Skip to content
Agents tracked: 492 Releases (7d): 545 Agent status: 1 with issuesSince last limit reset: Claude 9d · Codex 7d For AI agents: Markdown Updated Nov 20, 2026

AI agent releases: Nov 16 – Nov 22, 2026

431 stable releases from 120 tracked agents and frameworks in the week of Nov 16, 2026 (so far), newest first, with the headline changes from each set of release notes. Our lists keep each project's latest releases, so a very active project's older releases can be missing.

Nov 19, 2026

Magpie
  • Fixed gateway model listings showing Ultrafast speed for accounts that don’t actually support it.
  • Fixed routing responses to correctly show Ultrafast speed when a group member uses an Ultrafast model.
  • macOS (Apple Silicon): magpie-darwin-arm64.dmg
Loki Mode
  • An engine run on the Agent SDK route loads only the project's Claude settings, so a user whose sign-in lives in ~/.claude/settings.json…
  • Correction to the v11.159.0 headline: verify's default 900 second limit is still in place. 11.159.0 makes the hermes-agent issue 155462 checks finish well inside it (465 to 494 seconds).
  • release(langgraph): 1.2.23
  • release(prebuilt): 1.1.3
  • fix(langgraph,prebuilt): give Send-dispatched tool calls the same state as an edge
  • 609baba: Let FileBrowser hosts load experience metrics and override individual experience dashboard sections.
  • 704bec6: Add optional FileBrowser props for hosts to save and restore expanded folders independently of the selected resource.
  • 7a15a83: Make the AgentConnect full-instruction copy button more prominent and clarify its labels.
LiteLLM
  • build(ui): run the UI image on slim nginx so it no longer ships curl [stable/1.110.x]
  • fix(responses): restore GPT-6.2 prompt caching on 1.110 stable
  • build: refresh the Wolfi base image and bump litellm to 1.110.1 on stable/1.110.x
  • Cheaper than vector RAG: for both index and query.
  • Faster: trees are created from layout stats.
  • Unified SDK: vectorless RAG. Local or cloud, for index and chat.
  • 适配 DSH 0.2.2-rc.5(含 0.2.2-rc.4) — 主验证线从 0.2.2-rc.3 推进到 0.2.2-rc.5,peer 范围保留 0.2.2-rc.4、0.2.2-rc.3 及更早验证线,停在 rc.3/rc.4 的用户也能直接装这版,程序内 /update 会把 dsh 引擎对齐到 0.2.2-rc.5。rc.4 为 DeepSeek refusal 停止原因新增的…
  • present 交付文件卡 — present 工具不再渲染成原始参数 JSON 与纯文本 Presented 行,而是渲染为专用交付卡:折叠态按行显示四格类型徽章(或通过 /settings → 外观 → nerdFontIcons 开启 Nerd Font 图标)、文件名、文件大小与修改时间;展开(Ctrl+O…
  • 网页搜索默认改为「自动」 — 在 DSH 0.2.2-rc.5 上,/settings → 网页搜索中的「搜索」未设置时显示「自动」,选中会清除用户层覆盖并交回按路由选择。若下层组合仍固定 searchProvider(旧验证线 dsh-base 或部署自带的 base 层补丁),界面不提供「自动」且手输 automatic 会被拒绝,与官方 Web 菜单一致;用户 profile…
  • Fix detection of multimodal support for Claude 5.
  • Update snapshot and changelog for v1.15.42.
  • Experiential 0.7.256 adds optional, content-free evidence for online input checks to existing terminal ledger writes.
  • Collection is opt-in for supporting ledger writers. Terminal retries preserve the frozen evidence, and accepted requests that fail before accounting registration retain their check outcomes.
  • The release retains exp-gateway-native==0.3.201, published with 0.7.255.
  • ⏰ Automations are more flexible and powerful: choose which MCP services each run can use, share automations with others as read-only, and find those "Shared with me" tab.
  • 🔲 Session grids are on by default, so you can drag sidebar sessions into a conversation to watch up to four live chats side by side.
  • ⏪ Use /rewind in Claude SDK and Claude Code native chats to go back to an earlier prompt, edit it and resend it.
  • feat(pages): link the leaderboard to AACR-Bench
  • fix(llmloop): reserve complete compression context budget
  • fix(chatgptauth): drain OAuth callback responses before closing
  • Agent turns in Chat fold to their lead. Each turn in Chat is one block now. While the agent works, its tool calls show as muted lines above its text.
  • Sandboxed sessions on macOS. sessionSandbox now works on macOS as well as Linux. Each session runs under a Seatbelt profile Glimmervoid builds for it, with the same file and network limits as on…
  • Drag a frame to move everything inside it. Press a frame's title or edge on the board and drag: the frame moves with every box in it, and wires between those boxes keep their bends.
  • bash pip install claude-agent-sdk==0.2.185
  • Let Peers(network) join an InMemoryPeerNetwork under the agent's name, and rename PeerNetwork to PeerConnection
  • Add StavProvider for the Stav European inference gateway
  • Add Otari as an OpenAI-compatible gateway provider
  • fix(core,cli): preserve literal $ characters during template placeholder replacement
  • feat(cli): add /stats heatmap subcommand for daily token usage
  • fix(core): prevent GCLI_EXPAND_TARGET env collision in expandEnvVars
Phoenix
  • add Phoenix release notes (ed6f1d7)
  • skills: teach phoenix-harbor to build Harbor tasks and verifiers you can trust (77fe4b4)
  • Added auto mode for Claude Opus 4.6 and Sonnet 4.6 on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry and Claude apps gateway; interactive sessions with no permission mode start in it
  • Added "onFailure": "block" for mcp_tool hooks, as for command and HTTP hooks: a hook whose call fails blocks the action
  • Added a warning when a message from a session that skips permission prompts may be held and dropped by a cloud session
  • Each waiting agent gets its own line: which agent, which project, and the question it ended on. Claude in web-app asks: Ship it tonight? You know which window to open.
  • An approval reads as one: needs your approval (Bash). It counts once, and it goes away when that session is gone.
  • sno squad opens in about a second, with who is waiting for you at the top.
  • fix: keep env, secret, and workspace identity lookups on their real sources (29d9ba64f) — vaibhav upreti
  • chore: drop root config entries for removed modules (8beafc031) — vaibhav upreti
  • feat(search): import large permission snapshots in bounded parts
  • feat(library): Best AI Agent Platforms for University Student Services and IT
  • feat(library): What Do Enterprise AI Agent Platforms Charge for SSO and Audit Logs?
  • Added split view to the Remote Control browser view, so you can keep several conversations open side by side: choose Split Right or Split Down from the Split submenu of a conversation's menu in…
  • Added Gemini 4 Neon (Low, Medium, High, and XHigh) to /model and --model when you connect with a GEMINI_API_KEY.
  • Added a scroll position indicator to the full screen (altscreen) rendering mode: while you're scrolled up, the line above the prompt shows how much of the conversation is below what you're…
  • Added support for excluding custom messages from model context and compaction or branch summaries, while allowing queued steering and follow-up messages to continue correctly after excluded messages.
  • Added a prewarm stream option for Anthropic Messages requests to refresh prompt caches without generating billable output.
  • Claude usage reports now include the account subscription plan in metadata.subscriptionPlan, supporting Max 5x, Max 20x, Pro, Team, and Enterprise plans.

Nov 18, 2026

  • Loopmaxxing. A new developer landing at loopmaxxing.sh: one prompt in Claude Code, Codex, Cursor, OpenCode or pi adds a Software Factory to your repository.
