v 0.12.6

Release date: 2026-07-26

This release makes the Eye work with whatever model the investigator already has a key for, and turns imported evidence into a first-class, tamper-evident part of the case. You can now connect OpenRouter, NVIDIA, Groq, Mistral, xAI (Grok), DeepSeek, and Kimi alongside OpenAI/Anthropic/Gemini — a single API key each — and the Eye picks the most capable tool-calling model automatically. Every piece of external evidence you bring in — a database, a third-party report, an email export, browser-tool output — is now hashed, listed with a live integrity check, and readable by the Eye. Plus a broad responsiveness pass so the model menu and suggested actions feel instant.

🔌 New: Bring-your-own-model — seven more providers

Connect the AI backend you already use. Every new provider is a single-key, OpenAI-compatible setup in Settings.

• Added providers — OpenRouter, NVIDIA, Groq, Mistral, xAI (Grok), DeepSeek, and Kimi (Moonshot), alongside the existing OpenAI, Anthropic, and Gemini. OpenRouter alone fans out to essentially every model behind one key.
• Fixed Google AI Studio / Gemini key detection — the model list came back empty on the modern google-genai SDK, making a valid AIza… key look broken. It now discovers models correctly, and any provider falls back to a built-in "common models" list when live detection is unavailable.
• Smarter recommendations — because the Eye is agentic (it calls forensic tools), the recommended model for each provider is now the most powerful one that supports tool calling (e.g. DeepSeek-Chat over the non-tool Reasoner). Every provider ships an offline "common models" quick-pick.

🔀 New: Model switching that verifies itself

• Switching model or provider now validates the key and model on a background thread before committing. If it can't connect, the Eye reverts to the previous model and tells you exactly why — no silent hangs.
• The switch persists the correct backend configuration for every provider (previously only OpenAI/Anthropic/Gemini were mapped correctly).

🗂️ New: Imported Evidence window + chain of custody

• A dedicated Imported Evidence window (top bar and Case Settings) lists every imported item — databases and documents — with SHA-256 hashes of both the original file and the in-case copy, size, source path, and a live integrity verdict (Verified / mismatch / missing). Hashing runs off the UI thread, so multi-GB databases never freeze the app.
• Import reports, email, and browser-tool output verbatim — PDF, HTML, TXT/MD/LOG, and email exports (.eml/.mbox) are copied into the case without conversion (because third-party output is often a report, not a table), hashed, and readable by the Eye via the new read_imported_evidence tool.
• Imported evidence — databases and documents — now appears in the Compliance activity stream as IMPORT entries with their hashes, and the list auto-refreshes the moment an import finishes.

🧭 Improved: Narrative Map shows the real evidence

• Evidence cards now display the actual source rows — auto-loaded in the inspector and in the double-click detail popup — for both native and imported evidence, correctly resolving native-vs-imported databases that share a filename.
• Evidence produced by a non-database tool or a text-mode (non-function-calling) model now shows its full captured text instead of a dead-end "no source" message.
• Compliance entries deep-link to their Verdict / Narrative / Evidence card in the Narrative Map.

⚡ Improved: Responsiveness

• The model dropdown opens instantly. It used to freeze the UI on every open (a live network call plus up to ten OS-keychain reads on the GUI thread). It now opens immediately from a cached list and refreshes in the background.
• Suggested-action chips run on a single click with an instant running indicator; a small pencil icon inserts the action into the message box to edit first. Typing in long conversations no longer lags.

🔧 Fixes

• AI setup — "Validate & Save" now always works. Configuring a model backend could silently do nothing if you pressed Back and then Next in the setup wizard. The final step now reliably validates the connection and saves your configuration regardless of how you navigated the wizard.
• Local Ollama server on a LAN or custom address is now honored. When Ollama was configured as a Local Server, the Eye ignored the address you entered and always tried localhost:11434. It now connects to the endpoint you configured — a dedicated AI server on your LAN or a non-default port.
• Setup wizard shows a "Validating…" indicator. Clicking Validate & Save now shows a live progress indicator while the connection is checked, and the check runs off the UI thread so the window no longer freezes on a slow or unreachable backend.