Changelog

What's Changed

• Repin vmlx-swift to a983c6fc so the reasoning ceiling reaches the app (#2382) by @jjang-ai
• Repin vmlx-swift to 294c11e9 so LFM2.5-VL can ship (#2379) by @jjang-ai
• Repin vmlx-swift to 11a43493 (#2377) by @jjang-ai
• Repin vmlx-swift to 3c354fb0: the repetition guard now reaches this app (#2362) by @jjang-ai
• Fix #2327: a reasoning-only turn with tools offered was classified as a finished answer (#2344) by @jjang-ai
• Repin vmlx-swift to 751a5779 (#2361) by @jjang-ai
• Nemotron 3.5 Lightning: native MTP head + fix the fallback template eating its newlines (#2360) by @jjang-ai
• Let TTFT phase tracing be turned on in release (#2357) by @jjang-ai
• Repin vmlx: fix the Muse tool-recipient header leak (#2356) by @jjang-ai
• Repin vmlx: stop generation on a verbatim output cycle (#2353) by @jjang-ai
• Repin vmlx to the Muse Glimmer centered-norm fold (#2342) by @jjang-ai

🚀 Features

• add follow up question suggestions after a turn (#2384) by @RaajeevChandran
• added full screen preview for generated images (#2381) by @RaajeevChandran
• allow deleting individual assistant messages (#2371) by @RaajeevChandran
• added a delete all data option to agent settings (#2355) by @RaajeevChandran
• allow creating new folders in folder selection panels (#2349) by @RaajeevChandran
• pass full frontmatter through in read_knowledge output (#2338) by @RaajeevChandran

🐛 Bug Fixes

• fix knowledge search timeouts and improve what gets indexed (#2383) by @RaajeevChandran
• fix skill editor sheet clipping on small screens (#2380) by @RaajeevChandran
• open schema ahead chat history databases instead of refusing them (#2375) by @RaajeevChandran
• Give a quarantined plugin one retry after it or the app is upgraded (#2378) by @jjang-ai
• Stop agent plugin reads from depending on the Plugins UI being opened (#2372) by @jjang-ai
• Give host-folder artifact sharing the basename fallback it documents (#2370) by @jjang-ai
• Start the TTFT trace when the user hits send, not when generation does (#2369) by @jjang-ai
• Render $…$ inline math instead of showing its delimiters (#2368) by @jjang-ai
• minor UI polish (#2365) by @RaajeevChandran
• skip workspace baseline clones for oversized folders (#2359) by @RaajeevChandran
• Give the orphaned local-model lock a way out (#2363) by @jjang-ai
• Route a lone mac_query string to question when that field is absent (#2352) by @jjang-ai
• Accept a string contents on mac_query instead of rejecting it (#2351) by @jjang-ai
• Repin vmlx: parse GLM/DeepSeek tool calls that take no arguments (#2346) by @jjang-ai

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.22.20...0.22.21

What's Changed

• Repin vmlx to the Muse Glimmer vision fix (#2341) by @jjang-ai
• Fix the Muse Glimmer vision path: placeholders never rendered or matched (#2340) by @jjang-ai
• Prove all four Muse Glimmer reasoning strengths reach the template (#2339) by @jjang-ai
• Bound capability group-load schemas + repin vmlx (Muse Glimmer runtime) (#2336) by @jjang-ai

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.22.19...0.22.20

What's Changed

• Repin vmlx: capture cache boundaries during prefill (fixes a family that could never populate its cache) (#2335) by @jjang-ai
• Download large model shards as parallel HTTP Range requests (#1960) by @jjang-ai
• AppleScript subagent: tell the caller the repair that actually works (#2334) by @jjang-ai
• Repin vmlx: refuse truncated model shards, and stop the post-answer stall on every warm turn (#2333) by @jjang-ai
• Fix first-turn double prefill (warmup cache-salt mismatch), warmup tool-scope kill, cache-size settings hazards; repin vmlx (#2331) by @jjang-ai
• DSV4 Flash integration: vmlx repin (fastpaths + weight-leak fix), warmup triple-prefill fix, reload memory verdict (#2328) by @jjang-ai