  • Guided setup. The setup guide offers the template that fits your repository, updates an older CLI instead of keeping it, offers Slack so teams can give agents work from Slack, and connects Linear…
  • Your repository holds your Kortix setup. Setup proves your dev environment by running it, sizes the agents' computer to your stack, and resizes it from measured peaks.
Junie
No release notes published.
  • Approval settings now apply across routines, connected chat apps, and work influenced by outside content.
  • Creating event triggers or saving skills can require approval when outside content may have influenced the request.
  • Remote secret stores clean up unused credentials after interrupted writes, replaced sign-ins, and deletion of accounts or workspaces with connected services.
  • #3150 69c2440 Thanks @miguelg719! - Fix OpenAI strict-schema rejection of the nullable act action while still stripping unknown keys when parsing.
  • #3114 f6b3d85 Thanks @miguelg719! - Fix CAPTCHA awareness options for local Browserbase attachments and forward them through the API.
  • #3491 57b905e Thanks @seanmcguire12! - Fix clicks on partially occluded elements
Magpie
  • Fixed an issue where the magpie's end line could extend outside its designated area.
  • macOS (Apple Silicon): magpie-darwin-arm64.dmg
  • macOS (Intel): magpie-darwin-amd64.dmg
Loki Mode
  • A lint error that Loki's own change put in a file it changed no longer seals a working change as FAILED.
  • A lint error that Loki's own change put in a file it changed no longer seals a working change as FAILED.
  • The label now describes exactly what the pull request merges. Measured on a real run (loki start asklokesh/FireLater#19, pull request 70): the checkout was on a feature branch whose tip was not in…
HyperFrames
  • Capture: Accept a short-lived Vertex access token for image captioning (e0470b0add, #8215).
  • Studio: A script edit re-runs in place in a sealed preview, without a reload (b8bb8b3803, #8158).
LangChain
  • Changes since langchain==1.5.1
  • release(langchain): 1.5.2 feat(langchain): tag synthetic HITL tool messages with source metadata
  • The built-in Google Workspace connector is off by default. While Paperclip's Google app verification is pending, the unified Google Workspace connection introduced in v2026.1106.0 is gated behind…
  • Every subtask gets its own execution workspace. Child tasks no longer share the parent's checkout, branch, or unmerged changes, on every creation path…
  • Agent-created task trees stop at parent → child. When an agent asks for a grandchild, the server creates the task beside the requester under the requester's parent and returns parentAdjusted: true…
eve
  • Released 2026-11-18 · Compare with 1.0.1
Zed
  • Added an official Snap package for Linux
Cognee
  • Safer Postgres per-dataset connection pools: lower default overflow and bounded wait times to avoid huge connection spikes and long stalls.
  • Improved LLM call capture: records follow-up (nested) LLM calls and includes caller metadata for clearer debugging and usage analysis.
  • Chunking robustness: stricter validation and centralized stamping of chunks so stored chunks are reliably identified and reused when possible.
  • VLM Run: new livekit-plugins-vlmrun plugin for vision-language chat and System One decisions, including qwen/qwen3.8-27b.
  • AssemblyAI STT 3.7 Realtime Agent: supported through LiveKit Inference as assemblyai/assemblyai-stt-3-7-realtime-agent.
  • Sarvam: STTRealtime now supports saaras:v4.
Sno Station
  • When an agent finishes what it is doing, it reads the tasks waiting for it and keeps going. Nobody has to wake it.
  • Send one task to several agents. One of them takes it. The others step aside.
  • A task nobody touched gets a nudge later. If the agent holding it died, a new session picks it up from where it left off.
  • Agent: model replies are read off the main thread (board replies no longer wait behind them), and the app logs what the main thread runs when it stalls.
  • Agent: a list whose cards already show every fact the goal asks for is read with collect from the top and never opened item by item.
  • Agent: higher ceilings: model output 8000 tokens (a reply still cut at its ceiling is asked once more with 16000), each up to 300 items, notes up to 1000 characters and 5000 items, 12 write-up…
  • macOS (-sea assets): unsigned (Apple secrets not configured)
  • Linux (-sea assets): SHA256SUMS-sea published with the release
  • alerts: give every card options and one "Or tell me" box (0d42d34)
  • alerts: give held step and email reply cards options and Or tell me (2706d2f)
  • alerts: make held cards say who they wait on and why, and let Google Doc appends run (ad93df0)
  • bb desktop 0.50.1
Codex Security
  • GitHub Action for Codex Security 0.10.1.
  • Installs @openai/codex-security@0.10.1.
  • yaml uses: openai/codex-security@bf650710f36c16936324efe250a49cdd9f2da037 # action-v0.10.1
No release notes published.
  • e403988: add an interrupt option to steering so a steer can cancel a model call that has not produced output yet
  • b792b56: feat(ai): ship the AI SDK skill in the package at skills/ai-sdk for pnpm agent skills
  • Updated dependencies [700089c]
Deep Agents
  • Honor MCP hook approval policies.
  • Wait for background worker cleanup to complete during shutdown.
  • d61bf11 release(deepagents-code): 0.2.11
Omniget
  • Build v1.0.1. Publish with pnpm fetch:release and pnpm mirror on the site repository.
  • Make ChromaDB an optional extra
  • Default Knowledge to local Qdrant and decouple storage from ChromaDB
  • Pick tools for standalone agents and add them later
  • feat(cloud): pick the exact connected git provider for repositories
  • feat(verify-openhands): every run has a non-loopback plain-HTTP origin
  • fix(filters): close EnumFilterDropdown menus on Escape
  • Stop an agent avatar from taking clicks for the row above
  • Let an image step have a custom size and the 4:5 and 5:4 aspect ratios
  • Refuse a confirmation card in an external chat turn
Open Science
  • Scheduled research tasks. Set up recurring or one-time schedules that launch sessions and Notebook tasks on their own — each run opens in the workspace with its own record, and you stay in control…
  • Image generation and editing. Generate images from a prompt with your configured provider, then iterate: each edit keeps a new version, so you can compare, revisit, and reuse earlier variants…
  • Editable presentations. PowerPoint files open for editing: revise text with formatting — fonts, sizes, colors, alignment — and keep the file's structure intact, all inside the app.
Glimmervoid
  • Sane YOLO catches more risky commands and blocks fewer safe ones. The bundled cc-safety-net moves to 2.7.15.
  • The doorman logs a wake whose host never answers, and tags every failure line with its kind.
  • Terminals on a phone show all their text after you return to the dashboard. The phone browser can discard the glyphs a terminal draws from while the dashboard is in the background, and the old…
Opik
  • [OPIK-9255] [FE] fix: remove explainer hints from experiments feedback scores and prompt headers
  • chore(deps): Update base images (non-major)
  • [OPIK-7855] [BE] refactor: finish the Vertex AI move to the Gen AI SDK
  • 💻 Windows:双击下载的安装包,按提示完成安装。ARM64 需要 .NET Framework 4.8.1,较新的 Windows 11 已内置。
  • 🍎 Mac(macOS 14 及以上):双击下载的 .pkg,按提示继续安装;安装期间会自动下载并准备必要运行时,完成后从“应用程序”打开 Toonflow。
  • 🐧 Linux(Debian / Ubuntu):选择对应架构的 DEB,在下载目录执行 sudo apt install ./toonflow-版本号-linux-架构.deb。安装后可在「设置 → 关于」更新;使用 2.0.10 的用户需先手动覆盖安装一次。
  • Fixed Collapse turns hiding a finished reply when Claude answers a later background task notification
  • Fixed detached worktree commits getting lost when archiving a workspace
  • Fixed Desktop updates leaving an older daemon running when an earlier Desktop session started it
  • A new look: every screen after you sign in is redrawn. The sidebar groups them into Track, Find work and Your profiles.
