Security
--------
ComfyUI's PromptServer.add_routes() mirrors every registered route under a
second "/api" prefix, so /mjr/am/<mutation> is also served at
/api/mjr/am/<mutation>. The auth middleware matched only the canonical
prefix, so the mirrored path reached the same handler with the
authentication check skipped entirely, and without the hardening headers:
POST /mjr/am/assets/delete -> requires_auth = True
POST /api/mjr/am/assets/delete -> requires_auth = False
Replaced the prefix test with a shared _is_majoor_path() boundary covering
the canonical prefixes, ComfyUI's /api mirror, and the /api/v2/assets
compat surface. Verified it does not capture ComfyUI's own routes. CSRF,
rate limiting and safe-mode gates are per-handler and were unaffected.
Licensing
---------
PrimeVue 5 is no longer MIT: PrimeTek moved it to the commercial "PrimeUI"
license, which requires a key, shows a license notice without one, and
forbids redistributing it "as a component library or development tool"
without a separate OEM License. Dependabot had auto-merged 4.5.5 -> 5.0.1.
This project is GPL-3.0 and commits dist/ to the repo, so shipping it would
nag every user and conflict with the GPL's redistribution requirement.
Pinned back to 4.5.5 (last MIT release). The 5.x bump had touched only
package.json/package-lock.json, so no source depended on v5 APIs. The
revert drops @primeui/license-manager and its @noble/ed25519 key-signature
stack; the primevue chunk shrinks 555.70 -> 520.61 kB. Dependabot is now
configured to ignore primevue major bumps.
Correctness
-----------
- api_v2_assets: stop returning the DB's absolute filepath on this
network-reachable surface; subfolder + root_id already convey location
without disclosing local disk layout.
- watcher: enforce _RECENT_GENERATED_HARD_CAP after inserting a batch;
pruning only beforehand let a large batch overshoot the cap.
- routes/registry: track registered route groups individually so a group
failing part-way through does not re-register earlier groups on retry.
- fs_walker/scanner: apply the viewer's root check to symlinked files, so
the scanner stops indexing symlinks whose target escapes every allowed
root ("ghost assets" the viewer then refuses to open). Consulted only for
actual symlinks, so ordinary scans pay no extra I/O.
- routes/registry_app: skip the redundant app.add_routes() on the live
PromptServer app; ComfyUI adds the shared table itself after node
loading, so every core and Majoor route was registered twice at startup.
- Result: unwrap()/unwrap_or()/map() gated on `data is not None`, treating
a legitimate Ok(None) as failure; gate on `ok` instead.
- custom_roots: replace a leftover debug print() that wrote user file paths
to stderr on every move request with logger.debug.
- schema: drop the redundant UNIQUE(filepath, source, root_id) index
(migration m022). assets.filepath is already UNIQUE, so it could never
reject a row the column constraint does not, nor narrow a lookup - it
only cost a second B-tree write per asset on the scan hot path.
Frontend
--------
- entry.ts: app.registerExtension() throws on a duplicate name and ComfyUI
exposes no unregisterExtension, so a live re-import threw before setup()
ran and the reload-safe teardown never executed. Catch it and re-run
setup() directly.
- Rebuilt dist/ so both the reload fix and the PrimeVue revert ship.
CI / tooling
------------
- frontend-compat-nightly: the types step died with a bare "cp: cannot
stat" (upstream's build has not emitted dist/index.d.ts since Aug 26).
Check both candidate paths and dump dist/ with an actionable message.
- package.json: lenient checks that swallowed their exit code are now
suffixed :advisory, the plain names are strict, and quality:strict runs
the full set.
Verified: 1841 Python tests pass (5 skipped, symlink privileges), ruff
clean; tsc --noEmit clean, 831 JS tests pass, eslint clean, build succeeds.
Introduces a forward-only, additive, idempotent migration framework on top of the existing legacy implicit-repair path (which remains the source of truth for schema versions <= 16). Versioned migrations live in mjr_am_backend/adapters/db/migrations/ and are tracked in a new schema_migrations bookkeeping table. The runner is invoked on every startup right after migrate_schema(); pending migrations are applied in version order, already-applied ones short-circuit.
First migration v17 normalize_tags creates a parallel tags + asset_tags relational layer (case-insensitive UNIQUE on tag name, cascade-delete on asset side, covering index on tag_id). Backfill reads asset_metadata.tags JSON safely (malformed rows skipped) and uses INSERT OR IGNORE everywhere for full idempotency. The legacy JSON column is NOT touched; existing v16 readers keep working.
Adds a Repository layer under mjr_am_backend/data/repositories/ to isolate raw SQL from services. First repository TagsRepository exposes ensure / get_by_name / list_all / attach / detach / list_for_asset / list_assets_with_tag / count on the new normalized tables, all returning Result.
