v1.0.3-beta

[AronDaron]

Added

• Local LLM support — Ollama, LM Studio, llama.cpp, or any custom OpenAI-compatible endpoint. Pipeline now works fully offline.
• Auto-detect — one-click scan for local LLM endpoints on ports 11434 / 1234 / 8080.
• Mixed mode — pick generation and judge models from different providers per category (e.g. local Qwen3-14B for gen, cloud judge for quality control).
• Providers UI — new Settings → Providers tab with full CRUD, per-row test connection, and inline API-key edit.

Changed

• Settings → "API Keys" tab replaced by "Providers" (legacy /api/settings/api-key still works as a shim).
• Quality Report token column renamed to "Output Tokens by Category" and now counts only generated tokens (was double-counting prompt + completion).
• Default provider is auto-promoted when the current default is disabled, with a toast notification.

Fixed

• Cancel on local LLM jobs is now instant (~1 s) instead of waiting for the in-flight chat call to finish (could be several minutes on a 14B model).
• Reasoning models (Qwen3, DeepSeek-R1) on Ollama / LM Studio no longer falsely fail the token budget check — the pipeline normalises usage counts when providers don't break out reasoning tokens.
• Format C "reasoning starvation" (model burns the budget on <think> and emits empty content) now auto-retries with a 4× budget instead of skipping.
• Model picker refreshes immediately after changing the default provider in Settings (no more F5-equivalent restart needed in the desktop app).

Migrations

• v6 — new providers table; existing OpenRouter API key is backfilled silently into providers.openrouter-default.
• v7 — clears cached Quality Report snapshots so historical jobs recompute with the corrected token SQL.

Both run automatically on first launch. No manual action needed.

---

This discussion was created from the release v1.0.3-beta.