  • Positions, one library of postings: Find jobs kept a list for each search. It's now Positions, which keeps every posting your searches and the job agent find, once, with tabs for All open, New…
  • Every portal at once: Settings → Job search chose one source. It now lists every portal, each with its own switch, key and Test connection, and you switch on as many as you like.
Switch
  • The workspace switcher now shows on every build, not only one that can reach Switch Cloud — second workspaces, Create organization, New workspace and pending invitations are reachable from the…
  • Install failures show the whole error (and log it in full) instead of a truncated message
  • Sidebar reads a workspace's agents' rooms in one request rather than one per agent each refresh
  • Newly released models such as GPT-6.2 Sol appear in the model picker even when live catalog updates arrive while an agent is busy
  • Retain model catalog updates that arrive while agents are busy instead of dropping them
  • Update lettre, anyhow, event-listener, and tokio for new RustSec advisories
Mika
  • 移除输入栏的「推荐任务」星标入口(五角星),推荐任务 / 清单 / 工作流改从「设置 → 推荐任务」进入;复用常用任务仍在首页空状态的建议胶囊里。
  • TaskRun 巨型类拆解(1694 → 1238 行,拆出 TaskRunState / SteerCoordinator / TaskRunControl / RunTeardown / PhoneRuntime 五个类),纯重构、无行为变化。
  • 💬 From first word to final subtitle. Generate captions from your footage, bring in a subtitle file, or type them yourself.
  • 📤 One edit, seven places to take it. Ask your Agent for a vertical cut, a highlight reel, or a full-length adaptation.
  • ⚡ 96 new Skills. Pick a look and start making. Try animated notes over a talking head, a paper-collage explainer, a kinetic product launch, or a bold opening for an ad.
  • B1–B4 load artwork for the active shot and write generated sound effects in chunks to reduce peak memory.
  • B1–B7 hash large media files in chunks. B4 and B5 also avoid retaining or reading full images where lighter checks suffice. B6 streams additional media checks in its Python tools.
  • Targeted benchmarks showed lower peak memory, and tested output frames and audio matched the previous implementation. Overall render time still depends on the project and tools.
  • 💬 Continue Hermes conversations in Chat: Start Hermes from the New Tab menu, or choose Resume in New Chat on a Session History entry to continue a terminal conversation in its original workspace…
  • 🌐 Keep extensions beside your page: Chrome extensions that use side panels can now open in a resizable panel in Orca's browser. Click the extension's toolbar button to use its panel while browsing.
  • 🖥️ Reach your remote work: Managed SSH ports now have a Forward button. On Windows, projects opened through a mapped WSL drive letter can move onto their distro's WSL host.
  • Plugins. Edit, Canvas and the new Screening editor are plugins you turn on or off, and anyone can make one: a skill with a page, which can bring commands, a program and blocks laid out beside the…
  • The Community (early testing): find, add and publish plugins, skills, library packs and examples.
  • Motion in the film format. Clips and the elements of a page move by keyframes, with masks, text animated by character, word or line, 3D placement, and animatable blur, glow, brightness, contrast…
AgentOS
  • llm: carry a host's per-turn note in the application note (7e5b747)
  • api: pass the resume's note through the registered-tools helper; keep generic_completion's order for prompts with no host note (6baa725)
  • llm: a blank resume note sends none; count the note as it is sent (299ac68)
  • Knowledge alpha scope matching changed: $param now matches exactly one segment; raw : is no longer allowed inside a scope segment, requiring canonical encoding via encodeKnowledgeScopeId /…
  • @mastra/factory: newestOpenPullRequestNumber is no longer exported from @mastra/factory/boards; use workingPullRequest(item) + pullRequestSourceNumber.
  • @mastra/playground-ui: ListSearch and SearchFieldBlock are removed; migrate to SearchInput. Also removed SectionCard and CardHeading in favor of settings components.
  • Hand it a job; it comes back as a 事. delegate_task gives work to a DeepSeek Harness session in its
  • Your phone can use your Mac's agent. Pair them, flip one switch on the Mac, and the phone's Andee
  • 在办 is a door of its own on both bodies — a list, one page per task, and only you close a task.
  • 874897897fdd9ab8e26158521a0d14c1ffa2d452 MUL-7745 MUL-9463: fix(squads): run an @<!-- -->agent mention of the assigned squad's leader as a leader task (@ADanMan)
  • ed46a867d294e0faa00306ecc6bdc59293fbe42c MUL-8211: MUL-8209: feat(autopilots): payload conditions on webhook event filters (@gsayer)
  • 425cf08b6286260900f34b7dd8369597b414a747 MUL-8549 feat(weixin): add native QR-linked WeChat IM channel (@luxuze)
  • Reach approvers where they work. Configure email and an IM platform—Slack, Feishu/Lark, Microsoft Teams, DingTalk, or WeCom—to deliver workflow review requests through your team's existing channels.
  • Manage recipients in Contacts. Sync IM directories and select approvers from a shared contact list. Workflows can also address the initiator or use one-time and dynamic email recipients.
  • Approve in cards or forms. Recipients can review requests directly in supported IM cards or open standalone forms, with Dify login or email verification to verify their identity.
  • 128b63d: AgentConnect accepts isEnterprise to show English or Chinese team onboarding instructions for agents supported by the official installer, preserving API key masking and copying.
  • 196fad5: Add OpenViking CLI session, file download, and OVPack commands to Console. Local inputs use the browser file picker.
  • 81c42ac: Align the Console's built-in command options with the Volcengine terminal, including filesystem limits, search filters, ACL JSON, inline writes, resource import options, and task filters.
n8n
  • <!-- This is an auto-generated description by cubic. --> <a href=" target="_blank" rel="noopener noreferrer" data-no-image-dialog="true"><picture><source media="(prefers-color-scheme: dark)"…
  • New command bar with search, page actions, and bulk commands.
  • Copilot report analysis and actions with agent approval. [Enterprise, Beta]
  • Captain MCP connections and improved tool management. [Enterprise]
  • feat(agent): extract third-party agent into plugins
  • feat(console): quote response selections in follow-up drafts
  • fix(console): restore Safari message queue action icons
  • Ctrl+S stashes the input box draft, which comes back after you send another message
  • /goal <objective> sets a goal Vibe keeps working on across turns until it is done. /goal edit, /goal pause, /goal resume and /goal clear manage it
  • max_goal_token_budget in config.toml caps the token budget a goal may set, and is the budget of a goal set without one
  • iOS 版(测试)。 这一版起提供未签名的 iOS 安装包 dsh-go-1.2.0-ios-unsigned.ipa,需要用爱思助手、AltStore 等工具自己签名后安装。用免费 Apple ID 签名时 Bundle ID 会被改掉,部分能力(如 Siri、分享扩展、后台)可能用不了。iOS 版只在少数真机上试过,问题请开 issue。
  • 鸿蒙版合入主线。 鸿蒙版的代码已经和安卓、iOS 共用一套,这一版暂时不提供 HAP 安装包,签名就绪后再补到这个版本下。
  • 酒馆。 酒馆助手的接口补齐,支持原生 MVU 变量;能把 DeepSeek 网页里聊不下去的对话导入酒馆接着聊;导入大卡和长聊天时明显不卡了。
  • preview full shows the complete indexed function or Markdown section even when only one fragment matched. Short previews prioritize the matching lines.
  • PHP search recognizes classes, traits, enums, functions, methods, and properties, including their signatures and docblocks. Templates and top-level statements stay searchable.
  • zg recognizes Python scripts with uv run --script shebangs and indexes plain text in files with unfamiliar extensions.
  • Added dragging a subagent's tab from the side panel into the message box in the Remote Control browser view, which adds an @ mention of that subagent's conversation to your message.
