)
+        with patch.dict("sys.modules", {"library.models": _fake_models_module()}):
+            return self.client.post(
+                "/library/api/libraries/",
+                {"name": name, "library_type": "technical"},
+                format="json",
+            )
+
+    def test_token_create_stamps_token_name(self):
+        response = self._post(token=UserToken(name="Spelunker"))
+
+        self.assertEqual(response.status_code, 201)
+        self.assertEqual(_FakeLibrary.instances[0]["managed_by"], "Spelunker")
+        self.assertEqual(response.json()["managed_by"], "Spelunker")
+
+    def test_session_create_leaves_managed_by_null(self):
+        response = self._post(token=None)
+
+        self.assertEqual(response.status_code, 201)
+        self.assertIsNone(_FakeLibrary.instances[0]["managed_by"])
+
+    def test_duplicate_name_returns_409_name_conflict(self):
+        _FakeLibrary.existing = SimpleNamespace(
+            uid="lib-old", name="Docs", managed_by_display="Daedalus"
+        )
+        response = self._post(token=UserToken(name="Spelunker"))
+
+        self.assertEqual(response.status_code, 409)
+        body = response.json()
+        self.assertEqual(body["code"], "name_conflict")
+        self.assertIn("Docs", body["detail"])
+        self.assertEqual(body["uid"], "lib-old")
+        self.assertEqual(body["managed_by"], "Daedalus")
+        self.assertEqual(_FakeLibrary.instances, [])
+
+    def test_duplicate_check_is_case_insensitive(self):
+        """A case-variant name 409s and reports the existing spelling."""
+        _FakeLibrary.existing = SimpleNamespace(
+            uid="lib-old", name="Amazon Connect", managed_by_display="Spelunker"
+        )
+        response = self._post(token=None, name="amazon connect")
+
+        self.assertEqual(response.status_code, 409)
+        # The lookup received the posted name (case-folding happens in
+        # Cypher), and the response names the existing spelling.
+        self.assertEqual(_FakeLibrary.ci_queries, ["amazon connect"])
+        self.assertIn("Amazon Connect", response.json()["detail"])
+        self.assertEqual(_FakeLibrary.instances, [])
+
+    def test_duplicate_of_unmanaged_reports_null_manager(self):
+        _FakeLibrary.existing = SimpleNamespace(
+            uid="lib-old", name="Docs", managed_by_display=""
+        )
+        response = self._post(token=None)
+
+        self.assertEqual(response.status_code, 409)
+        self.assertIsNone(response.json()["managed_by"])
+
+    def test_save_race_returns_409_not_500(self):
+        """An exact-name twin landing between pre-check and save 409s."""
+        from neomodel.exceptions import UniqueProperty
+
+        _FakeLibrary.save_raises = UniqueProperty("name")
+        response = self._post(token=None)
+
+        self.assertEqual(response.status_code, 409)
+        self.assertEqual(response.json()["code"], "name_conflict")
+
+
+class WorkspaceStatusSerializerManagedByTests(TestCase):
+    """The workspace status payload carries ``managed_by`` (nullable)."""
+
+    BASE = {
+        "workspace_id": "ws_a",
+        "library_uid": "lib_1",
+        "name": "W",
+        "library_type": "technical",
+        "description": "",
+        "item_count": 0,
+        "chunk_count": 0,
+        "created_at": "2026-01-01T00:00:00Z",
+    }
+
+    def test_managed_by_value_round_trips(self):
+        s = WorkspaceStatusSerializer(data={**self.BASE, "managed_by": "Daedalus"})
+        self.assertTrue(s.is_valid(), s.errors)
+        self.assertEqual(s.validated_data["managed_by"], "Daedalus")
+
+    def test_managed_by_null_accepted(self):
+        s = WorkspaceStatusSerializer(data={**self.BASE, "managed_by": None})
+        self.assertTrue(s.is_valid(), s.errors)
+
+    def test_managed_by_absent_accepted(self):
+        s = WorkspaceStatusSerializer(data=self.BASE)
+        self.assertTrue(s.is_valid(), s.errors)

