⚡ Startup & Prompt Caching

Startup reaches the first token sooner — Embedding warm-up overlaps with chat restoration and startup watchdogs, reducing the reported median boot-to-first-token time by about 27%.
Chat switching does less waiting — Model and profile refreshes run independently of snapshot rendering instead of holding up the visible chat.
Long Responses conversations keep their cacheable prefixes — Completed turns retain masked extras at their original boundaries. A live test reached a 99.3% cache hit rate at 131K input tokens.
Existing instances receive the SDK upgrade automatically — Startup migration updates LiteLLM to 1.104.0 and the OpenAI SDK to 2.54.0, with installed-SDK HTTP test coverage.

🌐 Browser

Browser tools skip unnecessary work — Empty tab listings no longer launch Chromium, redundant load waits are removed, and the viewer reuses its native display size. Reported local navigation medians fall from 474 to 83 milliseconds, and content reads from 82 to 19 milliseconds.
Quick settings focus on everyday controls — Preset, Host/Container, zoom, and extensions stay close at hand; extension management and Chromium settings pages move into Browser config.
Screenshots and tooltips are clearer — Annotation screenshots become an opt-in setting, and overlapping quick-settings tooltips are removed.

Chats no longer disappear when Sidebar Folders is disabled — Importing the folders store from another plugin no longer hides the chat list.
Folder activity updates are faster — A shared parent-child index replaces repeated full-context scans, reducing the reported median from about 67 to 15 milliseconds.
Sidebar filters survive reloads — Project, Status, and Last Activity selections stay saved, while continuing a chat correctly updates its activity timestamp.
Deep delegation stays visible — Selecting a subagent reveals its children at every delegation depth.
Names and saved chats are better protected — Automatic naming preserves manual names even in "always" mode, custom names are validated before being marked manual, and pruning re-checks disk before removing a just-saved context.

Preset editing starts from fresh data — Presets reload before the editor opens so a stale draft cannot delete presets created in another tab or through the API. This fixes #1928.
Core providers use native LiteLLM routing — Nebius, Venice chat, Z.AI/Coding, and OpenRouter embeddings use built-in integrations while retaining existing overrides.
Claude guidance works across providers — Thoughts-prompt adaptation recognizes Claude models through OpenRouter, Vertex AI, Bedrock, and other providers, not only provider names containing "anthropic".
Context and pricing displays are configurable — Context-window display controls and price-currency selection join viewport-bounded preset and profile menus; currency conversion uses Frankfurter rates cached for 24 hours.
Cache routing and repeat checks are more consistent — Venice and Agent Zero API cache routing are enabled, custom providers can use a0_prompt_cache_mode, and native tool calls regain full-response repeat checks including commentary.

Reconnect without a manual disconnect — Rejected refresh tokens produce actionable feedback and a Reconnect option.
Device codes are easier to copy — Challenge codes appear in a copyable box with accessible success feedback.
Codex compatibility catches up — The client version is pinned to 0.160.1.

Subordinates own their execution state — Direct delegation runs through the child's context task, making child chats visible as Working/Paused and independently stoppable. Caller cancellation propagates, while stopping only the child is recoverable.
Terminal work no longer blocks shared event loops — POSIX PTY reads are nonblocking and writes cooperate with other tasks; native Windows TTY cleanup avoids POSIX-only SIGKILL handling.
Long SSH commands arrive intact — Complete commands, including large heredocs, are sent with sendall through asyncio.to_thread.
Host Computer Use starts when needed — The first action automatically starts the session within the same tool call.

User plugin settings keep precedence — Bundled values no longer overwrite community provider and profile definitions or subordinate profile settings during merges.
Scheduler prompts have room to breathe — Full-width resizable fields and an expanded modal make editing easier, with optional tasks.yaml files supported alongside JSON.
Local speech gains Paradee Heart — The default Local TTS voice joins Kokoro, using lazy CPU ONNX synthesis and cached model assets.
Folder downloads keep useful names — A single-folder download uses the folder's name instead of a generic timestamped archive name.

Dependency security floors are raised — Updates cover aiohttp, anyio, urllib3, filelock, fsspec, and multidict.
Normal container boots skip unnecessary APT refreshes — Package indexes are seeded only when absent.
Fallback behavior receives more coverage — Tests cover Chat Completions reasoning fallback, and the README's Launcher downloads now point to v1.9.