deucebucket · deucebucket · Jul 14, 2026 · Jun 12, 2026 · Jun 12, 2026 · Jun 12, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -2,6 +2,48 @@
 
 All notable changes to Library Manager will be documented in this file.
 
+## [0.9.0-beta.153] - 2026-06-12
+
+### Security
+- **#255: Validate and sanitize BookDB response data** — Added `_sanitize_api_response()` to strip null bytes, control characters, HTML tags, and truncate oversized fields from all Skaldleita/BookDB API responses. Prevents path traversal, XSS, and memory issues from malformed responses.
+- **#256: Signing nonce and tighter replay window** — Added UUID nonce (`X-LM-Nonce`) to signed request headers for replay attack prevention. Reduced timestamp tolerance from 300s to 120s. Backwards compatible with existing BookDB.
+- **#236: SSRF prevention on plugin endpoints** — (from 0.9.0-beta.152)
+
+### Fixed
+- **#252: Rate limiter aligned with actual BookDB tiers** — Increased min_delay from 3.6s (1000/hr) to 12.0s (300/hr) to match BookDB's API key tier. Added adaptive backoff from Retry-After headers and specific 429 handling in audio identification.
+- **#253: server_notice handled in audio identification** — BookDB's abort/notice signals are now processed in `identify_audio_with_bookdb()` (submit, poll, and result stages), matching existing coverage in `search_bookdb()`.
+
+### Enhanced
+- **#254: Differentiated BookDB source trust weighting** — `sl_source` field now maps to distinct source names (`bookdb_audio`, `bookdb_transcript`, `bookdb_scrape`) with graduated weights (65/55/45) instead of treating all BookDB audio results identically.
+
+### Documentation
+- **#257: ABS plugin architecture** — Added `docs/ABS-Plugin-Architecture.md` scoping Library Manager as an Audiobookshelf metadata provider plugin.
+
+---
+
+## [0.9.0-beta.152] - 2026-06-12
+
+### Security
+- **#219: Library boundary enforcement** on undo/replace/remove file operations
+- **#237: XSS fix** — escapeHtml covers all 5 HTML entities including single quotes
+- **#236: SSRF prevention** on custom plugin endpoints
+
+### Fixed
+- **#220: Atomic status guard** on apply_fix prevents race condition double-renames
+- **#221: Worker stop checks** in all 4 legacy batch loops
+- **#222-224: Rate limits and circuit breakers** for BookDB contribute, community consensus, Gemini language detect, OpenRouter identify, plus process endpoint re-entrancy guard
+- **#225-226: Undo parent mkdir** and path sanitize false positive fix
+- **#227: L1 preserves original metadata** on validation failure
+- **#228-229: Pipeline stuck prevention** — L3 trust_sl no longer deletes queue entry before L4, AI verify creates history on null drastic check
+- **#230: Custom source weights** now actually used in confidence calculations
+- **#231: Fixed broken route** in queue multi-edit modal
+- **#232: Settings save** now persists sl_trust_mode and enable_isbn_lookup, removed duplicate field
+- **#233: Double-click prevention** on all 11 history page action buttons
+- **#234: Famous numeric titles** (1984, 2001, etc.) no longer flagged as garbage
+- **#235: Confidence scale normalization** handles both 0-1 and 0-100 scales
+
+---
+
 ## [0.9.0-beta.151] - 2026-06-05
 
 ### Changed

diff --git a/README.md b/README.md
@@ -4,7 +4,7 @@
 
 **Smart Audiobook Library Organizer with Multi-Source Metadata & AI Verification**
 
-[![Version](https://img.shields.io/badge/version-0.9.0--beta.152-blue.svg)](CHANGELOG.md)
+[![Version](https://img.shields.io/badge/version-0.9.0--beta.153-blue.svg)](CHANGELOG.md)
 [![Docker](https://img.shields.io/badge/docker-ghcr.io-blue.svg)](https://ghcr.io/deucebucket/library-manager)
 [![License](https://img.shields.io/badge/license-AGPL--3.0-blue.svg)](LICENSE)
 

diff --git a/app.py b/app.py
@@ -11,7 +11,7 @@
 - Multi-provider AI (Gemini, OpenRouter, Ollama)
 """
 
-APP_VERSION = "0.9.0-beta.152"
+APP_VERSION = "0.9.0-beta.153"
 GITHUB_REPO = "deucebucket/library-manager"  # Your GitHub repo
 
 # Versioning Guide:
@@ -738,7 +738,7 @@
            try:
                with open(ERROR_REPORTS_PATH, 'r') as f:
                    reports = json.load(f)
            except:
                reports = []

