Changelog

What's Changed

• Pin Bonsai FP16 attention K/V with FP32 recurrent state (#2820) by @jjang-ai
• Consume merged Bonsai loading and bounded media prefill (#2818) by @jjang-ai

🐛 Bug Fixes

• Sandbox: attribute provisioning failures and stop the quit-time VM re-boot behind vmnet_in_use (#2819) by @tpae
• Preserve persisted document text in chat and warmup (#2816) by @jjang-ai
• fixed main thread hangs in the composer, markdown parsing and sidebar rows (#2817) by @RaajeevChandran
• Share parent handoff policy with images and compaction (#2798) by @jjang-ai
• Honor the loaded bundle sampling defaults with duplicate model names (#2814) by @jjang-ai
• Respect no-tool execution requests before prefill (#2815) by @jjang-ai
• Unify invoking-parent handoffs across local subagent kinds (#2796) by @jjang-ai
• Preserve user media for explicit Gemma tool selection (#2806) by @jjang-ai
• Propagate cancellation through buffered tool streams (#2813) by @jjang-ai

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.25.7...0.25.8

What's Changed

• Embed the Workspaces explainer in the tab's empty state; drop the intro dialog (#2811) by @tpae
• Support Bonsai2 runtime and preserve native Qwen tool templates (#2808) by @jjang-ai

🚀 Features

• added number_of_agents property to agent_created telemetry event (#2807) by @RaajeevChandran

🐛 Bug Fixes

• Preserve typed MCP image results in Chat and agent history (#2805) by @jjang-ai
• fixed tab strip jumping and overlapping the sidebar on window resize (#2802) by @RaajeevChandran
• added don't show this again to the ssd cache limit reached notice (#2801) by @RaajeevChandran

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.25.6...0.25.7

What's Changed

• Gate Sparkle update checks on chat, not Settings (#2794) by @tpae
• Unify file_read/file_write across document, image, and workbook formats (#2791) by @tpae

🐛 Bug Fixes

• warn before add model discards the agent creator draft (#2800) by @RaajeevChandran
• fix history dialog stacking when the toolbar button is tapped repeatedly (#2799) by @RaajeevChandran
• Show composer stuck chip only when a run is actually silent (#2795) by @tpae
• fix computer use failing with the privacy filter on and losing its step log (#2797) by @RaajeevChandran
• improved PII extraction flow (#2793) by @RaajeevChandran
• Preserve complete native local tool-call batches (#2792) by @jjang-ai
• fix: prevent Int overflow in CreditsTopUpSheet currentMicro (#2717) by @sentry[bot]
• Explain warning-pressure delegation refusal and qualify RAM opt-out (#2790) by @jjang-ai
• Enforce model manifest compatibility and detect model updates (#2789) by @jjang-ai

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.25.5...0.25.6

What's Changed

• Orchestrator polish and simplification (#2787) by @tpae
• Default onboarding to Raptor 0.6 (#2782) by @tpae

🐛 Bug Fixes

• Make model repair verify files and report download progress (#2788) by @jjang-ai
• Preserve local usage accounting and complete image chats cleanly (#2786) by @jjang-ai
• Keep MTP opt-in and enforce runtime settings across Chat and API (#2785) by @jjang-ai
• Fix resident child RAM admission and external model residency tracking (#2784) by @jjang-ai
• Validate vision processors and preserve complete image/cache input (#2774) by @jjang-ai
• Fix RAM admission measurement, capacity and token-budget contracts (#2752) by @jjang-ai
• Bound idle model residency and preserve utility ownership (#2771) by @jjang-ai
• Notify when SSD cache reaches its quota and clear indexed cache safely (#2783) by @jjang-ai