mnemosyne/library/tests/test_svg_raster.py

@@ -0,0 +1,195 @@
+"""
+Tests for SVG rasterization in the ingest pipeline.
+
+SVG is vector XML: Pillow cannot decode it and the vision stage cannot send it
+as a data URI, so it is rendered to PNG at parse time. The cases that matter
+are the ones that fail *silently* — Write notes stack their pages inside one
+document, and rendering that document whole yields a plausible-looking PNG that
+is either blank or an illegible tall strip.
+"""
+
+import io
+import os
+import tempfile
+
+from django.test import TestCase
+from PIL import Image
+
+from library.services.parsers import (
+    IMAGE_EXTENSIONS,
+    DocumentParser,
+)
+from library.services.svg_raster import (
+    SvgRenderError,
+    render_svg_pages,
+    split_svg_pages,
+)
+
+# Geometry copied from a real Write note; the page element is identical across
+# both root formats, which is what makes splitting format-agnostic.
+PAGE_WIDTH = 1094
+PAGE_HEIGHT = 1654
+PAGE_PITCH = 1674  # page height + inter-page gap
+
+
+def write_document(pages: int, root_size: bool = False) -> bytes:
+    """
+    Build a Write-style document.
+
+    :param pages: Number of stacked pages.
+    :param root_size: Emit width/height on the root <svg>. False models
+        pre-eeab021 files (the permanent majority), True models newer saves.
+    """
+    root_attrs = ""
+    if root_size:
+        total = 10 + pages * PAGE_PITCH
+        root_attrs = f' width="{PAGE_WIDTH + 20}" height="{total}"'
+
+    body = "".join(
+        f'<svg class="write-page" x="10" y="{10 + i * PAGE_PITCH}" '
+        f'width="{PAGE_WIDTH}px" height="{PAGE_HEIGHT}px" '
+        f'xmlns="http://www.w3.org/2000/svg">'
+        f'<path d="M 100 {100 + i * 40} L {600 + i * 120} {700 + i * 40}" '
+        f'stroke="#000000" stroke-width="12" fill="none"/></svg>'
+        for i in range(pages)
+    )
+    return (
+        f'<svg id="write-document"{root_attrs} '
+        f'xmlns="http://www.w3.org/2000/svg" '
+        f'xmlns:xlink="http://www.w3.org/1999/xlink">'
+        f'<rect id="write-doc-background" width="100%" height="100%" fill="#808080"/>'
+        f'<defs id="write-defs"><style>.write-flat-pen{{fill:none}}</style></defs>'
+        f"{body}</svg>"
+    ).encode()
+
+
+GENERIC_SVG = (
+    b'<svg xmlns="http://www.w3.org/2000/svg" width="400" height="200" '
+    b'viewBox="0 0 400 200">'
+    b'<path d="M 20 20 L 380 180" stroke="#000" stroke-width="10"/></svg>'
+)
+
+
+def ink_fraction(png: bytes) -> float:
+    """Fraction of non-white pixels — the blank-render canary."""
+    grey = Image.open(io.BytesIO(png)).convert("L")
+    histogram = grey.histogram()
+    return sum(histogram[:200]) / sum(histogram)
+
+
+class SvgSplitTests(TestCase):
+    """Splitting a Write document into per-page SVGs."""
+
+    def test_splits_one_document_per_page(self):
+        for pages in (1, 2, 5):
+            for root_size in (False, True):
+                with self.subTest(pages=pages, root_size=root_size):
+                    document = write_document(pages, root_size=root_size)
+                    self.assertEqual(len(split_svg_pages(document)), pages)
+
+    def test_both_root_formats_split_identically(self):
+        """Root width/height (Write eeab021) must not change the outcome.
+
+        Existing notes are never rewritten, so both formats persist
+        indefinitely and have to render the same.
+        """
+        without = split_svg_pages(write_document(3, root_size=False))
+        with_size = split_svg_pages(write_document(3, root_size=True))
+        self.assertEqual(len(without), len(with_size))
+        self.assertEqual(len(without), 3)
+
+    def test_generic_svg_is_a_single_page(self):
+        self.assertEqual(len(split_svg_pages(GENERIC_SVG)), 1)
+
+    def test_unsized_svg_is_rejected_rather_than_guessed(self):