        # Add new report (keep last 100 reports to avoid file bloat)
@@ -762,7 +762,7 @@
        try:
            with open(ERROR_REPORTS_PATH, 'r') as f:
                return json.load(f)
        except:
            return []
    return []

@@ -1717,7 +1717,7 @@
                    continue
                result = call_gemini(prompt, merged_config)
                if result:
                    logger.info(f"[PROVIDER CHAIN] Success with gemini")
                    return result

            elif provider == 'openrouter':
@@ -1726,13 +1726,13 @@
                    continue
                result = call_openrouter(prompt, merged_config)
                if result:
                    logger.info(f"[PROVIDER CHAIN] Success with openrouter")
                    return result

            elif provider == 'ollama':
                result = call_ollama(prompt, merged_config)
                if result:
                    logger.info(f"[PROVIDER CHAIN] Success with ollama")
                    return result

            else:
@@ -1834,7 +1834,7 @@
                        return result
                    elif result and result.get('transcript'):
                        # Got transcript but no match - still useful, return for potential AI fallback
                        logger.info(f"[AUDIO CHAIN] BookDB returned transcript only")
                        return result
                    elif result is None and attempt < max_retries - 1:
                        # Connection might be down, wait and retry
@@ -2166,11 +2166,11 @@
                device = "cuda"
                # int8 works on all CUDA devices including GTX 1080 (compute 6.1)
                # float16 only works on newer GPUs (compute 7.0+)
                logger.info(f"[WHISPER] Using CUDA GPU acceleration (10x faster)")
            else:
                logger.info(f"[WHISPER] Using CPU (no CUDA GPU detected)")
        except ImportError:
            logger.info(f"[WHISPER] Using CPU (ctranslate2 not available)")

        _whisper_model = WhisperModel(model_name, device=device, compute_type=compute_type)
        _whisper_model_name = model_name
@@ -2381,7 +2381,7 @@
        if sample_path and os.path.exists(sample_path):
            try:
                os.unlink(sample_path)
            except:
                pass