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.25.4...0.25.5

What's Changed

• Remote prompt-cache efficiency: session cache keys, cache-aware Router telemetry, BYOK cache parsing (#2779) by @tpae
• Guide n8n setup so pairing codes only issue when they can work (#2775) by @tpae
• QoL improvements (#2776) by @RaajeevChandran
• Audit installed vision bundles and repair discovery and admission parity (#2772) by @jjang-ai
• Multi-device identity: unique agent addresses, owner redeem, and iOS Keychain group (#2769) by @tpae

🚀 Features

• add recent folders to the + menu and agent working folder setting (#2765) by @RaajeevChandran

🐛 Bug Fixes

• Allow the release archive to trust SwiftPM build plugins (#2780) by @tpae
• fixed xcode 27 build errors (#2778) by @RaajeevChandran
• present the orchestrator config approval popup as a centered themed modal (#2773) by @RaajeevChandran
• Fix daily schedule double-fire and launch replay (#2770) by @tpae
• Reject unsupported embedding model requests instead of substituting Potion (#2768) by @jjang-ai
• Update MLX 0.32.2 pin and default only Flash Next MTP Off (#2745) by @jjang-ai
• fit the settings window to small screens (#2764) by @RaajeevChandran
• disable add folder for shared agent chats (#2763) by @RaajeevChandran

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.25.3...0.25.4

What's Changed

• Add n8n pairing code for remote Secure Channel setup (#2762) by @tpae
• qol improvements for providers and the chat sidebar (#2755) by @RaajeevChandran

🚀 Features

• persist open chat tabs across window close and relaunch (#2740) by @RaajeevChandran

🐛 Bug Fixes

• make sure upgraders see the workspaces intro after the whats new modal (#2760) by @RaajeevChandran
• show LAN discovered agents in the sidebar (#2758) by @RaajeevChandran
• Use installed config and weight evidence for local vision input (#2756) by @jjang-ai
• reduce main thread work behind app hang (#2757) by @RaajeevChandran
• Repair incomplete chat-history schemas so later turns persist (#2753) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.25.2...0.25.3

What's Changed

• Add first-party n8n agent channel with secret-verified webhook ingress (#2748) by @tpae
• Add Osaurus ID in Identity and fix Workspaces constant refresh (#2749) by @tpae
• Show execution surface (sandbox VM vs native Mac) on tool approval cards (#2734) by @tpae
• Computer Use audit: telemetry funnel, readiness gate, verified done, loop evals (#2731) by @tpae
• Chat sidebar QoL: per-agent drafts, new-agent highlight, badge counts, subagent language, spell check (#2727) by @tpae

🚀 Features

• added right click context menu to chat tabs (#2721) by @RaajeevChandran

🐛 Bug Fixes

• Make settings discoverable to the Orchestrator and Management search (#2750) by @tpae
• Repin vMLX for the Qwen AR scheduling checkpoint (#2741) by @jjang-ai
• restore import claude plugin from github after tools and plugins refactor (#2737) by @RaajeevChandran
• Fix Orchestrator row selection and fit the chat window to small screens (#2730) by @tpae
• Recover freed GPU buffers before final local child RAM refusal (#2733) by @jjang-ai
• Pin isolated Qwen Flash AR checkpoint and display authoritative final rates (#2726) by @jjang-ai

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.25.1...0.25.2

What's Changed

• Pin recurrent disk-cache state preservation (#2719) by @jjang-ai
• Pin Flash PLE per-gather row deduplication (#2712) by @jjang-ai
• Pin isolated Flash QSA optimization for AR and prefill (#2710) by @jjang-ai
• Make the chat folder chip persist as the agent's working folder (#2701) by @tpae
• Add host-side "Share my models for inference" toggle for paired peers (#2700) by @tpae
• Show model alignment preparation only on direct chat Send (#2697) by @jjang-ai

🚀 Features

• added filters in chat history modal (#2711) by @RaajeevChandran
• added interactive workspaces walkthrough (#2706) by @RaajeevChandran