+        unsized = b'<svg xmlns="http://www.w3.org/2000/svg"><rect/></svg>'
+        with self.assertRaises(SvgRenderError):
+            split_svg_pages(unsized)
+
+    def test_script_inside_defs_does_not_desync_the_split(self):
+        """Regression: a non-greedy <defs>...</defs> regex mis-parses these."""
+        document = write_document(2).replace(
+            b'<defs id="write-defs">',
+            b'<defs id="write-defs"><script><float value="770" /></script>',
+        )
+        self.assertEqual(len(split_svg_pages(document)), 2)
+
+    def test_unquoted_root_attributes_are_repaired(self):
+        """Write can emit ``width=auto``, which is not valid XML.
+
+        A malformation inside the root tag defeats recovery differently from
+        one in the body: libxml2 abandons the whole document and returns a
+        bare root, so every page silently disappears.
+        """
+        document = write_document(2).replace(
+            b'<svg id="write-document"',
+            b'<svg width=auto height=auto id="write-document"',
+        )
+        self.assertEqual(len(split_svg_pages(document)), 2)
+
+
+class SvgRenderTests(TestCase):
+    """Rasterizing pages to PNG."""
+
+    def test_renders_one_png_per_page(self):
+        pages = render_svg_pages(write_document(3))
+        self.assertEqual(len(pages), 3)
+        for png, width, height in pages:
+            self.assertEqual(Image.open(io.BytesIO(png)).format, "PNG")
+            self.assertEqual(max(width, height), 1568)
+
+    def test_rendered_pages_are_not_blank(self):
+        """The pre-patch trap: a whole-document render emits ruling, no ink."""
+        for root_size in (False, True):
+            with self.subTest(root_size=root_size):
+                for png, _, _ in render_svg_pages(
+                    write_document(3, root_size=root_size)
+                ):
+                    self.assertGreater(ink_fraction(png), 0.001)
+
+    def test_pages_are_page_shaped_not_a_stacked_strip(self):
+        """The post-patch trap: root dimensions span every stacked page.
+
+        Rendering that whole gives one tall strip which, once capped, squashes
+        a 5-page note to ~209px wide.
+        """
+        expected = PAGE_HEIGHT / PAGE_WIDTH
+        for root_size in (False, True):
+            with self.subTest(root_size=root_size):
+                for _, width, height in render_svg_pages(
+                    write_document(5, root_size=root_size)
+                ):
+                    self.assertAlmostEqual(height / width, expected, delta=0.05)
+
+    def test_page_cap_is_respected(self):
+        self.assertEqual(len(render_svg_pages(write_document(8), max_pages=5)), 5)
+
+
+class SvgParserIntegrationTests(TestCase):
+    """The parser dispatch — SVG must not reach the Pillow image path."""
+
+    def setUp(self):
+        self.parser = DocumentParser()
+
+    def test_svg_is_not_in_image_extensions(self):
+        # It was, and PIL.Image.open cannot decode SVG, so ingest always failed.
+        self.assertNotIn("svg", IMAGE_EXTENSIONS)
+
+    def test_parse_multipage_svg_yields_one_image_per_page(self):
+        with tempfile.NamedTemporaryFile(suffix=".svg", delete=False) as f:
+            f.write(write_document(3))
+            f.flush()
+            path = f.name
+        try:
+            result = self.parser.parse(path, "svg")
+        finally:
+            os.unlink(path)
+
+        self.assertEqual(len(result.images), 3)
+        self.assertEqual(result.metadata["page_count"], 3)
+        self.assertEqual(result.text_blocks, [])
+        for index, image in enumerate(result.images):
+            # PNG, not svg — the vision stage needs raster for its data URI.
+            self.assertEqual(image.ext, "png")
+            self.assertEqual(image.source_page, index)
+            self.assertGreater(ink_fraction(image.data), 0.001)