🐛 Bug Fixes

• fix context impact ability count chip wrapping to two lines (#2323) by @RaajeevChandran
• Refresh MCP catalogs and streamline Tools management (#2322) by @tpae
• fix tts drift on long text with per sentence synthesis (#2321) by @RaajeevChandran

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.22.18...0.22.19

What's Changed

🐛 Bug Fixes

• fix diff card header alignment (#2314) by @RaajeevChandran
• route skylight input through SLEventPostToPSN on macOS 26.4+ to prevent crashes (#2316) by @RaajeevChandran
• Fix MCP capability indexing and custom agent scope (#2308) by @tpae
• fixed main thread hangs (#2307) by @RaajeevChandran
• Raptor JANG_4M agent fixes + context_window discovery key (#2305) by @jjang-ai
• honor context windows advertised by custom provider /models endpoints (#2304) by @RaajeevChandran

🧰 Maintenance

• Fix remaining sandbox reachability gaps (#2313) by @tpae
• Fix sandbox plugin activation and dependency installs (#2311) by @tpae
• Clarify capability IDs and recover grouped tool calls (#2309) by @tpae
• Clarify model memory estimates and fit guidance (#2310) by @tpae
• Improve agent abilities context guidance (#2306) by @tpae
• Allow importing private Hugging Face models (#2303) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.22.17...0.22.18

What's Changed

• Prove LFM2.5-2.6B (JANG_6M + MXFP8) live and repin vMLX to the short-conv cache fixes (#2291) by @jjang-ai

🚀 Features

• add cmd+n shortcut option to start a new chat in the current window (#2296) by @RaajeevChandran
• Add native WhatsApp agent channel (WhatsApp Web bridge) (#2290) by @tpae
• Add "System" default theme card that follows macOS appearance (#2286) by @tpae

🐛 Bug Fixes

• Extend release timeout for slow hosted runners (#2302) by @tpae
• Fix custom agent capability dispatch (#2300) by @tpae
• surface specific keychain errors and recover from acl denied credential saves (#2295) by @RaajeevChandran
• race the warmup residency preflight against cancellation (#2297) by @RaajeevChandran
• Fix DSV4 thinking default, duplicate pre-send warmups, and 1M context resolution (#2293) by @jjang-ai
• fix agent recovery for deferred tools and knowledge reads (#2287) by @RaajeevChandran

🧰 Maintenance

• Fix production crashes and hangs from Sentry triage (#2285) by @tpae
• Improve activation telemetry: brain_source coverage, capability dimension, settings engagement (#2284) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.22.16...0.22.17

What's Changed

• chat import improvements (#2251) by @RaajeevChandran
• Add optional shadow style for chat input (#2256) by @hkay-dev
• Repin vMLX to the disk-store offset guard (#2276) by @jjang-ai
• Pin that an agent temperature of 0 forces argmax and makes top_p inert (#2274) by @jjang-ai
• Report why the disk cache was disabled instead of silently downgrading (#2273) by @jjang-ai

🚀 Features

• Add cloud image and video generation (#2281) by @tpae
• added /title slash command to name chats on demand (#2277) by @RaajeevChandran

🧰 Maintenance

• Add one-time post-onboarding chat history import prompt (#2283) by @tpae
• Expand onboarding tool picker with more curated tools (#2282) by @tpae
• Remove generative greetings (#2278) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.22.15...0.22.16

What's Changed

• Repin vMLX to the boundary-digest diagnostic (#2270) by @jjang-ai
• Repin vMLX to the DSV4 encoder/cache hardening fixes (#2269) by @jjang-ai
• Pin DSV4 disk-prefix fix and improve eval proof (#2264) by @jjang-ai
• Pin vMLX DSV4 warmup replay fix (#2262) by @jjang-ai
• Add DeepSeek V4 Flash 0731 app support (#2260) by @jjang-ai

🚀 Features

• Add LLM-powered context compaction (#2265) by @tpae
• Add live task status indicators to the History sidebar (#2263) by @tpae
• Add redeemable credit codes (#2259) by @tpae

🐛 Bug Fixes

• Wait for the localModelsChanged notification instead of assuming a sleep covered it (#2272) by @jjang-ai
• Stop cancellationDuringWave racing a fixed sleep against job completion (#2271) by @jjang-ai
• Redeem code follow-ups: settle first-action gate from Credits, 429 countdown, identity error copy (#2261) by @tpae