🐛 Bug Fixes

• Queue tool permission prompts and run sibling spawn calls as one wave (#2720) by @tpae
• keep unsent composer text across chat and agent switches (#2715) by @RaajeevChandran
• fix chat window sizing for the workspaces intro and restore the minimum size (#2716) by @RaajeevChandran
• MTP: pre-Send controls, bounded depths and preserved sampling (#2714) by @jjang-ai
• fix project page staying open after picking an agent from the sidebar (#2713) by @RaajeevChandran
• Let delegated agents write files in their own Working Folder (#2703) (#2707) by @tpae
• accept non-function tools on /v1/responses and surface decode errors (#2705) by @RaajeevChandran

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.25.0...0.25.1

What's Changed

• Workspace agents as orchestration and dispatch targets (#2692) by @tpae
• workspace improvements (#2687) by @RaajeevChandran
• Workspaces: shared agents, team-billed inference, single-plan trial billing (#2683) by @tpae
• Distinguish memory predictions from loading and resident warnings (#2672) by @jjang-ai
• Repin vmlx-swift and integrate native MiniCPM5 controls and tools (#2676) by @jjang-ai

🚀 Features

• added agent filter dropdown to the history dialog (#2662) by @RaajeevChandran
• browser style chat tabs with agents sidebar and history dialog (#2630) by @RaajeevChandran
• added update_skill tool so agents can edit user skills on request (#2648) by @RaajeevChandran

🐛 Bug Fixes

• Fix delegation model forwarding after workspace-target integration (#2695) by @jjang-ai
• refuse binary document reads on the sandbox file route with a read_knowledge hint (#2694) by @RaajeevChandran
• Honor agent-target model overrides through delegated dispatch (#2688) by @jjang-ai
• Keep one window owner for an open chat session (#2691) by @jjang-ai
• fix mcp integer arguments 0 and 1 being sent as booleans (#2690) by @RaajeevChandran
• Keep unavailable subagent token counts distinct from measured zero (#2686) by @jjang-ai
• Settings: distinguish inherited sampling from explicit overrides (#2685) by @jjang-ai
• Keep retired iMessage helper callbacks from invalidating replacement sessions (#2674) by @jjang-ai
• Keep Settings menus responsive and preserve explicit idle residency on relaunch (#2679) by @jjang-ai
• Avoid the unbounded Seatbelt post-exit wait and detach EOF readers (#2681) by @jjang-ai
• allow markdown and other parsable text files in the attach file picker (#2678) by @RaajeevChandran
• apply project working folder by turning the agent sandbox off (#2675) by @RaajeevChandran
• fix theme editor color picker jitter and hex input (#2673) by @RaajeevChandran
• persist memory consolidator last run and catch up on launch (#2671) by @RaajeevChandran
• Model load: never rewrite a bundle's config.json for the Gemma-4 audio stamp (#2633) (#2670) by @jjang-ai
• Chat: a run cancelled during its cold load must not roll back the retry that already owns the session (#2668) by @jjang-ai
• Swap-pressure banner: Unload Model through the guarded unload lifecycle, refusal reported, emulated target and button accessibility fixed (#2667) by @jjang-ai
• fix update_skill never reaching slash invoked skills (#2665) by @RaajeevChandran
• added x-opencode-session affinity header for opencode hosts (#2666) by @RaajeevChandran
• open chat windows full size and start the layout tour after first run dialogs (#2664) by @RaajeevChandran
• Lazy chat loading follow-up: residency-only chip tooltip, stale warm-up comment, activation semantics (#2663) by @jjang-ai
• Lazy chat model loading: selecting a model records the choice; the first Send loads and prefills the real request (#2660) by @jjang-ai
• fix history dialog row actions and keep the dialog open under nested alerts (#2661) by @RaajeevChandran
• Folder chip: accessibility name, value and identifier (#2659) by @jjang-ai
• Truthful termination: exhausted announcements and repeated retrieval failures end in one tool-free status step (#2657) by @jjang-ai
• Reopened sessions: a window opened from a history row loads the stored transcript instead of erasing it on the next send (#2658) by @jjang-ai
• Research completion: announce-only recovery names what the model can do; promised-work endings recovered; replayed retrieval failures escalate before stopping; window-close crash fixed (#2656) by @jjang-ai
• Tool flow: no replay of transient knowledge replies, paging signals survive compression, manual mode holds across surfaces, capped results say so (#2647) by @jjang-ai
• Folder and knowledge listings: paged file_search/list_knowledge with totals, truncated-tree prompt line, ungrounded-claim advisory (#2646) by @jjang-ai
• Chat: a queued steer lands before the pending assistant placeholder, not after it (#2653) by @jjang-ai
• recognize claude's export index and point users at the batch zips (#2650) by @RaajeevChandran
• Session tool scope follows the agent chip; advisory after repeated tool_not_found; canonicalise decorated tool names (#2645) by @jjang-ai
• Repetition penalty window: 64 tokens when a penalty is set (20 otherwise) (#2644) by @jjang-ai
• Watcher: run against the chosen folder, not the agent sandbox workspace (#2643) by @jjang-ai