  • Improved undo in Vim mode to work like Vim: u now undoes one command at a time instead of grouping edits made close together, a change that enters Insert mode, such as cw or o, undoes together…
  • Improved agy plugin validate so you can check a plugin before you install it: besides a local folder, it now accepts a plugin .zip file, an link to one, or a GitHub repository URL, using for a…
  • Backported GPT-6.2 Sol and Astra selection support to v2 when advertised by the service; automatic defaults are unchanged.
  • Restored support for legacy @github/copilot-sdk and @github/copilot-sdk/extension imports in ESM and CommonJS extensions alongside the @microsoft/copilot-sdk equivalents.
  • Restore support for older option formats in existing Node SDK extensions and direct RPC clients, without weakening v2 TypeScript types.
  • 2026-11-18
  • Fixes and changes
Metaharness
  • redblue defense <replay|verify|benchmark|help> adds the blue side: replay ARTEX findings or identity and edge telemetry against a baseline and a candidate detection policy, simulate a session…
  • README: a "Defense replay" section with an animated pipeline diagram, a CLI line, and six new keywords.
  • Long conversations and large session stores use bounded export and recovery work, reducing out-of-memory failures during chat, memory export, Doctor repair, and Gateway startup. Authored
  • Interrupted updates recover when their executor dies instead of leaving the Gateway stuck offline, while version-skew checks protect state still used by an installed service. Authored
  • Replies survive more restart and reconnect paths, preserving original prompts, recovering Telegram responses, and avoiding lost WhatsApp replies or skipped Signal messages.
  • feat(pages): add contact link to website footer
  • fix(session): keep finding identity fields separate
  • fix(telemetry): serialize lazy metric initialization
  • commit: b8a35999b2fec9e80430ad6f4b27a00d51323940
  • source tag: beta2026-11-18_11-09-26
  • regions: cn global
  • [client-v2] Fixed the sidebar collapse button covering links in the app switcher.
  • [Workflow: Loop node] Fixed scope variables being unavailable when configuring conditions on top-level loops.
  • [Block: template (deprecated)] Saving a legacy block template layout now requires permission to configure the interface.
  • factory run-issue <N> --continue-from <branch> [--handoff <file>] continues a budget-stopped run instead of re-planning.
  • CHECK respects an optional run deadline: set FACTORY_RUN_DEADLINE_AT (epoch ms or ISO-8601) and CHECK skips a flake re-run or rework round that would not finish before it, logs skipped…
  • factory run-issue parks before PLAN when the issue declares an open Depends on: / Prerequisite: issue, logs issue_deferred, and --json lists those issues in a new prerequisites array…
No release notes published.
  • What needs you comes first. On the home page the first request waiting on you is at the top, with the one orange button on the screen. Requests are one-line rows.
  • Shorter pages. On a real install: home is about 2,600px tall instead of 5,138; the List opens with 5 rows instead of 195; My Rules is about 4,400px instead of over 17,000.
  • Reply where you read. A request's drawer puts the reply box right under its progress line, in every state.
  • cli: add prompt tag command (5160fd5)
  • evaluators: run LLM evaluators on Anthropic prompts with extended thinking (593a776)
  • factory: pin commits and screen inputs before agent sessions (7d3a245)
  • 托管专家可选会话模式,并能使用连接器、知识库、专家回调和已启用的插件
  • 云服务器页的地域筛选改为按大区分组的下拉面板
  • 工具调用上限只拦截连续的相同调用
  • A scroll that moves nothing now fails — authoring scrolls, coordinate (vision) scrolls, and scrolls that never ran are reported as failures, the same way a locator click that misses one is…
  • Until-scrolls fail when the condition is never met — no more runs that scroll forever and then move on as if they succeeded.
  • Scroll detection reaches further — scrollable regions inside cross-origin (out-of-process) iframes, dialogs in front of the page, and <body> scrolling under a locked <html> are all detected and…
  • fix(gemini)!: answer 400 for an unsupported chat audio format, drop it under drop_params
  • fix(router): keep encrypted reasoning when a fallback hops between same-key deployments with no api_base
  • fix(anthropic)!: sync the beta header allowlist with what anthropic, vertex_ai, and azure_ai accept
OpenCode
  • GPT-6.2 Prompt Caching: GPT-6.2 models don't cache a system prompt sent in the Responses instructions field, so every step re-billed it as uncached input.
  • Fixed GPT-6.2 models re-billing the system prompt as uncached input on every step. On GitHub Copilot gpt-6.2-sol, uncached input per step dropped from about 14,000 tokens to 3.
  • webfetch structured output no longer repeats the url and format it was called with.
  • Compaction: drop the <analysis> scratchpad that Claude models refuse
  • Added the native Kiro provider: sign in with AWS IAM Identity Center, AWS Builder ID, Google, GitHub, or a Kiro API key (KIRO_API_KEY), stream responses with reasoning, tool calls, and images, and…
  • Added a definitive option to OAuthError; isDefinitiveOAuthFailure now takes the error object and disables a credential marked definitive without matching its message
  • buildAnthropicSearchHeaders takes extraBetas, request-specific betas sent beside the web search beta
Omnara
  • ghcr.io/omnara-ai/omnara-web@sha256:4c7a74a02cff03043654afaaa39df0bd75455c1717351d07bd6607f074e41426
  • ghcr.io/omnara-ai/omnara-api@sha256:500fb8137b28236612d682b2c06d56aeb49774980bf4325502d7a502e7e2bdde
  • ghcr.io/omnara-ai/omnara-worker@sha256:4403eefeea42242982166a8012385eff412a7113e72b36ec4317b3c2754826ad
  • fix: repair CI for main - Fixed the eight reproduced failures with a two (3345ad5b5) — vaibhav upreti
  • build(deps): bump litellm from 1.106.6 to 1.109.1 (628e86d15) — dependabot[bot]
  • build(deps): bump filelock from 4.2.4 to 4.2.10 (aff78dc1b) — dependabot[bot]
  • bash pip install claude-agent-sdk==0.2.184
  • Migrate Runloop workloads to another supported sandbox provider, such as Modal, or remain on 0.24.0 while migrating. Provider-specific sessions and snapshots require a separate data migration.
  • Resume nested agent-tool approvals with output guardrails using RunState schema 1.22. Pydantic checkpoints now preserve default values and in-place edits.
  • OpenAIResponsesCompactionSession now defaults to gpt-6-luna. Set model="gpt-4.1" to keep the previous choice. Agent model settings also follow the selected model more closely.

Nov 17, 2026

  • Say a rule to Claude Code, then say it again in other words. It is kept once, not twice.
  • Change your mind in Codex, and the old version stops reaching every agent, not just that one.
  • Each night the shared memory tidies itself. Repeats go, outdated lines go, and a line that drifted from what you actually said is put back right.
Loki Mode
  • In a repository with several packages, Loki no longer runs every package's tests on the base commit before the model starts.
  • Verify runs the tests of the packages a change touched and of the packages whose manifest depends on them, not every package.
  • Verify sizes itself from the time left. It may now use the time held back for a fix round (155 s became 335 s on the recorded run), each package suite gets what is left of the stage rather than a…
  • macOS (-sea assets): unsigned (Apple secrets not configured)
  • Linux (-sea assets): SHA256SUMS-sea published with the release
  • ⚡ Work across many chats: Chat UI uses lighter helpers on macOS and Linux. Hidden chat views pause their updates while retaining your draft, reading position and editor state for your return.
  • 🧹 Reclaim disk space: Settings → Workspace Cleanup can delete Orca, Claude Code and Codex chats from deleted workspaces. It is off by default.
  • ⌨️ Return to the same terminal folder: Panes remember where you last changed directory across restarts, including in folder workspaces.