🧰 Maintenance

• Revamp the default Osaurus agent: bundled guide, settings/watcher tools, expanded read scopes (#2268) by @tpae
• Harden capability readiness and authorization (#2258) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.22.14...0.22.15

What's Changed

• Fix explicit model unload and cache telemetry continuity (#2248) by @jjang-ai
• i18n(zh-Hans): call Osaurus's own server 服务端, not 服务器 (#2246) by @YspritanHyzygy
• i18n: split the "Server" key, which three unrelated UI elements share (#2243) by @YspritanHyzygy
• i18n(zh-Hans): keys that differ only in casing should not say different things (#2230) by @YspritanHyzygy
• Fix four reported runtime and model metadata issues (#2235) by @jjang-ai
• Unify subagent and server batching concurrency (#2234) by @jjang-ai

🐛 Bug Fixes

• Harden AgentDB imports and saved views (#2255) by @tpae
• Exclude warm-up calls from Insights (#2254) by @tpae
• Stabilize execution modes, prompts, and cache warmup (#2253) by @tpae
• Prevent stale whole-file rewrites after targeted edits (#2252) by @tpae
• fixed text selection in agent responses (#2247) by @RaajeevChandran
• fix tool permission panel sizing and card appearance issues (#2244) by @RaajeevChandran
• fix tool approval dialog overflow and add auto-allow all tools setting (#2241) by @RaajeevChandran

🧰 Maintenance

• Make the Mac harness leaner and more reliable (#2250) by @tpae
• Keep notch overlay pinned to the screen edge (#2240) by @tpae
• Eliminate app hangs: unblock MainActor, bounded cancellation, killable plugin host (#2239) by @tpae
• Add localized Channels What's New announcement (#2238) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.22.13...0.22.14

What's Changed

• fix(i18n): lint-zh-style.py refuses to run again — the catalog gained a final newline (#2229) by @YspritanHyzygy
• i18n: add 16 UI strings localized inside helper functions but never in the catalog (#2224) by @YspritanHyzygy
• i18n(zh-Hans): one word per concept, where the rest of the catalog already agrees (#2225) by @YspritanHyzygy
• fix(i18n): make lint-zh-style.py runnable, and stop it breaking shell commands (#2223) by @YspritanHyzygy
• Add safe heterogeneous subagent batching and delegation controls (#2221) by @jjang-ai
• i18n(zh-Hans): use one Chinese word for a UI tab (#2215) by @YspritanHyzygy
• i18n(zh-Hans): one second-person pronoun, and translate "Insights" (#2218) by @YspritanHyzygy
• Fix selected-chat idle warmup recovery (#2216) by @jjang-ai
• i18n(zh-Hans): use one Chinese word for "transcription" (#2207) by @YspritanHyzygy
• i18n(zh-Hans): add the missing space between Chinese and Latin runs (#2209) by @YspritanHyzygy
• i18n: add 22 UI strings that are localized indirectly but never entered the catalog (#2213) by @YspritanHyzygy

🚀 Features

• Add native iMessage agent channel with download-only imsg helper (#2227) by @tpae
• added chat import guide (#2222) by @RaajeevChandran
• import conversations from chatgpt, claude, grok and gemini (#2217) by @RaajeevChandran
• fix fireworks providers only showing 6 models (#2214) by @RaajeevChandran

🐛 Bug Fixes

• Make headless dispatches follow the agent's current default model (#2232) by @tpae
• Make web_search schema immutable so tool payloads never invalidate the KV-cache prefix (#2226) by @tpae
• fix app freeze when saving channel tokens in keychain (#2208) by @RaajeevChandran

🧰 Maintenance

• Make the model-selector dot reflect true model residency (#2231) by @tpae
• Fix plugin layout at compact widths (#2220) by @tpae
• Fix OsaurusAI model catalog cutoff (#2219) by @tpae
• Harden keychain reliability on relaunch; add identity restore from recovery phrase (#2212) by @tpae
• Redesign channels UX: name-first destinations, DM identity, intent-based setup (#2210) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.22.12...0.22.13