🧰 Maintenance

• Pin Spark2.5 runtime and native template handling (#2684) by @jjang-ai
• Repin vmlx-swift to 16379811: Ling 2.6 flash JANGTQ answers again (#2652), KDA fidelity, disk-cache integrity, awq sidecar skip, Bailing enable_thinking pass-through (#2655) by @jjang-ai

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.24.7...0.25.0

What's Changed

• Repin vmlx-swift to 577d9c63 (non-finite logits probe, Raptor top-level jang stamps) (#2642) by @jjang-ai
• Flash-Next MTP: layout advisory popup, explicit depth-3 default, stream-tail instrumentation, repin vmlx c8a4b66b (#2635) by @jjang-ai
• Repin vmlx-swift: proposal-head stamp runtime (vmlx#422) (#2631) by @jjang-ai
• Web pipeline: retrieval rides with discovery; hallucinated fetch calls get steered (#2625) by @jjang-ai

🚀 Features

• make cmd+n start a new chat inside the current project (#2634) by @RaajeevChandran

🐛 Bug Fixes

• Thinking default detection for top-level jang reasoning blocks and enable_thinking-undefined templates; compose agent-loop state notices (#2641) by @jjang-ai
• Agent loop: log when the invalid-args notice is staged (#2640) by @jjang-ai
• Capability grounding follow-ups: honest enum errors, Host Files scope copy, ungrounded file-side-effect notice (#2638) by @jjang-ai
• Delegation: enforce the RAM-safety sequence for every spawn_agent handoff (unload main → load delegate → run → unload delegate → reload main) (#2639) by @jjang-ai
• Watcher dispatch grounds itself; config apply says DONE (#2628) by @jjang-ai
• Agent loop: knowledge-tool dedupe, dynamic same-name run advisory, repeat visibility (#2632) by @jjang-ai
• Correct 0.24.6 release notes (#2627) by @jjang-ai

🧰 Maintenance

• Catalog: drop the Ling-2.6 Flash entries (Ling 2.6 runtime no longer shipped) (#2637) by @jjang-ai

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.24.6...0.24.7

What's Changed

• Live Activity: show what the adaptive MTP controller actually ran last turn (#2624) by @jjang-ai
• MTP detector + depth buttons appear during warmup, by weight, Flash-Next/27B only (#2620) by @jjang-ai

🐛 Bug Fixes

• Capability grounding: folder-mode agents get their manifest back; gateway honors its own contract (#2626) by @jjang-ai
• fixed main thread hangs (#2619) by @RaajeevChandran

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.24.5...0.24.6

What's Changed

🚀 Features

• added keyboard shortcuts for sidebar toggle and agent cycling (#2617) by @RaajeevChandran
• added project folder support (#2611) by @RaajeevChandran
• added clickable knowledge document links in chat (#2610) by @RaajeevChandran

