Changelog

What's Changed

• Harden plugin and registry archive extraction (#1987) by @mimeding
• HY3 official: repin vmlx + fix the load path that kept 94 GB packs from ever serving (#1984) by @jjang-ai
• evals(compat): leaderboard redesign — latest-run precedence, skip accounting, strengths (#1982) by @tpae

🐛 Bug Fixes

• Refuse background model loads inside the actor, not before it (#1994) by @jjang-ai
• Fix Product Hunt launch window to 24 hours (#2000) by @tpae
• Key the prefix cache to the weights, not just the model's name (#1989) by @jjang-ai
• Repin vmlx (for real this time) and push the memory-safety level to the cache-store gate (#1988) by @jjang-ai
• Background work must never evict the model the user is using (+ vmlx repin) (#1986) by @jjang-ai
• fix: support Codex Responses Lite for GPT-5.6 models (#1983) by @tpae

🧰 Maintenance

• Show the Product Hunt kitty next to the dinosaur in the launch dialog (#2011) by @tpae
• Make local-first onboarding immediately usable (#2008) by @tpae
• Add one-time Product Hunt launch dialog (July 2026) (#1998) by @tpae
• evals: hermetic run storage — eval runs can never touch the user's real contexts (#1992) by @tpae
• feat: catalog-driven reasoning controls + unified model options in the picker (#1991) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.22.1...0.22.2

What's Changed

• evals(community): OsaurusAI/Ornith-1.0-35B-MXFP8 on Apple M4 Max (#1979) by @mimeding
• added cli like @ filesystem menu to chat input (#1980) by @RaajeevChandran
• evals(community): OsaurusAI/gemma-4-12B-it-MXFP8 on Apple M4 Max (#1977) by @mimeding
• evals(community): OsaurusAI/Ornith-1.0-9B-MXFP8 on Apple M4 Max (#1975) by @mimeding
• Add measured per-model prefill step tuning (osaurus bench --tune-prefill) (#1905) by @mimeding
• Add ModelCapabilityLedger skeleton; migrate the production blocklist to it (#1913) by @mimeding
• Fix residency-default doc, expose batch diagnostics in /health, log effective policy (#1865) by @mimeding
• Add ChipProfile hardware capability detection, surface in /health (#1861) by @mimeding
• evals(community): OsaurusAI/gemma-4-E4B-it-8bit on Apple M4 Max (#1974) by @mimeding
• docs: refresh translator leaderboard and promote Traditional Chinese to community-maintained (#1972) by @tpae

🐛 Bug Fixes

• fix: Codex model discovery returned unusable gpt-5.5-*-wm slugs (#1981) by @tpae
• Update Kagi, You.com, and Tavily search providers to current APIs (#1978) by @tpae
• spawn_agent: carry the target agent's own tool policy into the child run (#1976) by @jjang-ai

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.22.0...0.22.1

What's Changed

• Improve Telegram channel recovery diagnostics (#1942) by @mimeding
• evals(community): OsaurusAI/gemma-4-E2B-it-8bit on Apple M4 Max (#1970) by @mimeding
• speed up onboarding model downloads via authenticated proxy (#1965) by @RaajeevChandran
• Update Simplified Chinese and Traditional Chinese translations (#1947) by @ftzahao
• Polish Russian localization (#1822) by @DrMaks22
• added get_current_time built-in tool (#1956) by @RaajeevChandran
• added in-app github token to fix plugin update rate limit errors (#1955) by @RaajeevChandran

🚀 Features

• added toggle for notch overlay placement on or below the menu bar (#1957) by @RaajeevChandran

🐛 Bug Fixes

• fixed crashes and hangs (#1971) by @RaajeevChandran
• fixed in-chat search bar focus, per occurrence navigation, and current match highlighting (#1969) by @RaajeevChandran
• Fix Router 409 IDEMPOTENCY_CONFLICT on refunded agent-loop iterations (#1968) by @tpae
• Size model fit against the GPU working set, not physical RAM (#1958) by @jjang-ai
• Flush pending stream text before tool-call frames; repin vmlx (stall + prose-cut fixes) (#1963) by @jjang-ai