  • Added a confirmation before /install-github-app changes a repository
  • Added claude mcp add-json <name> - to read the server's JSON from stdin, so tokens can stay off the command line; not available inside a Claude Code session
  • Added a line when a resumed session's earlier messages can't be loaded
  • SkillOperation and drop the Skills extension
  • Executable commands and SlashCommandOperation in oi-kit
  • Bang command operation in oi-kit
Junie
  • Session transfer: use /teleport and /continue to pick up a session on another machine
  • Context view: the new /context command, shows what fills the context window
  • /goal limit: now a hard cap, so credits don't go past the limit
Pydantic AI
  • GHSA-m4cr-m4qc-9hh3 (high): With strip_ansi=True, Pydantic AI Harness's ToolOutputLimits capability removed ANSI escape sequences with a pattern that took time quadratic in the length of a line…
  • GHSA-5m5q-5qj4-gx9h (high): The jwt pattern in Pydantic AI Harness's secret redaction backtracked in time quadratic in the length of the text it scanned, so a few hundred kilobytes of crafted text…
  • Forward service_tier to xAI and record the served tier in provider_details
  • Report agent status to terminals using OSC 7501.
  • Load MCP trust lists without blocking the server event loop.
  • Show a working upgrade command in the ACP import error message.
  • Azure: Fixed 400 when sending reasoning_effort to DeepSeek V4 Flash and Pro on azure_ai
  • AWS Bedrock: Fixed guarded_text parts in system and assistant messages being dropped instead of checked by the guardrail
  • AWS Bedrock: Fixed aborted Invoke streams on /v1/messages that kept generating billed output tokens upstream
  • Managed-agent tools refuse sessions this project did not create. The ownership ledger moved out of the workspace (~/.claude-flow/managed/), and terminate checks the provider's environment id.
  • The pre-bash root-delete guard reads $(...) substitutions and ${X:-/} parameter defaults. (#4848 via #5024;
  • Federation envelopes get a freshness check and replay guard. (#4849 via
Glimmervoid
  • Answer permission prompts from Chat. With the experimental Chat view switch on, a session launched without skip permissions holds each permission prompt from Claude Code, Codex, Kimi Code…
  • An experimental Chat view switch. Settings, Sessions, Chat view adds the switch for the Chat page being built as an alternative to a session's terminal. It is off by default and changes nothing yet.
  • Trace shows the conversation for every first-class agent. Codex, Kimi Code, OpenCode, pi and Antigravity sessions now fill the session's Trace with the prompt, thinking, replies and each tool call…
  • fix(nextplaid): restore 1.13 component template loading
  • fix: enforce custom component policy in background workflows
  • fix(auth): invalidate native sessions after password changes
Nod
  • Bug reports: Report bug is easier to find. Nod offers a report when something in Nod itself fails. You review what gets sent.
  • File previews: Right-click an image in the preview panel to save or open it.
  • Memory: Nod follows references in requests to remember earlier details. Notes you write keep their line breaks.
  • A model on an extension's custom api can take the Anthropic on-demand compaction lane by declaring remoteCompaction: { api: "anthropic-messages", enabled: true }.
  • Fixed provider-error handling so failed tool calls are no longer fed back to the model as refusal explanations.
  • Fixed one-shot retries and title generation for custom Zhipu Coding Plan providers so Beijing-time quota resets are honored.
  • This release improves performance and reliability for long Intelligence threads, adds support for distinct agent display names, and expands MCP Apps with model-context sharing.
No release notes published.
Phoenix
  • 5160fd5: Add px prompt tag <prompt-version-id> --tag <name> [--description <text>] to create a prompt version tag or move an existing one. Omitting --description clears the tag's existing description.
AI SDK
  • Updated dependencies [f886264]
  • Updated dependencies [d03d74a]
  • @ai-sdk/gateway@4.0.149
Claude Mem
  • Code Mode starts with claude-mem enabled
  • It resumes with a continuation prompt instead of the bare word "continue". Before clearing, Claude saves its work state one last time, naming the list it is working from, the task it just finished…
  • It starts at 40% of the context window instead of 33%. A CLAUDE_MEM_AUTO_CLEAR_PERCENT you set in settings.json or the environment still wins.
  • Themes and layouts: decks use new color themes and a set of layouts that are picked to fit each slide's content, so the same layout is not repeated slide after slide.
  • Live slide previews: slides appear in the chat as they are built, so you can watch the deck come together. Click a preview to see it full size.
  • Grounded content: figures, chart values and quotes are only kept when your outline, research or conversation contains them, and text is sized to fit each slide.
OpenMausBot
  • Team files: one shared folder for the team and its bots.
  • Team templates: a team library, templates from private GitHub repositories, and scripts that travel with their skills.
  • Bitwarden in the Vault: logins and secrets through Bitwarden, alongside 1Password.
Switch
  • Group chats — create from New room, open in the room view, manage agents and people from Configuration (group_chats, on by default)
  • Reworked Rooms experience — New Room in the sidebar, room cards with app-connection filter and archive, unread counts, leave a room, @-mention autocomplete
  • Room admins across every surface
Solidcraft
  • Windows: SolidCraft-0.44.1-windows-x64.exe runs as is. Or install the .msi, or unzip the
  • macOS 11+ (Apple silicon and Intel): open the .dmg and drag SolidCraft.app onto Applications.
  • Linux (x86_64 and aarch64, glibc 2.35+): the .AppImage (chmod +x, then run it), the
Sim
  • feat(library): Zapier Pricing at Scale: What Tasks Really Cost and When AI Agents Are Cheaper
  • feat(library): How Do You Build a Governed Workday AI Agent for Employee Onboarding and Offboarding?
  • feat(library): 6 Best StackAI Alternatives in 2026 After the Asana Acquisition
  • The Go SDK now lives at github.com/ai-openshell/sdk-go, with versions that follow OpenShell releases.
  • Wildcard network policy endpoints no longer allow private IP literals on their own. Add an exact endpoint or an allowed_ips grant for any private address your sandboxes still need to reach.
  • On Postgres, gateway migration 010 builds indexes on the objects table with a plain CREATE INDEX, which blocks writes to that table while it runs, so large deployments should plan for this during…
HyperFrames
  • CLI: check finds text over a face, measure cuts the names of resumed takes, and measure reports hands (820d63d7f, #8074).
  • Core: Hover and thumbnail captures never wait for a 4K clip's preview copy (f1f8ca31c, #8079).
  • Studio: Timeline rows in a sealed preview know their element, source file and audio group (cda94eb0f, #8078).
  • A0Cleaner — Automatic bulk cleanup of aged download caches, installer leftovers, retired browser profiles, inactive-chat screenshots, temporary captures, and unused Chromium versions.
  • Scheduled task catch-up after downtime — Recurring tasks that were missed while the system was offline now run exactly once on the next tick, coalescing missed occurrences.
  • Plugin maintenance tasks — Plugins can now own scheduled tasks that execute through the A0 task scheduler without creating an agent chat or calling a model.
  • Session Compactions: GET /sessions/{session_id}/compactions lists a session's compactions on SqliteDb and PostgresDb, newest first, with how many messages each replaced and the token count before…
  • Company Files: A Task's files and the shared notes can now be read over the API with GET /company/tasks/{task_id}/files and GET /company/notes, each with a /{path} route to read one file.
  • LinearTools, DiscordTools and ModelsLabTools: Now take a timeout (default 30 seconds, 120 for ModelsLab) that applies to every request.
  • The SDK has been rebranded from GitHub Copilot SDK to Microsoft Copilot SDK (see announcement). All package names and namespaces have been renamed accordidngly.
  • Starting with v2.0.0, the Copilot runtime runs inside your application process by default; previously, the SDK spawned an external child process to host the runtime.