What's Changed

• Match assistant and user chat text sizing (#2201) by @hkay-dev
• fix(i18n): two render sites that bypass localization (#2199) by @YspritanHyzygy
• Pin Nanbeige vMLX runtime (#2198) by @jjang-ai
• Wire remaining OsaurusCore network call sites into global proxy (#2169) by @peterlodri-sec
• i18n(zh-Hans): unify Computer Use naming, fix a label collision and four mistranslations (#2195) by @YspritanHyzygy
• Fix stale settings warmup and add SSD cache regression eval (#2194) by @jjang-ai
• i18n(zh): fix malformed format specifier in channel member count (#2176) by @YspritanHyzygy
• i18n: add 43 Settings-search strings that never entered the catalog (#2193) by @YspritanHyzygy

🐛 Bug Fixes

• fall back to lenient decode for nonstandard streaming chunks during tool calls (#2205) by @RaajeevChandran
• fixed app hangs (#2203) by @RaajeevChandran
• added guardrails for factory reset auto quit (#2200) by @RaajeevChandran

🧰 Maintenance

• Native channel presence, formatting, and reactions (#2202) by @tpae
• Clarify channel messaging UX with Replies/New Messages model and localize the channel surface (#2196) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.22.11...0.22.12

What's Changed

• Fix stale Todo agent-loop completion and strengthen cache proof (#2192) by @jjang-ai
• i18n(zh-Hans): quality pass — 859 fixes across eight defect classes + style linter (#2188) by @YspritanHyzygy
• Add scored agent-loop and disk-cache proof gates (#2179) by @jjang-ai
• Update Traditional Chinese translations (#2178) by @ftzahao
• Stop completed agent runs from reopening on stale todos (#2175) by @jjang-ai
• Update Simplified Chinese and Traditional Chinese translations (#2172) by @ftzahao
• Fix tracked tool completion and prompt cache stability (#2171) by @jjang-ai
• fix: don't install CLI symlink into Homebrew's managed bin directory (#2166) by @peterlodri-sec
• Add connectivity-driven offline mode for chat (#2170) by @tpae
• Unify runtime policy and bounded subagent delegation (#2165) by @jjang-ai
• Redact secret tool arguments across agent surfaces (#2157) by @mimeding
• Report MCP tool envelope failures (#2156) by @mimeding
• Preserve SSD warmup across chat lifecycle transitions (#2161) by @jjang-ai

🚀 Features

• inject current date and timezone into the latest user turn (#2173) by @RaajeevChandran

🐛 Bug Fixes

• fix factory reset hanging indefinitely and add reset progress ui (#2191) by @RaajeevChandran
• Fix Gemma and Qwen post-tool completion (#2167) by @jjang-ai

🧰 Maintenance

• Add proactive agent channel publishing with zero-config destinations (#2190) by @tpae
• Channels: focused setup UX, scalable access lists, and autosave (#2164) by @tpae
• improved chat UX (#2163) by @RaajeevChandran
• Channels: multi-agent routing and guided setup across all providers (#2162) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.22.10...0.22.11

What's Changed

• Remove binary prefill proof artifacts (#2135) by @jjang-ai
• Fix Bonsai DB recovery and SSD prefix reuse (#2153) by @jjang-ai
• Fix TextEdit AppleScript state and completion handling (#2151) by @jjang-ai
• Fix tool finalization and SSD cache warmup (#2146) by @jjang-ai

🚀 Features

• auto generate chat titles (#2144) by @RaajeevChandran
• group consecutive thinking and tool activity into one expandable row (#2142) by @RaajeevChandran

🐛 Bug Fixes

• fix settings window opening offscreen with stale saved frame (#2154) by @RaajeevChandran
• fix streaming error on deepseek and other openai compatible providers (#2150) by @RaajeevChandran
• fix mcp tool calls failing after invoking a skill via slash command (#2149) by @RaajeevChandran
• fix settings window not opening on sequoia (#2147) by @RaajeevChandran

