[BUG] Settings model picker only lists deprecated Gemini/Gemma models; Layer 4 fails and orphans queue items to needs_attention

[seannymurrs]

What happened?

The AI-provider model picker in Settings only offers deprecated Gemini/Gemma models. The currently-selected model gemma-3-27b-it returns 404 from the Google Gemini API because Google retired the Gemma 3 lineup when Gemma 4 launched (same pattern as when Gemma 2 was retired in favor of Gemma 3). None of the currently-available models — gemma-4-26b-a4b-it, gemma-4-31b-it, gemini-2.5-flash, gemini-2.5-pro, gemini-3-flash-preview, etc. — are selectable from the UI.

Side note: the configured fallback openrouter_model is google/gemma-3n-e4b-it:free, which is also a Gemma 3 variant and likely affected by the same deprecation cycle — so switching providers via the UI doesn't necessarily route around the bug.

The downstream impact is significant. When the configured AI model 404s, Layer 4 retries the batch 3 times, makes zero progress, then dumps every unresolved item into needs_attention. On my first scan, 532 of ~550 queue items ended up flagged that way purely because the AI fallback couldn't run — they look like "books library-manager can't identify" when they're actually "books library-manager never got to try AI on."

Worth noting: the Settings page reports Gemini: connected in API Connection Status even though generateContent is broken. The connection check isn't validating that the configured model actually works — it appears to be testing API key validity only.

Expected behavior

Either:

1. Update the hardcoded model lists (for both ai_provider: gemini and openrouter_model) to include current models, or
2. Dynamically populate the dropdowns by calling https://generativelanguage.googleapis.com/v1beta/models?key=<KEY> for Gemini and https://openrouter.ai/api/v1/models for OpenRouter against the user's configured keys. This is the more robust option since Google retires models on a fairly regular cadence — anything hardcoded will go stale again.

Additionally:

• The "connected" status check should call generateContent with the currently-configured model (not just validate the key) so this kind of misconfiguration is caught at the Settings page rather than at scan time.
• When every configured AI provider 404s with Model not found, surface that as a config error in the UI rather than silently orphaning hundreds of items as needs_attention. Right now there's no signal in the web UI that the problem is the model name and not the library content.

Steps to reproduce

1. Configure a valid Google Gemini API key in Settings → AI Providers.
2. Verify the Settings page reports Gemini: connected ✅.
3. Open the Gemini model dropdown — note that only legacy models (Gemma 3, older Gemini versions) are available.
4. Leave the default gemma-3-27b-it selected.
5. Run a deep library scan on a library with books that aren't cleanly identified by BookDB / Audnexus / OpenLibrary / Google Books / Hardcover (so Layer 4 actually gets invoked).
6. Layer 4 fails every batch with 404 Model not found; remaining items are marked needs_attention.

Environment

• App Version: 0.9.0-beta.134 (beta channel)
• Python: 3.11.15
• Platform: Docker (Podman on a Whatbox seedbox slot, Linux 6.12.40-gentoo-whatbox)
• Library size: ~500 books (1711 audio files)

Configuration (relevant fields)

{
  "ai_provider": "gemini",
  "enable_ai_verification": true,
  "enable_api_lookups": true,
  "openrouter_model": "google/gemma-3n-e4b-it:free",
  "ollama_model": "llama3.2:3b",
  "batch_size": 10,
  "max_requests_per_hour": 200,
  "update_channel": "beta"
}

API Connection Status reported by the UI:

• Gemini: connected (despite generateContent returning 404 — see above)
• BookDB: connected
• Google Books / OpenRouter / Audiobookshelf: not configured

Logs

2026-05-13 10:40:04,938 - WARNING - [PROVIDER CHAIN] All text providers failed
2026-05-13 10:40:04,939 - WARNING - No results from AI
2026-05-13 10:41:08,260 - WARNING - [PROVIDER CHAIN] All text providers failed
2026-05-13 10:41:08,261 - WARNING - No results from AI
2026-05-13 10:42:15,675 - WARNING - [PROVIDER CHAIN] All text providers failed
2026-05-13 10:42:15,676 - WARNING - No results from AI
2026-05-13 10:42:15,676 - INFO - AI providers circuit-broken (gemini) - waiting for recovery, not counting toward exhaustion

2026-05-13 10:43:18,897 - WARNING - Gemini: Model not found - the selected model may not exist
2026-05-13 10:43:18,897 - WARNING - Gemini detail: models/gemma-3-27b-it is not found for API version v1beta, or is not supported for generateContent. Call ListModels to see the list of available models and their supported methods.
2026-05-13 10:43:18,900 - WARNING - [PROVIDER CHAIN] All text providers failed
2026-05-13 10:43:18,903 - WARNING - No items processed but 538 remain (attempt 1/3)

2026-05-13 10:44:02,487 - WARNING - Gemini: Model not found - the selected model may not exist
2026-05-13 10:44:02,487 - WARNING - Gemini detail: models/gemma-3-27b-it is not found for API version v1beta, or is not supported for generateContent. Call ListModels to see the list of available models and their supported methods.
2026-05-13 10:44:02,491 - WARNING - No items processed but 538 remain (attempt 2/3)

2026-05-13 10:44:48,510 - WARNING - Gemini: Model not found - the selected model may not exist
2026-05-13 10:44:48,510 - WARNING - Gemini detail: models/gemma-3-27b-it is not found for API version v1beta, or is not supported for generateContent. Call ListModels to see the list of available models and their supported methods.
2026-05-13 10:44:48,513 - WARNING - No items processed but 538 remain (attempt 3/3)
2026-05-13 10:44:48,514 - INFO - Layer 4 cannot process remaining 538 items - marking as needs_attention
2026-05-13 10:44:48,526 - INFO - Marked 532 orphaned queue items as needs_attention
2026-05-13 10:44:48,529 - INFO - Layer 4 complete: 0 items processed, 0 fixed via folder fallback
2026-05-13 10:44:48,529 - INFO - === LAYERED PROCESSING COMPLETE: 550 processed, 13 fixed ===

Screenshots

---

References:

• Google's current Gemma-via-Gemini-API docs use gemma-4-26b-a4b-it in examples: https://ai.google.dev/gemma/docs/core/gemma_on_gemini_api
• List available models for a given API key: curl "https://generativelanguage.googleapis.com/v1beta/models?key=YOUR_API_KEY"

[deucebucket]

Thanks for bringing this to my attention, and thanks for the solid suggestion.

When I started this project I was still pretty green, and hardcoded model dropdowns were one of those "good enough at the time" choices that clearly doesn't hold up. I've also been using local models for the brunt of my own testing/work lately, so the hosted Gemini/OpenRouter model picker fell off my radar.

You're right that just updating the hardcoded list is the wrong fix. It would only go stale again. I've got a change now that replaces the static dropdowns with editable model fields backed by live provider model lists:

• Gemini models load from the Gemini models endpoint for the configured key.
• OpenRouter models load from OpenRouter's /models endpoint.
• The fields still allow manual model IDs, so users can paste a model name if the provider list is temporarily unavailable.
• The app no longer silently falls back to old hardcoded Gemini/OpenRouter model IDs.
• Connection tests now fail clearly if no model is configured.

I'm going to leave this issue open until you can verify it's fixed on your side. If I don't hear back after 3 days, I'll mark it complete and close it.

[deucebucket]

Implemented on main by commit 844bea7: Gemini/OpenRouter model choices are loaded dynamically, model IDs remain editable, hardcoded hosted-model fallbacks were removed, and connection tests now exercise the configured model. Closing per the maintainer's verification-window comment after no contrary report.