  • Added PAC-aware system proxy routing support in the runtime, improving proxy handling in environments that use proxy auto-configuration.
  • Administrators can disable built-in Browser Use with the managed setting features.browserUse, including in running sessions.
  • Stop eve init's terminal UI from flashing when it hands off to eve dev.
  • feat(cli): serve one CDP session over Chromium's remote debugging pipe
  • refactor(types): extract renderer Inspector commands and native replies #3038
  • feat(ax,fetch): show semantic link URLs using computed roles
Experiential
  • exp-code 0.1.8: the open-source coding and cowork agent (agent/). The sandbox runner is attached as one self-contained file with its sha256.
Opik
  • [OPIK-9189] [SDK] fix: restore DSPy LM span usage with dspy 3.5
  • [OPIK-9165] [FE] chore: move the frontend to TypeScript 7 and oxlint
  • [OPIK-9216] [INFRA] chore: remove the Java exclusion from .gitattributes
  • Docs: python the README, its translations, CONTRIBUTING, RELEASING, the bug-report template and the ruff/pyright targets now state Python 3.11 as the minimum, matching the floor that shipped in…
  • Docs: links the Always-on setup link in the README and all translations points straight at docs.graphify.com/open-source/vibe-coding instead of going through a redirect from the old…
  • Fix: java each overloaded method gets its own node keyed by arity and erased signature, instead of all overloads collapsing into one node that dropped members and misrouted calls.
AI
  • #2107 fde4e09 - withCompaction({ native }) now passes the tools, the model options, and the reasoning of the chat call into compact.
  • Updated dependencies \[fde4e09]
  • @tanstack/ai@1.5.0
  • #4994 2016917 Thanks @sailist! - 修复前台 Shell 命令输出重复且加载状态不消失的问题。
  • #4982 9d7f9af Thanks @sailist! - 修复了若干已知的问题。
  • #4999 9bf2891 Thanks @kimi-agent-bot! - 修复 Windows 上 WebBridge 守护进程运行时无法更新的问题。
  • alerts: drop the card menu's link to a person's page (15929ae)
  • alerts: make every card say what its button does (789f63f)
  • alerts: stop watch cards repeating themselves or handing out look-ups (d05dbd9)
  • feat(chat): install a missing CLI from the conversation failure
  • feat(git): support SSH repositories with per-host user keys
  • fix(security): confine agent CLI config writes to home
  • feat(cloud): gate legacy v3 surfaces behind cutoff switches
  • feat(api-keys): add Developer preset and rename Ingestion to Application
  • feat(native): normalize provider usage and supplied costs
  • Agent steps stream live activity: Claude Code and Codex tool calls, skills, MCP calls, text and results appear as agent-event log events, [factory] agent-event {json} stdout lines and a…
  • Stable tag publishes now create or update the GitHub release from CHANGELOG.md, and a drift check fails when npm latest and the GitHub latest release differ.
  • Per-package line budget: npm run loc-budget in verify and CI static checks loc-budget.json, which may only fall; raising it needs a human-applied loc-budget:override label.
  • Cheaper than vector RAG: for both index and query.
  • Faster: trees are created from layout stats.
  • Unified SDK: vectorless RAG. Local or cloud, for index and chat.
  • A rail with Home and Settings — A slim column on the left holds Home, your pinned Spaces, Settings and your account. Pin a Space to keep it one click away.
  • Home lists every Space and its Sessions — Open, start, pin, rename, end or delete a Session straight from Home.
  • Search everything — Press ⌘K or the Search button in the top bar to find Spaces, Sessions, Tasks, files, Automations and Dispatch, in any Space.
Qwen Code
  • feat(web-shell): add sidebar workspace filter
  • feat(managed-agent): add the DingTalk adapter on the managed channel path
  • fix(managed-agent): settle stale public-child recovery as terminal instead of retrying forever
  • Restart-safe conversations: Queued messages, send permissions, and final chat writes are retained across Gateway restarts. Authored
  • Safer updates: Failed Doctor or Gateway startup rolls back an activated update, while restored npm packages remain readable by non-root services. Thanks @RomneyDa, @obviyus, and @vincentkoc.
  • Telegram widgets: Interactive widgets now render in Telegram messages, including private-chat topics.. Related #197923 Thanks @obviyus and @dankarization.
  • api: a refused gmi step's text never stands in for its lifted value (f73c0a7)
  • api: preserve refusals in gmi session schema replies (8f847a0)
  • api: report a refused gmi turn's message and own text on send and stream (5e7584d)
  • Unified A1–A7 production: storyboard approval, standalone prompts, optional low-resolution trial, then confirmed final resolution and credit budget.
  • Flexible generation lengths and separate editing shots, independent narration/music/SFX timelines, and action/camera continuity checks.
  • B7 Pixel Theater: approved examples, portrait layout and short-caption guidance, packaged code helpers and acceptance gates.
  • 🎬 Full-stack video recreation. Code video with HTML. Your Agent breaks down the reference, rebuilds its shots and transitions, and generates missing assets.
  • 🧩 200+ starting points for your next video. Browse video references and templates by brand, use case and video type.
  • 🏷️ 190+ Brand Kits—and bring a website of your own. Explore brand identities or extract colors, typography and assets from a public website URL.
  • [PR] gate: two fixes — evidence file is now readable after the run completes + one retry added for the small-window round in | gate 两个修复:跑完后证据文件能读了 + R 车道小窗口轮次加一次重试
  • [PR] Gate performance overhaul: P/R lanes parallelized, wall clock 45 min → ~18 min | gate 性能改造:P/R 车道并行,墙钟 45min→~18min
  • [PR] chore(branch): land 20 single-purpose fixes as one history-preserving batch (wave 10)
  • [client-v2] Fixed hasMany record deletion when the target collection uses multiple record identifier fields, including bulk deletion in v1 and v2.
  • [Data source: Main] Fixed field synchronization for tables loaded from the main database and preserved existing field metadata when imported physical table names include the configured prefix
  • [Migration manager] Fixed PostgreSQL migrations losing lines that begin with -- in multiline fields, changing line endings, or failing to restore the exported data.
  • Godot 4.2+
  • Active Ziva account
  • npm: bash npm install -g base44@latest
  • Homebrew: bash brew install base44/tap/base44
  • Version: 0.3.6 NPM Tag: latest
  • OpenCode 1 now offers the upgrade to OpenCode 2.0. OpenCode 2 downloads in the background and installs only when you choose: Restart to upgrade, Upgrade on quit, or Later.
  • Sessions with many images no longer send oversized requests. When the images in a session's history pass 25 MiB, the oldest are replaced with a placeholder.
  • A provider error for a request body that is too large now starts compaction and a retry instead of failing the turn.
  • Her notes reach her before she answers. A model used to pick the notes for each message after
  • Settings → Memory. Choose how notes reach her and which model writes them. A change reaches
  • Every change to a memory note is kept. Each memory folder keeps log.jsonl: who changed a
  • Fixed Claude Bypass approving actions without a permission card when a permissions.ask rule or a hook that returns ask required one
  • Fixed controls being touch-sized on a touch-screen laptop when a trackpad or mouse is its primary input
  • Fixed Claude sessions failing until /login when Paseo refreshed the Claude model list while Claude Code was renewing an expired access token
  • commit: 9ffef6494dc9e2490cb68d669df61c8c0fdcc054
  • source tag: beta2026-11-17_14-24-05
  • regions: cn global
  • Added accept-edits mode to --remote-agent sessions: Shift+Tab now switches between the default mode and accept-edits instead of doing nothing.