    return result

diff --git a/docs/ABS-Plugin-Architecture.md b/docs/ABS-Plugin-Architecture.md
@@ -0,0 +1,252 @@
+# Audiobookshelf Plugin Architecture for Library Manager
+
+## Overview
+
+Library Manager as a metadata provider plugin for Audiobookshelf (ABS). When ABS needs to identify or enrich audiobook metadata, it queries Library Manager, which leverages the full Skaldleita pipeline: GPU Whisper transcription, 50M+ book database matching, multi-source API lookups, and AI-powered consensus verification.
+
+**Current state:** LM has an `abs_client.py` that pulls data FROM ABS (listening progress, library items, user management). This document covers the reverse direction — making LM available TO ABS as a metadata agent.
+
+---
+
+## Data Flow
+
+```
+Audiobookshelf                Library Manager              Skaldleita (BookDB)
+     │                              │                              │
+     │  1. "Identify this book"     │                              │
+     │  ─────────────────────────>  │                              │
+     │  (title, author, audio clip) │                              │
+     │                              │  2. Audio → Whisper queue    │
+     │                              │  ─────────────────────────>  │
+     │                              │                              │
+     │                              │  3. Metadata match (50M DB)  │
+     │                              │  <─────────────────────────  │
+     │                              │                              │
+     │                              │  4. API enrichment           │
+     │                              │  (Audnexus, Google Books,    │
+     │                              │   OpenLibrary, Hardcover)    │
+     │                              │                              │
+     │                              │  5. AI consensus verify      │
+     │                              │  (Gemini / OpenRouter)       │
+     │                              │                              │
+     │  6. Enriched metadata        │                              │
+     │  <─────────────────────────  │                              │
+     │  (author, title, narrator,   │                              │
+     │   series, year, confidence)  │                              │
+```
+
+---
+
+## Architecture Options
+
+### Option A: LM as ABS Metadata Provider (Recommended)
+
+ABS has a metadata provider plugin system. LM registers as a provider that ABS queries during its "Match" and "Quick Match" flows.
+
+**How it works:**
+- ABS sends search queries to LM's API
+- LM runs its full pipeline (BookDB + multi-source + AI verification)
+- LM returns structured metadata in ABS's expected format
+- ABS users see LM results alongside other providers (Google Books, Audible, etc.)
+
+**LM endpoints needed:**
+```
+GET  /api/abs/search?query=<title>&author=<author>
+     → Returns ABS-formatted search results
+
+GET  /api/abs/match?title=<title>&author=<author>&narrator=<narrator>
+     → Returns single best match with confidence
+
+POST /api/abs/identify-audio
+     → Accepts audio clip, returns identification via Skaldleita Whisper
+
+GET  /api/abs/cover?title=<title>&author=<author>
+     → Returns cover image URL if available
+```
+
+**Pros:** Native ABS integration, users see LM in their provider list
+**Cons:** ABS metadata provider API has a specific contract we must match exactly
+
+### Option B: LM as Standalone Enrichment Service
+
+LM runs alongside ABS and periodically scans the ABS library for books with missing/low-quality metadata, then pushes corrections back via ABS API.
+
+**How it works:**
+- LM uses existing `abs_client.py` to read ABS library
+- Identifies books with missing narrators, series info, etc.
+- Runs its pipeline to find correct metadata
+- Pushes updates back to ABS via API
+
+**Pros:** Works without ABS plugin system, LM controls the schedule
+**Cons:** Delayed updates, LM needs write access to ABS
+
+### Option C: Hybrid (Both A + B)
+
+Register as metadata provider AND run background enrichment. Provider handles new additions; background scan catches existing gaps.
+
+---
+
+## Recommended Approach: Option A (Metadata Provider)
+
+### ABS Metadata Provider Contract
+
+ABS expects providers to implement these operations:
+
+```json
+// Search request
+GET /search?query=Brandon+Sanderson+Mistborn
+
+// Search response
+{
+  "matches": [
+    {
+      "title": "The Final Empire",
+      "subtitle": "Mistborn Book 1",
+      "author": "Brandon Sanderson",
+      "narrator": "Michael Kramer",
+      "publisher": "Macmillan Audio",
+      "publishedYear": "2006",
+      "description": "...",
+      "cover": "https://...",
+      "isbn": "9780765350381",
+      "asin": "B002V1O7UE",
+      "series": [
+        { "series": "Mistborn", "sequence": "1" }
+      ],
+      "language": "en",
+      "duration": 95940
+    }
+  ]
+}
+```
+
+### LM → ABS Field Mapping
+
+| ABS Field | LM Source | Notes |
+|-----------|-----------|-------|
+| `title` | BookProfile.title | Highest confidence value |
+| `author` | BookProfile.author | Consensus from all sources |
+| `narrator` | BookProfile.narrator | Often from audio credits (L2) or Audnexus |
+| `series[].series` | BookProfile.series | From BookDB or API lookups |
+| `series[].sequence` | BookProfile.series_num | Position in series |
+| `publishedYear` | BookProfile.year | From API lookups |
+| `cover` | External API | Audnexus or Google Books cover URL |
+| `isbn` | API lookups | Google Books or OpenLibrary |
+| `asin` | Audnexus | Audible ASIN if available |
+| `duration` | Audio file analysis | From ffprobe or ABS library item |
+| `language` | Gemini detection | Audio language detection (L2) |
+| `description` | API enrichment | Google Books or OpenLibrary |
+
+### Confidence Translation
+
+LM uses 0-100 confidence scores. ABS doesn't have a native confidence field, but we can:
+1. Only return results above a configurable threshold (default: 60)
+2. Sort results by confidence (best first)
+3. Include confidence in a custom field for display
+
+### New LM Files
+
+```
+library_manager/abs_provider.py    # ABS metadata provider Blueprint
+  - /api/abs/search               # Search endpoint
+  - /api/abs/match                # Best-match endpoint
+  - /api/abs/identify             # Audio identification endpoint
+  - /api/abs/covers/<isbn>        # Cover proxy
+```
+
+### Authentication
+
+Options:
+1. **Shared secret** — LM generates an API key, user enters it in ABS custom provider config
+2. **Token exchange** — ABS sends its API token, LM validates via ABS `/api/me` endpoint
+3. **None (local only)** — If both run on same machine, trust localhost
+
+Recommend: Shared secret (simple, works across networks).
+
+### Configuration (LM Settings Page)
+
+Add to the Integrations tab:
+
+```
+ABS Metadata Provider
+├── Enable as ABS metadata source: [toggle]
+├── Provider API Key: [auto-generated, copyable]
+├── Minimum confidence threshold: [slider 0-100, default 60]
+├── Include audio identification: [toggle, default on]
+└── Provider URL for ABS config: http://<lm-host>:5757/api/abs
+```
+
+---
+
+## Security Considerations
+
+### What This Opens
+- A new API surface (`/api/abs/*`) accessible from the network
+- Audio clip uploads from ABS to LM
+- LM's metadata forwarded to ABS (already LM's core function)
+
+### Mitigations
+- API key authentication on all `/api/abs/*` endpoints
+- Rate limiting (reuse existing rate limiter infrastructure)
+- Audio clip size limits (same as existing Skaldleita submission limits)
+- Input validation on all query parameters (sanitize before pipeline)
+- SSRF protection already in place for plugin endpoints (issue #236)
+- No new filesystem access — ABS provider only reads/returns metadata
+
+### What NOT to Expose
+- Direct Skaldleita API passthrough (LM should be the intermediary, not a proxy)
+- BookDB API keys or signing credentials
+- Internal BookProfile objects (serialize to ABS format only)
+- Direct database access
+
+---
+
+## Implementation Phases
+
+### Phase 1: Search Provider
+- New Blueprint: `abs_provider.py`
+- `/api/abs/search` endpoint using existing `search_bookdb()` + API lookups
+- API key auth
+- Settings UI toggle
+- Basic test with ABS custom provider
+
+### Phase 2: Audio Identification
+- `/api/abs/identify` endpoint
+- Accept audio clips from ABS
+- Route through Skaldleita Whisper pipeline
+- Return structured results
+
+### Phase 3: Background Enrichment (Option B hybrid)
+- Scan ABS library via existing `abs_client.py`
+- Queue items with missing metadata
+- Push enrichments back to ABS
+- Configurable schedule (daily/weekly)
+
+### Phase 4: Cover Art + Descriptions
+- `/api/abs/covers/<isbn>` proxy endpoint
+- Cache covers locally
+- Pull descriptions from Google Books / OpenLibrary enrichment
+
+---
+
+## ABS Custom Provider Setup (User-Facing)
+
+In ABS Settings → Providers → Custom:
+```
+Name: Library Manager
+Base URL: http://<lm-ip>:5757/api/abs
+API Key: <paste from LM settings>
+```
+
+Then in ABS, when matching a book:
+1. Click "Match" on a library item
+2. Select "Library Manager" as provider
+3. See results from LM's multi-source pipeline
+4. One-click apply metadata
+
+---
+
+## Related Issues
+- #257 — This feature
+- #252-#256 — Integration fixes that improve the pipeline feeding this provider
+- abs_client.py — Existing read-only ABS integration (complements this)
diff --git a/library_manager/models/book_profile.py b/library_manager/models/book_profile.py
@@ -18,7 +18,10 @@
     'id3': 80,          # Embedded by producer/publisher
     'json': 75,         # Explicit metadata file
     'nfo': 70,          # Release info files
-    'bookdb': 65,       # Verified database match
+    'bookdb': 65,        # Verified database match
+    'bookdb_audio': 65,  # BookDB database match (same as bookdb)
+    'bookdb_transcript': 55,  # BookDB Whisper transcript only (unverified)
+    'bookdb_scrape': 45,     # BookDB live web scrape (unreliable)
     'ai': 60,           # AI verification
     'audnexus': 55,     # Audible data
     'googlebooks': 50,  # Google Books API