🐛 Bug Fixes

• FDA probe: require the real SQLite header, not merely a non-throwing read (#2613) (#2616) by @jjang-ai
• Repin vmlx-swift to 97676e19: qwen4_exp UAF + constrained-Mac wired-limit + docs-leak fixes (#2615) by @jjang-ai
• keep the main window hidden on login item and cli launches (#2614) by @RaajeevChandran

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.24.4...0.24.5

What's Changed

• Repin vmlx-swift to de0c26d3 (one canonical boundary record per tool-call cycle) (#2589) by @jjang-ai
• Mark internal utility generations as auxiliary cache intent (repin vmlx 86d23df8) (#2586) by @jjang-ai
• Evals: cache write/reuse scored per tool call + sustained-decode collapse gate (#2588) by @jjang-ai
• Repin vmlx-swift to 65ba6986 (index-union MTP evidence, hybrid cache dedup, DSV3 dtype, verify prefetch) (#2585) by @jjang-ai

🚀 Features

• added custom endpoint support to onboarding (#2599) by @RaajeevChandran
• added source filter to the model picker's local tab (#2597) by @RaajeevChandran
• added raptor v0.5 in what's new modal (#2577) by @RaajeevChandran

🐛 Bug Fixes

• fix /api/show for external models and report tool/thinking capabilities (#2606) by @RaajeevChandran
• Full Disk Access probe: anchor on the system TCC database, and read (#2601) (#2605) by @jjang-ai
• Watcher/dispatch agents can actually reach their target folder (#2603) by @jjang-ai
• fail image jobs on recovered mlx gpu errors instead of crashing and repin vmlx-swift (#2587) by @RaajeevChandran
• RAM-Safety preflight: price KV headroom by each family's real cache topology (#2595) by @jjang-ai
• fix unreadable approval content in tool approval modal (#2575) by @RaajeevChandran
• Delegated runs resolve their own model's reasoning, not the parent's toggle (#2592) by @jjang-ai
• Stop the engine at tool dispatch (kill zombie post-tool decode); repin vmlx de82613a (#2591) by @jjang-ai
• Model picker: resolve local effort capabilities live (None vs Extra-High contradiction) (#2590) by @jjang-ai
• fix theme editor freezes (#2578) by @RaajeevChandran

🧰 Maintenance

• Repin vmlx-swift to 5a63b9be: healer opt-in (#2604) + GLM-5.3 crash/wired/loader fixes (#2607) by @jjang-ai
• Repin vmlx-swift to 2422cfb8: fp32 cascade fixes, MLA/QSA/GDN decode perf, Qwen3VL media order (#2598) by @jjang-ai

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.24.3...0.24.4

What's Changed

• Advance the vmlx-swift pin to 1315fdcb (DeepseekV3 MoE f32 residual fix) (#2573) by @jjang-ai
• Add Claude Code CLI integration (#2257) by @Yao920127
• Warn when switching models mid-conversation (#2538) by @jjang-ai
• Repin vMLX for 90+ tok/s Ornith 35B decode (#2553) by @jjang-ai
• Repin vMLX for scoped Ornith 35B MTP safety (#2549) by @jjang-ai
• evals(community): OsaurusAI/Muse-Glimmer-30B-JANG_6M on Apple M5 Max (#2544) by @jax-0n-git
• Repin vMLX with Qwen 3.8 legacy-parser correction (#2543) by @jjang-ai