🧰 Maintenance

• minor tweak in favorite models UI (#1973) by @RaajeevChandran
• Evals harness overhaul: new domains, runtime fixes, community contribution flow (#1967) by @tpae
• Remote provider hardening: crash/hang fixes, endpoint trust, Router spend safety, streaming perf (#1961) by @tpae
• Credits UI improvements: composer wallet panel and refreshed Credits page (#1959) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.21.13...0.21.14

What's Changed

🚀 Features

• added an option to update all plugins in one go instead of individual update (#1949) by @RaajeevChandran

🐛 Bug Fixes

• Fix ChatGPT/Codex test-button model count mismatch; add grok-4.5 to xAI catalog (#1954) by @tpae
• Fix warm-up KV cache misses and sandbox tool-registration thrash (#1953) by @tpae
• Repin vmlx to 0d3444ca (gemma4 parser fix #128) + fix reasoning-toggle default display (#1950) by @jjang-ai
• fixed main thread hangs (#1948) by @RaajeevChandran

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.21.12...0.21.13

What's Changed

• Harden search diagnostics and extraction (#1944) by @mimeding
• Support installing plugin project directories (#1843) by @mimeding
• Harden plugin repository refresh diagnostics (#1845) by @mimeding
• Decouple startup completion from plugin loading (#1838) by @mimeding
• Add recoverable agent backup restore path (#1837) by @mimeding

🐛 Bug Fixes

• Fix Privacy Filter false "URLs were not scrubbed" send block (#1946) by @tpae

🧰 Maintenance

• Fix RAM banner false positives for resident and in-flight models (#1941) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.21.11...0.21.12

What's Changed

• added timestamps to user message cells (#1939) by @RaajeevChandran
• Onboarding: recommend Ornith MXFP8 + official OsaurusAI Gemma; drop Gemma QAT/MXFP4 (#1936) by @jjang-ai
• speed up Hugging Face model downloads (#1937) by @RaajeevChandran
• Polish release UX details (#1931) by @mimeding
• Add provider connectivity issue triage (#1895) by @mimeding
• Add osaurus bench: standardized TTFT/prefill/decode benchmark CLI (#1863) by @mimeding
• Add Agent Channels live proof readiness (#1882) by @mimeding
• Repin vmlx-swift: default-on hybrid-SSM prefix cache reuse (#1930) by @jjang-ai

🚀 Features

• Native web search: pluggable providers + Search settings tab (#1938) by @tpae

🐛 Bug Fixes

• Fix DeepSeek thinking-mode tool_choice HTTP 400 (#1940) by @tpae

🧰 Maintenance

• chat UI polish (#1932) by @RaajeevChandran

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.21.10...0.21.11

What's Changed

• Update Simplified Chinese and Traditional Chinese translations (#1909) by @ftzahao
• added ram banner dismiss button (#1924) by @RaajeevChandran

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.21.9...0.21.10

What's Changed

🐛 Bug Fixes

• fixed sentry hangs & crashes (#1920) by @RaajeevChandran
• Skip proactive model warm-up when the projected load exceeds the hard RAM ceiling (#1918) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.21.8...0.21.9

What's Changed

• improved ram tight fit banner UI (#1910) by @RaajeevChandran
• Add RAM tight-fit send gate and make model switches immediate (#1908) by @tpae
• Add Trendshift badge to README (#1906) by @tpae
• Run native-MTP cold warmup at model load, not on the first user request (#1903) by @mimeding
• Show live tool-call progress card during buffered tool generation (#1894) by @jjang-ai

🚀 Features

• Proactive model warm-up: prefill-ahead KV caching for faster TTFT (#1897) by @tpae