Tests: 6 migration-runner tests (version validation, idempotency, bookkeeping, failure propagation, registry application) and 7 TagsRepository tests (CRUD, case-insensitivity, cascade-delete) - 13/13 passing. Existing test suite remains green (regressions limited to 2 pre-existing health_routes failures unrelated to this change).
Phase 3.2 closes the DB hardening item of the v3 improvement plan.
- Added `_build_asset_filter_query` and `_query_allowed_asset_ids` functions to streamline asset filtering in hybrid search.
- Enhanced `_filter_semantic_hits` to utilize new asset filtering methods.
- Refactored vector search functions to improve K-means clustering logic and asset filtering.
- Introduced helper functions for parsing and validating query filters in `query_sanitizer.py`.
- Improved viewer context resolution in `viewer.py` with better error handling and path validation.
- Updated schema healing tests to use helper functions for column validation.
- Enhanced test coverage for vector indexing and filesystem entries.
- Added new test images for filesystem entry tests.
- Removed unnecessary imports from various test files to improve readability and maintainability.
- Added new settings for the Generated Feed display in the settingsFeed.js file.
- Introduced a new configuration file for Pyright type checking.
- Add AI alignment score display to asset metadata
- Implement image description fallback for non-workflow image assets
- Improve AI captioning UI and UX
- Optimize AI tagging suggestions with bootstrap generation
- Refactor collections controller to support AI toggling
- Enhance error handling and retry logic for asset indexing
- Updated vector_service.py to support SigLIP2, X-CLIP, and Florence-2 models for embedding images, text, and video keyframes.
- Introduced AI verbose logging settings, allowing dynamic configuration of logging verbosity for model loading and inference.
- Modified database maintenance to backfill missing vectors for both image and video assets.
- Added health check routes for managing AI logging settings.
- Implemented enhanced caption generation for images using Florence-2.
- Updated vector search routes to include enhanced prompt generation and improved error handling.
- Added tests for new features, including enhanced prompt generation and dimension fallback mechanisms.
- Move token from local to session storage on client side
- Set httpOnly cookie for server-side token on bootstrap/rotate
- Add token validation and auto-sanitization
- Implement retry logic for failed operations
- Add input validation for collection names
- Update toast durations and error handling
- Introduced `query_sanitizer.py` for inline query, date, and sort sanitization.
- Added `result_filter.py` for runtime filtering helpers, including asset exclusion and enrichment pause functionality.
- Implemented `result_hydrator.py` for database hydration helpers, ensuring assets are properly hydrated with metadata.
- Created unit tests in `test_db_refactor_helpers.py` to validate database lock detection and SQL write query identification.
- Organized and grouped import statements for better readability in health.py, metadata.py, releases.py, scan.py, search.py, version.py, viewer.py, and registry.py.
- Updated type hints to use the new union syntax (e.g., str | None) for better clarity in scan.py, settings.py, tool_detect.py, and result.py.
- Enhanced the handling of optional types in various functions to ensure more robust type checking.
- Removed unnecessary blank lines and improved formatting for consistency.
- Adjusted the logic in several test files to ensure compatibility with the updated import structure and type hints. (See <attachments> above for file contents. You may not need to search or read the file again.)
- Implement tests to ensure file paths are properly allowlisted in custom view and folder info handlers.
- Verify that collections mutations enforce CSRF protection and write access guards.
- Check that settings mutations require write access for specific routes.
- Implemented caching for tool availability and versions.
- Added functions to check for ExifTool and FFprobe availability.
- Included version satisfaction checks against minimum required versions.
- Created utility functions for parsing boolean and float values from environment variables.
- Introduced shared utilities for error message sanitization and logging with emoji indicators.
- Established a Result pattern for error handling without exceptions.
- Added scan throttling helpers to manage background scans efficiently.
- Implemented time utilities for timestamps and performance measurement.
- Defined shared types, enums, and constants for consistent error handling and file classification.
- Created version management utilities to retrieve version information from pyproject.toml.
- Developed installation script for runtime dependencies.
- Added tests for ExifTool batch processing and various configuration scenarios.
- Implemented tests for version route aliases and runtime output root configuration.
- Added DuplicatesService for background duplicate analysis and similarity detection.
- Introduced endpoints for analyzing duplicates, checking status, and merging tags.
- Created utility functions for file hashing and perceptual hashing.
- Implemented watcher settings for runtime tuning of debounce and deduplication.
- Added database maintenance tools for resetting and deleting the database.
- Enhanced error handling with sanitization for client responses.
- Added tests for duplicates service and error handling.
- Updated ESLint configuration for better code quality checks.
- Introduced filename validation and logging utilities.
Ewald