🐛 Bug Fixes

• Stop the memory-safety profile default from clamping the MLX allocator pool (#2572) by @jjang-ai
• Use Raptor for mainstream onboarding defaults (#2574) by @jjang-ai
• Preserve reusable cache checkpoints across tool dispatch (#2571) by @jjang-ai
• Fix embedding discovery in configured models directory (#2570) by @jjang-ai
• Fix stale agent completion and heal misaligned model bundles (#2567) by @jjang-ai
• Hide local memory warnings for cloud models (#2566) by @jjang-ai
• Show and cancel exact live inference work (#2563) by @jjang-ai
• Restore type tags on Responses Lite namespace tools (#2565) by @tpae
• Preserve the admitted allocator ceiling for native MTP (#2561) by @jjang-ai
• Recover media rejections and harden OpenAI Responses (#2562) by @tpae
• Remove hidden HTTP reasoning-budget coercion (#2560) by @jjang-ai
• Report native MTP fallback only when MTP did not run (#2554) by @jjang-ai
• Give loaded skills a directory anchor for bundled resources (#2555) by @jjang-ai
• Repin cancellable speculative prefill runtime (#2557) by @jjang-ai
• Make denied chat tool outcomes visible and bounded (#2539) by @jjang-ai
• Repin serialized MLX host reads for Qwen3.8 server crash (#2556) by @jjang-ai
• Recognize bundle-declared local reasoning channels (#2545) by @jjang-ai
• Evals: score grounded filesystem access across equivalent tools (#2548) by @jjang-ai
• Make batch evals architecture-aware and truthful (#2536) by @jjang-ai
• Evals: preserve agent-loop exit and architecture cache taxonomy (#2546) by @jjang-ai
• Preserve explicit reasoning on the first cold send (#2537) by @jjang-ai
• Remove the hidden Orchestrator tool kill switch (#2541) by @jjang-ai
• Fix sequential spawn capacity after active file-cache growth (#2535) by @jjang-ai

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.24.2...0.24.3

What's Changed

• Repin vmlx-swift to aeb5e21c19: QSA crash fix + end-of-output hang fix (#2532) by @jjang-ai
• Agent loop: record which branch ended the run (telemetry only) (#2520) by @jjang-ai

🚀 Features

• simplify tools and plugins in settings (#2534) by @RaajeevChandran

🐛 Bug Fixes

• Spawn admission: price a child by its bounded request, not the retention cap (#2533) by @jjang-ai
• Evals: explicit fail-closed native-MTP controls (--mtp off|auto|d1|d2|d3) (#2531) by @jjang-ai
• Agent loop: bound the truncated-read continuation steer (2 per file), keep the partial-read disclosure (#2528) by @jjang-ai

🧰 Maintenance

• Enable sandboxing by default with hardened runtime (#2529) by @tpae

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.24.1...0.24.2

What's Changed

🚀 Features

• revamped settings (#2517) by @RaajeevChandran

🐛 Bug Fixes

• fix permissions view showing full disk access as granted when it is not (#2524) by @RaajeevChandran
• fix settings crash on macos 15 when opening tabs with header sub tabs (#2522) by @RaajeevChandran
• swap-balloon: verified lifecycle (refuse-dirty, trap sweep, orphan self-exit) (#2518) by @jjang-ai
• Render the recorded cancelled state instead of an invisible empty turn (#2510) (#2516) by @jjang-ai
• Swap banner copy: name swap and show the measured GB (keeps #2512 design) (#2515) by @jjang-ai

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.24.0...0.24.1

What's Changed

• Swap-pressure warning banner + designer emulation controls (#2507) by @jjang-ai
• Promote the Orchestrator to a primary documented feature; prune stale docs (#2500) by @tpae
• Repin vmlx-swift to 447a6a2b: Qwen 3.8 Flash Next native runtime (#2504) by @jjang-ai
• fix(chat): render display math with stable SwiftMath images (#2417) by @yxf206-a11y
• fix leaked pacing timer on mid-stream processor dealloc (#2486) by @donmimo
• Introduce the Orchestrator: dedicated settings tab, identity, and launch polish (#2493) by @tpae
• Default-agent declarative config + delegation orchestrator (#2485) by @tpae
• Redesign onboarding as a 3-screen (#2481) by @tpae