diff --git a/library_manager/pipeline/layer_audio_id.py b/library_manager/pipeline/layer_audio_id.py
@@ -714,18 +714,20 @@ def process_layer_1_audio(
                 c = conn.cursor()
 
                 # Update book with audio-identified info
-                sl_source = result.get('sl_source', 'audio_transcription')
+                # Use the mapped source name (bookdb_audio/bookdb_transcript/bookdb_scrape)
+                # for proper trust weighting in BookProfile, falling back to sl_source
+                profile_source = result.get('source', result.get('sl_source', 'audio_transcription'))
                 base_confidence = 85 if confidence == 'high' else 70
                 profile = {
-                    'author': {'value': author, 'source': sl_source, 'confidence': base_confidence},
-                    'title': {'value': title, 'source': sl_source, 'confidence': base_confidence},
+                    'author': {'value': author, 'source': profile_source, 'confidence': base_confidence},
+                    'title': {'value': title, 'source': profile_source, 'confidence': base_confidence},
                 }
                 if narrator:
-                    profile['narrator'] = {'value': narrator, 'source': sl_source, 'confidence': 80}
+                    profile['narrator'] = {'value': narrator, 'source': profile_source, 'confidence': 80}
                 if series:
-                    profile['series'] = {'value': series, 'source': sl_source, 'confidence': 75}
+                    profile['series'] = {'value': series, 'source': profile_source, 'confidence': 75}
                 if series_num:
-                    profile['series_num'] = {'value': str(series_num), 'source': sl_source, 'confidence': 75}
+                    profile['series_num'] = {'value': str(series_num), 'source': profile_source, 'confidence': 75}
 
                 # Issue #227: Save original author/title BEFORE updating the DB.
                 # If validation fails later, we must revert these to prevent