🐛 Bug Fixes

• qwen3.5 agent-loop nudge + JANGTQ_K unblock + ZAYA tool-arg parser repin (#1907) by @jjang-ai
• Fix folder file-write bugs: firmlinked-root rejection + new-file diff count (#1900) by @jjang-ai
• Reflect tool-call failure in the collapsed card title (#1899) by @jjang-ai
• Fix 0.21.7 Sentry issues: notch display-reconfig hangs and TextKit downgrade crash (#1898) by @tpae
• improved favorite models UI (#1889) by @RaajeevChandran

🧰 Maintenance

• improved chat thread performance (#1891) by @RaajeevChandran
• Shorten chat-close unload grace from 60s to 15s (#1904) by @tpae
• Free model memory on chat close, respond to memory pressure, and cap unbounded caches (#1901) by @tpae
• added per-model API exposure controls with a Models tab in Server settings (#1893) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.21.7...0.21.8

What's Changed

• Backfill compatibility leaderboard with previously tested models (#1868) by @tpae
• Ornith 1.0 optimization loop: hybrid-family fix, recorded evals, community compatibility leaderboard (#1864) by @tpae
• Improve Telegram connection diagnostics (#1856) by @mimeding
• Cap sandbox plugin tool expansion (#1839) by @mimeding
• Pin agent run prompt injection contract (#1836) by @mimeding
• Add agent chat generation settings regression (#1835) by @mimeding
• Clarify memory ingestion diagnostics (#1834) by @mimeding
• Add support diagnostics bundle scaffold (#1829) by @mimeding
• Discover custom external model folders (#1850) by @mimeding
• updated chinese translations (#1858) by @ftzahao

🚀 Features

• added terminal style input history navigation with arrow keys in chat input text field (#1885) by @RaajeevChandran
• added full text conversation search and in-chat find with Cmd+F (#1883) by @RaajeevChandran
• added option to expand thinking while streaming (#1871) by @RaajeevChandran

🐛 Bug Fixes

• fixed plugin search matching unrelated plugins (#1880) by @RaajeevChandran
• fixed image attachments being blocked or silently dropped for remote provider model (#1879) by @RaajeevChandran
• fixed capabilities toggle initial render (#1878) by @RaajeevChandran
• fixed notch overlay covering the menu bar and improved its hover/ dismiss behavior (#1874) by @RaajeevChandran
• Repin vmlx-swift to merged main (class-killer #116 + norm #117 + Theme 3) — fixes Metal concurrency crash class (#1866) by @jjang-ai
• fixed sentry hangs (#1867) by @RaajeevChandran
• Fix ExternalPlugin shutdown teardown race and recurring app-side hangs (#1862) by @tpae
• Fix self-scheduled runs not auto-starting; fresh chat per wake (#1855) by @tpae

🧰 Maintenance

• Agent DB: sandbox file paths, db_export, and token-efficiency fixes (#1869) by @tpae
• Harden MCP surfaces: session recovery, OAuth single-flight, stdio death detection, external-surface enforcement (#1870) by @tpae
• Polish model catalog cards and promote Ornith Top Picks (#1860) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.21.6...0.21.7

What's Changed

• Persist user turns before chat streaming (#1851) by @mimeding
• Handle empty voice transcriptions (#1831) by @mimeding
• Persist user turns before chat streaming (#1830) by @mimeding
• Update Simplified Chinese and Traditional Chinese translations (#1820) by @ftzahao
• Redesign Agent Channels settings as the Channels tab (#1827) by @tpae
• Fix task progress overlay menu bar placement (#1825) by @mimeding
• Polish Slack and Telegram Agent Channels (#1808) by @mimeding

🐛 Bug Fixes

• fixed mistral remote provider reasoning support (#1847) by @RaajeevChandran
• fixed subagent feed hang (#1832) by @RaajeevChandran