🚀 Features

• chat UX improvements (#2499) by @RaajeevChandran
• Migrate credit system UI from dollars to credits (#2495) by @tpae
• let agents write knowledge directly with call-time approval (#2480) by @RaajeevChandran
• add bulk edit and on-device redaction tools for folder chats (#2474) by @RaajeevChandran
• api (ollama): add capabilities to /show response (#2462) by @RaajeevChandran

🐛 Bug Fixes

• improved memory swap banner (#2512) by @RaajeevChandran
• Single, surfaced greedy-while-MTP coercion site (corrective for #2506) (#2509) by @jjang-ai
• Speculative Depth buttons enforce a real manual MTP activation; greedy only while MTP runs (#2506) by @jjang-ai
• Delete the persisted global Disable Tools switch; tool availability is agent-scoped (#2505) by @jjang-ai
• The Speculative Depth row rendered on models with no MTP head (#2489) by @jjang-ai
• minor improvements in onboarding (#2487) by @RaajeevChandran
• Show the drafted width, and stop an unusable one from killing turns (#2473) by @jjang-ai
• fixed main thread hangs (#2475) by @RaajeevChandran
• Say why speculative decoding is not running (#2472) by @jjang-ai
• Adopt a bundle's declared presence/frequency penalties, and show them (#2470) by @jjang-ai
• fix wasted warm ups and the jumping context tooltip on channel bound agents (#2468) by @RaajeevChandran
• keep cross-chat prefill reusable by stabilizing the injected time block (#2467) by @RaajeevChandran
• attach granted plugin tools to channel dispatches (#2466) by @RaajeevChandran
• hardcode openai context windows as api never returns them (#2464) by @RaajeevChandran
• fix xai oauth models reporting a stale catalog and fallback context window (#2463) by @RaajeevChandran

🧰 Maintenance

• Onboarding polish: motion system, Figma speech bubble, click-through fixes; Top Picks to Ornith 1.5 (#2503) by @tpae
• Orchestrator-first delegation: default-on spawn pool, same-turn spawn activation, and delegated artifact pass-through (#2498) by @tpae
• Repin vmlx: Ling 3.0 (BailingMoeV3/KDA) runtime (#2491) by @jjang-ai
• Repin vmlx-swift to 1c17ce83 for the presence/frequency penalty fixes (#2469) by @jjang-ai