mnemosyne/library/tests/test_tasks.py

@@ -67,6 +67,114 @@ class ReembedItemTaskTests(TestCase):
         mock_pipeline.reprocess_item.assert_called_once()
 
 
+@override_settings(CELERY_TASK_ALWAYS_EAGER=True)
+class IngestFromDaedalusFailureClassificationTests(TestCase):
+    """A deterministic parse failure is terminal; a transient error retries.
+
+    Neo4j and S3 are mocked at the task's boundaries so the test exercises the
+    exception-classification branch (parsers.UnsupportedFileTypeError vs. any
+    other Exception) without a live graph or object store.
+    """
+
+    def setUp(self):
+        from library.tasks import ingest_from_daedalus
+
+        # Calling .run() bypasses Celery's request setup, but the task
+        # persists self.request.id into the NOT NULL celery_task_id column
+        # and reads self.request.retries — push a real request context.
+        ingest_from_daedalus.push_request(id="test-task-id", retries=0)
+        self.addCleanup(ingest_from_daedalus.pop_request)
+
+    def _make_job(self):
+        from library.models import IngestJob
+
+        return IngestJob.objects.create(
+            id="job_test_unsupported",
+            library_uid="lib-uid-123",
+            source="daedalus",
+            s3_key="incoming/bad.drawio",
+            file_type="vnd.jgraph.mxfile",
+            title="bad.drawio",
+            content_hash="abc123",
+        )
+
+    def _patched_boundaries(self, pipeline_side_effect):
+        """Patch every boundary the task hits before the pipeline runs.
+
+        Returns a context-manager list; the pipeline's ``process_item`` is set
+        to raise ``pipeline_side_effect``.
+        """
+        from library.services.parsers import UnsupportedFileTypeError  # noqa: F401
+
+        patchers = [
+            patch("library.tasks.db"),
+            patch("library.models.Library"),
+            patch("library.models.Item"),
+            patch("library.services.source_s3.fetch_from_source", return_value=b"data"),
+            patch("library.services.source_s3.copy_into_mnemosyne"),
+            patch("library.tasks._resolve_or_create_default_collection"),
+            patch("library.services.pipeline.EmbeddingPipeline"),
+        ]
+        mocks = [p.start() for p in patchers]
+        self.addCleanup(lambda: [p.stop() for p in patchers])
+
+        # db.cypher_query returns (rows, meta); no prior item to supersede.
+        mocks[0].cypher_query.return_value = ([], None)
+        # Library.nodes.get returns a stand-in library node.
+        mocks[1].nodes.get.return_value = MagicMock(uid="lib-uid-123")
+        # Item() instances carry a uid used for the S3 key.
+        item_instance = MagicMock(uid="item-uid-abc")
+        mocks[2].return_value = item_instance
+        # The pipeline raises the classification-relevant error.
+        pipeline_instance = MagicMock()
+        pipeline_instance.process_item.side_effect = pipeline_side_effect
+        mocks[6].return_value = pipeline_instance
+        return pipeline_instance
+
+    def test_unsupported_file_type_is_terminal_and_not_retried(self):
+        from library.models import IngestJob
+        from library.services.parsers import UnsupportedFileTypeError
+        from library.tasks import ingest_from_daedalus
+
+        job = self._make_job()
+        self._patched_boundaries(
+            UnsupportedFileTypeError("Unsupported file type 'vnd.jgraph.mxfile'.")
+        )
+
+        with patch.object(ingest_from_daedalus, "retry") as mock_retry:
+            result = ingest_from_daedalus.run(job.id)
+
+        # Never retried.
+        mock_retry.assert_not_called()
+        # Terminal failure with a machine-readable reason.
+        self.assertFalse(result["success"])
+        self.assertEqual(result["reason"], "unsupported_file_type")
+
+        job.refresh_from_db()
+        self.assertEqual(job.status, "failed")
+        self.assertEqual(job.retry_count, 0)
+        self.assertIsNotNone(job.completed_at)
+        self.assertIn("Unsupported file type", job.error)
+
+    def test_transient_error_takes_the_retry_path(self):
+        from library.tasks import ingest_from_daedalus
+
+        job = self._make_job()
+        self._patched_boundaries(ConnectionError("S3 hiccup"))
+
+        # self.retry raises Retry in real Celery; simulate that so the task
+        # doesn't fall through to the terminal branch.
+        from celery.exceptions import Retry
+
+        with patch.object(ingest_from_daedalus, "retry", side_effect=Retry()) as mock_retry:
+            with self.assertRaises(Retry):
+                ingest_from_daedalus.run(job.id)
+
+        mock_retry.assert_called_once()
+        job.refresh_from_db()
+        self.assertEqual(job.retry_count, 1)
+
+
 class ResolveUserTests(TestCase):
     """Tests for the _resolve_user helper."""
 