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.22.9...0.22.10

What's Changed

• Update Simplified Chinese and Traditional Chinese translations (#2141) by @ftzahao
• Fix stream finalization and model-picker thinking control (#2140) by @jjang-ai
• Drain cancelled prefill before the next chat (#2136) by @jjang-ai
• Pin LFM SSD prefix restore (#2134) by @jjang-ai
• auto categorize knowledge documents and soften the uncategorized warning (#2132) by @RaajeevChandran
• Fix Computer Use and AppleScript completion regressions (#2131) by @jjang-ai
• Context-optimization harness: attribution, experiment profiles, staged optimizer, strict promotion gate (#2130) by @tpae
• Wire Laguna S2.1 runtime and live cache controls (#2127) by @jjang-ai
• seatbelt sandbox fallback for macos 15 and earlier (#2126) by @RaajeevChandran
• Restore SSD warmup prefixes across new chats (#2124) by @jjang-ai
• Fix launch-time SSD prefix cache misses (#2119) by @jjang-ai
• Fix AppleScript completion and persistent SSD warmups (#2118) by @jjang-ai
• knowledge base improvements (#2117) by @RaajeevChandran

🚀 Features

• Native Browser Use: replace the osaurus.browser plugin with a built-in browser feature (#2123) by @tpae

🐛 Bug Fixes

• fixed app hangs (#2128) by @RaajeevChandran
• Attach SSD cache before publishing loaded models (#2122) by @jjang-ai
• fix settings window not opening on sequoia (#2121) by @RaajeevChandran

🧰 Maintenance

• Add Browser Use What's New and prune older release notes (#2139) by @tpae
• Expand BrowserUse eval suite to 10 cases (safety, tool coverage, error recovery) (#2125) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.22.8...0.22.9

What's Changed

• Pin explicit Gemma 4 paged cache support (#2109) by @jjang-ai
• Preserve visible generation telemetry across chat warmups (#2106) by @jjang-ai
• fix: honor reasoning toggle across agent tool loops (#2105) by @jjang-ai
• Preserve explicit Thinking mode across agent tool turns (#2101) by @jjang-ai

🚀 Features

• multi select chats to delete/archive in sidebar with Cmd+Click shortcut (#2114) by @RaajeevChandran
• pin chats in sidebar (#2113) by @RaajeevChandran
• add openai compatible tts server support (#2099) by @RaajeevChandran

🐛 Bug Fixes

• Continue cap-truncated file_read to the end of the file (#2112) by @tpae
• Fix truncated remote tool arguments (#2108) by @tpae
• fix prompt text deleted when escape dismisses the slash command popup (#2103) by @RaajeevChandran

🧰 Maintenance

• Adopt native macOS theme defaults with system accent following (#2115) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.22.7...0.22.8

What's Changed

• Pin paged cache eviction fix (#2097) by @jjang-ai
• Pin Bonsai hybrid cache replay and disk quota fixes (#2093) by @jjang-ai
• Fix Nemotron AIFF routing and hybrid cache restore (#2087) by @jjang-ai
• Pin ZAYA cache fix and enable Nemotron Omni multimodal routing (#2086) by @jjang-ai
• Publish plugin active version atomically (#1844) by @mimeding
• Bound skill archive subprocesses (#1840) by @mimeding
• Correct Gemma QAT cache settings and TQ telemetry (#2085) by @jjang-ai
• Fix built-in Data Visualizer chart tool activation (#2084) by @jjang-ai
• Update Simplified Chinese and Traditional Chinese translations (#2076) by @ftzahao
• Fix JANG AppleScript success finalization (#2075) by @jjang-ai

🚀 Features

• Add combined-mode file_copy bridge for binary files between workspace and sandbox (#2080) by @tpae
• Track chat file changes with undo, and add writable combined mode (#2078) by @tpae
• added knowledge base introduction modal (#2077) by @RaajeevChandran

🐛 Bug Fixes

• Fix release job: reuse sentry-cli installed by the dSYM step (#2096) by @tpae
• fixed app hangs (#2095) by @RaajeevChandran
• Fix Gemma cache telemetry and Bonsai chart parsing (#2091) by @jjang-ai
• Fix skill discovery metadata loss and MCP capability-load reliability seams (#2089) by @tpae
• detect mcp server urls pasted into the api provider form and offer a redirect (#2082) by @RaajeevChandran