  • Added the Orchestrator model to /model for Gemini Enterprise accounts and sign-ins with Application Default Credentials (AGY_ADC_AUTH): it plans with Gemini 4 Carbon (High) and hands the plan's…
  • Added a conversation navigator to the Remote Control browser view: a rail with a mark for each of your prompts previews the prompt when you hover over it and jumps to it when you click, and [ and…
  • feat(session): filter listed sessions by branch
  • feat(rules): add Move review rules
  • fix(session): compare sessions recorded under different path spellings
Omnara
  • feat: publish MCP server card and enrich tool metadata
  • feat: inject machine daemon tokens through sandbox provider proxies
  • fix: tighten IPv6 address classification for outbound requests
  • Reworked swarms plan jobs in phases and parallel lanes, automatically start workers as dependencies finish, and show the whole team's progress in one pinned view
  • jcode Cloud officially launches: move a session or coordinated swarm to a VM with /cloud, keep working after disconnecting, and bring the conversation and repository back with /local
  • A lighter terminal with lower CPU usage, fewer background wakeups, faster transcript rendering, and less rendered-text memory
  • Heal counts come from the real heal chain — the number of self-healing attempts in a run is read directly off the heal call chain, so the totals shown match the steps kane-cli actually took.
  • Replayed runs report the same stability numbers — telemetry for stability and self-healing is carried through replay, so a replayed run no longer shows different stats than the original.
  • If/else branches compare the right value — exported automation code now compares the operand from your step instead of an empty string, so conditional branches behave the same in exported code as…
  • fix(deps): require dnspython>=2.9.2 (GHSA-vvf4-9x63-8xvv) (d590fc1ca) — Vaibhav Upreti

Nov 16, 2026

  • Add AppSignal and Cloud 66 to the Apps marketplace
  • Tell the agent why a Discord history read is empty
  • Accept long GitHub skill ids when installing a skill
  • opencode-pty 0.6.0 (protocol 9) from packages/process-daemon, published to npm as @opencode/pty@0.6.0.
  • macOS (-sea assets): unsigned (Apple secrets not configured)
  • Linux (-sea assets): SHA256SUMS-sea published with the release
Glimmervoid
  • Run Google's Antigravity CLI (agy) as a session agent. Antigravity CLI shows up in the Add Session picker when agy is on your PATH.
  • Team review re-reviews an updated pull request on time on a Glimmervoid Cloud server that went to sleep.
  • A Glimmervoid Cloud server that wakes from sleep refreshes sandboxed agent logins and checks the GitHub connection at once.
  • bash pip install claude-agent-sdk==0.2.183
Loki Mode
  • The label is now a count, not a reading. Loki gives the session numbered acceptance lines and two check functions (loki_http for a request, loki_cmd for a command).
  • The same loop now covers the acceptance lines. Once Loki's own proof run passes, Loki checks the change against every acceptance line before the session is let go, and hands each unmet line back…
  • Loki now closes the loop on its own proof run. After the session says it is done, Loki runs the proof script itself in a new empty directory.
  • feat(library): AI Agent Identity, Credential Management, and Audit Trails: Enterprise Guide
  • feat(library): Snowflake and QuickBooks MCP Servers for AI Agents
  • feat(library): Glean vs Dust: Which Enterprise AI Platform Should You Choose?
eve
  • Keep the changelog URL in the package README from overflowing into the npm sidebar.
Junie
No release notes published.
  • Added support for turning on Code Mode while agent teams are enabled; if you already turned it on, sessions with teams on now start in Code Mode
  • Added a startup warning when an --allowedTools, --disallowedTools or --tools value has an entry that doesn't look like a complete rule
  • Added a message explaining why MCP servers stop loading when an administrator adds managed-mcp.json during a session
  • Vertex AI and Gemini: Fixed versioned claude-*@<version> models stopping at 4096 tokens when max_tokens is unset
  • Chat completions: Fixed 400 on async Codestral chat with a file part and streamed PDF URLs stalling the proxy worker
  • Cost tracking and pricing: Fixed $0 spend on deployments whose base_model is azure/eu/o3, o4-mini, o1 or o3-mini
Oh My Pi
  • Fixed slow streaming of Gemma tool calls with large arguments.
  • Moved binary analysis from the standalone ida and ghidra tools into the backend-neutral dsm namespace (xd://eval/dsm).
  • Replaced the ida.enabled, ida.maxOpen, and ida.idleCloseSec settings with disassembler.backend, disassembler.maxOpen, and disassembler.idleCloseSec.
  • Desktop chats created by an August build with a bare custom provider can send again after updating.
  • A Desktop or TUI reply that streamed beside a todo or memory call now settles instead of leaving the Stop button up.
  • A crash mid-stream on Telegram and other edit-streaming chats no longer shows a duplicate "Recovered reply"; a word holding a key prefix mid-word (program_id) is shown on time again.
  • feat(verify-openhands): scenarios beyond the happy path, map controls, browser timezone
  • feat(acp): choose the reasoning effort from what the agent reports
  • feat(hooks): name the workspace and user hooks files in the Available Hooks dialog
Deep Agents
  • Updated stable model recommendations.
  • Limited editable-SDK scanning to deepagents records.
  • 7db36f2 release(deepagents-code): 0.2.9
  • Cloud-agent tools enforce the bot's current integration assignments and allowed tools before using a connected account.
  • Inbound email needs signed recipients before it can be treated as the owner's instructions; an SPF pass reported by a relay is no longer enough.
  • Saving memory after reading outside content follows the run's approval mode and Always allow list, including after a run resumes (approvals).
  • ae0ce7a: Add the named compile workflow and its CLI guide. FileBrowser now reports selected tree entries so hosts can add directory references to VikingBot, whose send button uses a rounded square.
  • 62d5cd5: Support switching named compilation configurations through directory-style tabs, native Git snapshot history, and Wiki edits with explicit snapshots.
  • a08cec9: Fix Console showing a command-unavailable warning during execution. CommandConsole now accepts busy to lock input and submission while pending result cards show progress.
  • New bots on the Mac and Windows app work on This computer by default. The choices now say plainly what they mean: This computer uses your files and apps; a Private desktop can't see your files.
  • A bot offers the switch itself. When a task needs your own files and the bot isn't on This computer, it says so in one line and switches with one tap, then carries on.
  • Bots on API models (OpenRouter, Groq, local models…) can read and save files on This computer, not only drive the screen. Your approval level decides what is asked.
Nod
  • Long requests: Nod checks request size before sending it to a model. Large content stays available in workspace files, with shorter previews in the request.
  • Setup: A retry after account setup no longer repeats the account save when only the local device check failed.
  • Contacts: Shared contact previews now read quoted vCard parameters, including values with colons or semicolons.
  • fix(validate): detect missing overridden package parts
  • alerts: change a card's plan from a reply on the card (cd8e0db)
  • briefing: count everything that waits on Home (0fd96b9)
  • briefing: decide each Needs you card's words and buttons on the server (f1659ba)
  • [CLI] S3 Bucket Ownership Verification: The workflow CLI now asserts bucket ownership (your own AWS account) on every S3 operation against its derived default bucket…
  • Animesh Pandey <animep@amazon.com>
  • PR details, actions, comments and activity in the session inspector
  • Title bar on every pane to drag and unsplit
  • Persist lifecycle and respawn records for app and daemon
  • feat(datasets): filter dataset items by status
  • feat(ai-gateway): add Google Vertex AI as a gateway provider
  • feat(trace): show topics transcript on trace via three dots (internal)
HyperFrames
  • Studio: A sealed preview frame forwards Studio's playback keys (b741ecf13, #7844).
  • CLI: Show and open the page to manage or cancel a paid plan (b2e6076b7, #7784).
  • Studio: Export a sealed preview's checked element reads and stand-ins (a0848ad35, #7918).
