Anvira release notes

• Updates by themselves. When a newer Anvira is published, a small button appears at the top right. Click it to download, then again to restart into the new version. Nothing downloads or restarts until you click. Settings → About → Check for updates looks right now.
• Runtime updates. A second button appears when a newer Anvira Runtime is available. It updates the AI engine in place; your models and settings are kept.
• Optional account. Sign in or create an account from the sidebar or Settings → Account. Everything still works without one, and your chats and files stay on your PC.

• The window now always fits your screen and remembers where you left it (it could hang off the right or bottom edge).
• The composer "+" menu and dialogs are no longer cut off on small screens.
• The Search page quick-jump buttons work again.
• The start-up animation is smoother and the first-run dialog waits for it.
• Memory recall across chats works.
• A dropped connection to the AI engine now says what happened ("The AI engine stopped in the middle of the answer") with a hint, instead of a bare "terminated".
• Very long messages fold up behind Show more so one big paste does not take over the screen.

Anvira is an agent workspace that runs on your own PC. Chat with local AI models, point an agent at a project folder, and keep everything on your computer. It is built on Anvira Runtime and is a LocalHouseLLM product.

• Chat with streaming replies, markdown and code blocks, stop / regenerate / edit.
• Agent mode: link a project folder and let Anvira read, create and edit files and run commands in it. Every agent run shows a real before/after diff with Undo.
• Model library: search Hugging Face, filter by CPU / GPU / light / heavy / large, and download models with live progress.
• Projects, scheduled tasks, rules, skills and memory, all stored locally.
• Anvira Runtime included: the installer ships the runtime, so first launch works offline. You choose where it is installed. An optional NVIDIA GPU pack can be downloaded from the runtime release.

• Anvira 1.0 keeps its own data folder (%APPDATA%\Anvira Agent).
• Uninstalling keeps your conversations, Anvira Runtime and downloaded models.
• Runtime: Vansh-synthetica/Anvira-Runtime

Real web access for local models, and a lot more reliable tool-calling in general.

New: web_search (DuckDuckGo, falling back to Bing) and web_fetch (reads a page's real content and every link on it, so the agent can follow a link to a second, third, etc. page — genuine multi-hop crawling, not just one search). Both work fully offline-first: nothing reaches the internet unless the agent actually calls them, and they're gated by the same approval policy as running a command or pushing to git.

Fixed, all confirmed live against a real local model, not just unit tests:

• web_search/web_fetch were completely unreachable from any real run — a capability-wiring gap, now fixed.
• A routing bug sent "search the web for X" to the local-files-only search tool instead, because the phrase matched a local-search pattern first. Fixed with an explicit web/internet-signal check that wins the ambiguity.
• Small/quantized models could describe an answer in prose instead of ever calling a tool, even when tools were clearly needed and available. The response schema itself now removes that option until a real tool call has actually succeeded — not just attempted.
• A tool call that returns nothing (e.g. searching the wrong place) now gets an actionable hint attached, instead of silently looking like "there's nothing here."
• Tool name resolution now handles camelCase/PascalCase ("writeFile", "WriteFile") and many more wrapper-word variations, on top of the existing typo/alias resolver.
• A local file-search crash on a single unreadable file (e.g. a path exceeding Windows' 260-char limit) no longer aborts the whole search.

Study chat and the Notes "Ask AI" panel can now actually use tools (including the web) via a new "Web Research" toggle — previously both were pure text-completion with zero tool access at all.

Also: Model Library's size-tier filter (Light/CPU-friendly) now scans the real catalog instead of just whatever happened to be on the current page; softened the overly bright light theme in Study and Notes; fixed a rendering bug where AI-inserted content could show a literal <br> instead of a line break.

Fixes agentic tool-calling reliability for local models: small/quantized models were structurally able to choose describing an answer over actually calling a tool, even under JSON-schema-constrained decoding — now the schema itself removes that option until a real tool call has been made. Also fixes a false-positive 'task completed' bug (the model claiming success in prose with zero tool calls actually executed), a lost-context bug in multi-step task planning, and adds a tool-name alias/fuzzy resolver so a model calling e.g. "write_tool" instead of "write_file" still resolves correctly.

Bundles both CPU and CUDA llama-server builds — works on any Windows PC, with or without an NVIDIA GPU, fully offline on first model activation.

Continues from 1.6.1 (mmap crash fix, bundled CUDA runtime).