mnemosyne/library/tests/test_views.py

@@ -1,13 +1,13 @@
 """Tests for the library CRUD HTML views.
 
-Currently covers ``library_list``'s Daedalus-workspace scope filter. The
-view loads every ``Library`` node from Neo4j and narrows it by a ``scope``
-GET param (``all`` / ``global`` / ``daedalus``). These tests stub out
-Neo4j entirely — patching ``neo4j_available`` and injecting a fake
-``Library`` class via ``sys.modules`` — so they assert on the queryset
-``.filter(...)`` call the view makes and the context it renders, not on
-real graph behaviour. Mirrors the mocking style in
-``test_search_views_admin_scope.py``.
+Currently covers ``library_list``'s app-managed scope filter. The view
+loads every ``Library`` node from Neo4j and narrows it in Python by a
+``scope`` GET param (``all`` / ``unmanaged`` / ``managed``, with legacy
+``global`` / ``daedalus`` aliases). These tests stub out Neo4j entirely —
+patching ``neo4j_available`` and injecting a fake ``Library`` class via
+``sys.modules`` — so they assert on the filtering the view does and the
+context it renders, not on real graph behaviour. Mirrors the mocking
+style in ``test_search_views_admin_scope.py``.
 """
 
 from __future__ import annotations
@@ -22,6 +22,17 @@ from django.urls import reverse
 User = get_user_model()
 
 
+def _lib(name, managed_by_display):
+    return SimpleNamespace(
+        uid=f"uid-{name}",
+        name=name,
+        library_type="technical",
+        description="",
+        workspace_id=None,
+        managed_by_display=managed_by_display,
+    )
+
+
 class LibraryListScopeFilterTests(TestCase):
     """Cover the ``scope`` filter branches of ``library_list``."""
 
@@ -31,57 +42,64 @@ class LibraryListScopeFilterTests(TestCase):
         )
         self.client.force_login(self.user)
         self.url = reverse("library:library-list")
+        self.managed = _lib("Docs", "Spelunker")
+        self.unmanaged = _lib("Notes", "")
 
     def _fake_library_cls(self):
-        """Return (Library stub, nodes mock) where ``nodes`` chains fluently.
-
-        ``Library.nodes`` → ``.filter(...)`` → ``.order_by(...)`` all return
-        the same MagicMock so the view's queryset building works regardless
-        of which branch it takes, and ``.filter`` records its kwargs.
-        """
+        """Return a Library stub whose ``nodes.order_by`` yields two libraries."""
         fake_nodes = MagicMock()
-        fake_nodes.filter.return_value = fake_nodes
-        fake_nodes.order_by.return_value = []
-        return SimpleNamespace(nodes=fake_nodes), fake_nodes
+        fake_nodes.order_by.return_value = [self.managed, self.unmanaged]
+        return SimpleNamespace(nodes=fake_nodes)
 
-    def _get(self, fake_library_cls, **params):
+    def _get(self, **params):
         with patch("library.views.neo4j_available", return_value=True), \
              patch.dict(
                  "sys.modules",
-                 {"library.models": SimpleNamespace(Library=fake_library_cls)},
+                 {"library.models": SimpleNamespace(Library=self._fake_library_cls())},
              ):
             return self.client.get(self.url, params)
 