  • Local agent Bots add support for Antigravity (agy) and 24 more coding CLIs with automatic stdio MCP bridging, per-engine cwd isolation, reasoning-level restore, PATH directory filtering, and CLI…
  • Start a second live conversation part for a Bot while its earlier part is still working, and let a local agent Bot's own CLI and model act as its context steward for compacting or handing off to a…
  • Global Interface motion setting (Full, Reduced, Off) in Settings → General → Appearance, independent of per-Bot avatar animation.
  • pi-web-access 0.48.0. DuckDuckGo searches are paced and share a cooldown, so they no longer get blocked after a few searches in a row.
  • Automated: unit, regression and performance tests plus blackbox smoke in CI; release QA on the exact commit; a 14-case live Feishu channel smoke (Claude Code + Codex) on the release binary.
  • Upgrade path: a v1.5.0 instance with existing sessions and cron jobs was upgraded in place. Sessions, /list, /switch, long-conversation context, cron jobs and Codex sessions all survived.
  • Manual acceptance on Feishu
  • parlant lint lints a Parlant agent project through your coding agent. It
  • parlant init now installs a git pre-push hook (.githooks/pre-push, set
  • parlant dashboard agent cards AGENT_ID DESTINATION writes every file your
  • feat(cli): exit 75 when a failed request is safe to send again
  • Restrict Bach video generation to 1080p in the model package and HypiHub provider.
  • Require a first frame for frame-guided generation and accept an optional last frame.
  • Update Bach Surface definitions, examples and dependency versions.
  • Update the homepage copy, navigation, and project-star invitation.
  • Clarify Model Engine and Code Engine with a concise comparison and a three-step getting-started guide.
  • Explain A's unified workflow: approve the storyboard, receive standalone prompts, optionally test through a browser, then confirm resolution and credit budget before full production.
  • ⚡ Move between busy chats with less waiting: Long conversations load their history in pages, and background replies do less work in the sidebar and hidden chats when you switch conversations.
  • 🌐 Keep browser tabs without keeping every page loaded: Memory Saver frees eligible pages after four hours out of view by default, then reloads them when you return.
  • 🤖 Use Cursor in Chat: Start with the Cursor sign-in on your computer, choose a model, effort and Fast, approve commands, and reopen the conversation later.
  • 24ebdea7bed60633caaf6ef0b85dcd86e5d3bc16 MUL-6056: verify GitHub installation control before workspace binding (@Bohan-J)
  • c07c978407e4dd0b2b5081cf83478b49a20f707c MUL-6895, MUL-6586: fix(daemon,tasks): force a server-side re-check for woken runtimes to clear a stale empty-claim stall (@ADanMan)
  • 4c19b93e58b2e07959d3bdea5b2dd2142eb0d377 MUL-9191: remove MULTICA_THEME_STORE off switch; theme store is always on (@Bohan-J)
  • [OPIK-9147] [BE] fix: enforce Free plan span quota on OTel trace ingestion
  • [OPIK-9015] [BE] feat: add the guardrails validations API contract, answering 501
  • [issue-9443] [SDK] fix: warn when a thread has more traces than max_traces_per_thread
  • 回退「你好米卡」语音唤醒(#324):免 root 普通 App 拿不到系统热词入口,应用层常驻录音做不成系统级热词,功耗与后台限制补不回来。按住说话、胶囊长按说话等既有语音能力不受影响。
  • 模型列表把 403 + insufficient_quota 误判为 Key 无效,现正确判为额度用尽(#639)。
  • 个人助手回复无法长按选字/复制,补上选字 + 一键复制(#644、#661)。
  • Goals checked before they count as complete. Early in a Goal the agent proposes a checklist: each requirement paired with a command that proves it.
  • Sessions that keep their history. A Session that cannot be read is listed with the reason instead of being dropped. A stopped or failed Turn keeps what it already wrote.
  • A tighter sandbox. Agent commands no longer see the model keys DeepCode uses. Credential files and the private keys in ~/.ssh cannot be read, while git over SSH still works through your SSH agent.
  • <!-- This is an auto-generated description by cubic. --> <a href=" target="_blank" rel="noopener noreferrer" data-no-image-dialog="true"><picture><source media="(prefers-color-scheme: dark)"…
  • ci(ego-browser): move the real-browser E2E gate to pre-push and shorten it
  • fix(ego-browser): stop a pending navigation from evaluate and screenshot too
  • ci(codex-review): give the model direct shell tools instead of code mode
  • commit: 2bd4acad80376fd06e33b0b6fd12efe7733042bb
  • source tag: beta2026-11-16_14-43-43
  • regions: cn global
  • chore(version): bumping version to 2.0.12dev
  • fix(webui): include full calendar days in schedule date filter
  • fix(model): normalize nullable usage counters in adapters
  • Background tasks: See and stop background tasks above the composer for Claude Code, Codex and OpenCode v2.
  • Built-in plugin gallery: Browse, search and install plugins without leaving Paseo. Manage their settings and updates from the same page.
  • Command center tabs: Switch between Workspaces, Files and Actions, with recent workspaces shown first. Search now includes archived workspaces and matches agent titles within their workspace.
  • Databricks provider: drop thinking blocks bound to a different prefix
  • Release notes — formatted for people.
  • Changelog — plain Markdown for AI agents and tools.
  • npm package
  • Build: 0.12.10 Commit: 5b5b1a48cf Built at: 2026-11-16 03:25:25 UTC
  • Windows: SolidCraft-0.43.0-windows-x64.exe runs as is. Or install the .msi, or unzip the
  • macOS 11+ (Apple silicon and Intel): open the .dmg and drag SolidCraft.app onto Applications.
  • Linux (x86_64 and aarch64, glibc 2.35+): the .AppImage (chmod +x, then run it), the
  • Each conversation keeps the messages you sent from My Work and their replies, in order. Stop ends a running turn.
  • The box says which folder the turn runs in. When the recorded folder no longer exists (common after a worktree is cleaned up), the turn runs in the nearest existing folder above it.
  • "Copy resume command" copies claude --resume <id> or codex resume <id>. On a Mac, "Open in Claude" opens a Claude Code conversation in the Claude desktop app.
  • 工作台底部时钟可在“设置 → 通用”切换 12/24 小时制,保存后立即生效,并在重新打开时保留选择。
  • 降低原生环境轮询与画面传输开销,优化空闲内存回收。Windows 浏览器画面跟随查看设备的屏幕像素密度,保留清晰度和输入位置的一致性。
  • 桌面应用更新支持中断续传和重试,以无数据传输的时长判断停滞。Windows 更新支持可信发布者证书续期;Linux 发现新版时提供发行页入口,仍通过系统包手动安装。
  • fix(session): keep a hosted conversation on its repository and stop reading prose as one (f86668675) — vaibhav upreti
  • BU 4.0 is the default model. Agent(task) and ChatBrowserUse() now run bu-4.0 on the Browser Use gateway instead of bu-fast.
  • This costs more if you pass no llm. BU 4.0 tokens cost 15 times as much as bu-fast ($1.80 / $9.00 per 1M input / output tokens, against $0.12 / $0.60), and it runs at high instead of medium…
  • The first run without a key opens Browser Use login and continues after approval; cloud browsers print their live-view link.
  • Say "check the build tomorrow at 9" or "write the report every morning at 7" to Claude Code or Codex. At that time an agent starts and does it, with the conversation open or closed.
  • sno heartbeat --due shows what is scheduled. A task that could not start stays on the list with the reason beside it.
  • Your agents can now message each other across computers on the same network or over Tailscale.

Releases by week

Every stable release of the week, Monday to Sunday, for the last 26 weeks. Each agent's full history is on its own page.