🧰 Maintenance

• Eval matrix chat-model/subsystem split, delegation cases, and outcome-based frontier rubrics (#1853) by @tpae
• Start Sparkle update checks at app launch (#1849) by @tpae
• Local-model harness fixes from the gemma-4-E4B loop + full-coverage 3-way matrix (#1826) by @tpae
• Harden Slack/Telegram Agent Channels after the PR #1808 audit (#1824) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.21.5...0.21.6

What's Changed

• Polish Slack and Telegram Agent Channels (#1808) by @mimeding
• added transcription cleanup toggle and redesign the dictation overlay (#1816) by @RaajeevChandran

🚀 Features

• added favourite tab in the model selector (#1811) by @RaajeevChandran

🐛 Bug Fixes

• stopped image composer chips from wrapping in a narrow window (#1809) by @RaajeevChandran
• fixed Mistral reasoning format (#1804) by @RaajeevChandran
• skip gpu teardown on quit so exit does not cause race condition (#1805) by @RaajeevChandran

🧰 Maintenance

• Local-model harness fixes from the gemma-4-12B optimization loop (#1819) by @tpae
• Repin vmlx-swift: strip orphan tool-call closer tags (#115) (#1817) by @jjang-ai
• AppleScript subagent upgrade: capability reach, latency, consent UX, reliability (#1812) by @tpae
• Korean localization: more natural terminology for Providers, Watchers, Computer Use, Identity, and quick actions (#1807) by @tpae
• Onboarding model step: explicit resource costs, one pick per family, low-disk preflight (#1806) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.21.4...0.21.5

What's Changed

• Add Russian localization (#1782) by @DrMaks22
• Add safe custom JSON Agent Channel runner (#1652) by @mimeding
• Add Agent Channel inbox audit workbench (#1784) by @mimeding
• Repin vmlx-swift to c976fcbd — Ornith prefill counter advances (vmlx #105) (#1779) by @jjang-ai

🐛 Bug Fixes

• gated Rampart PII scan and media prep as exclusive GPU producers (#1799) by @RaajeevChandran
• Disconnect must not evict the BatchEngine (cold-load hangup crashed the server) (#1796) by @jjang-ai
• Generation params: wire seed into the sampler, accept string stop, native penalties, min_p (#1793) by @jjang-ai
• triage and fix sentry app hangs and crashes (#1781) by @RaajeevChandran
• Repin vmlx-swift b6eda04f — MLX Metal stream-map thread-safety (fixes concurrent-GPU EXC_BAD_ACCESS crashes) (#1780) by @jjang-ai

🧰 Maintenance

• Model catalog IA cleanup: family-grouped cards, variant picker, chat-only org fetch (#1800) by @tpae
• Cut per-turn prompt cost: remote prompt caching + local KV prefix reuse + measured text trims (#1798) by @tpae
• Subagent harness hardening: admission gate, mid-run interrupt, tool-capable workers, RAM coexistence, residency crash lane (#1797) by @tpae
• Repin vmlx-swift: deterministic RMSNorm shift + Mistral VLM full-resolution images (#1791) by @jjang-ai
• Settings IA cleanup: grouped sidebar, relocations, unified card primitives (#1792) by @tpae
• Upgrade evals harness: repeat/resume runs, judge audit + calibration, micro-perf and prompt-injection suites (#1789) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.21.3...0.21.4

What's Changed

• Add channel remote safety gate (#1768) by @mimeding
• updated chinese translations (#1762) by @ftzahao
• Add Computer Use web form proof lane (#1746) by @mimeding
• Add tool result grounding evals (#1740) by @mimeding
• Add model tool-use diagnostics (#1739) by @mimeding
• added Mistral provider support with reasoning effort control (#1777) by @RaajeevChandran
• Fix sweep: Mistral chat fix + hunyuan_v1_dense + Rampart PII (repin) + plugin auth (#1719) + MCP triage (#1761) by @jjang-ai
• cap tool rows in remote and plugin cards to prevent layout hang (#1767) by @RaajeevChandran

