Merge pull request #270 from deucebucket/fix/openai-compatible-provider

fix: make local AI providers model-agnostic

Author: deucebucket

.github/workflows/code-quality.yml

@@ -50,8 +50,11 @@ jobs:
           python -m pip install --upgrade pip
           pip install -r requirements.txt
 
-      - name: Run Naming Tests
-        run: python test-env/test-naming-issues.py
+      - name: Run Tests
+        run: |
+          python test-env/test-naming-issues.py
+          python test-env/test-ai-providers.py
+          python test-env/test-ai-provider-integration.py
 
   version-check:
     runs-on: ubuntu-latest

CHANGELOG.md

@@ -2,6 +2,17 @@
 
 All notable changes to Library Manager will be documented in this file.
 
+## [0.9.0-beta.155] - 2026-07-15
+
+### Added
+- **Generic local AI provider** - Added a user-configured OpenAI-compatible provider for llama.cpp, LM Studio, vLLM, LocalAI, and similar servers. Library Manager discovers live model IDs from `/v1/models`, supports optional bearer authentication, and uses `/v1/chat/completions` across text verification, localization, and transcript identification.
+
+### Fixed
+- **Ollama model selection is no longer locked to a hardcoded model** - Removed runtime and UI fallbacks to specific Ollama model names. Model discovery now accepts `name`, `model`, and string response shapes, excludes embedding-only entries, keeps the model field editable, and auto-selects only when the server exposes exactly one model. Native Ollama requests use bounded structured output, including an explicit array schema for multi-book batches, and both local-provider paths prevent runaway responses.
+- **Skaldleita transcripts now reach AI fallbacks** - When Skaldleita transcribes audio but cannot identify the book, the transcript continues through the configured provider chain instead of ending the audio-identification attempt early.
+- **Provider selection configures working fallback chains** - Setup and Settings automatically place the selected AI provider into the text and audio fallback chains while preserving Skaldleita as the first hosted audio-identification service.
+- **Provider secrets stay server-side** - New and updated `secrets.json` files enforce `0600` permissions, provider keys remain excluded from `config.json`, and Settings receives only configured/not-configured state instead of raw stored tokens.
+
 ## [0.9.0-beta.154] - 2026-07-14
 
 ### Fixed

README.md

@@ -16,6 +16,11 @@
 
 ## Recent Changes (stable)
 
+> **beta.155** - **Local AI Is Provider-Agnostic**
+> - Added llama.cpp and generic OpenAI-compatible API support with live `/v1/models` discovery.
+> - Removed hardcoded Ollama model fallbacks and fixed `undefined` model entries across API response shapes.
+> - Skaldleita transcripts now continue into the selected AI fallback instead of stopping early.
+
 > **beta.150** - **Fix: Hosted AI Model Picker Uses Live Provider Models** (Issue #216)
 > - Replaced stale hardcoded Gemini/OpenRouter model dropdowns with editable model fields backed by live provider model lists.
 > - Gemini models load from the configured Gemini key; OpenRouter models load from OpenRouter's `/models` endpoint.
@@ -419,6 +424,14 @@ See [docs/DOCKER.md](docs/DOCKER.md) for detailed setup guides.
 - Multiple model options
 - Free tier available
 
+**Ollama**
+- Self-hosted with live model discovery from the user's server
+- No hardcoded model requirement
+
+**llama.cpp / OpenAI-compatible**
+- Works with llama.cpp, LM Studio, vLLM, LocalAI, and compatible `/v1` APIs
+- Live model discovery and optional API-key authentication
+
 ---
 
 ## API Reference
@@ -561,7 +574,7 @@ The idea for **Skaldleita** (audio fingerprinting + narrator voice identificatio
 ## Contributing
 
 Pull requests welcome! Ideas:
-- [ ] Ollama/local LLM support
+- [x] Ollama and generic local LLM support
 - [ ] Cover art fetching
 - [x] Metadata embedding (added in v0.9.0-beta.20)
 - [ ] Movie/music library support