-    def test_default_scope_is_all_and_does_not_filter(self):
-        fake_cls, fake_nodes = self._fake_library_cls()
-        response = self._get(fake_cls)
+    def test_default_scope_is_all_and_returns_everything(self):
+        response = self._get()
 
         self.assertEqual(response.status_code, 200)
         self.assertEqual(response.context["scope"], "all")
-        fake_nodes.filter.assert_not_called()
-        fake_nodes.order_by.assert_called_once_with("name")
+        self.assertEqual(
+            list(response.context["libraries"]), [self.managed, self.unmanaged]
+        )
 
-    def test_global_scope_filters_workspace_isnull_true(self):
-        fake_cls, fake_nodes = self._fake_library_cls()
-        response = self._get(fake_cls, scope="global")
+    def test_managed_scope_keeps_only_managed(self):
+        response = self._get(scope="managed")
 
-        self.assertEqual(response.context["scope"], "global")
-        fake_nodes.filter.assert_called_once_with(workspace_id__isnull=True)
+        self.assertEqual(response.context["scope"], "managed")
+        self.assertEqual(list(response.context["libraries"]), [self.managed])
 
-    def test_daedalus_scope_filters_workspace_isnull_false(self):
-        fake_cls, fake_nodes = self._fake_library_cls()
-        response = self._get(fake_cls, scope="daedalus")
+    def test_unmanaged_scope_keeps_only_unmanaged(self):
+        response = self._get(scope="unmanaged")
 
-        self.assertEqual(response.context["scope"], "daedalus")
-        fake_nodes.filter.assert_called_once_with(workspace_id__isnull=False)
+        self.assertEqual(response.context["scope"], "unmanaged")
+        self.assertEqual(list(response.context["libraries"]), [self.unmanaged])
+
+    def test_legacy_daedalus_scope_aliases_to_managed(self):
+        response = self._get(scope="daedalus")
+
+        self.assertEqual(response.context["scope"], "managed")
+        self.assertEqual(list(response.context["libraries"]), [self.managed])
+
+    def test_legacy_global_scope_aliases_to_unmanaged(self):
+        response = self._get(scope="global")
+
+        self.assertEqual(response.context["scope"], "unmanaged")
+        self.assertEqual(list(response.context["libraries"]), [self.unmanaged])
 
     def test_unknown_scope_does_not_filter(self):
         """An unexpected scope value degrades to the unfiltered list."""
-        fake_cls, fake_nodes = self._fake_library_cls()
-        response = self._get(fake_cls, scope="bogus")
+        response = self._get(scope="bogus")
 
         self.assertEqual(response.context["scope"], "bogus")
-        fake_nodes.filter.assert_not_called()
+        self.assertEqual(
+            list(response.context["libraries"]), [self.managed, self.unmanaged]
+        )
 
     def test_neo4j_unavailable_sets_error_and_empty_list(self):
         with patch("library.views.neo4j_available", return_value=False):

mnemosyne/library/views.py

@@ -31,20 +31,23 @@ logger = logging.getLogger(__name__)
 
 @login_required
 def library_list(request):
-    """List libraries, optionally filtered by Daedalus-workspace scope."""
+    """List libraries, optionally filtered by app-managed scope."""
     scope = request.GET.get("scope", "all")
+    # Legacy bookmark values from before managed_by existed.
+    scope = {"daedalus": "managed", "global": "unmanaged"}.get(scope, scope)
     libraries = []
     error = None
     if neo4j_available():
         try:
             from .models import Library
 
-            qs = Library.nodes
-            if scope == "daedalus":
-                qs = qs.filter(workspace_id__isnull=False)
-            elif scope == "global":
-                qs = qs.filter(workspace_id__isnull=True)
-            libraries = qs.order_by("name")
+            libraries = list(Library.nodes.order_by("name"))
+            # managed_by_display covers legacy workspace libraries that
+            # predate the managed_by property, so filter in Python.
+            if scope == "managed":
+                libraries = [l for l in libraries if l.managed_by_display]
+            elif scope == "unmanaged":
+                libraries = [l for l in libraries if not l.managed_by_display]
         except Exception as e:
             error = f"Could not connect to Neo4j: {e}"
             logger.error(error)

mnemosyne/mcp_server/drf_auth.py

@@ -26,6 +26,20 @@ from rest_framework import authentication, exceptions
 from .auth import MCPAuthError, resolve_mcp_user
 