🚀 Features

• Add AppleScript Computer Use subagent with safe on-device execution (#1771) by @tpae

🐛 Bug Fixes

• fixed app hang sentry issues (#1776) by @RaajeevChandran

🧰 Maintenance

• Restore Ideogram in the image model download catalog (#1778) by @tpae
• minor UI tweaks (#1775) by @RaajeevChandran
• Unify the model-download tabs on a shared ModelListRow (#1774) by @tpae
• Standardize on "subagent" spelling (drop hyphenated "sub-agent") (#1769) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.21.2...0.21.3

What's Changed

• Add unified evidence report registry (#1629) by @mimeding
• Add Agent Channel security foundation (#1728) by @mimeding
• Add Agent Channel async substrate (#1651) by @mimeding

🚀 Features

• added Rampart as a lightweight on-device PII detection model (#1764) by @RaajeevChandran

🐛 Bug Fixes

• Fix KV-cache prefill reset after capabilities_load (#1760) by @tpae

🧰 Maintenance

• Make Osaurus Cloud available by default with a permanent opt-out (only during onboarding) (#1766) by @tpae
• Make local↔image model handoffs crash-free and seamless (#1765) by @tpae
• Make screen context a per-agent Computer Use option (#1763) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.21.1...0.21.2

What's Changed

🐛 Bug Fixes

• Make built-in Default agent non-configurable; fix sub-agent toggle persistence (#1756) by @tpae
• Fix chat history loss from non-idempotent schema migrations (#1755) by @tpae

🧰 Maintenance

• Improve image models tab: inline Hugging Face link and download sizes (#1759) by @tpae
• Rename Image Generation tab to Images (#1758) by @tpae
• Fix native image generation echoing artifact JSON instead of confirming (#1757) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.21.0...0.21.1

What's Changed

• reduced default agent prefill cost on small local models (#1748) by @RaajeevChandran
• fixed sentry hangs (#1747) by @RaajeevChandran
• Add skill policy diagnostics (#1744) by @mimeding
• Add redacted provider request diagnostics (#1743) by @mimeding
• Recover chat continuation from transcript fallback (#1742) by @mimeding
• Add agent collaboration protocol contract (#1630) by @mimeding
• Add plugin marketplace trust preview (#1647) by @mimeding
• Add Agent Channel action safety contract (#1646) by @mimeding
• Complete Simplified Chinese localization (#1612) by @jiajun-dev
• Add router account usage center (#1586) by @mimeding
• Reject out-of-range proxy ports in GlobalProxyConfiguration (#1693) by @jax-0n-git
• Avoid CLI crash on out-of-Int64-range model_info numbers in show (#1702) by @jax-0n-git
• Add osaurus serve --supervise keep-alive loop to survive app quit (#1712) by @jax-0n-git
• Accept an omitted args in desktop-format MCP bundle manifests (#1692) by @jax-0n-git
• Tolerate and reap corrupt coordinator lock files (#1690) by @jax-0n-git
• Fix MCP schema proxy miscasting JSON integers 0 and 1 as booleans (#1687) by @jax-0n-git
• Don't treat a bare tools uninstall name as a CWD path (avoid deleting the wrong directory) (#1713) by @jax-0n-git
• Add Computer Use forensic evidence runner (#1671) by @mimeding
• hide non-MLX models that local engine can't load (#1729) by @RaajeevChandran
• Fix same-turn capability tool activation (#1725) by @mimeding
• updated chinese translations (#1720) by @ftzahao
• fix: use veryShortWeekdaySymbols for schedule weekday display (#1722) by @ftzahao
• fix app hang watchdog stalls across folder, tool listing, permissions, and launch (#1699) by @RaajeevChandran
• improved plugin discovery for smaller local models (#1685) by @RaajeevChandran
• Repin vmlx-swift: fix Mistral3/Pixtral VLM model-load crash (#1686) by @jjang-ai
• Native image generation/editing + agent delegation (spawn / local_delegate) (#1682) by @jjang-ai