🧰 Maintenance

• Move Thinking toggle into the model picker's Model Options section (#2094) by @tpae
• Isolate working-folder context per chat session (#2092) by @tpae
• Fix Bonsai hybrid detection and add 16 GB eval coverage (#2090) by @tpae
• Harden Agent DB tool contracts for import modes and saved views (#2083) by @tpae
• Fix management-window sheet-swap crash and wire Sentry release telemetry (#2081) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.22.6...0.22.7

What's Changed

• Fix AppleScript Computer Use finalization and failure containment (#2071) by @jjang-ai
• Add eval watcher scoreboard artifacts (#1708) by @mimeding
• Stabilize Watchers pane layout (#1794) by @mimeding
• Add runtime installation doctor (#2043) by @mimeding
• Reliability audit: install integrity, loader/ABI validation, route timeouts, lifecycle ordering, CI coverage (#2061) by @tpae

🧰 Maintenance

• Fix singular and plural count labels (#2074) by @tpae
• Overhaul Tools management UX around an everyday-user catalog (#2073) by @tpae
• Redesign agent detail navigation and restyle the Tools tab (#2072) by @tpae
• Normalize Bonsai recommendations for onboarding (#2069) by @tpae
• Improve Models catalog browsability and tuck runtime diagnostics away (#2070) by @tpae
• Polish context budget popover (#2068) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.22.5...0.22.6

What's Changed

• Recover observed Bonsai chart argument drift (#2065) by @jjang-ai
• Fix AppleScript query completion outcomes (#2064) by @jjang-ai
• Fix spawned agent tool execution scope (#2062) by @jjang-ai
• Harden manual plugin receipt metadata (#1842) by @mimeding
• Handle empty voice transcriptions (#1852) by @mimeding
• Pin schema-1 JANG loader fix (#2047) by @jjang-ai

🚀 Features

• knowledge collections (#1815) by @RaajeevChandran
• added cmd+= and cmd+- font zoom shortcuts (#2054) by @RaajeevChandran
• Premium web search through the Osaurus router with search credits (#2055) by @tpae
• Redesign agent feature flags into an Abilities overview with live context estimate (#2053) by @tpae
• Universal skills library and tools-only agent UX (#2050) by @tpae

🐛 Bug Fixes

• updated knowledge empty state copy (#2067) by @RaajeevChandran
• mcp bearer 401 classification (#2048) by @RaajeevChandran
• Enforce selected memory safety limits during model loads (#2045) by @jjang-ai
• Fix Bonsai structured charts and hybrid memory safety (#2041) by @jjang-ai

🧰 Maintenance

• Redesign agent Database workspace and settings navigation UX (#2052) by @tpae
• Reduce redundant agent prompt context (#2051) by @tpae
• improved AppleScript model search UX (#2046) by @RaajeevChandran

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.22.4...0.22.5

What's Changed

• added bonsai models (#2040) by @RaajeevChandran
• Pin vmlx for native 1-bit JANG support (#2037) by @jjang-ai
• Update Simplified Chinese and Traditional Chinese translations (#2034) by @ftzahao
• fixed window maximize bug (#2028) by @RaajeevChandran
• updated readme (#2025) by @RaajeevChandran
• Update README.md (#2024) by @RaajeevChandran

🚀 Features

• Upgrade notch task navigation and follow-ups (#2036) by @tpae

🐛 Bug Fixes

• Fix spawn_model target contract (#2033) by @jjang-ai
• Computer Use: open apps foreground so the window is visible and the model doesn't misreport failure (#2032) by @jjang-ai
• fixed main thread hangs across launch, storage, tts, and streaming (#2030) by @RaajeevChandran
• fix scheduled runs never completing by evicting finished headless tasks (#2031) by @RaajeevChandran
• Computer Use AX bounding + AccessibilityBoundedElement test fix + vmlx repin (SpecDec tail flush #148) (#2029) by @jjang-ai
• read the hugging face keychain outside the token cache lock (#2027) by @RaajeevChandran

🧰 Maintenance

• make the chat window responsive down to narrow widths (#2038) by @RaajeevChandran
• Agent request lifecycles independent of the chat window (#2023) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.22.3...0.22.4