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.23.1...0.23.2

What's Changed

• Measure what coarsening the injected clock would be worth: 14.6x on a repeated document (#2461) by @jjang-ai
• Cross-chat reuse: the cause is a timestamp, not a scoping policy (#2460) by @jjang-ai
• Say when a configured disk cap is not the one being enforced (#2459) by @jjang-ai
• T15: flipping reasoning mid-conversation costs one re-prefill, then nothing (#2458) by @jjang-ai
• Correct the Step-3.7 row: text-only by design, not too big to run (#2457) by @jjang-ai
• T8 no longer holds as written — the 42× restart win did not reproduce (#2456) by @jjang-ai
• Step-3.7: attempted under a RAM guard rather than estimated away (#2455) by @jjang-ai
• T11: cold and cache-served answers are identical at three arbitrary prefix lengths (#2454) by @jjang-ai
• T9: the disk cap holds mid-conversation — and maxSizeGB is dead while maxSizePercent is set (#2453) by @jjang-ai
• T10: three images across three turns — the pipeline keeps them distinct (#2452) by @jjang-ai
• Prefix reuse is scoped to one conversation — a byte-identical prompt in a new chat reuses nothing (#2450) by @jjang-ai
• T13: three vision families reuse across a media prefix at 10k and 20k (#2449) by @jjang-ai
• Only an explicit Repair may rebuild a bundle the user changed (#2448) by @jjang-ai
• Audio coverage: third family proven, the depth limit that is not ours, and an audio badge on the model chip (#2447) by @jjang-ai
• Turning native MTP off did not stay off (#2446) by @jjang-ai
• Repin vmlx to the merged VL feature-order fix (#2445) by @jjang-ai
• An audio-capable model could not be given audio, and C7 was wrong about why (#2444) by @jjang-ai
• Sampling Defaults were inert on almost every model, and nothing showed what actually ran (#2442) by @jjang-ai
• Disk cache size is a percent in Settings, and every stale GB reader is fixed (#2441) by @jjang-ai
• Stop billing cold model load as TTFT, and say when the Mac is the bottleneck (#2440) by @jjang-ai
• Repin vmlx to the auto disk-cache size (10f27d03) (#2438) by @jjang-ai
• Disk cache: auto-size to 10% of disk, surface usage in chat, add Clear SSD Cache (#2436) by @jjang-ai
• Repin vmlx: adaptive depth re-arm + warmup-memo scope fix (#2432) by @jjang-ai
• Make the MTP Mode hint tell the truth about a selected DFlash drafter (#2430) by @jjang-ai
• Repin vmlx: DFlash 2 at 56-62 tok/s in-app (ring fix + q4 drafter + verify dispatch) (#2429) by @jjang-ai
• Images to remote providers: size for the wire, fix mime labels, never silently drop (#2428) by @jjang-ai
• Video attachments were silently dropped at send for bundles whose name lacks -vl (#2427) by @jjang-ai
• Unload image gen/edit models from Loaded Models, like LLMs (#2426) by @jjang-ai
• Fix the DFlash 2 drafter download link: point at the published incoai bundle (#2422) by @jjang-ai
• Settings: DFlash 2 drafter picker with download link and live folder validation (#2421) by @jjang-ai
• Repin vmlx: DFlash 2 drafter failures contained — AR fallback instead of a host crash (#2420) by @jjang-ai
• Repin vmlx to staged-verify MTP + crash fixes; picker rows gain an accessibility press (#2419) by @jjang-ai
• i18n: the Computer Use diagnostics panel was marked never-translate (#2329) by @YspritanHyzygy
• i18n: unflag the remaining strings that were marked never-translate by mistake (#2330) by @YspritanHyzygy

🚀 Features

• make sidebar resizable (#2413) by @RaajeevChandran

🐛 Bug Fixes

• fit fixed size sheets to the screen so action footers stay reachable (#2435) by @RaajeevChandran
• fix mcp connection docs to point to tools connections tab (#2434) by @RaajeevChandran
• warn when inbound dispatch is on but reply automatically is off (#2317) by @RaajeevChandran
• fix openai codex models reporting a fallback context window (#2416) by @RaajeevChandran
• fix external providers stuck disconnected after an update relaunch (#2415) by @RaajeevChandran
• minor adjustments to project pill for resizable sidebar (#2414) by @RaajeevChandran

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.23.0...0.23.1

What's Changed

• Attribute silent restarts: exit markers + a launch-time verdict file (#2401) by @jjang-ai
• Reserve answer room on API requests so thinking cannot eat the whole max_tokens (#2399) by @jjang-ai

🚀 Features

• added reasoning effort control for zai-glm models on mistral (#2412) by @RaajeevChandran
• added projects support to group chats with shared instructions, knowledge and memory (#2337) by @RaajeevChandran

🐛 Bug Fixes

• fix custom agents failing to spawn subagents by name (#2411) by @RaajeevChandran
• Treat a bare channel name as blank thinking so empty Harmony blocks stop rendering (#2407) by @jjang-ai
• added close button to router account sheet (#2406) by @RaajeevChandran
• fall back to the chat model when follow up generation hits a residency refusal (#2405) by @RaajeevChandran
• Derive evals-deterministic from floors.json so the CI lane cannot drift (#2404) by @jjang-ai
• Persist before stop in window cleanup so a close during load cannot destroy the send (#2402) by @jjang-ai
• Re-sync capability_search caseFloors with the suite and guard the pairing (#2400) by @jjang-ai
• Persist the cancelled marker when Stop beats the first delta (#2392) by @jjang-ai

Full Changelog: https://github.com/osaurus-ai/osaurus/compare/0.22.22...0.22.23