 
+def request_token_label(request):
+    """Name of the ``UserToken`` authenticating this request, or None.
+
+    Session-authenticated requests (``request.auth`` is None) and blank
+    token names return None — the caller treats both as "no managing app".
+    """
+    from .models import UserToken
+
+    token = getattr(request, "auth", None)
+    if isinstance(token, UserToken):
+        return token.name.strip() or None
+    return None
+
+
 class UserTokenAuthentication(authentication.BaseAuthentication):
     """Authenticate DRF requests with a ``UserToken`` bearer."""
 

mnemosyne/mcp_server/forms.py

@@ -57,6 +57,12 @@ class UserTokenCreateForm(forms.Form):
             "class": "input input-bordered w-full",
             "placeholder": "e.g. Claude Desktop, CI script",
         }),
+        help_text=(
+            "A friendly label so you can identify this token later. It also "
+            "labels any library the token creates (shown as “Managed by "
+            "<name>”) — for an app integration, use the app's name, e.g. "
+            "Daedalus, Kairos, Spelunker."
+        ),
     )
     expires_at = forms.DateTimeField(
         required=False,

mnemosyne/mcp_server/templates/mcp_server/tokens/create.html

@@ -21,7 +21,7 @@
                     </label>
                     {{ form.name }}
                     <label class="label">
-                        <span class="label-text-alt opacity-60">A friendly label so you can identify this token later (e.g. “Claude Desktop”).</span>
+                        <span class="label-text-alt opacity-60">{{ form.name.help_text }}</span>
                     </label>
                 </div>
                 <div class="form-control mt-4">

mnemosyne/mcp_server/tests/test_drf_auth.py

@@ -105,6 +105,44 @@ class UserTokenAuthenticationTest(TestCase):
         resp = self._get(f"Bearer {self.plaintext} extra")
         self.assertEqual(resp.status_code, status.HTTP_401_UNAUTHORIZED)
 
+    def test_request_token_label_reads_token_name(self):
+        from types import SimpleNamespace
+
+        from mcp_server.drf_auth import request_token_label
+
+        token = UserToken(name="  Spelunker  ")
+        self.assertEqual(
+            request_token_label(SimpleNamespace(auth=token)), "Spelunker"
+        )
+
+    def test_request_token_label_none_for_session(self):
+        from types import SimpleNamespace
+
+        from mcp_server.drf_auth import request_token_label
+
+        self.assertIsNone(request_token_label(SimpleNamespace(auth=None)))
+        # A request object with no auth attribute at all (plain Django).
+        self.assertIsNone(request_token_label(SimpleNamespace()))
+
+    def test_request_token_label_none_for_blank_name(self):
+        from types import SimpleNamespace
+
+        from mcp_server.drf_auth import request_token_label
+
+        self.assertIsNone(
+            request_token_label(SimpleNamespace(auth=UserToken(name="   ")))
+        )
+
+    def test_request_token_label_none_for_foreign_auth_object(self):
+        from types import SimpleNamespace
+
+        from mcp_server.drf_auth import request_token_label
+
+        # e.g. a JWT dict from another auth class — not a UserToken.
+        self.assertIsNone(
+            request_token_label(SimpleNamespace(auth={"iss": "daedalus"}))
+        )
+
     def test_request_auth_stashes_token(self):
         # The auth class returns (user, token); DRF places the token on
         # request.auth. Re-use a UserToken-aware endpoint to verify.

pyproject.toml

@@ -30,6 +30,9 @@ dependencies = [
     "semantic-text-splitter>=0.20,<1.0",
     "tokenizers>=0.20,<1.0",
     "Pillow>=10.0,<12.0",
+    # SVG page splitting — needs recover=True for Write notes that emit
+    # unescaped attribute content and aren't well-formed XML
+    "lxml>=5.3,<7",
     "requests>=2.31,<3.0",
     # Phase 5: MCP Server
     "fastmcp>=2.0,<3.0",