🚀 Features

• added file diff cards for folder and sandbox edits (#1683) by @RaajeevChandran

🐛 Bug Fixes

• Coalesce subagent image progress in place; show step/ETA + elapsed (#1753) by @tpae
• fixed main thread freezes (#1754) by @RaajeevChandran
• fix continuing and exporting conversations after the encryption update (#1749) by @RaajeevChandran
• Fix concurrent-GPU crash on sub-agent residency handoff (unload: drain before free) (#1736) by @jjang-ai
• fixed app hangs (#1730) by @RaajeevChandran
• fixed tools page layout hang (#1724) by @RaajeevChandran
• Fix LFM2 tool-calling: repin vmlx native 'List of tools' format (#1723) by @jjang-ai
• fix context budget popover crash on System Prompt drill-down (#1684) by @tpae
• Fix #1680: surface (not silently truncate) multi-step tool tasks on a premature-EOS empty turn (#1681) by @jjang-ai

🧰 Maintenance

• minor image generation ux improvements (#1752) by @RaajeevChandran
• Improve image generation: responsive cancel, live timer/estimate, dynamic OsaurusAI catalog (#1751) by @tpae
• Split spawn into spawn_agent + spawn_model with cross-model residency (#1750) by @tpae
• Expand and optimize the OsaurusEvals catalog + harness (local vs frontier) (#1741) by @tpae
• Add a standard per-agent model picker for chat-driven sub-agents (#1738) by @tpae
• Unify all subagent flows through the eval host and benchmark local vs frontier (#1735) by @tpae
• Promote Image Generation to its own Settings section (#1734) by @tpae
• Unify sub-agent delegation under one registry + per-agent Sub-agents tab (#1731) by @tpae
• improved settings ux (#1716) by @RaajeevChandran

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.20.9...0.20.10

What's Changed

• Add Computer Use regression scorecards (#1648) by @mimeding
• QoL improvements (#1655) by @RaajeevChandran
• updated chinese translations (#1659) by @ftzahao

🚀 Features

• add Korean (ko) localization (#1660) by @tpae

🐛 Bug Fixes

• make agent run honor agent temperature (#1678) by @RaajeevChandran
• stop wasted iterations on repeated failing capabilities_load calls (#1679) by @RaajeevChandran
• persist per tool disable state for MCP provider tools (#1677) by @RaajeevChandran
• fixed app hangs (#1672) by @RaajeevChandran

🧰 Maintenance

• Surface host-side remote-agent (p2p) activity (#1676) by @tpae
• Polish remote agent connect UX (#1657) by @tpae
• Harden Bonjour discovery (security, reliability, discoverability) (#1658) by @tpae
• unify local and remote agent detail views (#1656) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.20.8...0.20.9

What's Changed

• Add Computer Use evidence pack (#1631) by @mimeding
• Add Computer Use diagnostics (#1643) by @mimeding
• Add Screen Context capture lab (#1644) by @mimeding
• show pasted content preview in chat (#1645) by @RaajeevChandran
• Add Agent Channel Connection Center (#1616) by @mimeding
• updated chinese translations (#1633) by @ftzahao

🚀 Features

• Privacy filter: opt-in on-device model, no-regex rule builder, settings polish (#1641) by @tpae
• agent db upgrade (#1640) by @tpae

🐛 Bug Fixes

• fixed app hangs (#1638) by @RaajeevChandran
• Fix Computer Use main-thread hangs (#1634) by @tpae
• Fix and harden remote agent (Mode 2) runs (#1636) by @tpae
• fix /memory/ingest distillation and harden launch crashes (#1632) (#1635) by @tpae

🧰 Maintenance

• disable broken tools (#1650) by @tpae
• updated onboarding flow (#1649) by @tpae
• make plain-text answers the primary agent-loop exit (#1639) by @tpae
• qol improvements (#1637) by @RaajeevChandran

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.20.7...0.20.8